1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This coordinates the per-function state used while generating code.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CodeGenFunction.h"
15 #include "CGBlocks.h"
16 #include "CGCleanup.h"
17 #include "CGCUDARuntime.h"
18 #include "CGCXXABI.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;
43 
44 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
45 /// markers.
46 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
47                                       const LangOptions &LangOpts) {
48   if (CGOpts.DisableLifetimeMarkers)
49     return false;
50 
51   // Disable lifetime markers in msan builds.
52   // FIXME: Remove this when msan works with lifetime markers.
53   if (LangOpts.Sanitize.has(SanitizerKind::Memory))
54     return false;
55 
56   // Asan uses markers for use-after-scope checks.
57   if (CGOpts.SanitizeAddressUseAfterScope)
58     return true;
59 
60   // For now, only in optimized builds.
61   return CGOpts.OptimizationLevel != 0;
62 }
63 
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(Address::invalid()), 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();
89 
90   llvm::FastMathFlags FMF;
91   if (CGM.getLangOpts().FastMath)
92     FMF.setFast();
93   if (CGM.getLangOpts().FiniteMathOnly) {
94     FMF.setNoNaNs();
95     FMF.setNoInfs();
96   }
97   if (CGM.getCodeGenOpts().NoNaNsFPMath) {
98     FMF.setNoNaNs();
99   }
100   if (CGM.getCodeGenOpts().NoSignedZeros) {
101     FMF.setNoSignedZeros();
102   }
103   if (CGM.getCodeGenOpts().ReciprocalMath) {
104     FMF.setAllowReciprocal();
105   }
106   if (CGM.getCodeGenOpts().Reassociate) {
107     FMF.setAllowReassoc();
108   }
109   Builder.setFastMathFlags(FMF);
110 }
111 
112 CodeGenFunction::~CodeGenFunction() {
113   assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
114 
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
117   // something.
118   if (FirstBlockInfo)
119     destroyBlockInfos(FirstBlockInfo);
120 
121   if (getLangOpts().OpenMP && CurFn)
122     CGM.getOpenMPRuntime().functionFinished(*this);
123 }
124 
125 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
126                                                     LValueBaseInfo *BaseInfo,
127                                                     TBAAAccessInfo *TBAAInfo) {
128   return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo, TBAAInfo,
129                                  /* forPointeeType= */ true);
130 }
131 
132 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
133                                                    LValueBaseInfo *BaseInfo,
134                                                    TBAAAccessInfo *TBAAInfo,
135                                                    bool forPointeeType) {
136   if (TBAAInfo)
137     *TBAAInfo = CGM.getTBAAAccessInfo(T);
138 
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()) {
144       if (BaseInfo)
145         *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType);
146       return getContext().toCharUnitsFromBits(Align);
147     }
148   }
149 
150   if (BaseInfo)
151     *BaseInfo = LValueBaseInfo(AlignmentSource::Type);
152 
153   CharUnits Alignment;
154   if (T->isIncompleteType()) {
155     Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
156   } else {
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);
163     } else {
164       Alignment = getContext().getTypeAlignInChars(T);
165       if (T.getQualifiers().hasUnaligned())
166         Alignment = CharUnits::One();
167     }
168 
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);
175     }
176   }
177   return Alignment;
178 }
179 
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,
185                           TBAAInfo);
186 }
187 
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.
190 LValue
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);
197 }
198 
199 
200 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
201   return CGM.getTypes().ConvertTypeForMem(T);
202 }
203 
204 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
205   return CGM.getTypes().ConvertType(T);
206 }
207 
208 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
209   type = type.getCanonicalType();
210   while (true) {
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");
219 
220     case Type::Auto:
221     case Type::DeducedTemplateSpecialization:
222       llvm_unreachable("undeduced type in IR-generation");
223 
224     // Various scalar types.
225     case Type::Builtin:
226     case Type::Pointer:
227     case Type::BlockPointer:
228     case Type::LValueReference:
229     case Type::RValueReference:
230     case Type::MemberPointer:
231     case Type::Vector:
232     case Type::ExtVector:
233     case Type::FunctionProto:
234     case Type::FunctionNoProto:
235     case Type::Enum:
236     case Type::ObjCObjectPointer:
237     case Type::Pipe:
238       return TEK_Scalar;
239 
240     // Complexes.
241     case Type::Complex:
242       return TEK_Complex;
243 
244     // Arrays, records, and Objective-C objects.
245     case Type::ConstantArray:
246     case Type::IncompleteArray:
247     case Type::VariableArray:
248     case Type::Record:
249     case Type::ObjCObject:
250     case Type::ObjCInterface:
251       return TEK_Aggregate;
252 
253     // We operate on atomic values according to their underlying type.
254     case Type::Atomic:
255       type = cast<AtomicType>(type)->getValueType();
256       continue;
257     }
258     llvm_unreachable("unknown type kind!");
259   }
260 }
261 
262 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
263   // For cleanliness, we try to avoid emitting the return block for
264   // simple cases.
265   llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
266 
267   if (CurBB) {
268     assert(!CurBB->getTerminator() && "Unexpected terminated block.");
269 
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();
275     } else
276       EmitBlock(ReturnBlock.getBlock());
277     return llvm::DebugLoc();
278   }
279 
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();
294       return Loc;
295     }
296   }
297 
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.
301 
302   EmitBlock(ReturnBlock.getBlock());
303   return llvm::DebugLoc();
304 }
305 
306 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
307   if (!BB) return;
308   if (!BB->use_empty())
309     return CGF.CurFn->getBasicBlockList().push_back(BB);
310   delete BB;
311 }
312 
313 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
314   assert(BreakContinueStack.empty() &&
315          "mismatched push/pop in break/continue stack!");
316 
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.
328   //
329   // If there are multiple branches to the return block, the branch
330   // instructions will get the location of the return statements and
331   // all will be fine.
332   if (CGDebugInfo *DI = getDebugInfo()) {
333     if (OnlySimpleReturnStmts)
334       DI->EmitLocation(Builder, LastStopPoint);
335     else
336       DI->EmitLocation(Builder, EndLoc);
337   }
338 
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;
347   if (HasCleanups) {
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);
353 
354     PopCleanupBlocks(PrologueCleanupDepth);
355   }
356 
357   // Emit function epilog (to return).
358   llvm::DebugLoc Loc = EmitReturnBlock();
359 
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");
366   }
367 
368   // Emit debug descriptor for function end.
369   if (CGDebugInfo *DI = getDebugInfo())
370     DI->EmitFunctionEnd(Builder, CurFn);
371 
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);
377 
378   assert(EHStack.empty() &&
379          "did not remove all scopes from cleanup stack!");
380 
381   // If someone did an indirect goto, emit the indirect goto block at the end of
382   // the function.
383   if (IndirectBranch) {
384     EmitBlock(IndirectBranch->getParent());
385     Builder.ClearInsertionPoint();
386   }
387 
388   // If some of our locals escaped, insert a call to llvm.localescape in the
389   // entry block.
390   if (!EscapedLocals.empty()) {
391     // Invert the map from local to index into a simple vector. There should be
392     // no holes.
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);
400   }
401 
402   // Remove the AllocaInsertPt instruction, which is just a convenience for us.
403   llvm::Instruction *Ptr = AllocaInsertPt;
404   AllocaInsertPt = nullptr;
405   Ptr->eraseFromParent();
406 
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();
414     }
415   }
416 
417   EmitIfUsed(*this, EHResumeBlock);
418   EmitIfUsed(*this, TerminateLandingPad);
419   EmitIfUsed(*this, TerminateHandler);
420   EmitIfUsed(*this, UnreachableBlock);
421 
422   for (const auto &FuncletAndParent : TerminateFunclets)
423     EmitIfUsed(*this, FuncletAndParent.second);
424 
425   if (CGM.getCodeGenOpts().EmitDeclMetadata)
426     EmitDeclMetadata();
427 
428   for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
429            I = DeferredReplacements.begin(),
430            E = DeferredReplacements.end();
431        I != E; ++I) {
432     I->first->replaceAllUsesWith(I->second);
433     I->first->eraseFromParent();
434   }
435 
436   // Eliminate CleanupDestSlot alloca by replacing it with SSA values and
437   // PHIs if the current function is a coroutine. We don't do it for all
438   // functions as it may result in slight increase in numbers of instructions
439   // if compiled with no optimizations. We do it for coroutine as the lifetime
440   // of CleanupDestSlot alloca make correct coroutine frame building very
441   // difficult.
442   if (NormalCleanupDest.isValid() && isCoroutine()) {
443     llvm::DominatorTree DT(*CurFn);
444     llvm::PromoteMemToReg(
445         cast<llvm::AllocaInst>(NormalCleanupDest.getPointer()), DT);
446     NormalCleanupDest = Address::invalid();
447   }
448 }
449 
450 /// ShouldInstrumentFunction - Return true if the current function should be
451 /// instrumented with __cyg_profile_func_* calls
452 bool CodeGenFunction::ShouldInstrumentFunction() {
453   if (!CGM.getCodeGenOpts().InstrumentFunctions &&
454       !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining &&
455       !CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
456     return false;
457   if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
458     return false;
459   return true;
460 }
461 
462 /// ShouldXRayInstrument - Return true if the current function should be
463 /// instrumented with XRay nop sleds.
464 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
465   return CGM.getCodeGenOpts().XRayInstrumentFunctions;
466 }
467 
468 /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to
469 /// the __xray_customevent(...) builin calls, when doing XRay instrumentation.
470 bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const {
471   return CGM.getCodeGenOpts().XRayInstrumentFunctions &&
472          (CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents ||
473           CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask ==
474               XRayInstrKind::Custom);
475 }
476 
477 llvm::Constant *
478 CodeGenFunction::EncodeAddrForUseInPrologue(llvm::Function *F,
479                                             llvm::Constant *Addr) {
480   // Addresses stored in prologue data can't require run-time fixups and must
481   // be PC-relative. Run-time fixups are undesirable because they necessitate
482   // writable text segments, which are unsafe. And absolute addresses are
483   // undesirable because they break PIE mode.
484 
485   // Add a layer of indirection through a private global. Taking its address
486   // won't result in a run-time fixup, even if Addr has linkonce_odr linkage.
487   auto *GV = new llvm::GlobalVariable(CGM.getModule(), Addr->getType(),
488                                       /*isConstant=*/true,
489                                       llvm::GlobalValue::PrivateLinkage, Addr);
490 
491   // Create a PC-relative address.
492   auto *GOTAsInt = llvm::ConstantExpr::getPtrToInt(GV, IntPtrTy);
493   auto *FuncAsInt = llvm::ConstantExpr::getPtrToInt(F, IntPtrTy);
494   auto *PCRelAsInt = llvm::ConstantExpr::getSub(GOTAsInt, FuncAsInt);
495   return (IntPtrTy == Int32Ty)
496              ? PCRelAsInt
497              : llvm::ConstantExpr::getTrunc(PCRelAsInt, Int32Ty);
498 }
499 
500 llvm::Value *
501 CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F,
502                                           llvm::Value *EncodedAddr) {
503   // Reconstruct the address of the global.
504   auto *PCRelAsInt = Builder.CreateSExt(EncodedAddr, IntPtrTy);
505   auto *FuncAsInt = Builder.CreatePtrToInt(F, IntPtrTy, "func_addr.int");
506   auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int");
507   auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr");
508 
509   // Load the original pointer through the global.
510   return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()),
511                             "decoded_addr");
512 }
513 
514 static void removeImageAccessQualifier(std::string& TyName) {
515   std::string ReadOnlyQual("__read_only");
516   std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
517   if (ReadOnlyPos != std::string::npos)
518     // "+ 1" for the space after access qualifier.
519     TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
520   else {
521     std::string WriteOnlyQual("__write_only");
522     std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
523     if (WriteOnlyPos != std::string::npos)
524       TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
525     else {
526       std::string ReadWriteQual("__read_write");
527       std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
528       if (ReadWritePos != std::string::npos)
529         TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
530     }
531   }
532 }
533 
534 // Returns the address space id that should be produced to the
535 // kernel_arg_addr_space metadata. This is always fixed to the ids
536 // as specified in the SPIR 2.0 specification in order to differentiate
537 // for example in clGetKernelArgInfo() implementation between the address
538 // spaces with targets without unique mapping to the OpenCL address spaces
539 // (basically all single AS CPUs).
540 static unsigned ArgInfoAddressSpace(LangAS AS) {
541   switch (AS) {
542   case LangAS::opencl_global:   return 1;
543   case LangAS::opencl_constant: return 2;
544   case LangAS::opencl_local:    return 3;
545   case LangAS::opencl_generic:  return 4; // Not in SPIR 2.0 specs.
546   default:
547     return 0; // Assume private.
548   }
549 }
550 
551 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
552 // information in the program executable. The argument information stored
553 // includes the argument name, its type, the address and access qualifiers used.
554 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
555                                  CodeGenModule &CGM, llvm::LLVMContext &Context,
556                                  CGBuilderTy &Builder, ASTContext &ASTCtx) {
557   // Create MDNodes that represent the kernel arg metadata.
558   // Each MDNode is a list in the form of "key", N number of values which is
559   // the same number of values as their are kernel arguments.
560 
561   const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
562 
563   // MDNode for the kernel argument address space qualifiers.
564   SmallVector<llvm::Metadata *, 8> addressQuals;
565 
566   // MDNode for the kernel argument access qualifiers (images only).
567   SmallVector<llvm::Metadata *, 8> accessQuals;
568 
569   // MDNode for the kernel argument type names.
570   SmallVector<llvm::Metadata *, 8> argTypeNames;
571 
572   // MDNode for the kernel argument base type names.
573   SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
574 
575   // MDNode for the kernel argument type qualifiers.
576   SmallVector<llvm::Metadata *, 8> argTypeQuals;
577 
578   // MDNode for the kernel argument names.
579   SmallVector<llvm::Metadata *, 8> argNames;
580 
581   for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
582     const ParmVarDecl *parm = FD->getParamDecl(i);
583     QualType ty = parm->getType();
584     std::string typeQuals;
585 
586     if (ty->isPointerType()) {
587       QualType pointeeTy = ty->getPointeeType();
588 
589       // Get address qualifier.
590       addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
591         ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
592 
593       // Get argument type name.
594       std::string typeName =
595           pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
596 
597       // Turn "unsigned type" to "utype"
598       std::string::size_type pos = typeName.find("unsigned");
599       if (pointeeTy.isCanonical() && pos != std::string::npos)
600         typeName.erase(pos+1, 8);
601 
602       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
603 
604       std::string baseTypeName =
605           pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
606               Policy) +
607           "*";
608 
609       // Turn "unsigned type" to "utype"
610       pos = baseTypeName.find("unsigned");
611       if (pos != std::string::npos)
612         baseTypeName.erase(pos+1, 8);
613 
614       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
615 
616       // Get argument type qualifiers:
617       if (ty.isRestrictQualified())
618         typeQuals = "restrict";
619       if (pointeeTy.isConstQualified() ||
620           (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
621         typeQuals += typeQuals.empty() ? "const" : " const";
622       if (pointeeTy.isVolatileQualified())
623         typeQuals += typeQuals.empty() ? "volatile" : " volatile";
624     } else {
625       uint32_t AddrSpc = 0;
626       bool isPipe = ty->isPipeType();
627       if (ty->isImageType() || isPipe)
628         AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
629 
630       addressQuals.push_back(
631           llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
632 
633       // Get argument type name.
634       std::string typeName;
635       if (isPipe)
636         typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
637                      .getAsString(Policy);
638       else
639         typeName = ty.getUnqualifiedType().getAsString(Policy);
640 
641       // Turn "unsigned type" to "utype"
642       std::string::size_type pos = typeName.find("unsigned");
643       if (ty.isCanonical() && pos != std::string::npos)
644         typeName.erase(pos+1, 8);
645 
646       std::string baseTypeName;
647       if (isPipe)
648         baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
649                           ->getElementType().getCanonicalType()
650                           .getAsString(Policy);
651       else
652         baseTypeName =
653           ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
654 
655       // Remove access qualifiers on images
656       // (as they are inseparable from type in clang implementation,
657       // but OpenCL spec provides a special query to get access qualifier
658       // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
659       if (ty->isImageType()) {
660         removeImageAccessQualifier(typeName);
661         removeImageAccessQualifier(baseTypeName);
662       }
663 
664       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
665 
666       // Turn "unsigned type" to "utype"
667       pos = baseTypeName.find("unsigned");
668       if (pos != std::string::npos)
669         baseTypeName.erase(pos+1, 8);
670 
671       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
672 
673       if (isPipe)
674         typeQuals = "pipe";
675     }
676 
677     argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
678 
679     // Get image and pipe access qualifier:
680     if (ty->isImageType()|| ty->isPipeType()) {
681       const Decl *PDecl = parm;
682       if (auto *TD = dyn_cast<TypedefType>(ty))
683         PDecl = TD->getDecl();
684       const OpenCLAccessAttr *A = PDecl->getAttr<OpenCLAccessAttr>();
685       if (A && A->isWriteOnly())
686         accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
687       else if (A && A->isReadWrite())
688         accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
689       else
690         accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
691     } else
692       accessQuals.push_back(llvm::MDString::get(Context, "none"));
693 
694     // Get argument name.
695     argNames.push_back(llvm::MDString::get(Context, parm->getName()));
696   }
697 
698   Fn->setMetadata("kernel_arg_addr_space",
699                   llvm::MDNode::get(Context, addressQuals));
700   Fn->setMetadata("kernel_arg_access_qual",
701                   llvm::MDNode::get(Context, accessQuals));
702   Fn->setMetadata("kernel_arg_type",
703                   llvm::MDNode::get(Context, argTypeNames));
704   Fn->setMetadata("kernel_arg_base_type",
705                   llvm::MDNode::get(Context, argBaseTypeNames));
706   Fn->setMetadata("kernel_arg_type_qual",
707                   llvm::MDNode::get(Context, argTypeQuals));
708   if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
709     Fn->setMetadata("kernel_arg_name",
710                     llvm::MDNode::get(Context, argNames));
711 }
712 
713 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
714                                                llvm::Function *Fn)
715 {
716   if (!FD->hasAttr<OpenCLKernelAttr>())
717     return;
718 
719   llvm::LLVMContext &Context = getLLVMContext();
720 
721   GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
722 
723   if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
724     QualType HintQTy = A->getTypeHint();
725     const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
726     bool IsSignedInteger =
727         HintQTy->isSignedIntegerType() ||
728         (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
729     llvm::Metadata *AttrMDArgs[] = {
730         llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
731             CGM.getTypes().ConvertType(A->getTypeHint()))),
732         llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
733             llvm::IntegerType::get(Context, 32),
734             llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
735     Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
736   }
737 
738   if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
739     llvm::Metadata *AttrMDArgs[] = {
740         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
741         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
742         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
743     Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
744   }
745 
746   if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
747     llvm::Metadata *AttrMDArgs[] = {
748         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
749         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
750         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
751     Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
752   }
753 
754   if (const OpenCLIntelReqdSubGroupSizeAttr *A =
755           FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
756     llvm::Metadata *AttrMDArgs[] = {
757         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
758     Fn->setMetadata("intel_reqd_sub_group_size",
759                     llvm::MDNode::get(Context, AttrMDArgs));
760   }
761 }
762 
763 /// Determine whether the function F ends with a return stmt.
764 static bool endsWithReturn(const Decl* F) {
765   const Stmt *Body = nullptr;
766   if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
767     Body = FD->getBody();
768   else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
769     Body = OMD->getBody();
770 
771   if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
772     auto LastStmt = CS->body_rbegin();
773     if (LastStmt != CS->body_rend())
774       return isa<ReturnStmt>(*LastStmt);
775   }
776   return false;
777 }
778 
779 static void markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
780   Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
781   Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
782 }
783 
784 static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) {
785   auto *MD = dyn_cast_or_null<CXXMethodDecl>(D);
786   if (!MD || !MD->getDeclName().getAsIdentifierInfo() ||
787       !MD->getDeclName().getAsIdentifierInfo()->isStr("allocate") ||
788       (MD->getNumParams() != 1 && MD->getNumParams() != 2))
789     return false;
790 
791   if (MD->parameters()[0]->getType().getCanonicalType() != Ctx.getSizeType())
792     return false;
793 
794   if (MD->getNumParams() == 2) {
795     auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>();
796     if (!PT || !PT->isVoidPointerType() ||
797         !PT->getPointeeType().isConstQualified())
798       return false;
799   }
800 
801   return true;
802 }
803 
804 /// Return the UBSan prologue signature for \p FD if one is available.
805 static llvm::Constant *getPrologueSignature(CodeGenModule &CGM,
806                                             const FunctionDecl *FD) {
807   if (const auto *MD = dyn_cast<CXXMethodDecl>(FD))
808     if (!MD->isStatic())
809       return nullptr;
810   return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM);
811 }
812 
813 void CodeGenFunction::StartFunction(GlobalDecl GD,
814                                     QualType RetTy,
815                                     llvm::Function *Fn,
816                                     const CGFunctionInfo &FnInfo,
817                                     const FunctionArgList &Args,
818                                     SourceLocation Loc,
819                                     SourceLocation StartLoc) {
820   assert(!CurFn &&
821          "Do not use a CodeGenFunction object for more than one function");
822 
823   const Decl *D = GD.getDecl();
824 
825   DidCallStackSave = false;
826   CurCodeDecl = D;
827   if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
828     if (FD->usesSEHTry())
829       CurSEHParent = FD;
830   CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
831   FnRetTy = RetTy;
832   CurFn = Fn;
833   CurFnInfo = &FnInfo;
834   assert(CurFn->isDeclaration() && "Function already has body?");
835 
836   // If this function has been blacklisted for any of the enabled sanitizers,
837   // disable the sanitizer for the function.
838   do {
839 #define SANITIZER(NAME, ID)                                                    \
840   if (SanOpts.empty())                                                         \
841     break;                                                                     \
842   if (SanOpts.has(SanitizerKind::ID))                                          \
843     if (CGM.isInSanitizerBlacklist(SanitizerKind::ID, Fn, Loc))                \
844       SanOpts.set(SanitizerKind::ID, false);
845 
846 #include "clang/Basic/Sanitizers.def"
847 #undef SANITIZER
848   } while (0);
849 
850   if (D) {
851     // Apply the no_sanitize* attributes to SanOpts.
852     for (auto Attr : D->specific_attrs<NoSanitizeAttr>()) {
853       SanitizerMask mask = Attr->getMask();
854       SanOpts.Mask &= ~mask;
855       if (mask & SanitizerKind::Address)
856         SanOpts.set(SanitizerKind::KernelAddress, false);
857       if (mask & SanitizerKind::KernelAddress)
858         SanOpts.set(SanitizerKind::Address, false);
859       if (mask & SanitizerKind::HWAddress)
860         SanOpts.set(SanitizerKind::KernelHWAddress, false);
861       if (mask & SanitizerKind::KernelHWAddress)
862         SanOpts.set(SanitizerKind::HWAddress, false);
863     }
864   }
865 
866   // Apply sanitizer attributes to the function.
867   if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
868     Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
869   if (SanOpts.hasOneOf(SanitizerKind::HWAddress | SanitizerKind::KernelHWAddress))
870     Fn->addFnAttr(llvm::Attribute::SanitizeHWAddress);
871   if (SanOpts.has(SanitizerKind::Thread))
872     Fn->addFnAttr(llvm::Attribute::SanitizeThread);
873   if (SanOpts.has(SanitizerKind::Memory))
874     Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
875   if (SanOpts.has(SanitizerKind::SafeStack))
876     Fn->addFnAttr(llvm::Attribute::SafeStack);
877   if (SanOpts.has(SanitizerKind::ShadowCallStack))
878     Fn->addFnAttr(llvm::Attribute::ShadowCallStack);
879 
880   // Apply fuzzing attribute to the function.
881   if (SanOpts.hasOneOf(SanitizerKind::Fuzzer | SanitizerKind::FuzzerNoLink))
882     Fn->addFnAttr(llvm::Attribute::OptForFuzzing);
883 
884   // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
885   // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
886   if (SanOpts.has(SanitizerKind::Thread)) {
887     if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
888       IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
889       if (OMD->getMethodFamily() == OMF_dealloc ||
890           OMD->getMethodFamily() == OMF_initialize ||
891           (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
892         markAsIgnoreThreadCheckingAtRuntime(Fn);
893       }
894     } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) {
895       IdentifierInfo *II = FD->getIdentifier();
896       if (II && II->isStr("__destroy_helper_block_"))
897         markAsIgnoreThreadCheckingAtRuntime(Fn);
898     }
899   }
900 
901   // Ignore unrelated casts in STL allocate() since the allocator must cast
902   // from void* to T* before object initialization completes. Don't match on the
903   // namespace because not all allocators are in std::
904   if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
905     if (matchesStlAllocatorFn(D, getContext()))
906       SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast;
907   }
908 
909   // Apply xray attributes to the function (as a string, for now)
910   bool InstrumentXray = ShouldXRayInstrumentFunction() &&
911                         CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
912                             XRayInstrKind::Function);
913   if (D && InstrumentXray) {
914     if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
915       if (XRayAttr->alwaysXRayInstrument())
916         Fn->addFnAttr("function-instrument", "xray-always");
917       if (XRayAttr->neverXRayInstrument())
918         Fn->addFnAttr("function-instrument", "xray-never");
919       if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) {
920         Fn->addFnAttr("xray-log-args",
921                       llvm::utostr(LogArgs->getArgumentCount()));
922       }
923     } else {
924       if (!CGM.imbueXRayAttrs(Fn, Loc))
925         Fn->addFnAttr(
926             "xray-instruction-threshold",
927             llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
928     }
929   }
930 
931   // Add no-jump-tables value.
932   Fn->addFnAttr("no-jump-tables",
933                 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
934 
935   // Add profile-sample-accurate value.
936   if (CGM.getCodeGenOpts().ProfileSampleAccurate)
937     Fn->addFnAttr("profile-sample-accurate");
938 
939   if (getLangOpts().OpenCL) {
940     // Add metadata for a kernel function.
941     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
942       EmitOpenCLKernelMetadata(FD, Fn);
943   }
944 
945   // If we are checking function types, emit a function type signature as
946   // prologue data.
947   if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
948     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
949       if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
950         // Remove any (C++17) exception specifications, to allow calling e.g. a
951         // noexcept function through a non-noexcept pointer.
952         auto ProtoTy =
953           getContext().getFunctionTypeWithExceptionSpec(FD->getType(),
954                                                         EST_None);
955         llvm::Constant *FTRTTIConst =
956             CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true);
957         llvm::Constant *FTRTTIConstEncoded =
958             EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
959         llvm::Constant *PrologueStructElems[] = {PrologueSig,
960                                                  FTRTTIConstEncoded};
961         llvm::Constant *PrologueStructConst =
962             llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
963         Fn->setPrologueData(PrologueStructConst);
964       }
965     }
966   }
967 
968   // If we're checking nullability, we need to know whether we can check the
969   // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
970   if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
971     auto Nullability = FnRetTy->getNullability(getContext());
972     if (Nullability && *Nullability == NullabilityKind::NonNull) {
973       if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
974             CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
975         RetValNullabilityPrecondition =
976             llvm::ConstantInt::getTrue(getLLVMContext());
977     }
978   }
979 
980   // If we're in C++ mode and the function name is "main", it is guaranteed
981   // to be norecurse by the standard (3.6.1.3 "The function main shall not be
982   // used within a program").
983   if (getLangOpts().CPlusPlus)
984     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
985       if (FD->isMain())
986         Fn->addFnAttr(llvm::Attribute::NoRecurse);
987 
988   llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
989 
990   // Create a marker to make it easy to insert allocas into the entryblock
991   // later.  Don't create this with the builder, because we don't want it
992   // folded.
993   llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
994   AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
995 
996   ReturnBlock = getJumpDestInCurrentScope("return");
997 
998   Builder.SetInsertPoint(EntryBB);
999 
1000   // If we're checking the return value, allocate space for a pointer to a
1001   // precise source location of the checked return statement.
1002   if (requiresReturnValueCheck()) {
1003     ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
1004     InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
1005   }
1006 
1007   // Emit subprogram debug descriptor.
1008   if (CGDebugInfo *DI = getDebugInfo()) {
1009     // Reconstruct the type from the argument list so that implicit parameters,
1010     // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
1011     // convention.
1012     CallingConv CC = CallingConv::CC_C;
1013     if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
1014       if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
1015         CC = SrcFnTy->getCallConv();
1016     SmallVector<QualType, 16> ArgTypes;
1017     for (const VarDecl *VD : Args)
1018       ArgTypes.push_back(VD->getType());
1019     QualType FnType = getContext().getFunctionType(
1020         RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
1021     DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, CurFuncIsThunk,
1022                           Builder);
1023   }
1024 
1025   if (ShouldInstrumentFunction()) {
1026     if (CGM.getCodeGenOpts().InstrumentFunctions)
1027       CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
1028     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
1029       CurFn->addFnAttr("instrument-function-entry-inlined",
1030                        "__cyg_profile_func_enter");
1031     if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
1032       CurFn->addFnAttr("instrument-function-entry-inlined",
1033                        "__cyg_profile_func_enter_bare");
1034   }
1035 
1036   // Since emitting the mcount call here impacts optimizations such as function
1037   // inlining, we just add an attribute to insert a mcount call in backend.
1038   // The attribute "counting-function" is set to mcount function name which is
1039   // architecture dependent.
1040   if (CGM.getCodeGenOpts().InstrumentForProfiling) {
1041     // Calls to fentry/mcount should not be generated if function has
1042     // the no_instrument_function attribute.
1043     if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
1044       if (CGM.getCodeGenOpts().CallFEntry)
1045         Fn->addFnAttr("fentry-call", "true");
1046       else {
1047         Fn->addFnAttr("instrument-function-entry-inlined",
1048                       getTarget().getMCountName());
1049       }
1050     }
1051   }
1052 
1053   if (RetTy->isVoidType()) {
1054     // Void type; nothing to return.
1055     ReturnValue = Address::invalid();
1056 
1057     // Count the implicit return.
1058     if (!endsWithReturn(D))
1059       ++NumReturnExprs;
1060   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
1061              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1062     // Indirect aggregate return; emit returned value directly into sret slot.
1063     // This reduces code size, and affects correctness in C++.
1064     auto AI = CurFn->arg_begin();
1065     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
1066       ++AI;
1067     ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
1068   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
1069              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1070     // Load the sret pointer from the argument struct and return into that.
1071     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
1072     llvm::Function::arg_iterator EI = CurFn->arg_end();
1073     --EI;
1074     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
1075     Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
1076     ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
1077   } else {
1078     ReturnValue = CreateIRTemp(RetTy, "retval");
1079 
1080     // Tell the epilog emitter to autorelease the result.  We do this
1081     // now so that various specialized functions can suppress it
1082     // during their IR-generation.
1083     if (getLangOpts().ObjCAutoRefCount &&
1084         !CurFnInfo->isReturnsRetained() &&
1085         RetTy->isObjCRetainableType())
1086       AutoreleaseResult = true;
1087   }
1088 
1089   EmitStartEHSpec(CurCodeDecl);
1090 
1091   PrologueCleanupDepth = EHStack.stable_begin();
1092 
1093   // Emit OpenMP specific initialization of the device functions.
1094   if (getLangOpts().OpenMP && CurCodeDecl)
1095     CGM.getOpenMPRuntime().emitFunctionProlog(*this, CurCodeDecl);
1096 
1097   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
1098 
1099   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
1100     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
1101     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
1102     if (MD->getParent()->isLambda() &&
1103         MD->getOverloadedOperator() == OO_Call) {
1104       // We're in a lambda; figure out the captures.
1105       MD->getParent()->getCaptureFields(LambdaCaptureFields,
1106                                         LambdaThisCaptureField);
1107       if (LambdaThisCaptureField) {
1108         // If the lambda captures the object referred to by '*this' - either by
1109         // value or by reference, make sure CXXThisValue points to the correct
1110         // object.
1111 
1112         // Get the lvalue for the field (which is a copy of the enclosing object
1113         // or contains the address of the enclosing object).
1114         LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
1115         if (!LambdaThisCaptureField->getType()->isPointerType()) {
1116           // If the enclosing object was captured by value, just use its address.
1117           CXXThisValue = ThisFieldLValue.getAddress().getPointer();
1118         } else {
1119           // Load the lvalue pointed to by the field, since '*this' was captured
1120           // by reference.
1121           CXXThisValue =
1122               EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
1123         }
1124       }
1125       for (auto *FD : MD->getParent()->fields()) {
1126         if (FD->hasCapturedVLAType()) {
1127           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
1128                                            SourceLocation()).getScalarVal();
1129           auto VAT = FD->getCapturedVLAType();
1130           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
1131         }
1132       }
1133     } else {
1134       // Not in a lambda; just use 'this' from the method.
1135       // FIXME: Should we generate a new load for each use of 'this'?  The
1136       // fast register allocator would be happier...
1137       CXXThisValue = CXXABIThisValue;
1138     }
1139 
1140     // Check the 'this' pointer once per function, if it's available.
1141     if (CXXABIThisValue) {
1142       SanitizerSet SkippedChecks;
1143       SkippedChecks.set(SanitizerKind::ObjectSize, true);
1144       QualType ThisTy = MD->getThisType(getContext());
1145 
1146       // If this is the call operator of a lambda with no capture-default, it
1147       // may have a static invoker function, which may call this operator with
1148       // a null 'this' pointer.
1149       if (isLambdaCallOperator(MD) &&
1150           MD->getParent()->getLambdaCaptureDefault() == LCD_None)
1151         SkippedChecks.set(SanitizerKind::Null, true);
1152 
1153       EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1154                                                 : TCK_MemberCall,
1155                     Loc, CXXABIThisValue, ThisTy,
1156                     getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1157                     SkippedChecks);
1158     }
1159   }
1160 
1161   // If any of the arguments have a variably modified type, make sure to
1162   // emit the type size.
1163   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1164        i != e; ++i) {
1165     const VarDecl *VD = *i;
1166 
1167     // Dig out the type as written from ParmVarDecls; it's unclear whether
1168     // the standard (C99 6.9.1p10) requires this, but we're following the
1169     // precedent set by gcc.
1170     QualType Ty;
1171     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1172       Ty = PVD->getOriginalType();
1173     else
1174       Ty = VD->getType();
1175 
1176     if (Ty->isVariablyModifiedType())
1177       EmitVariablyModifiedType(Ty);
1178   }
1179   // Emit a location at the end of the prologue.
1180   if (CGDebugInfo *DI = getDebugInfo())
1181     DI->EmitLocation(Builder, StartLoc);
1182 }
1183 
1184 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1185                                        const Stmt *Body) {
1186   incrementProfileCounter(Body);
1187   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1188     EmitCompoundStmtWithoutScope(*S);
1189   else
1190     EmitStmt(Body);
1191 }
1192 
1193 /// When instrumenting to collect profile data, the counts for some blocks
1194 /// such as switch cases need to not include the fall-through counts, so
1195 /// emit a branch around the instrumentation code. When not instrumenting,
1196 /// this just calls EmitBlock().
1197 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1198                                                const Stmt *S) {
1199   llvm::BasicBlock *SkipCountBB = nullptr;
1200   if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1201     // When instrumenting for profiling, the fallthrough to certain
1202     // statements needs to skip over the instrumentation code so that we
1203     // get an accurate count.
1204     SkipCountBB = createBasicBlock("skipcount");
1205     EmitBranch(SkipCountBB);
1206   }
1207   EmitBlock(BB);
1208   uint64_t CurrentCount = getCurrentProfileCount();
1209   incrementProfileCounter(S);
1210   setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1211   if (SkipCountBB)
1212     EmitBlock(SkipCountBB);
1213 }
1214 
1215 /// Tries to mark the given function nounwind based on the
1216 /// non-existence of any throwing calls within it.  We believe this is
1217 /// lightweight enough to do at -O0.
1218 static void TryMarkNoThrow(llvm::Function *F) {
1219   // LLVM treats 'nounwind' on a function as part of the type, so we
1220   // can't do this on functions that can be overwritten.
1221   if (F->isInterposable()) return;
1222 
1223   for (llvm::BasicBlock &BB : *F)
1224     for (llvm::Instruction &I : BB)
1225       if (I.mayThrow())
1226         return;
1227 
1228   F->setDoesNotThrow();
1229 }
1230 
1231 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1232                                                FunctionArgList &Args) {
1233   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1234   QualType ResTy = FD->getReturnType();
1235 
1236   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1237   if (MD && MD->isInstance()) {
1238     if (CGM.getCXXABI().HasThisReturn(GD))
1239       ResTy = MD->getThisType(getContext());
1240     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1241       ResTy = CGM.getContext().VoidPtrTy;
1242     CGM.getCXXABI().buildThisParam(*this, Args);
1243   }
1244 
1245   // The base version of an inheriting constructor whose constructed base is a
1246   // virtual base is not passed any arguments (because it doesn't actually call
1247   // the inherited constructor).
1248   bool PassedParams = true;
1249   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1250     if (auto Inherited = CD->getInheritedConstructor())
1251       PassedParams =
1252           getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1253 
1254   if (PassedParams) {
1255     for (auto *Param : FD->parameters()) {
1256       Args.push_back(Param);
1257       if (!Param->hasAttr<PassObjectSizeAttr>())
1258         continue;
1259 
1260       auto *Implicit = ImplicitParamDecl::Create(
1261           getContext(), Param->getDeclContext(), Param->getLocation(),
1262           /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1263       SizeArguments[Param] = Implicit;
1264       Args.push_back(Implicit);
1265     }
1266   }
1267 
1268   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1269     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1270 
1271   return ResTy;
1272 }
1273 
1274 static bool
1275 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1276                                              const ASTContext &Context) {
1277   QualType T = FD->getReturnType();
1278   // Avoid the optimization for functions that return a record type with a
1279   // trivial destructor or another trivially copyable type.
1280   if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1281     if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1282       return !ClassDecl->hasTrivialDestructor();
1283   }
1284   return !T.isTriviallyCopyableType(Context);
1285 }
1286 
1287 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1288                                    const CGFunctionInfo &FnInfo) {
1289   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1290   CurGD = GD;
1291 
1292   FunctionArgList Args;
1293   QualType ResTy = BuildFunctionArgList(GD, Args);
1294 
1295   // Check if we should generate debug info for this function.
1296   if (FD->hasAttr<NoDebugAttr>())
1297     DebugInfo = nullptr; // disable debug info indefinitely for this function
1298 
1299   // The function might not have a body if we're generating thunks for a
1300   // function declaration.
1301   SourceRange BodyRange;
1302   if (Stmt *Body = FD->getBody())
1303     BodyRange = Body->getSourceRange();
1304   else
1305     BodyRange = FD->getLocation();
1306   CurEHLocation = BodyRange.getEnd();
1307 
1308   // Use the location of the start of the function to determine where
1309   // the function definition is located. By default use the location
1310   // of the declaration as the location for the subprogram. A function
1311   // may lack a declaration in the source code if it is created by code
1312   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1313   SourceLocation Loc = FD->getLocation();
1314 
1315   // If this is a function specialization then use the pattern body
1316   // as the location for the function.
1317   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1318     if (SpecDecl->hasBody(SpecDecl))
1319       Loc = SpecDecl->getLocation();
1320 
1321   Stmt *Body = FD->getBody();
1322 
1323   // Initialize helper which will detect jumps which can cause invalid lifetime
1324   // markers.
1325   if (Body && ShouldEmitLifetimeMarkers)
1326     Bypasses.Init(Body);
1327 
1328   // Emit the standard function prologue.
1329   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1330 
1331   // Generate the body of the function.
1332   PGO.assignRegionCounters(GD, CurFn);
1333   if (isa<CXXDestructorDecl>(FD))
1334     EmitDestructorBody(Args);
1335   else if (isa<CXXConstructorDecl>(FD))
1336     EmitConstructorBody(Args);
1337   else if (getLangOpts().CUDA &&
1338            !getLangOpts().CUDAIsDevice &&
1339            FD->hasAttr<CUDAGlobalAttr>())
1340     CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1341   else if (isa<CXXMethodDecl>(FD) &&
1342            cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1343     // The lambda static invoker function is special, because it forwards or
1344     // clones the body of the function call operator (but is actually static).
1345     EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1346   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1347              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1348               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1349     // Implicit copy-assignment gets the same special treatment as implicit
1350     // copy-constructors.
1351     emitImplicitAssignmentOperatorBody(Args);
1352   } else if (Body) {
1353     EmitFunctionBody(Args, Body);
1354   } else
1355     llvm_unreachable("no definition for emitted function");
1356 
1357   // C++11 [stmt.return]p2:
1358   //   Flowing off the end of a function [...] results in undefined behavior in
1359   //   a value-returning function.
1360   // C11 6.9.1p12:
1361   //   If the '}' that terminates a function is reached, and the value of the
1362   //   function call is used by the caller, the behavior is undefined.
1363   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1364       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1365     bool ShouldEmitUnreachable =
1366         CGM.getCodeGenOpts().StrictReturn ||
1367         shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1368     if (SanOpts.has(SanitizerKind::Return)) {
1369       SanitizerScope SanScope(this);
1370       llvm::Value *IsFalse = Builder.getFalse();
1371       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1372                 SanitizerHandler::MissingReturn,
1373                 EmitCheckSourceLocation(FD->getLocation()), None);
1374     } else if (ShouldEmitUnreachable) {
1375       if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1376         EmitTrapCall(llvm::Intrinsic::trap);
1377     }
1378     if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1379       Builder.CreateUnreachable();
1380       Builder.ClearInsertionPoint();
1381     }
1382   }
1383 
1384   // Emit the standard function epilogue.
1385   FinishFunction(BodyRange.getEnd());
1386 
1387   // If we haven't marked the function nothrow through other means, do
1388   // a quick pass now to see if we can.
1389   if (!CurFn->doesNotThrow())
1390     TryMarkNoThrow(CurFn);
1391 }
1392 
1393 /// ContainsLabel - Return true if the statement contains a label in it.  If
1394 /// this statement is not executed normally, it not containing a label means
1395 /// that we can just remove the code.
1396 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1397   // Null statement, not a label!
1398   if (!S) return false;
1399 
1400   // If this is a label, we have to emit the code, consider something like:
1401   // if (0) {  ...  foo:  bar(); }  goto foo;
1402   //
1403   // TODO: If anyone cared, we could track __label__'s, since we know that you
1404   // can't jump to one from outside their declared region.
1405   if (isa<LabelStmt>(S))
1406     return true;
1407 
1408   // If this is a case/default statement, and we haven't seen a switch, we have
1409   // to emit the code.
1410   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1411     return true;
1412 
1413   // If this is a switch statement, we want to ignore cases below it.
1414   if (isa<SwitchStmt>(S))
1415     IgnoreCaseStmts = true;
1416 
1417   // Scan subexpressions for verboten labels.
1418   for (const Stmt *SubStmt : S->children())
1419     if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1420       return true;
1421 
1422   return false;
1423 }
1424 
1425 /// containsBreak - Return true if the statement contains a break out of it.
1426 /// If the statement (recursively) contains a switch or loop with a break
1427 /// inside of it, this is fine.
1428 bool CodeGenFunction::containsBreak(const Stmt *S) {
1429   // Null statement, not a label!
1430   if (!S) return false;
1431 
1432   // If this is a switch or loop that defines its own break scope, then we can
1433   // include it and anything inside of it.
1434   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1435       isa<ForStmt>(S))
1436     return false;
1437 
1438   if (isa<BreakStmt>(S))
1439     return true;
1440 
1441   // Scan subexpressions for verboten breaks.
1442   for (const Stmt *SubStmt : S->children())
1443     if (containsBreak(SubStmt))
1444       return true;
1445 
1446   return false;
1447 }
1448 
1449 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1450   if (!S) return false;
1451 
1452   // Some statement kinds add a scope and thus never add a decl to the current
1453   // scope. Note, this list is longer than the list of statements that might
1454   // have an unscoped decl nested within them, but this way is conservatively
1455   // correct even if more statement kinds are added.
1456   if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1457       isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1458       isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1459       isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1460     return false;
1461 
1462   if (isa<DeclStmt>(S))
1463     return true;
1464 
1465   for (const Stmt *SubStmt : S->children())
1466     if (mightAddDeclToScope(SubStmt))
1467       return true;
1468 
1469   return false;
1470 }
1471 
1472 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1473 /// to a constant, or if it does but contains a label, return false.  If it
1474 /// constant folds return true and set the boolean result in Result.
1475 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1476                                                    bool &ResultBool,
1477                                                    bool AllowLabels) {
1478   llvm::APSInt ResultInt;
1479   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1480     return false;
1481 
1482   ResultBool = ResultInt.getBoolValue();
1483   return true;
1484 }
1485 
1486 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1487 /// to a constant, or if it does but contains a label, return false.  If it
1488 /// constant folds return true and set the folded value.
1489 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1490                                                    llvm::APSInt &ResultInt,
1491                                                    bool AllowLabels) {
1492   // FIXME: Rename and handle conversion of other evaluatable things
1493   // to bool.
1494   llvm::APSInt Int;
1495   if (!Cond->EvaluateAsInt(Int, getContext()))
1496     return false;  // Not foldable, not integer or not fully evaluatable.
1497 
1498   if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1499     return false;  // Contains a label.
1500 
1501   ResultInt = Int;
1502   return true;
1503 }
1504 
1505 
1506 
1507 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1508 /// statement) to the specified blocks.  Based on the condition, this might try
1509 /// to simplify the codegen of the conditional based on the branch.
1510 ///
1511 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1512                                            llvm::BasicBlock *TrueBlock,
1513                                            llvm::BasicBlock *FalseBlock,
1514                                            uint64_t TrueCount) {
1515   Cond = Cond->IgnoreParens();
1516 
1517   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1518 
1519     // Handle X && Y in a condition.
1520     if (CondBOp->getOpcode() == BO_LAnd) {
1521       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1522       // folded if the case was simple enough.
1523       bool ConstantBool = false;
1524       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1525           ConstantBool) {
1526         // br(1 && X) -> br(X).
1527         incrementProfileCounter(CondBOp);
1528         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1529                                     TrueCount);
1530       }
1531 
1532       // If we have "X && 1", simplify the code to use an uncond branch.
1533       // "X && 0" would have been constant folded to 0.
1534       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1535           ConstantBool) {
1536         // br(X && 1) -> br(X).
1537         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1538                                     TrueCount);
1539       }
1540 
1541       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1542       // want to jump to the FalseBlock.
1543       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1544       // The counter tells us how often we evaluate RHS, and all of TrueCount
1545       // can be propagated to that branch.
1546       uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1547 
1548       ConditionalEvaluation eval(*this);
1549       {
1550         ApplyDebugLocation DL(*this, Cond);
1551         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1552         EmitBlock(LHSTrue);
1553       }
1554 
1555       incrementProfileCounter(CondBOp);
1556       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1557 
1558       // Any temporaries created here are conditional.
1559       eval.begin(*this);
1560       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1561       eval.end(*this);
1562 
1563       return;
1564     }
1565 
1566     if (CondBOp->getOpcode() == BO_LOr) {
1567       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1568       // folded if the case was simple enough.
1569       bool ConstantBool = false;
1570       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1571           !ConstantBool) {
1572         // br(0 || X) -> br(X).
1573         incrementProfileCounter(CondBOp);
1574         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1575                                     TrueCount);
1576       }
1577 
1578       // If we have "X || 0", simplify the code to use an uncond branch.
1579       // "X || 1" would have been constant folded to 1.
1580       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1581           !ConstantBool) {
1582         // br(X || 0) -> br(X).
1583         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1584                                     TrueCount);
1585       }
1586 
1587       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1588       // want to jump to the TrueBlock.
1589       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1590       // We have the count for entry to the RHS and for the whole expression
1591       // being true, so we can divy up True count between the short circuit and
1592       // the RHS.
1593       uint64_t LHSCount =
1594           getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1595       uint64_t RHSCount = TrueCount - LHSCount;
1596 
1597       ConditionalEvaluation eval(*this);
1598       {
1599         ApplyDebugLocation DL(*this, Cond);
1600         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1601         EmitBlock(LHSFalse);
1602       }
1603 
1604       incrementProfileCounter(CondBOp);
1605       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1606 
1607       // Any temporaries created here are conditional.
1608       eval.begin(*this);
1609       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1610 
1611       eval.end(*this);
1612 
1613       return;
1614     }
1615   }
1616 
1617   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1618     // br(!x, t, f) -> br(x, f, t)
1619     if (CondUOp->getOpcode() == UO_LNot) {
1620       // Negate the count.
1621       uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1622       // Negate the condition and swap the destination blocks.
1623       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1624                                   FalseCount);
1625     }
1626   }
1627 
1628   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1629     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1630     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1631     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1632 
1633     ConditionalEvaluation cond(*this);
1634     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1635                          getProfileCount(CondOp));
1636 
1637     // When computing PGO branch weights, we only know the overall count for
1638     // the true block. This code is essentially doing tail duplication of the
1639     // naive code-gen, introducing new edges for which counts are not
1640     // available. Divide the counts proportionally between the LHS and RHS of
1641     // the conditional operator.
1642     uint64_t LHSScaledTrueCount = 0;
1643     if (TrueCount) {
1644       double LHSRatio =
1645           getProfileCount(CondOp) / (double)getCurrentProfileCount();
1646       LHSScaledTrueCount = TrueCount * LHSRatio;
1647     }
1648 
1649     cond.begin(*this);
1650     EmitBlock(LHSBlock);
1651     incrementProfileCounter(CondOp);
1652     {
1653       ApplyDebugLocation DL(*this, Cond);
1654       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1655                            LHSScaledTrueCount);
1656     }
1657     cond.end(*this);
1658 
1659     cond.begin(*this);
1660     EmitBlock(RHSBlock);
1661     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1662                          TrueCount - LHSScaledTrueCount);
1663     cond.end(*this);
1664 
1665     return;
1666   }
1667 
1668   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1669     // Conditional operator handling can give us a throw expression as a
1670     // condition for a case like:
1671     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1672     // Fold this to:
1673     //   br(c, throw x, br(y, t, f))
1674     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1675     return;
1676   }
1677 
1678   // If the branch has a condition wrapped by __builtin_unpredictable,
1679   // create metadata that specifies that the branch is unpredictable.
1680   // Don't bother if not optimizing because that metadata would not be used.
1681   llvm::MDNode *Unpredictable = nullptr;
1682   auto *Call = dyn_cast<CallExpr>(Cond);
1683   if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1684     auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1685     if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1686       llvm::MDBuilder MDHelper(getLLVMContext());
1687       Unpredictable = MDHelper.createUnpredictable();
1688     }
1689   }
1690 
1691   // Create branch weights based on the number of times we get here and the
1692   // number of times the condition should be true.
1693   uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1694   llvm::MDNode *Weights =
1695       createProfileWeights(TrueCount, CurrentCount - TrueCount);
1696 
1697   // Emit the code with the fully general case.
1698   llvm::Value *CondV;
1699   {
1700     ApplyDebugLocation DL(*this, Cond);
1701     CondV = EvaluateExprAsBool(Cond);
1702   }
1703   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1704 }
1705 
1706 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1707 /// specified stmt yet.
1708 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1709   CGM.ErrorUnsupported(S, Type);
1710 }
1711 
1712 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1713 /// variable-length array whose elements have a non-zero bit-pattern.
1714 ///
1715 /// \param baseType the inner-most element type of the array
1716 /// \param src - a char* pointing to the bit-pattern for a single
1717 /// base element of the array
1718 /// \param sizeInChars - the total size of the VLA, in chars
1719 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1720                                Address dest, Address src,
1721                                llvm::Value *sizeInChars) {
1722   CGBuilderTy &Builder = CGF.Builder;
1723 
1724   CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1725   llvm::Value *baseSizeInChars
1726     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1727 
1728   Address begin =
1729     Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1730   llvm::Value *end =
1731     Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1732 
1733   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1734   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1735   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1736 
1737   // Make a loop over the VLA.  C99 guarantees that the VLA element
1738   // count must be nonzero.
1739   CGF.EmitBlock(loopBB);
1740 
1741   llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1742   cur->addIncoming(begin.getPointer(), originBB);
1743 
1744   CharUnits curAlign =
1745     dest.getAlignment().alignmentOfArrayElement(baseSize);
1746 
1747   // memcpy the individual element bit-pattern.
1748   Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1749                        /*volatile*/ false);
1750 
1751   // Go to the next element.
1752   llvm::Value *next =
1753     Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1754 
1755   // Leave if that's the end of the VLA.
1756   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1757   Builder.CreateCondBr(done, contBB, loopBB);
1758   cur->addIncoming(next, loopBB);
1759 
1760   CGF.EmitBlock(contBB);
1761 }
1762 
1763 void
1764 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1765   // Ignore empty classes in C++.
1766   if (getLangOpts().CPlusPlus) {
1767     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1768       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1769         return;
1770     }
1771   }
1772 
1773   // Cast the dest ptr to the appropriate i8 pointer type.
1774   if (DestPtr.getElementType() != Int8Ty)
1775     DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1776 
1777   // Get size and alignment info for this aggregate.
1778   CharUnits size = getContext().getTypeSizeInChars(Ty);
1779 
1780   llvm::Value *SizeVal;
1781   const VariableArrayType *vla;
1782 
1783   // Don't bother emitting a zero-byte memset.
1784   if (size.isZero()) {
1785     // But note that getTypeInfo returns 0 for a VLA.
1786     if (const VariableArrayType *vlaType =
1787           dyn_cast_or_null<VariableArrayType>(
1788                                           getContext().getAsArrayType(Ty))) {
1789       auto VlaSize = getVLASize(vlaType);
1790       SizeVal = VlaSize.NumElts;
1791       CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type);
1792       if (!eltSize.isOne())
1793         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1794       vla = vlaType;
1795     } else {
1796       return;
1797     }
1798   } else {
1799     SizeVal = CGM.getSize(size);
1800     vla = nullptr;
1801   }
1802 
1803   // If the type contains a pointer to data member we can't memset it to zero.
1804   // Instead, create a null constant and copy it to the destination.
1805   // TODO: there are other patterns besides zero that we can usefully memset,
1806   // like -1, which happens to be the pattern used by member-pointers.
1807   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1808     // For a VLA, emit a single element, then splat that over the VLA.
1809     if (vla) Ty = getContext().getBaseElementType(vla);
1810 
1811     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1812 
1813     llvm::GlobalVariable *NullVariable =
1814       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1815                                /*isConstant=*/true,
1816                                llvm::GlobalVariable::PrivateLinkage,
1817                                NullConstant, Twine());
1818     CharUnits NullAlign = DestPtr.getAlignment();
1819     NullVariable->setAlignment(NullAlign.getQuantity());
1820     Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1821                    NullAlign);
1822 
1823     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1824 
1825     // Get and call the appropriate llvm.memcpy overload.
1826     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1827     return;
1828   }
1829 
1830   // Otherwise, just memset the whole thing to zero.  This is legal
1831   // because in LLVM, all default initializers (other than the ones we just
1832   // handled above) are guaranteed to have a bit pattern of all zeros.
1833   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1834 }
1835 
1836 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1837   // Make sure that there is a block for the indirect goto.
1838   if (!IndirectBranch)
1839     GetIndirectGotoBlock();
1840 
1841   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1842 
1843   // Make sure the indirect branch includes all of the address-taken blocks.
1844   IndirectBranch->addDestination(BB);
1845   return llvm::BlockAddress::get(CurFn, BB);
1846 }
1847 
1848 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1849   // If we already made the indirect branch for indirect goto, return its block.
1850   if (IndirectBranch) return IndirectBranch->getParent();
1851 
1852   CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1853 
1854   // Create the PHI node that indirect gotos will add entries to.
1855   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1856                                               "indirect.goto.dest");
1857 
1858   // Create the indirect branch instruction.
1859   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1860   return IndirectBranch->getParent();
1861 }
1862 
1863 /// Computes the length of an array in elements, as well as the base
1864 /// element type and a properly-typed first element pointer.
1865 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1866                                               QualType &baseType,
1867                                               Address &addr) {
1868   const ArrayType *arrayType = origArrayType;
1869 
1870   // If it's a VLA, we have to load the stored size.  Note that
1871   // this is the size of the VLA in bytes, not its size in elements.
1872   llvm::Value *numVLAElements = nullptr;
1873   if (isa<VariableArrayType>(arrayType)) {
1874     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts;
1875 
1876     // Walk into all VLAs.  This doesn't require changes to addr,
1877     // which has type T* where T is the first non-VLA element type.
1878     do {
1879       QualType elementType = arrayType->getElementType();
1880       arrayType = getContext().getAsArrayType(elementType);
1881 
1882       // If we only have VLA components, 'addr' requires no adjustment.
1883       if (!arrayType) {
1884         baseType = elementType;
1885         return numVLAElements;
1886       }
1887     } while (isa<VariableArrayType>(arrayType));
1888 
1889     // We get out here only if we find a constant array type
1890     // inside the VLA.
1891   }
1892 
1893   // We have some number of constant-length arrays, so addr should
1894   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1895   // down to the first element of addr.
1896   SmallVector<llvm::Value*, 8> gepIndices;
1897 
1898   // GEP down to the array type.
1899   llvm::ConstantInt *zero = Builder.getInt32(0);
1900   gepIndices.push_back(zero);
1901 
1902   uint64_t countFromCLAs = 1;
1903   QualType eltType;
1904 
1905   llvm::ArrayType *llvmArrayType =
1906     dyn_cast<llvm::ArrayType>(addr.getElementType());
1907   while (llvmArrayType) {
1908     assert(isa<ConstantArrayType>(arrayType));
1909     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1910              == llvmArrayType->getNumElements());
1911 
1912     gepIndices.push_back(zero);
1913     countFromCLAs *= llvmArrayType->getNumElements();
1914     eltType = arrayType->getElementType();
1915 
1916     llvmArrayType =
1917       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1918     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1919     assert((!llvmArrayType || arrayType) &&
1920            "LLVM and Clang types are out-of-synch");
1921   }
1922 
1923   if (arrayType) {
1924     // From this point onwards, the Clang array type has been emitted
1925     // as some other type (probably a packed struct). Compute the array
1926     // size, and just emit the 'begin' expression as a bitcast.
1927     while (arrayType) {
1928       countFromCLAs *=
1929           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1930       eltType = arrayType->getElementType();
1931       arrayType = getContext().getAsArrayType(eltType);
1932     }
1933 
1934     llvm::Type *baseType = ConvertType(eltType);
1935     addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1936   } else {
1937     // Create the actual GEP.
1938     addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1939                                              gepIndices, "array.begin"),
1940                    addr.getAlignment());
1941   }
1942 
1943   baseType = eltType;
1944 
1945   llvm::Value *numElements
1946     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1947 
1948   // If we had any VLA dimensions, factor them in.
1949   if (numVLAElements)
1950     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1951 
1952   return numElements;
1953 }
1954 
1955 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) {
1956   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1957   assert(vla && "type was not a variable array type!");
1958   return getVLASize(vla);
1959 }
1960 
1961 CodeGenFunction::VlaSizePair
1962 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1963   // The number of elements so far; always size_t.
1964   llvm::Value *numElements = nullptr;
1965 
1966   QualType elementType;
1967   do {
1968     elementType = type->getElementType();
1969     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1970     assert(vlaSize && "no size for VLA!");
1971     assert(vlaSize->getType() == SizeTy);
1972 
1973     if (!numElements) {
1974       numElements = vlaSize;
1975     } else {
1976       // It's undefined behavior if this wraps around, so mark it that way.
1977       // FIXME: Teach -fsanitize=undefined to trap this.
1978       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1979     }
1980   } while ((type = getContext().getAsVariableArrayType(elementType)));
1981 
1982   return { numElements, elementType };
1983 }
1984 
1985 CodeGenFunction::VlaSizePair
1986 CodeGenFunction::getVLAElements1D(QualType type) {
1987   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1988   assert(vla && "type was not a variable array type!");
1989   return getVLAElements1D(vla);
1990 }
1991 
1992 CodeGenFunction::VlaSizePair
1993 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) {
1994   llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()];
1995   assert(VlaSize && "no size for VLA!");
1996   assert(VlaSize->getType() == SizeTy);
1997   return { VlaSize, Vla->getElementType() };
1998 }
1999 
2000 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
2001   assert(type->isVariablyModifiedType() &&
2002          "Must pass variably modified type to EmitVLASizes!");
2003 
2004   EnsureInsertPoint();
2005 
2006   // We're going to walk down into the type and look for VLA
2007   // expressions.
2008   do {
2009     assert(type->isVariablyModifiedType());
2010 
2011     const Type *ty = type.getTypePtr();
2012     switch (ty->getTypeClass()) {
2013 
2014 #define TYPE(Class, Base)
2015 #define ABSTRACT_TYPE(Class, Base)
2016 #define NON_CANONICAL_TYPE(Class, Base)
2017 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
2018 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
2019 #include "clang/AST/TypeNodes.def"
2020       llvm_unreachable("unexpected dependent type!");
2021 
2022     // These types are never variably-modified.
2023     case Type::Builtin:
2024     case Type::Complex:
2025     case Type::Vector:
2026     case Type::ExtVector:
2027     case Type::Record:
2028     case Type::Enum:
2029     case Type::Elaborated:
2030     case Type::TemplateSpecialization:
2031     case Type::ObjCTypeParam:
2032     case Type::ObjCObject:
2033     case Type::ObjCInterface:
2034     case Type::ObjCObjectPointer:
2035       llvm_unreachable("type class is never variably-modified!");
2036 
2037     case Type::Adjusted:
2038       type = cast<AdjustedType>(ty)->getAdjustedType();
2039       break;
2040 
2041     case Type::Decayed:
2042       type = cast<DecayedType>(ty)->getPointeeType();
2043       break;
2044 
2045     case Type::Pointer:
2046       type = cast<PointerType>(ty)->getPointeeType();
2047       break;
2048 
2049     case Type::BlockPointer:
2050       type = cast<BlockPointerType>(ty)->getPointeeType();
2051       break;
2052 
2053     case Type::LValueReference:
2054     case Type::RValueReference:
2055       type = cast<ReferenceType>(ty)->getPointeeType();
2056       break;
2057 
2058     case Type::MemberPointer:
2059       type = cast<MemberPointerType>(ty)->getPointeeType();
2060       break;
2061 
2062     case Type::ConstantArray:
2063     case Type::IncompleteArray:
2064       // Losing element qualification here is fine.
2065       type = cast<ArrayType>(ty)->getElementType();
2066       break;
2067 
2068     case Type::VariableArray: {
2069       // Losing element qualification here is fine.
2070       const VariableArrayType *vat = cast<VariableArrayType>(ty);
2071 
2072       // Unknown size indication requires no size computation.
2073       // Otherwise, evaluate and record it.
2074       if (const Expr *size = vat->getSizeExpr()) {
2075         // It's possible that we might have emitted this already,
2076         // e.g. with a typedef and a pointer to it.
2077         llvm::Value *&entry = VLASizeMap[size];
2078         if (!entry) {
2079           llvm::Value *Size = EmitScalarExpr(size);
2080 
2081           // C11 6.7.6.2p5:
2082           //   If the size is an expression that is not an integer constant
2083           //   expression [...] each time it is evaluated it shall have a value
2084           //   greater than zero.
2085           if (SanOpts.has(SanitizerKind::VLABound) &&
2086               size->getType()->isSignedIntegerType()) {
2087             SanitizerScope SanScope(this);
2088             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
2089             llvm::Constant *StaticArgs[] = {
2090               EmitCheckSourceLocation(size->getLocStart()),
2091               EmitCheckTypeDescriptor(size->getType())
2092             };
2093             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
2094                                      SanitizerKind::VLABound),
2095                       SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
2096           }
2097 
2098           // Always zexting here would be wrong if it weren't
2099           // undefined behavior to have a negative bound.
2100           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
2101         }
2102       }
2103       type = vat->getElementType();
2104       break;
2105     }
2106 
2107     case Type::FunctionProto:
2108     case Type::FunctionNoProto:
2109       type = cast<FunctionType>(ty)->getReturnType();
2110       break;
2111 
2112     case Type::Paren:
2113     case Type::TypeOf:
2114     case Type::UnaryTransform:
2115     case Type::Attributed:
2116     case Type::SubstTemplateTypeParm:
2117     case Type::PackExpansion:
2118       // Keep walking after single level desugaring.
2119       type = type.getSingleStepDesugaredType(getContext());
2120       break;
2121 
2122     case Type::Typedef:
2123     case Type::Decltype:
2124     case Type::Auto:
2125     case Type::DeducedTemplateSpecialization:
2126       // Stop walking: nothing to do.
2127       return;
2128 
2129     case Type::TypeOfExpr:
2130       // Stop walking: emit typeof expression.
2131       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
2132       return;
2133 
2134     case Type::Atomic:
2135       type = cast<AtomicType>(ty)->getValueType();
2136       break;
2137 
2138     case Type::Pipe:
2139       type = cast<PipeType>(ty)->getElementType();
2140       break;
2141     }
2142   } while (type->isVariablyModifiedType());
2143 }
2144 
2145 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
2146   if (getContext().getBuiltinVaListType()->isArrayType())
2147     return EmitPointerWithAlignment(E);
2148   return EmitLValue(E).getAddress();
2149 }
2150 
2151 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
2152   return EmitLValue(E).getAddress();
2153 }
2154 
2155 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2156                                               const APValue &Init) {
2157   assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
2158   if (CGDebugInfo *Dbg = getDebugInfo())
2159     if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2160       Dbg->EmitGlobalVariable(E->getDecl(), Init);
2161 }
2162 
2163 CodeGenFunction::PeepholeProtection
2164 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2165   // At the moment, the only aggressive peephole we do in IR gen
2166   // is trunc(zext) folding, but if we add more, we can easily
2167   // extend this protection.
2168 
2169   if (!rvalue.isScalar()) return PeepholeProtection();
2170   llvm::Value *value = rvalue.getScalarVal();
2171   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2172 
2173   // Just make an extra bitcast.
2174   assert(HaveInsertPoint());
2175   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2176                                                   Builder.GetInsertBlock());
2177 
2178   PeepholeProtection protection;
2179   protection.Inst = inst;
2180   return protection;
2181 }
2182 
2183 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2184   if (!protection.Inst) return;
2185 
2186   // In theory, we could try to duplicate the peepholes now, but whatever.
2187   protection.Inst->eraseFromParent();
2188 }
2189 
2190 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2191                                                  llvm::Value *AnnotatedVal,
2192                                                  StringRef AnnotationStr,
2193                                                  SourceLocation Location) {
2194   llvm::Value *Args[4] = {
2195     AnnotatedVal,
2196     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2197     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2198     CGM.EmitAnnotationLineNo(Location)
2199   };
2200   return Builder.CreateCall(AnnotationFn, Args);
2201 }
2202 
2203 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2204   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2205   // FIXME We create a new bitcast for every annotation because that's what
2206   // llvm-gcc was doing.
2207   for (const auto *I : D->specific_attrs<AnnotateAttr>())
2208     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2209                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2210                        I->getAnnotation(), D->getLocation());
2211 }
2212 
2213 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2214                                               Address Addr) {
2215   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2216   llvm::Value *V = Addr.getPointer();
2217   llvm::Type *VTy = V->getType();
2218   llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2219                                     CGM.Int8PtrTy);
2220 
2221   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2222     // FIXME Always emit the cast inst so we can differentiate between
2223     // annotation on the first field of a struct and annotation on the struct
2224     // itself.
2225     if (VTy != CGM.Int8PtrTy)
2226       V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2227     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2228     V = Builder.CreateBitCast(V, VTy);
2229   }
2230 
2231   return Address(V, Addr.getAlignment());
2232 }
2233 
2234 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2235 
2236 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2237     : CGF(CGF) {
2238   assert(!CGF->IsSanitizerScope);
2239   CGF->IsSanitizerScope = true;
2240 }
2241 
2242 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2243   CGF->IsSanitizerScope = false;
2244 }
2245 
2246 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2247                                    const llvm::Twine &Name,
2248                                    llvm::BasicBlock *BB,
2249                                    llvm::BasicBlock::iterator InsertPt) const {
2250   LoopStack.InsertHelper(I);
2251   if (IsSanitizerScope)
2252     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2253 }
2254 
2255 void CGBuilderInserter::InsertHelper(
2256     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2257     llvm::BasicBlock::iterator InsertPt) const {
2258   llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2259   if (CGF)
2260     CGF->InsertHelper(I, Name, BB, InsertPt);
2261 }
2262 
2263 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2264                                 CodeGenModule &CGM, const FunctionDecl *FD,
2265                                 std::string &FirstMissing) {
2266   // If there aren't any required features listed then go ahead and return.
2267   if (ReqFeatures.empty())
2268     return false;
2269 
2270   // Now build up the set of caller features and verify that all the required
2271   // features are there.
2272   llvm::StringMap<bool> CallerFeatureMap;
2273   CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2274 
2275   // If we have at least one of the features in the feature list return
2276   // true, otherwise return false.
2277   return std::all_of(
2278       ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2279         SmallVector<StringRef, 1> OrFeatures;
2280         Feature.split(OrFeatures, "|");
2281         return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2282                            [&](StringRef Feature) {
2283                              if (!CallerFeatureMap.lookup(Feature)) {
2284                                FirstMissing = Feature.str();
2285                                return false;
2286                              }
2287                              return true;
2288                            });
2289       });
2290 }
2291 
2292 // Emits an error if we don't have a valid set of target features for the
2293 // called function.
2294 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2295                                           const FunctionDecl *TargetDecl) {
2296   // Early exit if this is an indirect call.
2297   if (!TargetDecl)
2298     return;
2299 
2300   // Get the current enclosing function if it exists. If it doesn't
2301   // we can't check the target features anyhow.
2302   const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2303   if (!FD)
2304     return;
2305 
2306   // Grab the required features for the call. For a builtin this is listed in
2307   // the td file with the default cpu, for an always_inline function this is any
2308   // listed cpu and any listed features.
2309   unsigned BuiltinID = TargetDecl->getBuiltinID();
2310   std::string MissingFeature;
2311   if (BuiltinID) {
2312     SmallVector<StringRef, 1> ReqFeatures;
2313     const char *FeatureList =
2314         CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2315     // Return if the builtin doesn't have any required features.
2316     if (!FeatureList || StringRef(FeatureList) == "")
2317       return;
2318     StringRef(FeatureList).split(ReqFeatures, ",");
2319     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2320       CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2321           << TargetDecl->getDeclName()
2322           << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2323 
2324   } else if (TargetDecl->hasAttr<TargetAttr>()) {
2325     // Get the required features for the callee.
2326     SmallVector<StringRef, 1> ReqFeatures;
2327     llvm::StringMap<bool> CalleeFeatureMap;
2328     CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2329     for (const auto &F : CalleeFeatureMap) {
2330       // Only positive features are "required".
2331       if (F.getValue())
2332         ReqFeatures.push_back(F.getKey());
2333     }
2334     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2335       CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2336           << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2337   }
2338 }
2339 
2340 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2341   if (!CGM.getCodeGenOpts().SanitizeStats)
2342     return;
2343 
2344   llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2345   IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2346   CGM.getSanStats().create(IRB, SSK);
2347 }
2348 
2349 llvm::Value *
2350 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
2351   llvm::Value *TrueCondition = nullptr;
2352   if (!RO.ParsedAttribute.Architecture.empty())
2353     TrueCondition = EmitX86CpuIs(RO.ParsedAttribute.Architecture);
2354 
2355   if (!RO.ParsedAttribute.Features.empty()) {
2356     SmallVector<StringRef, 8> FeatureList;
2357     llvm::for_each(RO.ParsedAttribute.Features,
2358                    [&FeatureList](const std::string &Feature) {
2359                      FeatureList.push_back(StringRef{Feature}.substr(1));
2360                    });
2361     llvm::Value *FeatureCmp = EmitX86CpuSupports(FeatureList);
2362     TrueCondition = TrueCondition ? Builder.CreateAnd(TrueCondition, FeatureCmp)
2363                                   : FeatureCmp;
2364   }
2365   return TrueCondition;
2366 }
2367 
2368 void CodeGenFunction::EmitMultiVersionResolver(
2369     llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
2370   assert((getContext().getTargetInfo().getTriple().getArch() ==
2371               llvm::Triple::x86 ||
2372           getContext().getTargetInfo().getTriple().getArch() ==
2373               llvm::Triple::x86_64) &&
2374          "Only implemented for x86 targets");
2375 
2376   // Main function's basic block.
2377   llvm::BasicBlock *CurBlock = createBasicBlock("entry", Resolver);
2378   Builder.SetInsertPoint(CurBlock);
2379   EmitX86CpuInit();
2380 
2381   llvm::Function *DefaultFunc = nullptr;
2382   for (const MultiVersionResolverOption &RO : Options) {
2383     Builder.SetInsertPoint(CurBlock);
2384     llvm::Value *TrueCondition = FormResolverCondition(RO);
2385 
2386     if (!TrueCondition) {
2387       DefaultFunc = RO.Function;
2388     } else {
2389       llvm::BasicBlock *RetBlock = createBasicBlock("ro_ret", Resolver);
2390       llvm::IRBuilder<> RetBuilder(RetBlock);
2391       RetBuilder.CreateRet(RO.Function);
2392       CurBlock = createBasicBlock("ro_else", Resolver);
2393       Builder.CreateCondBr(TrueCondition, RetBlock, CurBlock);
2394     }
2395   }
2396 
2397   assert(DefaultFunc && "No default version?");
2398   // Emit return from the 'else-ist' block.
2399   Builder.SetInsertPoint(CurBlock);
2400   Builder.CreateRet(DefaultFunc);
2401 }
2402 
2403 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2404   if (CGDebugInfo *DI = getDebugInfo())
2405     return DI->SourceLocToDebugLoc(Location);
2406 
2407   return llvm::DebugLoc();
2408 }
2409