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, Builder);
1022   }
1023 
1024   if (ShouldInstrumentFunction()) {
1025     if (CGM.getCodeGenOpts().InstrumentFunctions)
1026       CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
1027     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
1028       CurFn->addFnAttr("instrument-function-entry-inlined",
1029                        "__cyg_profile_func_enter");
1030     if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
1031       CurFn->addFnAttr("instrument-function-entry-inlined",
1032                        "__cyg_profile_func_enter_bare");
1033   }
1034 
1035   // Since emitting the mcount call here impacts optimizations such as function
1036   // inlining, we just add an attribute to insert a mcount call in backend.
1037   // The attribute "counting-function" is set to mcount function name which is
1038   // architecture dependent.
1039   if (CGM.getCodeGenOpts().InstrumentForProfiling) {
1040     // Calls to fentry/mcount should not be generated if function has
1041     // the no_instrument_function attribute.
1042     if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
1043       if (CGM.getCodeGenOpts().CallFEntry)
1044         Fn->addFnAttr("fentry-call", "true");
1045       else {
1046         Fn->addFnAttr("instrument-function-entry-inlined",
1047                       getTarget().getMCountName());
1048       }
1049     }
1050   }
1051 
1052   if (RetTy->isVoidType()) {
1053     // Void type; nothing to return.
1054     ReturnValue = Address::invalid();
1055 
1056     // Count the implicit return.
1057     if (!endsWithReturn(D))
1058       ++NumReturnExprs;
1059   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
1060              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1061     // Indirect aggregate return; emit returned value directly into sret slot.
1062     // This reduces code size, and affects correctness in C++.
1063     auto AI = CurFn->arg_begin();
1064     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
1065       ++AI;
1066     ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
1067   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
1068              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1069     // Load the sret pointer from the argument struct and return into that.
1070     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
1071     llvm::Function::arg_iterator EI = CurFn->arg_end();
1072     --EI;
1073     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
1074     Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
1075     ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
1076   } else {
1077     ReturnValue = CreateIRTemp(RetTy, "retval");
1078 
1079     // Tell the epilog emitter to autorelease the result.  We do this
1080     // now so that various specialized functions can suppress it
1081     // during their IR-generation.
1082     if (getLangOpts().ObjCAutoRefCount &&
1083         !CurFnInfo->isReturnsRetained() &&
1084         RetTy->isObjCRetainableType())
1085       AutoreleaseResult = true;
1086   }
1087 
1088   EmitStartEHSpec(CurCodeDecl);
1089 
1090   PrologueCleanupDepth = EHStack.stable_begin();
1091 
1092   // Emit OpenMP specific initialization of the device functions.
1093   if (getLangOpts().OpenMP && CurCodeDecl)
1094     CGM.getOpenMPRuntime().emitFunctionProlog(*this, CurCodeDecl);
1095 
1096   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
1097 
1098   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
1099     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
1100     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
1101     if (MD->getParent()->isLambda() &&
1102         MD->getOverloadedOperator() == OO_Call) {
1103       // We're in a lambda; figure out the captures.
1104       MD->getParent()->getCaptureFields(LambdaCaptureFields,
1105                                         LambdaThisCaptureField);
1106       if (LambdaThisCaptureField) {
1107         // If the lambda captures the object referred to by '*this' - either by
1108         // value or by reference, make sure CXXThisValue points to the correct
1109         // object.
1110 
1111         // Get the lvalue for the field (which is a copy of the enclosing object
1112         // or contains the address of the enclosing object).
1113         LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
1114         if (!LambdaThisCaptureField->getType()->isPointerType()) {
1115           // If the enclosing object was captured by value, just use its address.
1116           CXXThisValue = ThisFieldLValue.getAddress().getPointer();
1117         } else {
1118           // Load the lvalue pointed to by the field, since '*this' was captured
1119           // by reference.
1120           CXXThisValue =
1121               EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
1122         }
1123       }
1124       for (auto *FD : MD->getParent()->fields()) {
1125         if (FD->hasCapturedVLAType()) {
1126           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
1127                                            SourceLocation()).getScalarVal();
1128           auto VAT = FD->getCapturedVLAType();
1129           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
1130         }
1131       }
1132     } else {
1133       // Not in a lambda; just use 'this' from the method.
1134       // FIXME: Should we generate a new load for each use of 'this'?  The
1135       // fast register allocator would be happier...
1136       CXXThisValue = CXXABIThisValue;
1137     }
1138 
1139     // Check the 'this' pointer once per function, if it's available.
1140     if (CXXABIThisValue) {
1141       SanitizerSet SkippedChecks;
1142       SkippedChecks.set(SanitizerKind::ObjectSize, true);
1143       QualType ThisTy = MD->getThisType(getContext());
1144 
1145       // If this is the call operator of a lambda with no capture-default, it
1146       // may have a static invoker function, which may call this operator with
1147       // a null 'this' pointer.
1148       if (isLambdaCallOperator(MD) &&
1149           MD->getParent()->getLambdaCaptureDefault() == LCD_None)
1150         SkippedChecks.set(SanitizerKind::Null, true);
1151 
1152       EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1153                                                 : TCK_MemberCall,
1154                     Loc, CXXABIThisValue, ThisTy,
1155                     getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1156                     SkippedChecks);
1157     }
1158   }
1159 
1160   // If any of the arguments have a variably modified type, make sure to
1161   // emit the type size.
1162   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1163        i != e; ++i) {
1164     const VarDecl *VD = *i;
1165 
1166     // Dig out the type as written from ParmVarDecls; it's unclear whether
1167     // the standard (C99 6.9.1p10) requires this, but we're following the
1168     // precedent set by gcc.
1169     QualType Ty;
1170     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1171       Ty = PVD->getOriginalType();
1172     else
1173       Ty = VD->getType();
1174 
1175     if (Ty->isVariablyModifiedType())
1176       EmitVariablyModifiedType(Ty);
1177   }
1178   // Emit a location at the end of the prologue.
1179   if (CGDebugInfo *DI = getDebugInfo())
1180     DI->EmitLocation(Builder, StartLoc);
1181 }
1182 
1183 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1184                                        const Stmt *Body) {
1185   incrementProfileCounter(Body);
1186   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1187     EmitCompoundStmtWithoutScope(*S);
1188   else
1189     EmitStmt(Body);
1190 }
1191 
1192 /// When instrumenting to collect profile data, the counts for some blocks
1193 /// such as switch cases need to not include the fall-through counts, so
1194 /// emit a branch around the instrumentation code. When not instrumenting,
1195 /// this just calls EmitBlock().
1196 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1197                                                const Stmt *S) {
1198   llvm::BasicBlock *SkipCountBB = nullptr;
1199   if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1200     // When instrumenting for profiling, the fallthrough to certain
1201     // statements needs to skip over the instrumentation code so that we
1202     // get an accurate count.
1203     SkipCountBB = createBasicBlock("skipcount");
1204     EmitBranch(SkipCountBB);
1205   }
1206   EmitBlock(BB);
1207   uint64_t CurrentCount = getCurrentProfileCount();
1208   incrementProfileCounter(S);
1209   setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1210   if (SkipCountBB)
1211     EmitBlock(SkipCountBB);
1212 }
1213 
1214 /// Tries to mark the given function nounwind based on the
1215 /// non-existence of any throwing calls within it.  We believe this is
1216 /// lightweight enough to do at -O0.
1217 static void TryMarkNoThrow(llvm::Function *F) {
1218   // LLVM treats 'nounwind' on a function as part of the type, so we
1219   // can't do this on functions that can be overwritten.
1220   if (F->isInterposable()) return;
1221 
1222   for (llvm::BasicBlock &BB : *F)
1223     for (llvm::Instruction &I : BB)
1224       if (I.mayThrow())
1225         return;
1226 
1227   F->setDoesNotThrow();
1228 }
1229 
1230 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1231                                                FunctionArgList &Args) {
1232   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1233   QualType ResTy = FD->getReturnType();
1234 
1235   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1236   if (MD && MD->isInstance()) {
1237     if (CGM.getCXXABI().HasThisReturn(GD))
1238       ResTy = MD->getThisType(getContext());
1239     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1240       ResTy = CGM.getContext().VoidPtrTy;
1241     CGM.getCXXABI().buildThisParam(*this, Args);
1242   }
1243 
1244   // The base version of an inheriting constructor whose constructed base is a
1245   // virtual base is not passed any arguments (because it doesn't actually call
1246   // the inherited constructor).
1247   bool PassedParams = true;
1248   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1249     if (auto Inherited = CD->getInheritedConstructor())
1250       PassedParams =
1251           getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1252 
1253   if (PassedParams) {
1254     for (auto *Param : FD->parameters()) {
1255       Args.push_back(Param);
1256       if (!Param->hasAttr<PassObjectSizeAttr>())
1257         continue;
1258 
1259       auto *Implicit = ImplicitParamDecl::Create(
1260           getContext(), Param->getDeclContext(), Param->getLocation(),
1261           /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1262       SizeArguments[Param] = Implicit;
1263       Args.push_back(Implicit);
1264     }
1265   }
1266 
1267   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1268     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1269 
1270   return ResTy;
1271 }
1272 
1273 static bool
1274 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1275                                              const ASTContext &Context) {
1276   QualType T = FD->getReturnType();
1277   // Avoid the optimization for functions that return a record type with a
1278   // trivial destructor or another trivially copyable type.
1279   if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1280     if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1281       return !ClassDecl->hasTrivialDestructor();
1282   }
1283   return !T.isTriviallyCopyableType(Context);
1284 }
1285 
1286 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1287                                    const CGFunctionInfo &FnInfo) {
1288   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1289   CurGD = GD;
1290 
1291   FunctionArgList Args;
1292   QualType ResTy = BuildFunctionArgList(GD, Args);
1293 
1294   // Check if we should generate debug info for this function.
1295   if (FD->hasAttr<NoDebugAttr>())
1296     DebugInfo = nullptr; // disable debug info indefinitely for this function
1297 
1298   // The function might not have a body if we're generating thunks for a
1299   // function declaration.
1300   SourceRange BodyRange;
1301   if (Stmt *Body = FD->getBody())
1302     BodyRange = Body->getSourceRange();
1303   else
1304     BodyRange = FD->getLocation();
1305   CurEHLocation = BodyRange.getEnd();
1306 
1307   // Use the location of the start of the function to determine where
1308   // the function definition is located. By default use the location
1309   // of the declaration as the location for the subprogram. A function
1310   // may lack a declaration in the source code if it is created by code
1311   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1312   SourceLocation Loc = FD->getLocation();
1313 
1314   // If this is a function specialization then use the pattern body
1315   // as the location for the function.
1316   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1317     if (SpecDecl->hasBody(SpecDecl))
1318       Loc = SpecDecl->getLocation();
1319 
1320   Stmt *Body = FD->getBody();
1321 
1322   // Initialize helper which will detect jumps which can cause invalid lifetime
1323   // markers.
1324   if (Body && ShouldEmitLifetimeMarkers)
1325     Bypasses.Init(Body);
1326 
1327   // Emit the standard function prologue.
1328   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1329 
1330   // Generate the body of the function.
1331   PGO.assignRegionCounters(GD, CurFn);
1332   if (isa<CXXDestructorDecl>(FD))
1333     EmitDestructorBody(Args);
1334   else if (isa<CXXConstructorDecl>(FD))
1335     EmitConstructorBody(Args);
1336   else if (getLangOpts().CUDA &&
1337            !getLangOpts().CUDAIsDevice &&
1338            FD->hasAttr<CUDAGlobalAttr>())
1339     CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1340   else if (isa<CXXMethodDecl>(FD) &&
1341            cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1342     // The lambda static invoker function is special, because it forwards or
1343     // clones the body of the function call operator (but is actually static).
1344     EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1345   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1346              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1347               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1348     // Implicit copy-assignment gets the same special treatment as implicit
1349     // copy-constructors.
1350     emitImplicitAssignmentOperatorBody(Args);
1351   } else if (Body) {
1352     EmitFunctionBody(Args, Body);
1353   } else
1354     llvm_unreachable("no definition for emitted function");
1355 
1356   // C++11 [stmt.return]p2:
1357   //   Flowing off the end of a function [...] results in undefined behavior in
1358   //   a value-returning function.
1359   // C11 6.9.1p12:
1360   //   If the '}' that terminates a function is reached, and the value of the
1361   //   function call is used by the caller, the behavior is undefined.
1362   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1363       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1364     bool ShouldEmitUnreachable =
1365         CGM.getCodeGenOpts().StrictReturn ||
1366         shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1367     if (SanOpts.has(SanitizerKind::Return)) {
1368       SanitizerScope SanScope(this);
1369       llvm::Value *IsFalse = Builder.getFalse();
1370       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1371                 SanitizerHandler::MissingReturn,
1372                 EmitCheckSourceLocation(FD->getLocation()), None);
1373     } else if (ShouldEmitUnreachable) {
1374       if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1375         EmitTrapCall(llvm::Intrinsic::trap);
1376     }
1377     if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1378       Builder.CreateUnreachable();
1379       Builder.ClearInsertionPoint();
1380     }
1381   }
1382 
1383   // Emit the standard function epilogue.
1384   FinishFunction(BodyRange.getEnd());
1385 
1386   // If we haven't marked the function nothrow through other means, do
1387   // a quick pass now to see if we can.
1388   if (!CurFn->doesNotThrow())
1389     TryMarkNoThrow(CurFn);
1390 }
1391 
1392 /// ContainsLabel - Return true if the statement contains a label in it.  If
1393 /// this statement is not executed normally, it not containing a label means
1394 /// that we can just remove the code.
1395 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1396   // Null statement, not a label!
1397   if (!S) return false;
1398 
1399   // If this is a label, we have to emit the code, consider something like:
1400   // if (0) {  ...  foo:  bar(); }  goto foo;
1401   //
1402   // TODO: If anyone cared, we could track __label__'s, since we know that you
1403   // can't jump to one from outside their declared region.
1404   if (isa<LabelStmt>(S))
1405     return true;
1406 
1407   // If this is a case/default statement, and we haven't seen a switch, we have
1408   // to emit the code.
1409   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1410     return true;
1411 
1412   // If this is a switch statement, we want to ignore cases below it.
1413   if (isa<SwitchStmt>(S))
1414     IgnoreCaseStmts = true;
1415 
1416   // Scan subexpressions for verboten labels.
1417   for (const Stmt *SubStmt : S->children())
1418     if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1419       return true;
1420 
1421   return false;
1422 }
1423 
1424 /// containsBreak - Return true if the statement contains a break out of it.
1425 /// If the statement (recursively) contains a switch or loop with a break
1426 /// inside of it, this is fine.
1427 bool CodeGenFunction::containsBreak(const Stmt *S) {
1428   // Null statement, not a label!
1429   if (!S) return false;
1430 
1431   // If this is a switch or loop that defines its own break scope, then we can
1432   // include it and anything inside of it.
1433   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1434       isa<ForStmt>(S))
1435     return false;
1436 
1437   if (isa<BreakStmt>(S))
1438     return true;
1439 
1440   // Scan subexpressions for verboten breaks.
1441   for (const Stmt *SubStmt : S->children())
1442     if (containsBreak(SubStmt))
1443       return true;
1444 
1445   return false;
1446 }
1447 
1448 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1449   if (!S) return false;
1450 
1451   // Some statement kinds add a scope and thus never add a decl to the current
1452   // scope. Note, this list is longer than the list of statements that might
1453   // have an unscoped decl nested within them, but this way is conservatively
1454   // correct even if more statement kinds are added.
1455   if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1456       isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1457       isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1458       isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1459     return false;
1460 
1461   if (isa<DeclStmt>(S))
1462     return true;
1463 
1464   for (const Stmt *SubStmt : S->children())
1465     if (mightAddDeclToScope(SubStmt))
1466       return true;
1467 
1468   return false;
1469 }
1470 
1471 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1472 /// to a constant, or if it does but contains a label, return false.  If it
1473 /// constant folds return true and set the boolean result in Result.
1474 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1475                                                    bool &ResultBool,
1476                                                    bool AllowLabels) {
1477   llvm::APSInt ResultInt;
1478   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1479     return false;
1480 
1481   ResultBool = ResultInt.getBoolValue();
1482   return true;
1483 }
1484 
1485 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1486 /// to a constant, or if it does but contains a label, return false.  If it
1487 /// constant folds return true and set the folded value.
1488 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1489                                                    llvm::APSInt &ResultInt,
1490                                                    bool AllowLabels) {
1491   // FIXME: Rename and handle conversion of other evaluatable things
1492   // to bool.
1493   llvm::APSInt Int;
1494   if (!Cond->EvaluateAsInt(Int, getContext()))
1495     return false;  // Not foldable, not integer or not fully evaluatable.
1496 
1497   if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1498     return false;  // Contains a label.
1499 
1500   ResultInt = Int;
1501   return true;
1502 }
1503 
1504 
1505 
1506 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1507 /// statement) to the specified blocks.  Based on the condition, this might try
1508 /// to simplify the codegen of the conditional based on the branch.
1509 ///
1510 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1511                                            llvm::BasicBlock *TrueBlock,
1512                                            llvm::BasicBlock *FalseBlock,
1513                                            uint64_t TrueCount) {
1514   Cond = Cond->IgnoreParens();
1515 
1516   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1517 
1518     // Handle X && Y in a condition.
1519     if (CondBOp->getOpcode() == BO_LAnd) {
1520       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1521       // folded if the case was simple enough.
1522       bool ConstantBool = false;
1523       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1524           ConstantBool) {
1525         // br(1 && X) -> br(X).
1526         incrementProfileCounter(CondBOp);
1527         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1528                                     TrueCount);
1529       }
1530 
1531       // If we have "X && 1", simplify the code to use an uncond branch.
1532       // "X && 0" would have been constant folded to 0.
1533       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1534           ConstantBool) {
1535         // br(X && 1) -> br(X).
1536         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1537                                     TrueCount);
1538       }
1539 
1540       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1541       // want to jump to the FalseBlock.
1542       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1543       // The counter tells us how often we evaluate RHS, and all of TrueCount
1544       // can be propagated to that branch.
1545       uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1546 
1547       ConditionalEvaluation eval(*this);
1548       {
1549         ApplyDebugLocation DL(*this, Cond);
1550         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1551         EmitBlock(LHSTrue);
1552       }
1553 
1554       incrementProfileCounter(CondBOp);
1555       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1556 
1557       // Any temporaries created here are conditional.
1558       eval.begin(*this);
1559       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1560       eval.end(*this);
1561 
1562       return;
1563     }
1564 
1565     if (CondBOp->getOpcode() == BO_LOr) {
1566       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1567       // folded if the case was simple enough.
1568       bool ConstantBool = false;
1569       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1570           !ConstantBool) {
1571         // br(0 || X) -> br(X).
1572         incrementProfileCounter(CondBOp);
1573         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1574                                     TrueCount);
1575       }
1576 
1577       // If we have "X || 0", simplify the code to use an uncond branch.
1578       // "X || 1" would have been constant folded to 1.
1579       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1580           !ConstantBool) {
1581         // br(X || 0) -> br(X).
1582         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1583                                     TrueCount);
1584       }
1585 
1586       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1587       // want to jump to the TrueBlock.
1588       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1589       // We have the count for entry to the RHS and for the whole expression
1590       // being true, so we can divy up True count between the short circuit and
1591       // the RHS.
1592       uint64_t LHSCount =
1593           getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1594       uint64_t RHSCount = TrueCount - LHSCount;
1595 
1596       ConditionalEvaluation eval(*this);
1597       {
1598         ApplyDebugLocation DL(*this, Cond);
1599         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1600         EmitBlock(LHSFalse);
1601       }
1602 
1603       incrementProfileCounter(CondBOp);
1604       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1605 
1606       // Any temporaries created here are conditional.
1607       eval.begin(*this);
1608       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1609 
1610       eval.end(*this);
1611 
1612       return;
1613     }
1614   }
1615 
1616   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1617     // br(!x, t, f) -> br(x, f, t)
1618     if (CondUOp->getOpcode() == UO_LNot) {
1619       // Negate the count.
1620       uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1621       // Negate the condition and swap the destination blocks.
1622       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1623                                   FalseCount);
1624     }
1625   }
1626 
1627   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1628     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1629     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1630     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1631 
1632     ConditionalEvaluation cond(*this);
1633     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1634                          getProfileCount(CondOp));
1635 
1636     // When computing PGO branch weights, we only know the overall count for
1637     // the true block. This code is essentially doing tail duplication of the
1638     // naive code-gen, introducing new edges for which counts are not
1639     // available. Divide the counts proportionally between the LHS and RHS of
1640     // the conditional operator.
1641     uint64_t LHSScaledTrueCount = 0;
1642     if (TrueCount) {
1643       double LHSRatio =
1644           getProfileCount(CondOp) / (double)getCurrentProfileCount();
1645       LHSScaledTrueCount = TrueCount * LHSRatio;
1646     }
1647 
1648     cond.begin(*this);
1649     EmitBlock(LHSBlock);
1650     incrementProfileCounter(CondOp);
1651     {
1652       ApplyDebugLocation DL(*this, Cond);
1653       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1654                            LHSScaledTrueCount);
1655     }
1656     cond.end(*this);
1657 
1658     cond.begin(*this);
1659     EmitBlock(RHSBlock);
1660     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1661                          TrueCount - LHSScaledTrueCount);
1662     cond.end(*this);
1663 
1664     return;
1665   }
1666 
1667   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1668     // Conditional operator handling can give us a throw expression as a
1669     // condition for a case like:
1670     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1671     // Fold this to:
1672     //   br(c, throw x, br(y, t, f))
1673     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1674     return;
1675   }
1676 
1677   // If the branch has a condition wrapped by __builtin_unpredictable,
1678   // create metadata that specifies that the branch is unpredictable.
1679   // Don't bother if not optimizing because that metadata would not be used.
1680   llvm::MDNode *Unpredictable = nullptr;
1681   auto *Call = dyn_cast<CallExpr>(Cond);
1682   if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1683     auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1684     if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1685       llvm::MDBuilder MDHelper(getLLVMContext());
1686       Unpredictable = MDHelper.createUnpredictable();
1687     }
1688   }
1689 
1690   // Create branch weights based on the number of times we get here and the
1691   // number of times the condition should be true.
1692   uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1693   llvm::MDNode *Weights =
1694       createProfileWeights(TrueCount, CurrentCount - TrueCount);
1695 
1696   // Emit the code with the fully general case.
1697   llvm::Value *CondV;
1698   {
1699     ApplyDebugLocation DL(*this, Cond);
1700     CondV = EvaluateExprAsBool(Cond);
1701   }
1702   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1703 }
1704 
1705 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1706 /// specified stmt yet.
1707 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1708   CGM.ErrorUnsupported(S, Type);
1709 }
1710 
1711 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1712 /// variable-length array whose elements have a non-zero bit-pattern.
1713 ///
1714 /// \param baseType the inner-most element type of the array
1715 /// \param src - a char* pointing to the bit-pattern for a single
1716 /// base element of the array
1717 /// \param sizeInChars - the total size of the VLA, in chars
1718 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1719                                Address dest, Address src,
1720                                llvm::Value *sizeInChars) {
1721   CGBuilderTy &Builder = CGF.Builder;
1722 
1723   CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1724   llvm::Value *baseSizeInChars
1725     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1726 
1727   Address begin =
1728     Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1729   llvm::Value *end =
1730     Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1731 
1732   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1733   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1734   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1735 
1736   // Make a loop over the VLA.  C99 guarantees that the VLA element
1737   // count must be nonzero.
1738   CGF.EmitBlock(loopBB);
1739 
1740   llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1741   cur->addIncoming(begin.getPointer(), originBB);
1742 
1743   CharUnits curAlign =
1744     dest.getAlignment().alignmentOfArrayElement(baseSize);
1745 
1746   // memcpy the individual element bit-pattern.
1747   Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1748                        /*volatile*/ false);
1749 
1750   // Go to the next element.
1751   llvm::Value *next =
1752     Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1753 
1754   // Leave if that's the end of the VLA.
1755   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1756   Builder.CreateCondBr(done, contBB, loopBB);
1757   cur->addIncoming(next, loopBB);
1758 
1759   CGF.EmitBlock(contBB);
1760 }
1761 
1762 void
1763 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1764   // Ignore empty classes in C++.
1765   if (getLangOpts().CPlusPlus) {
1766     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1767       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1768         return;
1769     }
1770   }
1771 
1772   // Cast the dest ptr to the appropriate i8 pointer type.
1773   if (DestPtr.getElementType() != Int8Ty)
1774     DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1775 
1776   // Get size and alignment info for this aggregate.
1777   CharUnits size = getContext().getTypeSizeInChars(Ty);
1778 
1779   llvm::Value *SizeVal;
1780   const VariableArrayType *vla;
1781 
1782   // Don't bother emitting a zero-byte memset.
1783   if (size.isZero()) {
1784     // But note that getTypeInfo returns 0 for a VLA.
1785     if (const VariableArrayType *vlaType =
1786           dyn_cast_or_null<VariableArrayType>(
1787                                           getContext().getAsArrayType(Ty))) {
1788       auto VlaSize = getVLASize(vlaType);
1789       SizeVal = VlaSize.NumElts;
1790       CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type);
1791       if (!eltSize.isOne())
1792         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1793       vla = vlaType;
1794     } else {
1795       return;
1796     }
1797   } else {
1798     SizeVal = CGM.getSize(size);
1799     vla = nullptr;
1800   }
1801 
1802   // If the type contains a pointer to data member we can't memset it to zero.
1803   // Instead, create a null constant and copy it to the destination.
1804   // TODO: there are other patterns besides zero that we can usefully memset,
1805   // like -1, which happens to be the pattern used by member-pointers.
1806   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1807     // For a VLA, emit a single element, then splat that over the VLA.
1808     if (vla) Ty = getContext().getBaseElementType(vla);
1809 
1810     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1811 
1812     llvm::GlobalVariable *NullVariable =
1813       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1814                                /*isConstant=*/true,
1815                                llvm::GlobalVariable::PrivateLinkage,
1816                                NullConstant, Twine());
1817     CharUnits NullAlign = DestPtr.getAlignment();
1818     NullVariable->setAlignment(NullAlign.getQuantity());
1819     Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1820                    NullAlign);
1821 
1822     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1823 
1824     // Get and call the appropriate llvm.memcpy overload.
1825     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1826     return;
1827   }
1828 
1829   // Otherwise, just memset the whole thing to zero.  This is legal
1830   // because in LLVM, all default initializers (other than the ones we just
1831   // handled above) are guaranteed to have a bit pattern of all zeros.
1832   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1833 }
1834 
1835 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1836   // Make sure that there is a block for the indirect goto.
1837   if (!IndirectBranch)
1838     GetIndirectGotoBlock();
1839 
1840   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1841 
1842   // Make sure the indirect branch includes all of the address-taken blocks.
1843   IndirectBranch->addDestination(BB);
1844   return llvm::BlockAddress::get(CurFn, BB);
1845 }
1846 
1847 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1848   // If we already made the indirect branch for indirect goto, return its block.
1849   if (IndirectBranch) return IndirectBranch->getParent();
1850 
1851   CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1852 
1853   // Create the PHI node that indirect gotos will add entries to.
1854   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1855                                               "indirect.goto.dest");
1856 
1857   // Create the indirect branch instruction.
1858   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1859   return IndirectBranch->getParent();
1860 }
1861 
1862 /// Computes the length of an array in elements, as well as the base
1863 /// element type and a properly-typed first element pointer.
1864 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1865                                               QualType &baseType,
1866                                               Address &addr) {
1867   const ArrayType *arrayType = origArrayType;
1868 
1869   // If it's a VLA, we have to load the stored size.  Note that
1870   // this is the size of the VLA in bytes, not its size in elements.
1871   llvm::Value *numVLAElements = nullptr;
1872   if (isa<VariableArrayType>(arrayType)) {
1873     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts;
1874 
1875     // Walk into all VLAs.  This doesn't require changes to addr,
1876     // which has type T* where T is the first non-VLA element type.
1877     do {
1878       QualType elementType = arrayType->getElementType();
1879       arrayType = getContext().getAsArrayType(elementType);
1880 
1881       // If we only have VLA components, 'addr' requires no adjustment.
1882       if (!arrayType) {
1883         baseType = elementType;
1884         return numVLAElements;
1885       }
1886     } while (isa<VariableArrayType>(arrayType));
1887 
1888     // We get out here only if we find a constant array type
1889     // inside the VLA.
1890   }
1891 
1892   // We have some number of constant-length arrays, so addr should
1893   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1894   // down to the first element of addr.
1895   SmallVector<llvm::Value*, 8> gepIndices;
1896 
1897   // GEP down to the array type.
1898   llvm::ConstantInt *zero = Builder.getInt32(0);
1899   gepIndices.push_back(zero);
1900 
1901   uint64_t countFromCLAs = 1;
1902   QualType eltType;
1903 
1904   llvm::ArrayType *llvmArrayType =
1905     dyn_cast<llvm::ArrayType>(addr.getElementType());
1906   while (llvmArrayType) {
1907     assert(isa<ConstantArrayType>(arrayType));
1908     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1909              == llvmArrayType->getNumElements());
1910 
1911     gepIndices.push_back(zero);
1912     countFromCLAs *= llvmArrayType->getNumElements();
1913     eltType = arrayType->getElementType();
1914 
1915     llvmArrayType =
1916       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1917     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1918     assert((!llvmArrayType || arrayType) &&
1919            "LLVM and Clang types are out-of-synch");
1920   }
1921 
1922   if (arrayType) {
1923     // From this point onwards, the Clang array type has been emitted
1924     // as some other type (probably a packed struct). Compute the array
1925     // size, and just emit the 'begin' expression as a bitcast.
1926     while (arrayType) {
1927       countFromCLAs *=
1928           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1929       eltType = arrayType->getElementType();
1930       arrayType = getContext().getAsArrayType(eltType);
1931     }
1932 
1933     llvm::Type *baseType = ConvertType(eltType);
1934     addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1935   } else {
1936     // Create the actual GEP.
1937     addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1938                                              gepIndices, "array.begin"),
1939                    addr.getAlignment());
1940   }
1941 
1942   baseType = eltType;
1943 
1944   llvm::Value *numElements
1945     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1946 
1947   // If we had any VLA dimensions, factor them in.
1948   if (numVLAElements)
1949     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1950 
1951   return numElements;
1952 }
1953 
1954 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) {
1955   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1956   assert(vla && "type was not a variable array type!");
1957   return getVLASize(vla);
1958 }
1959 
1960 CodeGenFunction::VlaSizePair
1961 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1962   // The number of elements so far; always size_t.
1963   llvm::Value *numElements = nullptr;
1964 
1965   QualType elementType;
1966   do {
1967     elementType = type->getElementType();
1968     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1969     assert(vlaSize && "no size for VLA!");
1970     assert(vlaSize->getType() == SizeTy);
1971 
1972     if (!numElements) {
1973       numElements = vlaSize;
1974     } else {
1975       // It's undefined behavior if this wraps around, so mark it that way.
1976       // FIXME: Teach -fsanitize=undefined to trap this.
1977       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1978     }
1979   } while ((type = getContext().getAsVariableArrayType(elementType)));
1980 
1981   return { numElements, elementType };
1982 }
1983 
1984 CodeGenFunction::VlaSizePair
1985 CodeGenFunction::getVLAElements1D(QualType type) {
1986   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1987   assert(vla && "type was not a variable array type!");
1988   return getVLAElements1D(vla);
1989 }
1990 
1991 CodeGenFunction::VlaSizePair
1992 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) {
1993   llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()];
1994   assert(VlaSize && "no size for VLA!");
1995   assert(VlaSize->getType() == SizeTy);
1996   return { VlaSize, Vla->getElementType() };
1997 }
1998 
1999 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
2000   assert(type->isVariablyModifiedType() &&
2001          "Must pass variably modified type to EmitVLASizes!");
2002 
2003   EnsureInsertPoint();
2004 
2005   // We're going to walk down into the type and look for VLA
2006   // expressions.
2007   do {
2008     assert(type->isVariablyModifiedType());
2009 
2010     const Type *ty = type.getTypePtr();
2011     switch (ty->getTypeClass()) {
2012 
2013 #define TYPE(Class, Base)
2014 #define ABSTRACT_TYPE(Class, Base)
2015 #define NON_CANONICAL_TYPE(Class, Base)
2016 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
2017 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
2018 #include "clang/AST/TypeNodes.def"
2019       llvm_unreachable("unexpected dependent type!");
2020 
2021     // These types are never variably-modified.
2022     case Type::Builtin:
2023     case Type::Complex:
2024     case Type::Vector:
2025     case Type::ExtVector:
2026     case Type::Record:
2027     case Type::Enum:
2028     case Type::Elaborated:
2029     case Type::TemplateSpecialization:
2030     case Type::ObjCTypeParam:
2031     case Type::ObjCObject:
2032     case Type::ObjCInterface:
2033     case Type::ObjCObjectPointer:
2034       llvm_unreachable("type class is never variably-modified!");
2035 
2036     case Type::Adjusted:
2037       type = cast<AdjustedType>(ty)->getAdjustedType();
2038       break;
2039 
2040     case Type::Decayed:
2041       type = cast<DecayedType>(ty)->getPointeeType();
2042       break;
2043 
2044     case Type::Pointer:
2045       type = cast<PointerType>(ty)->getPointeeType();
2046       break;
2047 
2048     case Type::BlockPointer:
2049       type = cast<BlockPointerType>(ty)->getPointeeType();
2050       break;
2051 
2052     case Type::LValueReference:
2053     case Type::RValueReference:
2054       type = cast<ReferenceType>(ty)->getPointeeType();
2055       break;
2056 
2057     case Type::MemberPointer:
2058       type = cast<MemberPointerType>(ty)->getPointeeType();
2059       break;
2060 
2061     case Type::ConstantArray:
2062     case Type::IncompleteArray:
2063       // Losing element qualification here is fine.
2064       type = cast<ArrayType>(ty)->getElementType();
2065       break;
2066 
2067     case Type::VariableArray: {
2068       // Losing element qualification here is fine.
2069       const VariableArrayType *vat = cast<VariableArrayType>(ty);
2070 
2071       // Unknown size indication requires no size computation.
2072       // Otherwise, evaluate and record it.
2073       if (const Expr *size = vat->getSizeExpr()) {
2074         // It's possible that we might have emitted this already,
2075         // e.g. with a typedef and a pointer to it.
2076         llvm::Value *&entry = VLASizeMap[size];
2077         if (!entry) {
2078           llvm::Value *Size = EmitScalarExpr(size);
2079 
2080           // C11 6.7.6.2p5:
2081           //   If the size is an expression that is not an integer constant
2082           //   expression [...] each time it is evaluated it shall have a value
2083           //   greater than zero.
2084           if (SanOpts.has(SanitizerKind::VLABound) &&
2085               size->getType()->isSignedIntegerType()) {
2086             SanitizerScope SanScope(this);
2087             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
2088             llvm::Constant *StaticArgs[] = {
2089               EmitCheckSourceLocation(size->getLocStart()),
2090               EmitCheckTypeDescriptor(size->getType())
2091             };
2092             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
2093                                      SanitizerKind::VLABound),
2094                       SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
2095           }
2096 
2097           // Always zexting here would be wrong if it weren't
2098           // undefined behavior to have a negative bound.
2099           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
2100         }
2101       }
2102       type = vat->getElementType();
2103       break;
2104     }
2105 
2106     case Type::FunctionProto:
2107     case Type::FunctionNoProto:
2108       type = cast<FunctionType>(ty)->getReturnType();
2109       break;
2110 
2111     case Type::Paren:
2112     case Type::TypeOf:
2113     case Type::UnaryTransform:
2114     case Type::Attributed:
2115     case Type::SubstTemplateTypeParm:
2116     case Type::PackExpansion:
2117       // Keep walking after single level desugaring.
2118       type = type.getSingleStepDesugaredType(getContext());
2119       break;
2120 
2121     case Type::Typedef:
2122     case Type::Decltype:
2123     case Type::Auto:
2124     case Type::DeducedTemplateSpecialization:
2125       // Stop walking: nothing to do.
2126       return;
2127 
2128     case Type::TypeOfExpr:
2129       // Stop walking: emit typeof expression.
2130       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
2131       return;
2132 
2133     case Type::Atomic:
2134       type = cast<AtomicType>(ty)->getValueType();
2135       break;
2136 
2137     case Type::Pipe:
2138       type = cast<PipeType>(ty)->getElementType();
2139       break;
2140     }
2141   } while (type->isVariablyModifiedType());
2142 }
2143 
2144 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
2145   if (getContext().getBuiltinVaListType()->isArrayType())
2146     return EmitPointerWithAlignment(E);
2147   return EmitLValue(E).getAddress();
2148 }
2149 
2150 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
2151   return EmitLValue(E).getAddress();
2152 }
2153 
2154 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2155                                               const APValue &Init) {
2156   assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
2157   if (CGDebugInfo *Dbg = getDebugInfo())
2158     if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2159       Dbg->EmitGlobalVariable(E->getDecl(), Init);
2160 }
2161 
2162 CodeGenFunction::PeepholeProtection
2163 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2164   // At the moment, the only aggressive peephole we do in IR gen
2165   // is trunc(zext) folding, but if we add more, we can easily
2166   // extend this protection.
2167 
2168   if (!rvalue.isScalar()) return PeepholeProtection();
2169   llvm::Value *value = rvalue.getScalarVal();
2170   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2171 
2172   // Just make an extra bitcast.
2173   assert(HaveInsertPoint());
2174   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2175                                                   Builder.GetInsertBlock());
2176 
2177   PeepholeProtection protection;
2178   protection.Inst = inst;
2179   return protection;
2180 }
2181 
2182 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2183   if (!protection.Inst) return;
2184 
2185   // In theory, we could try to duplicate the peepholes now, but whatever.
2186   protection.Inst->eraseFromParent();
2187 }
2188 
2189 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2190                                                  llvm::Value *AnnotatedVal,
2191                                                  StringRef AnnotationStr,
2192                                                  SourceLocation Location) {
2193   llvm::Value *Args[4] = {
2194     AnnotatedVal,
2195     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2196     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2197     CGM.EmitAnnotationLineNo(Location)
2198   };
2199   return Builder.CreateCall(AnnotationFn, Args);
2200 }
2201 
2202 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2203   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2204   // FIXME We create a new bitcast for every annotation because that's what
2205   // llvm-gcc was doing.
2206   for (const auto *I : D->specific_attrs<AnnotateAttr>())
2207     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2208                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2209                        I->getAnnotation(), D->getLocation());
2210 }
2211 
2212 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2213                                               Address Addr) {
2214   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2215   llvm::Value *V = Addr.getPointer();
2216   llvm::Type *VTy = V->getType();
2217   llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2218                                     CGM.Int8PtrTy);
2219 
2220   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2221     // FIXME Always emit the cast inst so we can differentiate between
2222     // annotation on the first field of a struct and annotation on the struct
2223     // itself.
2224     if (VTy != CGM.Int8PtrTy)
2225       V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2226     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2227     V = Builder.CreateBitCast(V, VTy);
2228   }
2229 
2230   return Address(V, Addr.getAlignment());
2231 }
2232 
2233 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2234 
2235 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2236     : CGF(CGF) {
2237   assert(!CGF->IsSanitizerScope);
2238   CGF->IsSanitizerScope = true;
2239 }
2240 
2241 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2242   CGF->IsSanitizerScope = false;
2243 }
2244 
2245 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2246                                    const llvm::Twine &Name,
2247                                    llvm::BasicBlock *BB,
2248                                    llvm::BasicBlock::iterator InsertPt) const {
2249   LoopStack.InsertHelper(I);
2250   if (IsSanitizerScope)
2251     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2252 }
2253 
2254 void CGBuilderInserter::InsertHelper(
2255     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2256     llvm::BasicBlock::iterator InsertPt) const {
2257   llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2258   if (CGF)
2259     CGF->InsertHelper(I, Name, BB, InsertPt);
2260 }
2261 
2262 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2263                                 CodeGenModule &CGM, const FunctionDecl *FD,
2264                                 std::string &FirstMissing) {
2265   // If there aren't any required features listed then go ahead and return.
2266   if (ReqFeatures.empty())
2267     return false;
2268 
2269   // Now build up the set of caller features and verify that all the required
2270   // features are there.
2271   llvm::StringMap<bool> CallerFeatureMap;
2272   CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2273 
2274   // If we have at least one of the features in the feature list return
2275   // true, otherwise return false.
2276   return std::all_of(
2277       ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2278         SmallVector<StringRef, 1> OrFeatures;
2279         Feature.split(OrFeatures, "|");
2280         return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2281                            [&](StringRef Feature) {
2282                              if (!CallerFeatureMap.lookup(Feature)) {
2283                                FirstMissing = Feature.str();
2284                                return false;
2285                              }
2286                              return true;
2287                            });
2288       });
2289 }
2290 
2291 // Emits an error if we don't have a valid set of target features for the
2292 // called function.
2293 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2294                                           const FunctionDecl *TargetDecl) {
2295   // Early exit if this is an indirect call.
2296   if (!TargetDecl)
2297     return;
2298 
2299   // Get the current enclosing function if it exists. If it doesn't
2300   // we can't check the target features anyhow.
2301   const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2302   if (!FD)
2303     return;
2304 
2305   // Grab the required features for the call. For a builtin this is listed in
2306   // the td file with the default cpu, for an always_inline function this is any
2307   // listed cpu and any listed features.
2308   unsigned BuiltinID = TargetDecl->getBuiltinID();
2309   std::string MissingFeature;
2310   if (BuiltinID) {
2311     SmallVector<StringRef, 1> ReqFeatures;
2312     const char *FeatureList =
2313         CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2314     // Return if the builtin doesn't have any required features.
2315     if (!FeatureList || StringRef(FeatureList) == "")
2316       return;
2317     StringRef(FeatureList).split(ReqFeatures, ",");
2318     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2319       CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2320           << TargetDecl->getDeclName()
2321           << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2322 
2323   } else if (TargetDecl->hasAttr<TargetAttr>()) {
2324     // Get the required features for the callee.
2325     SmallVector<StringRef, 1> ReqFeatures;
2326     llvm::StringMap<bool> CalleeFeatureMap;
2327     CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2328     for (const auto &F : CalleeFeatureMap) {
2329       // Only positive features are "required".
2330       if (F.getValue())
2331         ReqFeatures.push_back(F.getKey());
2332     }
2333     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2334       CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2335           << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2336   }
2337 }
2338 
2339 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2340   if (!CGM.getCodeGenOpts().SanitizeStats)
2341     return;
2342 
2343   llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2344   IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2345   CGM.getSanStats().create(IRB, SSK);
2346 }
2347 
2348 llvm::Value *
2349 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
2350   llvm::Value *TrueCondition = nullptr;
2351   if (!RO.ParsedAttribute.Architecture.empty())
2352     TrueCondition = EmitX86CpuIs(RO.ParsedAttribute.Architecture);
2353 
2354   if (!RO.ParsedAttribute.Features.empty()) {
2355     SmallVector<StringRef, 8> FeatureList;
2356     llvm::for_each(RO.ParsedAttribute.Features,
2357                    [&FeatureList](const std::string &Feature) {
2358                      FeatureList.push_back(StringRef{Feature}.substr(1));
2359                    });
2360     llvm::Value *FeatureCmp = EmitX86CpuSupports(FeatureList);
2361     TrueCondition = TrueCondition ? Builder.CreateAnd(TrueCondition, FeatureCmp)
2362                                   : FeatureCmp;
2363   }
2364   return TrueCondition;
2365 }
2366 
2367 void CodeGenFunction::EmitMultiVersionResolver(
2368     llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
2369   assert((getContext().getTargetInfo().getTriple().getArch() ==
2370               llvm::Triple::x86 ||
2371           getContext().getTargetInfo().getTriple().getArch() ==
2372               llvm::Triple::x86_64) &&
2373          "Only implemented for x86 targets");
2374 
2375   // Main function's basic block.
2376   llvm::BasicBlock *CurBlock = createBasicBlock("entry", Resolver);
2377   Builder.SetInsertPoint(CurBlock);
2378   EmitX86CpuInit();
2379 
2380   llvm::Function *DefaultFunc = nullptr;
2381   for (const MultiVersionResolverOption &RO : Options) {
2382     Builder.SetInsertPoint(CurBlock);
2383     llvm::Value *TrueCondition = FormResolverCondition(RO);
2384 
2385     if (!TrueCondition) {
2386       DefaultFunc = RO.Function;
2387     } else {
2388       llvm::BasicBlock *RetBlock = createBasicBlock("ro_ret", Resolver);
2389       llvm::IRBuilder<> RetBuilder(RetBlock);
2390       RetBuilder.CreateRet(RO.Function);
2391       CurBlock = createBasicBlock("ro_else", Resolver);
2392       Builder.CreateCondBr(TrueCondition, RetBlock, CurBlock);
2393     }
2394   }
2395 
2396   assert(DefaultFunc && "No default version?");
2397   // Emit return from the 'else-ist' block.
2398   Builder.SetInsertPoint(CurBlock);
2399   Builder.CreateRet(DefaultFunc);
2400 }
2401 
2402 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2403   if (CGDebugInfo *DI = getDebugInfo())
2404     return DI->SourceLocToDebugLoc(Location);
2405 
2406   return llvm::DebugLoc();
2407 }
2408