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