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   // Apply fuzzing attribute to the function.
866   if (SanOpts.hasOneOf(SanitizerKind::Fuzzer | SanitizerKind::FuzzerNoLink))
867     Fn->addFnAttr(llvm::Attribute::OptForFuzzing);
868 
869   // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
870   // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
871   if (SanOpts.has(SanitizerKind::Thread)) {
872     if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
873       IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
874       if (OMD->getMethodFamily() == OMF_dealloc ||
875           OMD->getMethodFamily() == OMF_initialize ||
876           (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
877         markAsIgnoreThreadCheckingAtRuntime(Fn);
878       }
879     } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) {
880       IdentifierInfo *II = FD->getIdentifier();
881       if (II && II->isStr("__destroy_helper_block_"))
882         markAsIgnoreThreadCheckingAtRuntime(Fn);
883     }
884   }
885 
886   // Ignore unrelated casts in STL allocate() since the allocator must cast
887   // from void* to T* before object initialization completes. Don't match on the
888   // namespace because not all allocators are in std::
889   if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
890     if (matchesStlAllocatorFn(D, getContext()))
891       SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast;
892   }
893 
894   // Apply xray attributes to the function (as a string, for now)
895   bool InstrumentXray = ShouldXRayInstrumentFunction();
896   if (D && InstrumentXray) {
897     if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
898       if (XRayAttr->alwaysXRayInstrument())
899         Fn->addFnAttr("function-instrument", "xray-always");
900       if (XRayAttr->neverXRayInstrument())
901         Fn->addFnAttr("function-instrument", "xray-never");
902       if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) {
903         Fn->addFnAttr("xray-log-args",
904                       llvm::utostr(LogArgs->getArgumentCount()));
905       }
906     } else {
907       if (!CGM.imbueXRayAttrs(Fn, Loc))
908         Fn->addFnAttr(
909             "xray-instruction-threshold",
910             llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
911     }
912   }
913 
914   // Add no-jump-tables value.
915   Fn->addFnAttr("no-jump-tables",
916                 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
917 
918   // Add profile-sample-accurate value.
919   if (CGM.getCodeGenOpts().ProfileSampleAccurate)
920     Fn->addFnAttr("profile-sample-accurate");
921 
922   if (getLangOpts().OpenCL) {
923     // Add metadata for a kernel function.
924     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
925       EmitOpenCLKernelMetadata(FD, Fn);
926   }
927 
928   // If we are checking function types, emit a function type signature as
929   // prologue data.
930   if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
931     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
932       if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
933         // Remove any (C++17) exception specifications, to allow calling e.g. a
934         // noexcept function through a non-noexcept pointer.
935         auto ProtoTy =
936           getContext().getFunctionTypeWithExceptionSpec(FD->getType(),
937                                                         EST_None);
938         llvm::Constant *FTRTTIConst =
939             CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true);
940         llvm::Constant *FTRTTIConstEncoded =
941             EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
942         llvm::Constant *PrologueStructElems[] = {PrologueSig,
943                                                  FTRTTIConstEncoded};
944         llvm::Constant *PrologueStructConst =
945             llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
946         Fn->setPrologueData(PrologueStructConst);
947       }
948     }
949   }
950 
951   // If we're checking nullability, we need to know whether we can check the
952   // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
953   if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
954     auto Nullability = FnRetTy->getNullability(getContext());
955     if (Nullability && *Nullability == NullabilityKind::NonNull) {
956       if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
957             CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
958         RetValNullabilityPrecondition =
959             llvm::ConstantInt::getTrue(getLLVMContext());
960     }
961   }
962 
963   // If we're in C++ mode and the function name is "main", it is guaranteed
964   // to be norecurse by the standard (3.6.1.3 "The function main shall not be
965   // used within a program").
966   if (getLangOpts().CPlusPlus)
967     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
968       if (FD->isMain())
969         Fn->addFnAttr(llvm::Attribute::NoRecurse);
970 
971   llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
972 
973   // Create a marker to make it easy to insert allocas into the entryblock
974   // later.  Don't create this with the builder, because we don't want it
975   // folded.
976   llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
977   AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
978 
979   ReturnBlock = getJumpDestInCurrentScope("return");
980 
981   Builder.SetInsertPoint(EntryBB);
982 
983   // If we're checking the return value, allocate space for a pointer to a
984   // precise source location of the checked return statement.
985   if (requiresReturnValueCheck()) {
986     ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
987     InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
988   }
989 
990   // Emit subprogram debug descriptor.
991   if (CGDebugInfo *DI = getDebugInfo()) {
992     // Reconstruct the type from the argument list so that implicit parameters,
993     // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
994     // convention.
995     CallingConv CC = CallingConv::CC_C;
996     if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
997       if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
998         CC = SrcFnTy->getCallConv();
999     SmallVector<QualType, 16> ArgTypes;
1000     for (const VarDecl *VD : Args)
1001       ArgTypes.push_back(VD->getType());
1002     QualType FnType = getContext().getFunctionType(
1003         RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
1004     DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
1005   }
1006 
1007   if (ShouldInstrumentFunction()) {
1008     if (CGM.getCodeGenOpts().InstrumentFunctions)
1009       CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
1010     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
1011       CurFn->addFnAttr("instrument-function-entry-inlined",
1012                        "__cyg_profile_func_enter");
1013     if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
1014       CurFn->addFnAttr("instrument-function-entry-inlined",
1015                        "__cyg_profile_func_enter_bare");
1016   }
1017 
1018   // Since emitting the mcount call here impacts optimizations such as function
1019   // inlining, we just add an attribute to insert a mcount call in backend.
1020   // The attribute "counting-function" is set to mcount function name which is
1021   // architecture dependent.
1022   if (CGM.getCodeGenOpts().InstrumentForProfiling) {
1023     // Calls to fentry/mcount should not be generated if function has
1024     // the no_instrument_function attribute.
1025     if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
1026       if (CGM.getCodeGenOpts().CallFEntry)
1027         Fn->addFnAttr("fentry-call", "true");
1028       else {
1029         Fn->addFnAttr("instrument-function-entry-inlined",
1030                       getTarget().getMCountName());
1031       }
1032     }
1033   }
1034 
1035   if (RetTy->isVoidType()) {
1036     // Void type; nothing to return.
1037     ReturnValue = Address::invalid();
1038 
1039     // Count the implicit return.
1040     if (!endsWithReturn(D))
1041       ++NumReturnExprs;
1042   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
1043              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1044     // Indirect aggregate return; emit returned value directly into sret slot.
1045     // This reduces code size, and affects correctness in C++.
1046     auto AI = CurFn->arg_begin();
1047     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
1048       ++AI;
1049     ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
1050   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
1051              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1052     // Load the sret pointer from the argument struct and return into that.
1053     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
1054     llvm::Function::arg_iterator EI = CurFn->arg_end();
1055     --EI;
1056     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
1057     Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
1058     ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
1059   } else {
1060     ReturnValue = CreateIRTemp(RetTy, "retval");
1061 
1062     // Tell the epilog emitter to autorelease the result.  We do this
1063     // now so that various specialized functions can suppress it
1064     // during their IR-generation.
1065     if (getLangOpts().ObjCAutoRefCount &&
1066         !CurFnInfo->isReturnsRetained() &&
1067         RetTy->isObjCRetainableType())
1068       AutoreleaseResult = true;
1069   }
1070 
1071   EmitStartEHSpec(CurCodeDecl);
1072 
1073   PrologueCleanupDepth = EHStack.stable_begin();
1074 
1075   // Emit OpenMP specific initialization of the device functions.
1076   if (getLangOpts().OpenMP && CurCodeDecl)
1077     CGM.getOpenMPRuntime().emitFunctionProlog(*this, CurCodeDecl);
1078 
1079   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
1080 
1081   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
1082     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
1083     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
1084     if (MD->getParent()->isLambda() &&
1085         MD->getOverloadedOperator() == OO_Call) {
1086       // We're in a lambda; figure out the captures.
1087       MD->getParent()->getCaptureFields(LambdaCaptureFields,
1088                                         LambdaThisCaptureField);
1089       if (LambdaThisCaptureField) {
1090         // If the lambda captures the object referred to by '*this' - either by
1091         // value or by reference, make sure CXXThisValue points to the correct
1092         // object.
1093 
1094         // Get the lvalue for the field (which is a copy of the enclosing object
1095         // or contains the address of the enclosing object).
1096         LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
1097         if (!LambdaThisCaptureField->getType()->isPointerType()) {
1098           // If the enclosing object was captured by value, just use its address.
1099           CXXThisValue = ThisFieldLValue.getAddress().getPointer();
1100         } else {
1101           // Load the lvalue pointed to by the field, since '*this' was captured
1102           // by reference.
1103           CXXThisValue =
1104               EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
1105         }
1106       }
1107       for (auto *FD : MD->getParent()->fields()) {
1108         if (FD->hasCapturedVLAType()) {
1109           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
1110                                            SourceLocation()).getScalarVal();
1111           auto VAT = FD->getCapturedVLAType();
1112           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
1113         }
1114       }
1115     } else {
1116       // Not in a lambda; just use 'this' from the method.
1117       // FIXME: Should we generate a new load for each use of 'this'?  The
1118       // fast register allocator would be happier...
1119       CXXThisValue = CXXABIThisValue;
1120     }
1121 
1122     // Check the 'this' pointer once per function, if it's available.
1123     if (CXXABIThisValue) {
1124       SanitizerSet SkippedChecks;
1125       SkippedChecks.set(SanitizerKind::ObjectSize, true);
1126       QualType ThisTy = MD->getThisType(getContext());
1127 
1128       // If this is the call operator of a lambda with no capture-default, it
1129       // may have a static invoker function, which may call this operator with
1130       // a null 'this' pointer.
1131       if (isLambdaCallOperator(MD) &&
1132           MD->getParent()->getLambdaCaptureDefault() == LCD_None)
1133         SkippedChecks.set(SanitizerKind::Null, true);
1134 
1135       EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1136                                                 : TCK_MemberCall,
1137                     Loc, CXXABIThisValue, ThisTy,
1138                     getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1139                     SkippedChecks);
1140     }
1141   }
1142 
1143   // If any of the arguments have a variably modified type, make sure to
1144   // emit the type size.
1145   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1146        i != e; ++i) {
1147     const VarDecl *VD = *i;
1148 
1149     // Dig out the type as written from ParmVarDecls; it's unclear whether
1150     // the standard (C99 6.9.1p10) requires this, but we're following the
1151     // precedent set by gcc.
1152     QualType Ty;
1153     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1154       Ty = PVD->getOriginalType();
1155     else
1156       Ty = VD->getType();
1157 
1158     if (Ty->isVariablyModifiedType())
1159       EmitVariablyModifiedType(Ty);
1160   }
1161   // Emit a location at the end of the prologue.
1162   if (CGDebugInfo *DI = getDebugInfo())
1163     DI->EmitLocation(Builder, StartLoc);
1164 }
1165 
1166 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1167                                        const Stmt *Body) {
1168   incrementProfileCounter(Body);
1169   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1170     EmitCompoundStmtWithoutScope(*S);
1171   else
1172     EmitStmt(Body);
1173 }
1174 
1175 /// When instrumenting to collect profile data, the counts for some blocks
1176 /// such as switch cases need to not include the fall-through counts, so
1177 /// emit a branch around the instrumentation code. When not instrumenting,
1178 /// this just calls EmitBlock().
1179 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1180                                                const Stmt *S) {
1181   llvm::BasicBlock *SkipCountBB = nullptr;
1182   if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1183     // When instrumenting for profiling, the fallthrough to certain
1184     // statements needs to skip over the instrumentation code so that we
1185     // get an accurate count.
1186     SkipCountBB = createBasicBlock("skipcount");
1187     EmitBranch(SkipCountBB);
1188   }
1189   EmitBlock(BB);
1190   uint64_t CurrentCount = getCurrentProfileCount();
1191   incrementProfileCounter(S);
1192   setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1193   if (SkipCountBB)
1194     EmitBlock(SkipCountBB);
1195 }
1196 
1197 /// Tries to mark the given function nounwind based on the
1198 /// non-existence of any throwing calls within it.  We believe this is
1199 /// lightweight enough to do at -O0.
1200 static void TryMarkNoThrow(llvm::Function *F) {
1201   // LLVM treats 'nounwind' on a function as part of the type, so we
1202   // can't do this on functions that can be overwritten.
1203   if (F->isInterposable()) return;
1204 
1205   for (llvm::BasicBlock &BB : *F)
1206     for (llvm::Instruction &I : BB)
1207       if (I.mayThrow())
1208         return;
1209 
1210   F->setDoesNotThrow();
1211 }
1212 
1213 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1214                                                FunctionArgList &Args) {
1215   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1216   QualType ResTy = FD->getReturnType();
1217 
1218   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1219   if (MD && MD->isInstance()) {
1220     if (CGM.getCXXABI().HasThisReturn(GD))
1221       ResTy = MD->getThisType(getContext());
1222     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1223       ResTy = CGM.getContext().VoidPtrTy;
1224     CGM.getCXXABI().buildThisParam(*this, Args);
1225   }
1226 
1227   // The base version of an inheriting constructor whose constructed base is a
1228   // virtual base is not passed any arguments (because it doesn't actually call
1229   // the inherited constructor).
1230   bool PassedParams = true;
1231   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1232     if (auto Inherited = CD->getInheritedConstructor())
1233       PassedParams =
1234           getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1235 
1236   if (PassedParams) {
1237     for (auto *Param : FD->parameters()) {
1238       Args.push_back(Param);
1239       if (!Param->hasAttr<PassObjectSizeAttr>())
1240         continue;
1241 
1242       auto *Implicit = ImplicitParamDecl::Create(
1243           getContext(), Param->getDeclContext(), Param->getLocation(),
1244           /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1245       SizeArguments[Param] = Implicit;
1246       Args.push_back(Implicit);
1247     }
1248   }
1249 
1250   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1251     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1252 
1253   return ResTy;
1254 }
1255 
1256 static bool
1257 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1258                                              const ASTContext &Context) {
1259   QualType T = FD->getReturnType();
1260   // Avoid the optimization for functions that return a record type with a
1261   // trivial destructor or another trivially copyable type.
1262   if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1263     if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1264       return !ClassDecl->hasTrivialDestructor();
1265   }
1266   return !T.isTriviallyCopyableType(Context);
1267 }
1268 
1269 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1270                                    const CGFunctionInfo &FnInfo) {
1271   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1272   CurGD = GD;
1273 
1274   FunctionArgList Args;
1275   QualType ResTy = BuildFunctionArgList(GD, Args);
1276 
1277   // Check if we should generate debug info for this function.
1278   if (FD->hasAttr<NoDebugAttr>())
1279     DebugInfo = nullptr; // disable debug info indefinitely for this function
1280 
1281   // The function might not have a body if we're generating thunks for a
1282   // function declaration.
1283   SourceRange BodyRange;
1284   if (Stmt *Body = FD->getBody())
1285     BodyRange = Body->getSourceRange();
1286   else
1287     BodyRange = FD->getLocation();
1288   CurEHLocation = BodyRange.getEnd();
1289 
1290   // Use the location of the start of the function to determine where
1291   // the function definition is located. By default use the location
1292   // of the declaration as the location for the subprogram. A function
1293   // may lack a declaration in the source code if it is created by code
1294   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1295   SourceLocation Loc = FD->getLocation();
1296 
1297   // If this is a function specialization then use the pattern body
1298   // as the location for the function.
1299   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1300     if (SpecDecl->hasBody(SpecDecl))
1301       Loc = SpecDecl->getLocation();
1302 
1303   Stmt *Body = FD->getBody();
1304 
1305   // Initialize helper which will detect jumps which can cause invalid lifetime
1306   // markers.
1307   if (Body && ShouldEmitLifetimeMarkers)
1308     Bypasses.Init(Body);
1309 
1310   // Emit the standard function prologue.
1311   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1312 
1313   // Generate the body of the function.
1314   PGO.assignRegionCounters(GD, CurFn);
1315   if (isa<CXXDestructorDecl>(FD))
1316     EmitDestructorBody(Args);
1317   else if (isa<CXXConstructorDecl>(FD))
1318     EmitConstructorBody(Args);
1319   else if (getLangOpts().CUDA &&
1320            !getLangOpts().CUDAIsDevice &&
1321            FD->hasAttr<CUDAGlobalAttr>())
1322     CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1323   else if (isa<CXXMethodDecl>(FD) &&
1324            cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1325     // The lambda static invoker function is special, because it forwards or
1326     // clones the body of the function call operator (but is actually static).
1327     EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1328   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1329              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1330               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1331     // Implicit copy-assignment gets the same special treatment as implicit
1332     // copy-constructors.
1333     emitImplicitAssignmentOperatorBody(Args);
1334   } else if (Body) {
1335     EmitFunctionBody(Args, Body);
1336   } else
1337     llvm_unreachable("no definition for emitted function");
1338 
1339   // C++11 [stmt.return]p2:
1340   //   Flowing off the end of a function [...] results in undefined behavior in
1341   //   a value-returning function.
1342   // C11 6.9.1p12:
1343   //   If the '}' that terminates a function is reached, and the value of the
1344   //   function call is used by the caller, the behavior is undefined.
1345   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1346       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1347     bool ShouldEmitUnreachable =
1348         CGM.getCodeGenOpts().StrictReturn ||
1349         shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1350     if (SanOpts.has(SanitizerKind::Return)) {
1351       SanitizerScope SanScope(this);
1352       llvm::Value *IsFalse = Builder.getFalse();
1353       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1354                 SanitizerHandler::MissingReturn,
1355                 EmitCheckSourceLocation(FD->getLocation()), None);
1356     } else if (ShouldEmitUnreachable) {
1357       if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1358         EmitTrapCall(llvm::Intrinsic::trap);
1359     }
1360     if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1361       Builder.CreateUnreachable();
1362       Builder.ClearInsertionPoint();
1363     }
1364   }
1365 
1366   // Emit the standard function epilogue.
1367   FinishFunction(BodyRange.getEnd());
1368 
1369   // If we haven't marked the function nothrow through other means, do
1370   // a quick pass now to see if we can.
1371   if (!CurFn->doesNotThrow())
1372     TryMarkNoThrow(CurFn);
1373 }
1374 
1375 /// ContainsLabel - Return true if the statement contains a label in it.  If
1376 /// this statement is not executed normally, it not containing a label means
1377 /// that we can just remove the code.
1378 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1379   // Null statement, not a label!
1380   if (!S) return false;
1381 
1382   // If this is a label, we have to emit the code, consider something like:
1383   // if (0) {  ...  foo:  bar(); }  goto foo;
1384   //
1385   // TODO: If anyone cared, we could track __label__'s, since we know that you
1386   // can't jump to one from outside their declared region.
1387   if (isa<LabelStmt>(S))
1388     return true;
1389 
1390   // If this is a case/default statement, and we haven't seen a switch, we have
1391   // to emit the code.
1392   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1393     return true;
1394 
1395   // If this is a switch statement, we want to ignore cases below it.
1396   if (isa<SwitchStmt>(S))
1397     IgnoreCaseStmts = true;
1398 
1399   // Scan subexpressions for verboten labels.
1400   for (const Stmt *SubStmt : S->children())
1401     if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1402       return true;
1403 
1404   return false;
1405 }
1406 
1407 /// containsBreak - Return true if the statement contains a break out of it.
1408 /// If the statement (recursively) contains a switch or loop with a break
1409 /// inside of it, this is fine.
1410 bool CodeGenFunction::containsBreak(const Stmt *S) {
1411   // Null statement, not a label!
1412   if (!S) return false;
1413 
1414   // If this is a switch or loop that defines its own break scope, then we can
1415   // include it and anything inside of it.
1416   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1417       isa<ForStmt>(S))
1418     return false;
1419 
1420   if (isa<BreakStmt>(S))
1421     return true;
1422 
1423   // Scan subexpressions for verboten breaks.
1424   for (const Stmt *SubStmt : S->children())
1425     if (containsBreak(SubStmt))
1426       return true;
1427 
1428   return false;
1429 }
1430 
1431 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1432   if (!S) return false;
1433 
1434   // Some statement kinds add a scope and thus never add a decl to the current
1435   // scope. Note, this list is longer than the list of statements that might
1436   // have an unscoped decl nested within them, but this way is conservatively
1437   // correct even if more statement kinds are added.
1438   if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1439       isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1440       isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1441       isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1442     return false;
1443 
1444   if (isa<DeclStmt>(S))
1445     return true;
1446 
1447   for (const Stmt *SubStmt : S->children())
1448     if (mightAddDeclToScope(SubStmt))
1449       return true;
1450 
1451   return false;
1452 }
1453 
1454 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1455 /// to a constant, or if it does but contains a label, return false.  If it
1456 /// constant folds return true and set the boolean result in Result.
1457 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1458                                                    bool &ResultBool,
1459                                                    bool AllowLabels) {
1460   llvm::APSInt ResultInt;
1461   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1462     return false;
1463 
1464   ResultBool = ResultInt.getBoolValue();
1465   return true;
1466 }
1467 
1468 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1469 /// to a constant, or if it does but contains a label, return false.  If it
1470 /// constant folds return true and set the folded value.
1471 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1472                                                    llvm::APSInt &ResultInt,
1473                                                    bool AllowLabels) {
1474   // FIXME: Rename and handle conversion of other evaluatable things
1475   // to bool.
1476   llvm::APSInt Int;
1477   if (!Cond->EvaluateAsInt(Int, getContext()))
1478     return false;  // Not foldable, not integer or not fully evaluatable.
1479 
1480   if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1481     return false;  // Contains a label.
1482 
1483   ResultInt = Int;
1484   return true;
1485 }
1486 
1487 
1488 
1489 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1490 /// statement) to the specified blocks.  Based on the condition, this might try
1491 /// to simplify the codegen of the conditional based on the branch.
1492 ///
1493 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1494                                            llvm::BasicBlock *TrueBlock,
1495                                            llvm::BasicBlock *FalseBlock,
1496                                            uint64_t TrueCount) {
1497   Cond = Cond->IgnoreParens();
1498 
1499   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1500 
1501     // Handle X && Y in a condition.
1502     if (CondBOp->getOpcode() == BO_LAnd) {
1503       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1504       // folded if the case was simple enough.
1505       bool ConstantBool = false;
1506       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1507           ConstantBool) {
1508         // br(1 && X) -> br(X).
1509         incrementProfileCounter(CondBOp);
1510         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1511                                     TrueCount);
1512       }
1513 
1514       // If we have "X && 1", simplify the code to use an uncond branch.
1515       // "X && 0" would have been constant folded to 0.
1516       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1517           ConstantBool) {
1518         // br(X && 1) -> br(X).
1519         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1520                                     TrueCount);
1521       }
1522 
1523       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1524       // want to jump to the FalseBlock.
1525       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1526       // The counter tells us how often we evaluate RHS, and all of TrueCount
1527       // can be propagated to that branch.
1528       uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1529 
1530       ConditionalEvaluation eval(*this);
1531       {
1532         ApplyDebugLocation DL(*this, Cond);
1533         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1534         EmitBlock(LHSTrue);
1535       }
1536 
1537       incrementProfileCounter(CondBOp);
1538       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1539 
1540       // Any temporaries created here are conditional.
1541       eval.begin(*this);
1542       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1543       eval.end(*this);
1544 
1545       return;
1546     }
1547 
1548     if (CondBOp->getOpcode() == BO_LOr) {
1549       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1550       // folded if the case was simple enough.
1551       bool ConstantBool = false;
1552       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1553           !ConstantBool) {
1554         // br(0 || X) -> br(X).
1555         incrementProfileCounter(CondBOp);
1556         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1557                                     TrueCount);
1558       }
1559 
1560       // If we have "X || 0", simplify the code to use an uncond branch.
1561       // "X || 1" would have been constant folded to 1.
1562       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1563           !ConstantBool) {
1564         // br(X || 0) -> br(X).
1565         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1566                                     TrueCount);
1567       }
1568 
1569       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1570       // want to jump to the TrueBlock.
1571       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1572       // We have the count for entry to the RHS and for the whole expression
1573       // being true, so we can divy up True count between the short circuit and
1574       // the RHS.
1575       uint64_t LHSCount =
1576           getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1577       uint64_t RHSCount = TrueCount - LHSCount;
1578 
1579       ConditionalEvaluation eval(*this);
1580       {
1581         ApplyDebugLocation DL(*this, Cond);
1582         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1583         EmitBlock(LHSFalse);
1584       }
1585 
1586       incrementProfileCounter(CondBOp);
1587       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1588 
1589       // Any temporaries created here are conditional.
1590       eval.begin(*this);
1591       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1592 
1593       eval.end(*this);
1594 
1595       return;
1596     }
1597   }
1598 
1599   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1600     // br(!x, t, f) -> br(x, f, t)
1601     if (CondUOp->getOpcode() == UO_LNot) {
1602       // Negate the count.
1603       uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1604       // Negate the condition and swap the destination blocks.
1605       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1606                                   FalseCount);
1607     }
1608   }
1609 
1610   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1611     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1612     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1613     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1614 
1615     ConditionalEvaluation cond(*this);
1616     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1617                          getProfileCount(CondOp));
1618 
1619     // When computing PGO branch weights, we only know the overall count for
1620     // the true block. This code is essentially doing tail duplication of the
1621     // naive code-gen, introducing new edges for which counts are not
1622     // available. Divide the counts proportionally between the LHS and RHS of
1623     // the conditional operator.
1624     uint64_t LHSScaledTrueCount = 0;
1625     if (TrueCount) {
1626       double LHSRatio =
1627           getProfileCount(CondOp) / (double)getCurrentProfileCount();
1628       LHSScaledTrueCount = TrueCount * LHSRatio;
1629     }
1630 
1631     cond.begin(*this);
1632     EmitBlock(LHSBlock);
1633     incrementProfileCounter(CondOp);
1634     {
1635       ApplyDebugLocation DL(*this, Cond);
1636       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1637                            LHSScaledTrueCount);
1638     }
1639     cond.end(*this);
1640 
1641     cond.begin(*this);
1642     EmitBlock(RHSBlock);
1643     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1644                          TrueCount - LHSScaledTrueCount);
1645     cond.end(*this);
1646 
1647     return;
1648   }
1649 
1650   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1651     // Conditional operator handling can give us a throw expression as a
1652     // condition for a case like:
1653     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1654     // Fold this to:
1655     //   br(c, throw x, br(y, t, f))
1656     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1657     return;
1658   }
1659 
1660   // If the branch has a condition wrapped by __builtin_unpredictable,
1661   // create metadata that specifies that the branch is unpredictable.
1662   // Don't bother if not optimizing because that metadata would not be used.
1663   llvm::MDNode *Unpredictable = nullptr;
1664   auto *Call = dyn_cast<CallExpr>(Cond);
1665   if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1666     auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1667     if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1668       llvm::MDBuilder MDHelper(getLLVMContext());
1669       Unpredictable = MDHelper.createUnpredictable();
1670     }
1671   }
1672 
1673   // Create branch weights based on the number of times we get here and the
1674   // number of times the condition should be true.
1675   uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1676   llvm::MDNode *Weights =
1677       createProfileWeights(TrueCount, CurrentCount - TrueCount);
1678 
1679   // Emit the code with the fully general case.
1680   llvm::Value *CondV;
1681   {
1682     ApplyDebugLocation DL(*this, Cond);
1683     CondV = EvaluateExprAsBool(Cond);
1684   }
1685   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1686 }
1687 
1688 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1689 /// specified stmt yet.
1690 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1691   CGM.ErrorUnsupported(S, Type);
1692 }
1693 
1694 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1695 /// variable-length array whose elements have a non-zero bit-pattern.
1696 ///
1697 /// \param baseType the inner-most element type of the array
1698 /// \param src - a char* pointing to the bit-pattern for a single
1699 /// base element of the array
1700 /// \param sizeInChars - the total size of the VLA, in chars
1701 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1702                                Address dest, Address src,
1703                                llvm::Value *sizeInChars) {
1704   CGBuilderTy &Builder = CGF.Builder;
1705 
1706   CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1707   llvm::Value *baseSizeInChars
1708     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1709 
1710   Address begin =
1711     Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1712   llvm::Value *end =
1713     Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1714 
1715   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1716   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1717   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1718 
1719   // Make a loop over the VLA.  C99 guarantees that the VLA element
1720   // count must be nonzero.
1721   CGF.EmitBlock(loopBB);
1722 
1723   llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1724   cur->addIncoming(begin.getPointer(), originBB);
1725 
1726   CharUnits curAlign =
1727     dest.getAlignment().alignmentOfArrayElement(baseSize);
1728 
1729   // memcpy the individual element bit-pattern.
1730   Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1731                        /*volatile*/ false);
1732 
1733   // Go to the next element.
1734   llvm::Value *next =
1735     Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1736 
1737   // Leave if that's the end of the VLA.
1738   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1739   Builder.CreateCondBr(done, contBB, loopBB);
1740   cur->addIncoming(next, loopBB);
1741 
1742   CGF.EmitBlock(contBB);
1743 }
1744 
1745 void
1746 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1747   // Ignore empty classes in C++.
1748   if (getLangOpts().CPlusPlus) {
1749     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1750       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1751         return;
1752     }
1753   }
1754 
1755   // Cast the dest ptr to the appropriate i8 pointer type.
1756   if (DestPtr.getElementType() != Int8Ty)
1757     DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1758 
1759   // Get size and alignment info for this aggregate.
1760   CharUnits size = getContext().getTypeSizeInChars(Ty);
1761 
1762   llvm::Value *SizeVal;
1763   const VariableArrayType *vla;
1764 
1765   // Don't bother emitting a zero-byte memset.
1766   if (size.isZero()) {
1767     // But note that getTypeInfo returns 0 for a VLA.
1768     if (const VariableArrayType *vlaType =
1769           dyn_cast_or_null<VariableArrayType>(
1770                                           getContext().getAsArrayType(Ty))) {
1771       auto VlaSize = getVLASize(vlaType);
1772       SizeVal = VlaSize.NumElts;
1773       CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type);
1774       if (!eltSize.isOne())
1775         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1776       vla = vlaType;
1777     } else {
1778       return;
1779     }
1780   } else {
1781     SizeVal = CGM.getSize(size);
1782     vla = nullptr;
1783   }
1784 
1785   // If the type contains a pointer to data member we can't memset it to zero.
1786   // Instead, create a null constant and copy it to the destination.
1787   // TODO: there are other patterns besides zero that we can usefully memset,
1788   // like -1, which happens to be the pattern used by member-pointers.
1789   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1790     // For a VLA, emit a single element, then splat that over the VLA.
1791     if (vla) Ty = getContext().getBaseElementType(vla);
1792 
1793     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1794 
1795     llvm::GlobalVariable *NullVariable =
1796       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1797                                /*isConstant=*/true,
1798                                llvm::GlobalVariable::PrivateLinkage,
1799                                NullConstant, Twine());
1800     CharUnits NullAlign = DestPtr.getAlignment();
1801     NullVariable->setAlignment(NullAlign.getQuantity());
1802     Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1803                    NullAlign);
1804 
1805     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1806 
1807     // Get and call the appropriate llvm.memcpy overload.
1808     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1809     return;
1810   }
1811 
1812   // Otherwise, just memset the whole thing to zero.  This is legal
1813   // because in LLVM, all default initializers (other than the ones we just
1814   // handled above) are guaranteed to have a bit pattern of all zeros.
1815   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1816 }
1817 
1818 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1819   // Make sure that there is a block for the indirect goto.
1820   if (!IndirectBranch)
1821     GetIndirectGotoBlock();
1822 
1823   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1824 
1825   // Make sure the indirect branch includes all of the address-taken blocks.
1826   IndirectBranch->addDestination(BB);
1827   return llvm::BlockAddress::get(CurFn, BB);
1828 }
1829 
1830 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1831   // If we already made the indirect branch for indirect goto, return its block.
1832   if (IndirectBranch) return IndirectBranch->getParent();
1833 
1834   CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1835 
1836   // Create the PHI node that indirect gotos will add entries to.
1837   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1838                                               "indirect.goto.dest");
1839 
1840   // Create the indirect branch instruction.
1841   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1842   return IndirectBranch->getParent();
1843 }
1844 
1845 /// Computes the length of an array in elements, as well as the base
1846 /// element type and a properly-typed first element pointer.
1847 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1848                                               QualType &baseType,
1849                                               Address &addr) {
1850   const ArrayType *arrayType = origArrayType;
1851 
1852   // If it's a VLA, we have to load the stored size.  Note that
1853   // this is the size of the VLA in bytes, not its size in elements.
1854   llvm::Value *numVLAElements = nullptr;
1855   if (isa<VariableArrayType>(arrayType)) {
1856     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts;
1857 
1858     // Walk into all VLAs.  This doesn't require changes to addr,
1859     // which has type T* where T is the first non-VLA element type.
1860     do {
1861       QualType elementType = arrayType->getElementType();
1862       arrayType = getContext().getAsArrayType(elementType);
1863 
1864       // If we only have VLA components, 'addr' requires no adjustment.
1865       if (!arrayType) {
1866         baseType = elementType;
1867         return numVLAElements;
1868       }
1869     } while (isa<VariableArrayType>(arrayType));
1870 
1871     // We get out here only if we find a constant array type
1872     // inside the VLA.
1873   }
1874 
1875   // We have some number of constant-length arrays, so addr should
1876   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1877   // down to the first element of addr.
1878   SmallVector<llvm::Value*, 8> gepIndices;
1879 
1880   // GEP down to the array type.
1881   llvm::ConstantInt *zero = Builder.getInt32(0);
1882   gepIndices.push_back(zero);
1883 
1884   uint64_t countFromCLAs = 1;
1885   QualType eltType;
1886 
1887   llvm::ArrayType *llvmArrayType =
1888     dyn_cast<llvm::ArrayType>(addr.getElementType());
1889   while (llvmArrayType) {
1890     assert(isa<ConstantArrayType>(arrayType));
1891     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1892              == llvmArrayType->getNumElements());
1893 
1894     gepIndices.push_back(zero);
1895     countFromCLAs *= llvmArrayType->getNumElements();
1896     eltType = arrayType->getElementType();
1897 
1898     llvmArrayType =
1899       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1900     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1901     assert((!llvmArrayType || arrayType) &&
1902            "LLVM and Clang types are out-of-synch");
1903   }
1904 
1905   if (arrayType) {
1906     // From this point onwards, the Clang array type has been emitted
1907     // as some other type (probably a packed struct). Compute the array
1908     // size, and just emit the 'begin' expression as a bitcast.
1909     while (arrayType) {
1910       countFromCLAs *=
1911           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1912       eltType = arrayType->getElementType();
1913       arrayType = getContext().getAsArrayType(eltType);
1914     }
1915 
1916     llvm::Type *baseType = ConvertType(eltType);
1917     addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1918   } else {
1919     // Create the actual GEP.
1920     addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1921                                              gepIndices, "array.begin"),
1922                    addr.getAlignment());
1923   }
1924 
1925   baseType = eltType;
1926 
1927   llvm::Value *numElements
1928     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1929 
1930   // If we had any VLA dimensions, factor them in.
1931   if (numVLAElements)
1932     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1933 
1934   return numElements;
1935 }
1936 
1937 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) {
1938   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1939   assert(vla && "type was not a variable array type!");
1940   return getVLASize(vla);
1941 }
1942 
1943 CodeGenFunction::VlaSizePair
1944 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1945   // The number of elements so far; always size_t.
1946   llvm::Value *numElements = nullptr;
1947 
1948   QualType elementType;
1949   do {
1950     elementType = type->getElementType();
1951     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1952     assert(vlaSize && "no size for VLA!");
1953     assert(vlaSize->getType() == SizeTy);
1954 
1955     if (!numElements) {
1956       numElements = vlaSize;
1957     } else {
1958       // It's undefined behavior if this wraps around, so mark it that way.
1959       // FIXME: Teach -fsanitize=undefined to trap this.
1960       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1961     }
1962   } while ((type = getContext().getAsVariableArrayType(elementType)));
1963 
1964   return { numElements, elementType };
1965 }
1966 
1967 CodeGenFunction::VlaSizePair
1968 CodeGenFunction::getVLAElements1D(QualType type) {
1969   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1970   assert(vla && "type was not a variable array type!");
1971   return getVLAElements1D(vla);
1972 }
1973 
1974 CodeGenFunction::VlaSizePair
1975 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) {
1976   llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()];
1977   assert(VlaSize && "no size for VLA!");
1978   assert(VlaSize->getType() == SizeTy);
1979   return { VlaSize, Vla->getElementType() };
1980 }
1981 
1982 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1983   assert(type->isVariablyModifiedType() &&
1984          "Must pass variably modified type to EmitVLASizes!");
1985 
1986   EnsureInsertPoint();
1987 
1988   // We're going to walk down into the type and look for VLA
1989   // expressions.
1990   do {
1991     assert(type->isVariablyModifiedType());
1992 
1993     const Type *ty = type.getTypePtr();
1994     switch (ty->getTypeClass()) {
1995 
1996 #define TYPE(Class, Base)
1997 #define ABSTRACT_TYPE(Class, Base)
1998 #define NON_CANONICAL_TYPE(Class, Base)
1999 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
2000 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
2001 #include "clang/AST/TypeNodes.def"
2002       llvm_unreachable("unexpected dependent type!");
2003 
2004     // These types are never variably-modified.
2005     case Type::Builtin:
2006     case Type::Complex:
2007     case Type::Vector:
2008     case Type::ExtVector:
2009     case Type::Record:
2010     case Type::Enum:
2011     case Type::Elaborated:
2012     case Type::TemplateSpecialization:
2013     case Type::ObjCTypeParam:
2014     case Type::ObjCObject:
2015     case Type::ObjCInterface:
2016     case Type::ObjCObjectPointer:
2017       llvm_unreachable("type class is never variably-modified!");
2018 
2019     case Type::Adjusted:
2020       type = cast<AdjustedType>(ty)->getAdjustedType();
2021       break;
2022 
2023     case Type::Decayed:
2024       type = cast<DecayedType>(ty)->getPointeeType();
2025       break;
2026 
2027     case Type::Pointer:
2028       type = cast<PointerType>(ty)->getPointeeType();
2029       break;
2030 
2031     case Type::BlockPointer:
2032       type = cast<BlockPointerType>(ty)->getPointeeType();
2033       break;
2034 
2035     case Type::LValueReference:
2036     case Type::RValueReference:
2037       type = cast<ReferenceType>(ty)->getPointeeType();
2038       break;
2039 
2040     case Type::MemberPointer:
2041       type = cast<MemberPointerType>(ty)->getPointeeType();
2042       break;
2043 
2044     case Type::ConstantArray:
2045     case Type::IncompleteArray:
2046       // Losing element qualification here is fine.
2047       type = cast<ArrayType>(ty)->getElementType();
2048       break;
2049 
2050     case Type::VariableArray: {
2051       // Losing element qualification here is fine.
2052       const VariableArrayType *vat = cast<VariableArrayType>(ty);
2053 
2054       // Unknown size indication requires no size computation.
2055       // Otherwise, evaluate and record it.
2056       if (const Expr *size = vat->getSizeExpr()) {
2057         // It's possible that we might have emitted this already,
2058         // e.g. with a typedef and a pointer to it.
2059         llvm::Value *&entry = VLASizeMap[size];
2060         if (!entry) {
2061           llvm::Value *Size = EmitScalarExpr(size);
2062 
2063           // C11 6.7.6.2p5:
2064           //   If the size is an expression that is not an integer constant
2065           //   expression [...] each time it is evaluated it shall have a value
2066           //   greater than zero.
2067           if (SanOpts.has(SanitizerKind::VLABound) &&
2068               size->getType()->isSignedIntegerType()) {
2069             SanitizerScope SanScope(this);
2070             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
2071             llvm::Constant *StaticArgs[] = {
2072               EmitCheckSourceLocation(size->getLocStart()),
2073               EmitCheckTypeDescriptor(size->getType())
2074             };
2075             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
2076                                      SanitizerKind::VLABound),
2077                       SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
2078           }
2079 
2080           // Always zexting here would be wrong if it weren't
2081           // undefined behavior to have a negative bound.
2082           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
2083         }
2084       }
2085       type = vat->getElementType();
2086       break;
2087     }
2088 
2089     case Type::FunctionProto:
2090     case Type::FunctionNoProto:
2091       type = cast<FunctionType>(ty)->getReturnType();
2092       break;
2093 
2094     case Type::Paren:
2095     case Type::TypeOf:
2096     case Type::UnaryTransform:
2097     case Type::Attributed:
2098     case Type::SubstTemplateTypeParm:
2099     case Type::PackExpansion:
2100       // Keep walking after single level desugaring.
2101       type = type.getSingleStepDesugaredType(getContext());
2102       break;
2103 
2104     case Type::Typedef:
2105     case Type::Decltype:
2106     case Type::Auto:
2107     case Type::DeducedTemplateSpecialization:
2108       // Stop walking: nothing to do.
2109       return;
2110 
2111     case Type::TypeOfExpr:
2112       // Stop walking: emit typeof expression.
2113       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
2114       return;
2115 
2116     case Type::Atomic:
2117       type = cast<AtomicType>(ty)->getValueType();
2118       break;
2119 
2120     case Type::Pipe:
2121       type = cast<PipeType>(ty)->getElementType();
2122       break;
2123     }
2124   } while (type->isVariablyModifiedType());
2125 }
2126 
2127 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
2128   if (getContext().getBuiltinVaListType()->isArrayType())
2129     return EmitPointerWithAlignment(E);
2130   return EmitLValue(E).getAddress();
2131 }
2132 
2133 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
2134   return EmitLValue(E).getAddress();
2135 }
2136 
2137 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2138                                               const APValue &Init) {
2139   assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
2140   if (CGDebugInfo *Dbg = getDebugInfo())
2141     if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2142       Dbg->EmitGlobalVariable(E->getDecl(), Init);
2143 }
2144 
2145 CodeGenFunction::PeepholeProtection
2146 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2147   // At the moment, the only aggressive peephole we do in IR gen
2148   // is trunc(zext) folding, but if we add more, we can easily
2149   // extend this protection.
2150 
2151   if (!rvalue.isScalar()) return PeepholeProtection();
2152   llvm::Value *value = rvalue.getScalarVal();
2153   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2154 
2155   // Just make an extra bitcast.
2156   assert(HaveInsertPoint());
2157   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2158                                                   Builder.GetInsertBlock());
2159 
2160   PeepholeProtection protection;
2161   protection.Inst = inst;
2162   return protection;
2163 }
2164 
2165 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2166   if (!protection.Inst) return;
2167 
2168   // In theory, we could try to duplicate the peepholes now, but whatever.
2169   protection.Inst->eraseFromParent();
2170 }
2171 
2172 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2173                                                  llvm::Value *AnnotatedVal,
2174                                                  StringRef AnnotationStr,
2175                                                  SourceLocation Location) {
2176   llvm::Value *Args[4] = {
2177     AnnotatedVal,
2178     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2179     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2180     CGM.EmitAnnotationLineNo(Location)
2181   };
2182   return Builder.CreateCall(AnnotationFn, Args);
2183 }
2184 
2185 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2186   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2187   // FIXME We create a new bitcast for every annotation because that's what
2188   // llvm-gcc was doing.
2189   for (const auto *I : D->specific_attrs<AnnotateAttr>())
2190     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2191                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2192                        I->getAnnotation(), D->getLocation());
2193 }
2194 
2195 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2196                                               Address Addr) {
2197   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2198   llvm::Value *V = Addr.getPointer();
2199   llvm::Type *VTy = V->getType();
2200   llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2201                                     CGM.Int8PtrTy);
2202 
2203   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2204     // FIXME Always emit the cast inst so we can differentiate between
2205     // annotation on the first field of a struct and annotation on the struct
2206     // itself.
2207     if (VTy != CGM.Int8PtrTy)
2208       V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2209     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2210     V = Builder.CreateBitCast(V, VTy);
2211   }
2212 
2213   return Address(V, Addr.getAlignment());
2214 }
2215 
2216 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2217 
2218 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2219     : CGF(CGF) {
2220   assert(!CGF->IsSanitizerScope);
2221   CGF->IsSanitizerScope = true;
2222 }
2223 
2224 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2225   CGF->IsSanitizerScope = false;
2226 }
2227 
2228 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2229                                    const llvm::Twine &Name,
2230                                    llvm::BasicBlock *BB,
2231                                    llvm::BasicBlock::iterator InsertPt) const {
2232   LoopStack.InsertHelper(I);
2233   if (IsSanitizerScope)
2234     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2235 }
2236 
2237 void CGBuilderInserter::InsertHelper(
2238     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2239     llvm::BasicBlock::iterator InsertPt) const {
2240   llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2241   if (CGF)
2242     CGF->InsertHelper(I, Name, BB, InsertPt);
2243 }
2244 
2245 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2246                                 CodeGenModule &CGM, const FunctionDecl *FD,
2247                                 std::string &FirstMissing) {
2248   // If there aren't any required features listed then go ahead and return.
2249   if (ReqFeatures.empty())
2250     return false;
2251 
2252   // Now build up the set of caller features and verify that all the required
2253   // features are there.
2254   llvm::StringMap<bool> CallerFeatureMap;
2255   CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2256 
2257   // If we have at least one of the features in the feature list return
2258   // true, otherwise return false.
2259   return std::all_of(
2260       ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2261         SmallVector<StringRef, 1> OrFeatures;
2262         Feature.split(OrFeatures, "|");
2263         return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2264                            [&](StringRef Feature) {
2265                              if (!CallerFeatureMap.lookup(Feature)) {
2266                                FirstMissing = Feature.str();
2267                                return false;
2268                              }
2269                              return true;
2270                            });
2271       });
2272 }
2273 
2274 // Emits an error if we don't have a valid set of target features for the
2275 // called function.
2276 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2277                                           const FunctionDecl *TargetDecl) {
2278   // Early exit if this is an indirect call.
2279   if (!TargetDecl)
2280     return;
2281 
2282   // Get the current enclosing function if it exists. If it doesn't
2283   // we can't check the target features anyhow.
2284   const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2285   if (!FD)
2286     return;
2287 
2288   // Grab the required features for the call. For a builtin this is listed in
2289   // the td file with the default cpu, for an always_inline function this is any
2290   // listed cpu and any listed features.
2291   unsigned BuiltinID = TargetDecl->getBuiltinID();
2292   std::string MissingFeature;
2293   if (BuiltinID) {
2294     SmallVector<StringRef, 1> ReqFeatures;
2295     const char *FeatureList =
2296         CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2297     // Return if the builtin doesn't have any required features.
2298     if (!FeatureList || StringRef(FeatureList) == "")
2299       return;
2300     StringRef(FeatureList).split(ReqFeatures, ",");
2301     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2302       CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2303           << TargetDecl->getDeclName()
2304           << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2305 
2306   } else if (TargetDecl->hasAttr<TargetAttr>()) {
2307     // Get the required features for the callee.
2308     SmallVector<StringRef, 1> ReqFeatures;
2309     llvm::StringMap<bool> CalleeFeatureMap;
2310     CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2311     for (const auto &F : CalleeFeatureMap) {
2312       // Only positive features are "required".
2313       if (F.getValue())
2314         ReqFeatures.push_back(F.getKey());
2315     }
2316     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2317       CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2318           << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2319   }
2320 }
2321 
2322 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2323   if (!CGM.getCodeGenOpts().SanitizeStats)
2324     return;
2325 
2326   llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2327   IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2328   CGM.getSanStats().create(IRB, SSK);
2329 }
2330 
2331 llvm::Value *
2332 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
2333   llvm::Value *TrueCondition = nullptr;
2334   if (!RO.ParsedAttribute.Architecture.empty())
2335     TrueCondition = EmitX86CpuIs(RO.ParsedAttribute.Architecture);
2336 
2337   if (!RO.ParsedAttribute.Features.empty()) {
2338     SmallVector<StringRef, 8> FeatureList;
2339     llvm::for_each(RO.ParsedAttribute.Features,
2340                    [&FeatureList](const std::string &Feature) {
2341                      FeatureList.push_back(StringRef{Feature}.substr(1));
2342                    });
2343     llvm::Value *FeatureCmp = EmitX86CpuSupports(FeatureList);
2344     TrueCondition = TrueCondition ? Builder.CreateAnd(TrueCondition, FeatureCmp)
2345                                   : FeatureCmp;
2346   }
2347   return TrueCondition;
2348 }
2349 
2350 void CodeGenFunction::EmitMultiVersionResolver(
2351     llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
2352   assert((getContext().getTargetInfo().getTriple().getArch() ==
2353               llvm::Triple::x86 ||
2354           getContext().getTargetInfo().getTriple().getArch() ==
2355               llvm::Triple::x86_64) &&
2356          "Only implemented for x86 targets");
2357 
2358   // Main function's basic block.
2359   llvm::BasicBlock *CurBlock = createBasicBlock("entry", Resolver);
2360   Builder.SetInsertPoint(CurBlock);
2361   EmitX86CpuInit();
2362 
2363   llvm::Function *DefaultFunc = nullptr;
2364   for (const MultiVersionResolverOption &RO : Options) {
2365     Builder.SetInsertPoint(CurBlock);
2366     llvm::Value *TrueCondition = FormResolverCondition(RO);
2367 
2368     if (!TrueCondition) {
2369       DefaultFunc = RO.Function;
2370     } else {
2371       llvm::BasicBlock *RetBlock = createBasicBlock("ro_ret", Resolver);
2372       llvm::IRBuilder<> RetBuilder(RetBlock);
2373       RetBuilder.CreateRet(RO.Function);
2374       CurBlock = createBasicBlock("ro_else", Resolver);
2375       Builder.CreateCondBr(TrueCondition, RetBlock, CurBlock);
2376     }
2377   }
2378 
2379   assert(DefaultFunc && "No default version?");
2380   // Emit return from the 'else-ist' block.
2381   Builder.SetInsertPoint(CurBlock);
2382   Builder.CreateRet(DefaultFunc);
2383 }
2384 
2385 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2386   if (CGDebugInfo *DI = getDebugInfo())
2387     return DI->SourceLocToDebugLoc(Location);
2388 
2389   return llvm::DebugLoc();
2390 }
2391