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/Decl.h"
26 #include "clang/AST/DeclCXX.h"
27 #include "clang/AST/StmtCXX.h"
28 #include "clang/Basic/Builtins.h"
29 #include "clang/Basic/TargetInfo.h"
30 #include "clang/CodeGen/CGFunctionInfo.h"
31 #include "clang/Frontend/CodeGenOptions.h"
32 #include "clang/Sema/SemaDiagnostic.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Operator.h"
37 using namespace clang;
38 using namespace CodeGen;
39 
40 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
41     : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
42       Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
43               CGBuilderInserterTy(this)),
44       CurFn(nullptr), ReturnValue(Address::invalid()),
45       CapturedStmtInfo(nullptr),
46       SanOpts(CGM.getLangOpts().Sanitize), IsSanitizerScope(false),
47       CurFuncIsThunk(false), AutoreleaseResult(false), SawAsmBlock(false),
48       IsOutlinedSEHHelper(false),
49       BlockInfo(nullptr), BlockPointer(nullptr),
50       LambdaThisCaptureField(nullptr), NormalCleanupDest(nullptr),
51       NextCleanupDestIndex(1), FirstBlockInfo(nullptr), EHResumeBlock(nullptr),
52       ExceptionSlot(nullptr), EHSelectorSlot(nullptr),
53       DebugInfo(CGM.getModuleDebugInfo()),
54       DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
55       PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
56       CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
57       NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
58       CXXABIThisValue(nullptr), CXXThisValue(nullptr),
59       CXXStructorImplicitParamDecl(nullptr),
60       CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
61       CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
62       TerminateHandler(nullptr), TrapBB(nullptr) {
63   if (!suppressNewContext)
64     CGM.getCXXABI().getMangleContext().startNewFunction();
65 
66   llvm::FastMathFlags FMF;
67   if (CGM.getLangOpts().FastMath)
68     FMF.setUnsafeAlgebra();
69   if (CGM.getLangOpts().FiniteMathOnly) {
70     FMF.setNoNaNs();
71     FMF.setNoInfs();
72   }
73   if (CGM.getCodeGenOpts().NoNaNsFPMath) {
74     FMF.setNoNaNs();
75   }
76   if (CGM.getCodeGenOpts().NoSignedZeros) {
77     FMF.setNoSignedZeros();
78   }
79   if (CGM.getCodeGenOpts().ReciprocalMath) {
80     FMF.setAllowReciprocal();
81   }
82   Builder.setFastMathFlags(FMF);
83 }
84 
85 CodeGenFunction::~CodeGenFunction() {
86   assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
87 
88   // If there are any unclaimed block infos, go ahead and destroy them
89   // now.  This can happen if IR-gen gets clever and skips evaluating
90   // something.
91   if (FirstBlockInfo)
92     destroyBlockInfos(FirstBlockInfo);
93 
94   if (getLangOpts().OpenMP) {
95     CGM.getOpenMPRuntime().functionFinished(*this);
96   }
97 }
98 
99 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
100                                                      AlignmentSource *Source) {
101   return getNaturalTypeAlignment(T->getPointeeType(), Source,
102                                  /*forPointee*/ true);
103 }
104 
105 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
106                                                    AlignmentSource *Source,
107                                                    bool forPointeeType) {
108   // Honor alignment typedef attributes even on incomplete types.
109   // We also honor them straight for C++ class types, even as pointees;
110   // there's an expressivity gap here.
111   if (auto TT = T->getAs<TypedefType>()) {
112     if (auto Align = TT->getDecl()->getMaxAlignment()) {
113       if (Source) *Source = AlignmentSource::AttributedType;
114       return getContext().toCharUnitsFromBits(Align);
115     }
116   }
117 
118   if (Source) *Source = AlignmentSource::Type;
119 
120   CharUnits Alignment;
121   if (T->isIncompleteType()) {
122     Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
123   } else {
124     // For C++ class pointees, we don't know whether we're pointing at a
125     // base or a complete object, so we generally need to use the
126     // non-virtual alignment.
127     const CXXRecordDecl *RD;
128     if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
129       Alignment = CGM.getClassPointerAlignment(RD);
130     } else {
131       Alignment = getContext().getTypeAlignInChars(T);
132     }
133 
134     // Cap to the global maximum type alignment unless the alignment
135     // was somehow explicit on the type.
136     if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
137       if (Alignment.getQuantity() > MaxAlign &&
138           !getContext().isAlignmentRequired(T))
139         Alignment = CharUnits::fromQuantity(MaxAlign);
140     }
141   }
142   return Alignment;
143 }
144 
145 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
146   AlignmentSource AlignSource;
147   CharUnits Alignment = getNaturalTypeAlignment(T, &AlignSource);
148   return LValue::MakeAddr(Address(V, Alignment), T, getContext(), AlignSource,
149                           CGM.getTBAAInfo(T));
150 }
151 
152 /// Given a value of type T* that may not be to a complete object,
153 /// construct an l-value with the natural pointee alignment of T.
154 LValue
155 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
156   AlignmentSource AlignSource;
157   CharUnits Align = getNaturalTypeAlignment(T, &AlignSource, /*pointee*/ true);
158   return MakeAddrLValue(Address(V, Align), T, AlignSource);
159 }
160 
161 
162 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
163   return CGM.getTypes().ConvertTypeForMem(T);
164 }
165 
166 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
167   return CGM.getTypes().ConvertType(T);
168 }
169 
170 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
171   type = type.getCanonicalType();
172   while (true) {
173     switch (type->getTypeClass()) {
174 #define TYPE(name, parent)
175 #define ABSTRACT_TYPE(name, parent)
176 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
177 #define DEPENDENT_TYPE(name, parent) case Type::name:
178 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
179 #include "clang/AST/TypeNodes.def"
180       llvm_unreachable("non-canonical or dependent type in IR-generation");
181 
182     case Type::Auto:
183       llvm_unreachable("undeduced auto type in IR-generation");
184 
185     // Various scalar types.
186     case Type::Builtin:
187     case Type::Pointer:
188     case Type::BlockPointer:
189     case Type::LValueReference:
190     case Type::RValueReference:
191     case Type::MemberPointer:
192     case Type::Vector:
193     case Type::ExtVector:
194     case Type::FunctionProto:
195     case Type::FunctionNoProto:
196     case Type::Enum:
197     case Type::ObjCObjectPointer:
198     case Type::Pipe:
199       return TEK_Scalar;
200 
201     // Complexes.
202     case Type::Complex:
203       return TEK_Complex;
204 
205     // Arrays, records, and Objective-C objects.
206     case Type::ConstantArray:
207     case Type::IncompleteArray:
208     case Type::VariableArray:
209     case Type::Record:
210     case Type::ObjCObject:
211     case Type::ObjCInterface:
212       return TEK_Aggregate;
213 
214     // We operate on atomic values according to their underlying type.
215     case Type::Atomic:
216       type = cast<AtomicType>(type)->getValueType();
217       continue;
218     }
219     llvm_unreachable("unknown type kind!");
220   }
221 }
222 
223 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
224   // For cleanliness, we try to avoid emitting the return block for
225   // simple cases.
226   llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
227 
228   if (CurBB) {
229     assert(!CurBB->getTerminator() && "Unexpected terminated block.");
230 
231     // We have a valid insert point, reuse it if it is empty or there are no
232     // explicit jumps to the return block.
233     if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
234       ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
235       delete ReturnBlock.getBlock();
236     } else
237       EmitBlock(ReturnBlock.getBlock());
238     return llvm::DebugLoc();
239   }
240 
241   // Otherwise, if the return block is the target of a single direct
242   // branch then we can just put the code in that block instead. This
243   // cleans up functions which started with a unified return block.
244   if (ReturnBlock.getBlock()->hasOneUse()) {
245     llvm::BranchInst *BI =
246       dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
247     if (BI && BI->isUnconditional() &&
248         BI->getSuccessor(0) == ReturnBlock.getBlock()) {
249       // Record/return the DebugLoc of the simple 'return' expression to be used
250       // later by the actual 'ret' instruction.
251       llvm::DebugLoc Loc = BI->getDebugLoc();
252       Builder.SetInsertPoint(BI->getParent());
253       BI->eraseFromParent();
254       delete ReturnBlock.getBlock();
255       return Loc;
256     }
257   }
258 
259   // FIXME: We are at an unreachable point, there is no reason to emit the block
260   // unless it has uses. However, we still need a place to put the debug
261   // region.end for now.
262 
263   EmitBlock(ReturnBlock.getBlock());
264   return llvm::DebugLoc();
265 }
266 
267 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
268   if (!BB) return;
269   if (!BB->use_empty())
270     return CGF.CurFn->getBasicBlockList().push_back(BB);
271   delete BB;
272 }
273 
274 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
275   assert(BreakContinueStack.empty() &&
276          "mismatched push/pop in break/continue stack!");
277 
278   bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
279     && NumSimpleReturnExprs == NumReturnExprs
280     && ReturnBlock.getBlock()->use_empty();
281   // Usually the return expression is evaluated before the cleanup
282   // code.  If the function contains only a simple return statement,
283   // such as a constant, the location before the cleanup code becomes
284   // the last useful breakpoint in the function, because the simple
285   // return expression will be evaluated after the cleanup code. To be
286   // safe, set the debug location for cleanup code to the location of
287   // the return statement.  Otherwise the cleanup code should be at the
288   // end of the function's lexical scope.
289   //
290   // If there are multiple branches to the return block, the branch
291   // instructions will get the location of the return statements and
292   // all will be fine.
293   if (CGDebugInfo *DI = getDebugInfo()) {
294     if (OnlySimpleReturnStmts)
295       DI->EmitLocation(Builder, LastStopPoint);
296     else
297       DI->EmitLocation(Builder, EndLoc);
298   }
299 
300   // Pop any cleanups that might have been associated with the
301   // parameters.  Do this in whatever block we're currently in; it's
302   // important to do this before we enter the return block or return
303   // edges will be *really* confused.
304   bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
305   bool HasOnlyLifetimeMarkers =
306       HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
307   bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
308   if (HasCleanups) {
309     // Make sure the line table doesn't jump back into the body for
310     // the ret after it's been at EndLoc.
311     if (CGDebugInfo *DI = getDebugInfo())
312       if (OnlySimpleReturnStmts)
313         DI->EmitLocation(Builder, EndLoc);
314 
315     PopCleanupBlocks(PrologueCleanupDepth);
316   }
317 
318   // Emit function epilog (to return).
319   llvm::DebugLoc Loc = EmitReturnBlock();
320 
321   if (ShouldInstrumentFunction())
322     EmitFunctionInstrumentation("__cyg_profile_func_exit");
323 
324   // Emit debug descriptor for function end.
325   if (CGDebugInfo *DI = getDebugInfo())
326     DI->EmitFunctionEnd(Builder);
327 
328   // Reset the debug location to that of the simple 'return' expression, if any
329   // rather than that of the end of the function's scope '}'.
330   ApplyDebugLocation AL(*this, Loc);
331   EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
332   EmitEndEHSpec(CurCodeDecl);
333 
334   assert(EHStack.empty() &&
335          "did not remove all scopes from cleanup stack!");
336 
337   // If someone did an indirect goto, emit the indirect goto block at the end of
338   // the function.
339   if (IndirectBranch) {
340     EmitBlock(IndirectBranch->getParent());
341     Builder.ClearInsertionPoint();
342   }
343 
344   // If some of our locals escaped, insert a call to llvm.localescape in the
345   // entry block.
346   if (!EscapedLocals.empty()) {
347     // Invert the map from local to index into a simple vector. There should be
348     // no holes.
349     SmallVector<llvm::Value *, 4> EscapeArgs;
350     EscapeArgs.resize(EscapedLocals.size());
351     for (auto &Pair : EscapedLocals)
352       EscapeArgs[Pair.second] = Pair.first;
353     llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
354         &CGM.getModule(), llvm::Intrinsic::localescape);
355     CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
356   }
357 
358   // Remove the AllocaInsertPt instruction, which is just a convenience for us.
359   llvm::Instruction *Ptr = AllocaInsertPt;
360   AllocaInsertPt = nullptr;
361   Ptr->eraseFromParent();
362 
363   // If someone took the address of a label but never did an indirect goto, we
364   // made a zero entry PHI node, which is illegal, zap it now.
365   if (IndirectBranch) {
366     llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
367     if (PN->getNumIncomingValues() == 0) {
368       PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
369       PN->eraseFromParent();
370     }
371   }
372 
373   EmitIfUsed(*this, EHResumeBlock);
374   EmitIfUsed(*this, TerminateLandingPad);
375   EmitIfUsed(*this, TerminateHandler);
376   EmitIfUsed(*this, UnreachableBlock);
377 
378   if (CGM.getCodeGenOpts().EmitDeclMetadata)
379     EmitDeclMetadata();
380 
381   for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
382            I = DeferredReplacements.begin(),
383            E = DeferredReplacements.end();
384        I != E; ++I) {
385     I->first->replaceAllUsesWith(I->second);
386     I->first->eraseFromParent();
387   }
388 }
389 
390 /// ShouldInstrumentFunction - Return true if the current function should be
391 /// instrumented with __cyg_profile_func_* calls
392 bool CodeGenFunction::ShouldInstrumentFunction() {
393   if (!CGM.getCodeGenOpts().InstrumentFunctions)
394     return false;
395   if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
396     return false;
397   return true;
398 }
399 
400 /// ShouldXRayInstrument - Return true if the current function should be
401 /// instrumented with XRay nop sleds.
402 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
403   return CGM.getCodeGenOpts().XRayInstrumentFunctions;
404 }
405 
406 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
407 /// instrumentation function with the current function and the call site, if
408 /// function instrumentation is enabled.
409 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
410   auto NL = ApplyDebugLocation::CreateArtificial(*this);
411   // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
412   llvm::PointerType *PointerTy = Int8PtrTy;
413   llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
414   llvm::FunctionType *FunctionTy =
415     llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
416 
417   llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
418   llvm::CallInst *CallSite = Builder.CreateCall(
419     CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
420     llvm::ConstantInt::get(Int32Ty, 0),
421     "callsite");
422 
423   llvm::Value *args[] = {
424     llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
425     CallSite
426   };
427 
428   EmitNounwindRuntimeCall(F, args);
429 }
430 
431 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
432 // information in the program executable. The argument information stored
433 // includes the argument name, its type, the address and access qualifiers used.
434 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
435                                  CodeGenModule &CGM, llvm::LLVMContext &Context,
436                                  CGBuilderTy &Builder, ASTContext &ASTCtx) {
437   // Create MDNodes that represent the kernel arg metadata.
438   // Each MDNode is a list in the form of "key", N number of values which is
439   // the same number of values as their are kernel arguments.
440 
441   const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
442 
443   // MDNode for the kernel argument address space qualifiers.
444   SmallVector<llvm::Metadata *, 8> addressQuals;
445 
446   // MDNode for the kernel argument access qualifiers (images only).
447   SmallVector<llvm::Metadata *, 8> accessQuals;
448 
449   // MDNode for the kernel argument type names.
450   SmallVector<llvm::Metadata *, 8> argTypeNames;
451 
452   // MDNode for the kernel argument base type names.
453   SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
454 
455   // MDNode for the kernel argument type qualifiers.
456   SmallVector<llvm::Metadata *, 8> argTypeQuals;
457 
458   // MDNode for the kernel argument names.
459   SmallVector<llvm::Metadata *, 8> argNames;
460 
461   for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
462     const ParmVarDecl *parm = FD->getParamDecl(i);
463     QualType ty = parm->getType();
464     std::string typeQuals;
465 
466     if (ty->isPointerType()) {
467       QualType pointeeTy = ty->getPointeeType();
468 
469       // Get address qualifier.
470       addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
471           ASTCtx.getTargetAddressSpace(pointeeTy.getAddressSpace()))));
472 
473       // Get argument type name.
474       std::string typeName =
475           pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
476 
477       // Turn "unsigned type" to "utype"
478       std::string::size_type pos = typeName.find("unsigned");
479       if (pointeeTy.isCanonical() && pos != std::string::npos)
480         typeName.erase(pos+1, 8);
481 
482       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
483 
484       std::string baseTypeName =
485           pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
486               Policy) +
487           "*";
488 
489       // Turn "unsigned type" to "utype"
490       pos = baseTypeName.find("unsigned");
491       if (pos != std::string::npos)
492         baseTypeName.erase(pos+1, 8);
493 
494       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
495 
496       // Get argument type qualifiers:
497       if (ty.isRestrictQualified())
498         typeQuals = "restrict";
499       if (pointeeTy.isConstQualified() ||
500           (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
501         typeQuals += typeQuals.empty() ? "const" : " const";
502       if (pointeeTy.isVolatileQualified())
503         typeQuals += typeQuals.empty() ? "volatile" : " volatile";
504     } else {
505       uint32_t AddrSpc = 0;
506       bool isPipe = ty->isPipeType();
507       if (ty->isImageType() || isPipe)
508         AddrSpc =
509           CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
510 
511       addressQuals.push_back(
512           llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
513 
514       // Get argument type name.
515       std::string typeName;
516       if (isPipe)
517         typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
518                      .getAsString(Policy);
519       else
520         typeName = ty.getUnqualifiedType().getAsString(Policy);
521 
522       // Turn "unsigned type" to "utype"
523       std::string::size_type pos = typeName.find("unsigned");
524       if (ty.isCanonical() && pos != std::string::npos)
525         typeName.erase(pos+1, 8);
526 
527       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
528 
529       std::string baseTypeName;
530       if (isPipe)
531         baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
532                           ->getElementType().getCanonicalType()
533                           .getAsString(Policy);
534       else
535         baseTypeName =
536           ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
537 
538       // Turn "unsigned type" to "utype"
539       pos = baseTypeName.find("unsigned");
540       if (pos != std::string::npos)
541         baseTypeName.erase(pos+1, 8);
542 
543       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
544 
545       // Get argument type qualifiers:
546       if (ty.isConstQualified())
547         typeQuals = "const";
548       if (ty.isVolatileQualified())
549         typeQuals += typeQuals.empty() ? "volatile" : " volatile";
550       if (isPipe)
551         typeQuals = "pipe";
552     }
553 
554     argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
555 
556     // Get image and pipe access qualifier:
557     if (ty->isImageType()|| ty->isPipeType()) {
558       const OpenCLAccessAttr *A = parm->getAttr<OpenCLAccessAttr>();
559       if (A && A->isWriteOnly())
560         accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
561       else if (A && A->isReadWrite())
562         accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
563       else
564         accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
565     } else
566       accessQuals.push_back(llvm::MDString::get(Context, "none"));
567 
568     // Get argument name.
569     argNames.push_back(llvm::MDString::get(Context, parm->getName()));
570   }
571 
572   Fn->setMetadata("kernel_arg_addr_space",
573                   llvm::MDNode::get(Context, addressQuals));
574   Fn->setMetadata("kernel_arg_access_qual",
575                   llvm::MDNode::get(Context, accessQuals));
576   Fn->setMetadata("kernel_arg_type",
577                   llvm::MDNode::get(Context, argTypeNames));
578   Fn->setMetadata("kernel_arg_base_type",
579                   llvm::MDNode::get(Context, argBaseTypeNames));
580   Fn->setMetadata("kernel_arg_type_qual",
581                   llvm::MDNode::get(Context, argTypeQuals));
582   if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
583     Fn->setMetadata("kernel_arg_name",
584                     llvm::MDNode::get(Context, argNames));
585 }
586 
587 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
588                                                llvm::Function *Fn)
589 {
590   if (!FD->hasAttr<OpenCLKernelAttr>())
591     return;
592 
593   llvm::LLVMContext &Context = getLLVMContext();
594 
595   GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
596 
597   if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
598     QualType hintQTy = A->getTypeHint();
599     const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
600     bool isSignedInteger =
601         hintQTy->isSignedIntegerType() ||
602         (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
603     llvm::Metadata *attrMDArgs[] = {
604         llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
605             CGM.getTypes().ConvertType(A->getTypeHint()))),
606         llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
607             llvm::IntegerType::get(Context, 32),
608             llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
609     Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, attrMDArgs));
610   }
611 
612   if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
613     llvm::Metadata *attrMDArgs[] = {
614         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
615         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
616         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
617     Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, attrMDArgs));
618   }
619 
620   if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
621     llvm::Metadata *attrMDArgs[] = {
622         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
623         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
624         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
625     Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, attrMDArgs));
626   }
627 }
628 
629 /// Determine whether the function F ends with a return stmt.
630 static bool endsWithReturn(const Decl* F) {
631   const Stmt *Body = nullptr;
632   if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
633     Body = FD->getBody();
634   else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
635     Body = OMD->getBody();
636 
637   if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
638     auto LastStmt = CS->body_rbegin();
639     if (LastStmt != CS->body_rend())
640       return isa<ReturnStmt>(*LastStmt);
641   }
642   return false;
643 }
644 
645 void CodeGenFunction::StartFunction(GlobalDecl GD,
646                                     QualType RetTy,
647                                     llvm::Function *Fn,
648                                     const CGFunctionInfo &FnInfo,
649                                     const FunctionArgList &Args,
650                                     SourceLocation Loc,
651                                     SourceLocation StartLoc) {
652   assert(!CurFn &&
653          "Do not use a CodeGenFunction object for more than one function");
654 
655   const Decl *D = GD.getDecl();
656 
657   DidCallStackSave = false;
658   CurCodeDecl = D;
659   if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
660     if (FD->usesSEHTry())
661       CurSEHParent = FD;
662   CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
663   FnRetTy = RetTy;
664   CurFn = Fn;
665   CurFnInfo = &FnInfo;
666   assert(CurFn->isDeclaration() && "Function already has body?");
667 
668   if (CGM.isInSanitizerBlacklist(Fn, Loc))
669     SanOpts.clear();
670 
671   if (D) {
672     // Apply the no_sanitize* attributes to SanOpts.
673     for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
674       SanOpts.Mask &= ~Attr->getMask();
675   }
676 
677   // Apply sanitizer attributes to the function.
678   if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
679     Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
680   if (SanOpts.has(SanitizerKind::Thread))
681     Fn->addFnAttr(llvm::Attribute::SanitizeThread);
682   if (SanOpts.has(SanitizerKind::Memory))
683     Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
684   if (SanOpts.has(SanitizerKind::SafeStack))
685     Fn->addFnAttr(llvm::Attribute::SafeStack);
686 
687   // Apply xray attributes to the function (as a string, for now)
688   if (D && ShouldXRayInstrumentFunction()) {
689     if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
690       if (XRayAttr->alwaysXRayInstrument())
691         Fn->addFnAttr("function-instrument", "xray-always");
692       if (XRayAttr->neverXRayInstrument())
693         Fn->addFnAttr("function-instrument", "xray-never");
694     } else {
695       Fn->addFnAttr(
696           "xray-instruction-threshold",
697           llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
698     }
699   }
700 
701   // Pass inline keyword to optimizer if it appears explicitly on any
702   // declaration. Also, in the case of -fno-inline attach NoInline
703   // attribute to all functions that are not marked AlwaysInline, or
704   // to all functions that are not marked inline or implicitly inline
705   // in the case of -finline-hint-functions.
706   if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
707     const CodeGenOptions& CodeGenOpts = CGM.getCodeGenOpts();
708     if (!CodeGenOpts.NoInline) {
709       for (auto RI : FD->redecls())
710         if (RI->isInlineSpecified()) {
711           Fn->addFnAttr(llvm::Attribute::InlineHint);
712           break;
713         }
714       if (CodeGenOpts.getInlining() == CodeGenOptions::OnlyHintInlining &&
715           !FD->isInlined() && !Fn->hasFnAttribute(llvm::Attribute::InlineHint))
716         Fn->addFnAttr(llvm::Attribute::NoInline);
717     } else if (!FD->hasAttr<AlwaysInlineAttr>())
718       Fn->addFnAttr(llvm::Attribute::NoInline);
719     if (CGM.getLangOpts().OpenMP && FD->hasAttr<OMPDeclareSimdDeclAttr>())
720       CGM.getOpenMPRuntime().emitDeclareSimdFunction(FD, Fn);
721   }
722 
723   // Add no-jump-tables value.
724   Fn->addFnAttr("no-jump-tables",
725                 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
726 
727   if (getLangOpts().OpenCL) {
728     // Add metadata for a kernel function.
729     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
730       EmitOpenCLKernelMetadata(FD, Fn);
731   }
732 
733   // If we are checking function types, emit a function type signature as
734   // prologue data.
735   if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
736     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
737       if (llvm::Constant *PrologueSig =
738               CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
739         llvm::Constant *FTRTTIConst =
740             CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
741         llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
742         llvm::Constant *PrologueStructConst =
743             llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
744         Fn->setPrologueData(PrologueStructConst);
745       }
746     }
747   }
748 
749   // If we're in C++ mode and the function name is "main", it is guaranteed
750   // to be norecurse by the standard (3.6.1.3 "The function main shall not be
751   // used within a program").
752   if (getLangOpts().CPlusPlus)
753     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
754       if (FD->isMain())
755         Fn->addFnAttr(llvm::Attribute::NoRecurse);
756 
757   llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
758 
759   // Create a marker to make it easy to insert allocas into the entryblock
760   // later.  Don't create this with the builder, because we don't want it
761   // folded.
762   llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
763   AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
764 
765   ReturnBlock = getJumpDestInCurrentScope("return");
766 
767   Builder.SetInsertPoint(EntryBB);
768 
769   // Emit subprogram debug descriptor.
770   if (CGDebugInfo *DI = getDebugInfo()) {
771     // Reconstruct the type from the argument list so that implicit parameters,
772     // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
773     // convention.
774     CallingConv CC = CallingConv::CC_C;
775     if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
776       if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
777         CC = SrcFnTy->getCallConv();
778     SmallVector<QualType, 16> ArgTypes;
779     for (const VarDecl *VD : Args)
780       ArgTypes.push_back(VD->getType());
781     QualType FnType = getContext().getFunctionType(
782         RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
783     DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
784   }
785 
786   if (ShouldInstrumentFunction())
787     EmitFunctionInstrumentation("__cyg_profile_func_enter");
788 
789   // Since emitting the mcount call here impacts optimizations such as function
790   // inlining, we just add an attribute to insert a mcount call in backend.
791   // The attribute "counting-function" is set to mcount function name which is
792   // architecture dependent.
793   if (CGM.getCodeGenOpts().InstrumentForProfiling)
794     Fn->addFnAttr("counting-function", getTarget().getMCountName());
795 
796   if (RetTy->isVoidType()) {
797     // Void type; nothing to return.
798     ReturnValue = Address::invalid();
799 
800     // Count the implicit return.
801     if (!endsWithReturn(D))
802       ++NumReturnExprs;
803   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
804              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
805     // Indirect aggregate return; emit returned value directly into sret slot.
806     // This reduces code size, and affects correctness in C++.
807     auto AI = CurFn->arg_begin();
808     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
809       ++AI;
810     ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
811   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
812              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
813     // Load the sret pointer from the argument struct and return into that.
814     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
815     llvm::Function::arg_iterator EI = CurFn->arg_end();
816     --EI;
817     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
818     Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
819     ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
820   } else {
821     ReturnValue = CreateIRTemp(RetTy, "retval");
822 
823     // Tell the epilog emitter to autorelease the result.  We do this
824     // now so that various specialized functions can suppress it
825     // during their IR-generation.
826     if (getLangOpts().ObjCAutoRefCount &&
827         !CurFnInfo->isReturnsRetained() &&
828         RetTy->isObjCRetainableType())
829       AutoreleaseResult = true;
830   }
831 
832   EmitStartEHSpec(CurCodeDecl);
833 
834   PrologueCleanupDepth = EHStack.stable_begin();
835   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
836 
837   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
838     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
839     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
840     if (MD->getParent()->isLambda() &&
841         MD->getOverloadedOperator() == OO_Call) {
842       // We're in a lambda; figure out the captures.
843       MD->getParent()->getCaptureFields(LambdaCaptureFields,
844                                         LambdaThisCaptureField);
845       if (LambdaThisCaptureField) {
846         // If the lambda captures the object referred to by '*this' - either by
847         // value or by reference, make sure CXXThisValue points to the correct
848         // object.
849 
850         // Get the lvalue for the field (which is a copy of the enclosing object
851         // or contains the address of the enclosing object).
852         LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
853         if (!LambdaThisCaptureField->getType()->isPointerType()) {
854           // If the enclosing object was captured by value, just use its address.
855           CXXThisValue = ThisFieldLValue.getAddress().getPointer();
856         } else {
857           // Load the lvalue pointed to by the field, since '*this' was captured
858           // by reference.
859           CXXThisValue =
860               EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
861         }
862       }
863       for (auto *FD : MD->getParent()->fields()) {
864         if (FD->hasCapturedVLAType()) {
865           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
866                                            SourceLocation()).getScalarVal();
867           auto VAT = FD->getCapturedVLAType();
868           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
869         }
870       }
871     } else {
872       // Not in a lambda; just use 'this' from the method.
873       // FIXME: Should we generate a new load for each use of 'this'?  The
874       // fast register allocator would be happier...
875       CXXThisValue = CXXABIThisValue;
876     }
877   }
878 
879   // If any of the arguments have a variably modified type, make sure to
880   // emit the type size.
881   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
882        i != e; ++i) {
883     const VarDecl *VD = *i;
884 
885     // Dig out the type as written from ParmVarDecls; it's unclear whether
886     // the standard (C99 6.9.1p10) requires this, but we're following the
887     // precedent set by gcc.
888     QualType Ty;
889     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
890       Ty = PVD->getOriginalType();
891     else
892       Ty = VD->getType();
893 
894     if (Ty->isVariablyModifiedType())
895       EmitVariablyModifiedType(Ty);
896   }
897   // Emit a location at the end of the prologue.
898   if (CGDebugInfo *DI = getDebugInfo())
899     DI->EmitLocation(Builder, StartLoc);
900 }
901 
902 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
903                                        const Stmt *Body) {
904   incrementProfileCounter(Body);
905   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
906     EmitCompoundStmtWithoutScope(*S);
907   else
908     EmitStmt(Body);
909 }
910 
911 /// When instrumenting to collect profile data, the counts for some blocks
912 /// such as switch cases need to not include the fall-through counts, so
913 /// emit a branch around the instrumentation code. When not instrumenting,
914 /// this just calls EmitBlock().
915 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
916                                                const Stmt *S) {
917   llvm::BasicBlock *SkipCountBB = nullptr;
918   if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
919     // When instrumenting for profiling, the fallthrough to certain
920     // statements needs to skip over the instrumentation code so that we
921     // get an accurate count.
922     SkipCountBB = createBasicBlock("skipcount");
923     EmitBranch(SkipCountBB);
924   }
925   EmitBlock(BB);
926   uint64_t CurrentCount = getCurrentProfileCount();
927   incrementProfileCounter(S);
928   setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
929   if (SkipCountBB)
930     EmitBlock(SkipCountBB);
931 }
932 
933 /// Tries to mark the given function nounwind based on the
934 /// non-existence of any throwing calls within it.  We believe this is
935 /// lightweight enough to do at -O0.
936 static void TryMarkNoThrow(llvm::Function *F) {
937   // LLVM treats 'nounwind' on a function as part of the type, so we
938   // can't do this on functions that can be overwritten.
939   if (F->isInterposable()) return;
940 
941   for (llvm::BasicBlock &BB : *F)
942     for (llvm::Instruction &I : BB)
943       if (I.mayThrow())
944         return;
945 
946   F->setDoesNotThrow();
947 }
948 
949 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
950                                                FunctionArgList &Args) {
951   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
952   QualType ResTy = FD->getReturnType();
953 
954   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
955   if (MD && MD->isInstance()) {
956     if (CGM.getCXXABI().HasThisReturn(GD))
957       ResTy = MD->getThisType(getContext());
958     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
959       ResTy = CGM.getContext().VoidPtrTy;
960     CGM.getCXXABI().buildThisParam(*this, Args);
961   }
962 
963   // The base version of an inheriting constructor whose constructed base is a
964   // virtual base is not passed any arguments (because it doesn't actually call
965   // the inherited constructor).
966   bool PassedParams = true;
967   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
968     if (auto Inherited = CD->getInheritedConstructor())
969       PassedParams =
970           getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
971 
972   if (PassedParams) {
973     for (auto *Param : FD->parameters()) {
974       Args.push_back(Param);
975       if (!Param->hasAttr<PassObjectSizeAttr>())
976         continue;
977 
978       IdentifierInfo *NoID = nullptr;
979       auto *Implicit = ImplicitParamDecl::Create(
980           getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
981           getContext().getSizeType());
982       SizeArguments[Param] = Implicit;
983       Args.push_back(Implicit);
984     }
985   }
986 
987   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
988     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
989 
990   return ResTy;
991 }
992 
993 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
994                                    const CGFunctionInfo &FnInfo) {
995   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
996   CurGD = GD;
997 
998   FunctionArgList Args;
999   QualType ResTy = BuildFunctionArgList(GD, Args);
1000 
1001   // Check if we should generate debug info for this function.
1002   if (FD->hasAttr<NoDebugAttr>())
1003     DebugInfo = nullptr; // disable debug info indefinitely for this function
1004 
1005   SourceRange BodyRange;
1006   if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
1007   CurEHLocation = BodyRange.getEnd();
1008 
1009   // Use the location of the start of the function to determine where
1010   // the function definition is located. By default use the location
1011   // of the declaration as the location for the subprogram. A function
1012   // may lack a declaration in the source code if it is created by code
1013   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1014   SourceLocation Loc = FD->getLocation();
1015 
1016   // If this is a function specialization then use the pattern body
1017   // as the location for the function.
1018   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1019     if (SpecDecl->hasBody(SpecDecl))
1020       Loc = SpecDecl->getLocation();
1021 
1022   // Emit the standard function prologue.
1023   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1024 
1025   // Generate the body of the function.
1026   PGO.assignRegionCounters(GD, CurFn);
1027   if (isa<CXXDestructorDecl>(FD))
1028     EmitDestructorBody(Args);
1029   else if (isa<CXXConstructorDecl>(FD))
1030     EmitConstructorBody(Args);
1031   else if (getLangOpts().CUDA &&
1032            !getLangOpts().CUDAIsDevice &&
1033            FD->hasAttr<CUDAGlobalAttr>())
1034     CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1035   else if (isa<CXXConversionDecl>(FD) &&
1036            cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
1037     // The lambda conversion to block pointer is special; the semantics can't be
1038     // expressed in the AST, so IRGen needs to special-case it.
1039     EmitLambdaToBlockPointerBody(Args);
1040   } else if (isa<CXXMethodDecl>(FD) &&
1041              cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1042     // The lambda static invoker function is special, because it forwards or
1043     // clones the body of the function call operator (but is actually static).
1044     EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
1045   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1046              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1047               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1048     // Implicit copy-assignment gets the same special treatment as implicit
1049     // copy-constructors.
1050     emitImplicitAssignmentOperatorBody(Args);
1051   } else if (Stmt *Body = FD->getBody()) {
1052     EmitFunctionBody(Args, Body);
1053   } else
1054     llvm_unreachable("no definition for emitted function");
1055 
1056   // C++11 [stmt.return]p2:
1057   //   Flowing off the end of a function [...] results in undefined behavior in
1058   //   a value-returning function.
1059   // C11 6.9.1p12:
1060   //   If the '}' that terminates a function is reached, and the value of the
1061   //   function call is used by the caller, the behavior is undefined.
1062   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1063       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1064     if (SanOpts.has(SanitizerKind::Return)) {
1065       SanitizerScope SanScope(this);
1066       llvm::Value *IsFalse = Builder.getFalse();
1067       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1068                 "missing_return", EmitCheckSourceLocation(FD->getLocation()),
1069                 None);
1070     } else if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
1071       EmitTrapCall(llvm::Intrinsic::trap);
1072     }
1073     Builder.CreateUnreachable();
1074     Builder.ClearInsertionPoint();
1075   }
1076 
1077   // Emit the standard function epilogue.
1078   FinishFunction(BodyRange.getEnd());
1079 
1080   // If we haven't marked the function nothrow through other means, do
1081   // a quick pass now to see if we can.
1082   if (!CurFn->doesNotThrow())
1083     TryMarkNoThrow(CurFn);
1084 }
1085 
1086 /// ContainsLabel - Return true if the statement contains a label in it.  If
1087 /// this statement is not executed normally, it not containing a label means
1088 /// that we can just remove the code.
1089 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1090   // Null statement, not a label!
1091   if (!S) return false;
1092 
1093   // If this is a label, we have to emit the code, consider something like:
1094   // if (0) {  ...  foo:  bar(); }  goto foo;
1095   //
1096   // TODO: If anyone cared, we could track __label__'s, since we know that you
1097   // can't jump to one from outside their declared region.
1098   if (isa<LabelStmt>(S))
1099     return true;
1100 
1101   // If this is a case/default statement, and we haven't seen a switch, we have
1102   // to emit the code.
1103   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1104     return true;
1105 
1106   // If this is a switch statement, we want to ignore cases below it.
1107   if (isa<SwitchStmt>(S))
1108     IgnoreCaseStmts = true;
1109 
1110   // Scan subexpressions for verboten labels.
1111   for (const Stmt *SubStmt : S->children())
1112     if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1113       return true;
1114 
1115   return false;
1116 }
1117 
1118 /// containsBreak - Return true if the statement contains a break out of it.
1119 /// If the statement (recursively) contains a switch or loop with a break
1120 /// inside of it, this is fine.
1121 bool CodeGenFunction::containsBreak(const Stmt *S) {
1122   // Null statement, not a label!
1123   if (!S) return false;
1124 
1125   // If this is a switch or loop that defines its own break scope, then we can
1126   // include it and anything inside of it.
1127   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1128       isa<ForStmt>(S))
1129     return false;
1130 
1131   if (isa<BreakStmt>(S))
1132     return true;
1133 
1134   // Scan subexpressions for verboten breaks.
1135   for (const Stmt *SubStmt : S->children())
1136     if (containsBreak(SubStmt))
1137       return true;
1138 
1139   return false;
1140 }
1141 
1142 
1143 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1144 /// to a constant, or if it does but contains a label, return false.  If it
1145 /// constant folds return true and set the boolean result in Result.
1146 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1147                                                    bool &ResultBool,
1148                                                    bool AllowLabels) {
1149   llvm::APSInt ResultInt;
1150   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1151     return false;
1152 
1153   ResultBool = ResultInt.getBoolValue();
1154   return true;
1155 }
1156 
1157 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1158 /// to a constant, or if it does but contains a label, return false.  If it
1159 /// constant folds return true and set the folded value.
1160 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1161                                                    llvm::APSInt &ResultInt,
1162                                                    bool AllowLabels) {
1163   // FIXME: Rename and handle conversion of other evaluatable things
1164   // to bool.
1165   llvm::APSInt Int;
1166   if (!Cond->EvaluateAsInt(Int, getContext()))
1167     return false;  // Not foldable, not integer or not fully evaluatable.
1168 
1169   if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1170     return false;  // Contains a label.
1171 
1172   ResultInt = Int;
1173   return true;
1174 }
1175 
1176 
1177 
1178 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1179 /// statement) to the specified blocks.  Based on the condition, this might try
1180 /// to simplify the codegen of the conditional based on the branch.
1181 ///
1182 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1183                                            llvm::BasicBlock *TrueBlock,
1184                                            llvm::BasicBlock *FalseBlock,
1185                                            uint64_t TrueCount) {
1186   Cond = Cond->IgnoreParens();
1187 
1188   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1189 
1190     // Handle X && Y in a condition.
1191     if (CondBOp->getOpcode() == BO_LAnd) {
1192       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1193       // folded if the case was simple enough.
1194       bool ConstantBool = false;
1195       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1196           ConstantBool) {
1197         // br(1 && X) -> br(X).
1198         incrementProfileCounter(CondBOp);
1199         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1200                                     TrueCount);
1201       }
1202 
1203       // If we have "X && 1", simplify the code to use an uncond branch.
1204       // "X && 0" would have been constant folded to 0.
1205       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1206           ConstantBool) {
1207         // br(X && 1) -> br(X).
1208         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1209                                     TrueCount);
1210       }
1211 
1212       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1213       // want to jump to the FalseBlock.
1214       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1215       // The counter tells us how often we evaluate RHS, and all of TrueCount
1216       // can be propagated to that branch.
1217       uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1218 
1219       ConditionalEvaluation eval(*this);
1220       {
1221         ApplyDebugLocation DL(*this, Cond);
1222         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1223         EmitBlock(LHSTrue);
1224       }
1225 
1226       incrementProfileCounter(CondBOp);
1227       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1228 
1229       // Any temporaries created here are conditional.
1230       eval.begin(*this);
1231       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1232       eval.end(*this);
1233 
1234       return;
1235     }
1236 
1237     if (CondBOp->getOpcode() == BO_LOr) {
1238       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1239       // folded if the case was simple enough.
1240       bool ConstantBool = false;
1241       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1242           !ConstantBool) {
1243         // br(0 || X) -> br(X).
1244         incrementProfileCounter(CondBOp);
1245         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1246                                     TrueCount);
1247       }
1248 
1249       // If we have "X || 0", simplify the code to use an uncond branch.
1250       // "X || 1" would have been constant folded to 1.
1251       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1252           !ConstantBool) {
1253         // br(X || 0) -> br(X).
1254         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1255                                     TrueCount);
1256       }
1257 
1258       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1259       // want to jump to the TrueBlock.
1260       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1261       // We have the count for entry to the RHS and for the whole expression
1262       // being true, so we can divy up True count between the short circuit and
1263       // the RHS.
1264       uint64_t LHSCount =
1265           getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1266       uint64_t RHSCount = TrueCount - LHSCount;
1267 
1268       ConditionalEvaluation eval(*this);
1269       {
1270         ApplyDebugLocation DL(*this, Cond);
1271         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1272         EmitBlock(LHSFalse);
1273       }
1274 
1275       incrementProfileCounter(CondBOp);
1276       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1277 
1278       // Any temporaries created here are conditional.
1279       eval.begin(*this);
1280       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1281 
1282       eval.end(*this);
1283 
1284       return;
1285     }
1286   }
1287 
1288   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1289     // br(!x, t, f) -> br(x, f, t)
1290     if (CondUOp->getOpcode() == UO_LNot) {
1291       // Negate the count.
1292       uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1293       // Negate the condition and swap the destination blocks.
1294       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1295                                   FalseCount);
1296     }
1297   }
1298 
1299   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1300     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1301     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1302     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1303 
1304     ConditionalEvaluation cond(*this);
1305     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1306                          getProfileCount(CondOp));
1307 
1308     // When computing PGO branch weights, we only know the overall count for
1309     // the true block. This code is essentially doing tail duplication of the
1310     // naive code-gen, introducing new edges for which counts are not
1311     // available. Divide the counts proportionally between the LHS and RHS of
1312     // the conditional operator.
1313     uint64_t LHSScaledTrueCount = 0;
1314     if (TrueCount) {
1315       double LHSRatio =
1316           getProfileCount(CondOp) / (double)getCurrentProfileCount();
1317       LHSScaledTrueCount = TrueCount * LHSRatio;
1318     }
1319 
1320     cond.begin(*this);
1321     EmitBlock(LHSBlock);
1322     incrementProfileCounter(CondOp);
1323     {
1324       ApplyDebugLocation DL(*this, Cond);
1325       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1326                            LHSScaledTrueCount);
1327     }
1328     cond.end(*this);
1329 
1330     cond.begin(*this);
1331     EmitBlock(RHSBlock);
1332     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1333                          TrueCount - LHSScaledTrueCount);
1334     cond.end(*this);
1335 
1336     return;
1337   }
1338 
1339   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1340     // Conditional operator handling can give us a throw expression as a
1341     // condition for a case like:
1342     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1343     // Fold this to:
1344     //   br(c, throw x, br(y, t, f))
1345     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1346     return;
1347   }
1348 
1349   // If the branch has a condition wrapped by __builtin_unpredictable,
1350   // create metadata that specifies that the branch is unpredictable.
1351   // Don't bother if not optimizing because that metadata would not be used.
1352   llvm::MDNode *Unpredictable = nullptr;
1353   auto *Call = dyn_cast<CallExpr>(Cond);
1354   if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1355     auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1356     if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1357       llvm::MDBuilder MDHelper(getLLVMContext());
1358       Unpredictable = MDHelper.createUnpredictable();
1359     }
1360   }
1361 
1362   // Create branch weights based on the number of times we get here and the
1363   // number of times the condition should be true.
1364   uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1365   llvm::MDNode *Weights =
1366       createProfileWeights(TrueCount, CurrentCount - TrueCount);
1367 
1368   // Emit the code with the fully general case.
1369   llvm::Value *CondV;
1370   {
1371     ApplyDebugLocation DL(*this, Cond);
1372     CondV = EvaluateExprAsBool(Cond);
1373   }
1374   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1375 }
1376 
1377 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1378 /// specified stmt yet.
1379 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1380   CGM.ErrorUnsupported(S, Type);
1381 }
1382 
1383 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1384 /// variable-length array whose elements have a non-zero bit-pattern.
1385 ///
1386 /// \param baseType the inner-most element type of the array
1387 /// \param src - a char* pointing to the bit-pattern for a single
1388 /// base element of the array
1389 /// \param sizeInChars - the total size of the VLA, in chars
1390 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1391                                Address dest, Address src,
1392                                llvm::Value *sizeInChars) {
1393   CGBuilderTy &Builder = CGF.Builder;
1394 
1395   CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1396   llvm::Value *baseSizeInChars
1397     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1398 
1399   Address begin =
1400     Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1401   llvm::Value *end =
1402     Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1403 
1404   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1405   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1406   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1407 
1408   // Make a loop over the VLA.  C99 guarantees that the VLA element
1409   // count must be nonzero.
1410   CGF.EmitBlock(loopBB);
1411 
1412   llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1413   cur->addIncoming(begin.getPointer(), originBB);
1414 
1415   CharUnits curAlign =
1416     dest.getAlignment().alignmentOfArrayElement(baseSize);
1417 
1418   // memcpy the individual element bit-pattern.
1419   Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1420                        /*volatile*/ false);
1421 
1422   // Go to the next element.
1423   llvm::Value *next =
1424     Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1425 
1426   // Leave if that's the end of the VLA.
1427   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1428   Builder.CreateCondBr(done, contBB, loopBB);
1429   cur->addIncoming(next, loopBB);
1430 
1431   CGF.EmitBlock(contBB);
1432 }
1433 
1434 void
1435 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1436   // Ignore empty classes in C++.
1437   if (getLangOpts().CPlusPlus) {
1438     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1439       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1440         return;
1441     }
1442   }
1443 
1444   // Cast the dest ptr to the appropriate i8 pointer type.
1445   if (DestPtr.getElementType() != Int8Ty)
1446     DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1447 
1448   // Get size and alignment info for this aggregate.
1449   CharUnits size = getContext().getTypeSizeInChars(Ty);
1450 
1451   llvm::Value *SizeVal;
1452   const VariableArrayType *vla;
1453 
1454   // Don't bother emitting a zero-byte memset.
1455   if (size.isZero()) {
1456     // But note that getTypeInfo returns 0 for a VLA.
1457     if (const VariableArrayType *vlaType =
1458           dyn_cast_or_null<VariableArrayType>(
1459                                           getContext().getAsArrayType(Ty))) {
1460       QualType eltType;
1461       llvm::Value *numElts;
1462       std::tie(numElts, eltType) = getVLASize(vlaType);
1463 
1464       SizeVal = numElts;
1465       CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1466       if (!eltSize.isOne())
1467         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1468       vla = vlaType;
1469     } else {
1470       return;
1471     }
1472   } else {
1473     SizeVal = CGM.getSize(size);
1474     vla = nullptr;
1475   }
1476 
1477   // If the type contains a pointer to data member we can't memset it to zero.
1478   // Instead, create a null constant and copy it to the destination.
1479   // TODO: there are other patterns besides zero that we can usefully memset,
1480   // like -1, which happens to be the pattern used by member-pointers.
1481   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1482     // For a VLA, emit a single element, then splat that over the VLA.
1483     if (vla) Ty = getContext().getBaseElementType(vla);
1484 
1485     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1486 
1487     llvm::GlobalVariable *NullVariable =
1488       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1489                                /*isConstant=*/true,
1490                                llvm::GlobalVariable::PrivateLinkage,
1491                                NullConstant, Twine());
1492     CharUnits NullAlign = DestPtr.getAlignment();
1493     NullVariable->setAlignment(NullAlign.getQuantity());
1494     Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1495                    NullAlign);
1496 
1497     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1498 
1499     // Get and call the appropriate llvm.memcpy overload.
1500     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1501     return;
1502   }
1503 
1504   // Otherwise, just memset the whole thing to zero.  This is legal
1505   // because in LLVM, all default initializers (other than the ones we just
1506   // handled above) are guaranteed to have a bit pattern of all zeros.
1507   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1508 }
1509 
1510 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1511   // Make sure that there is a block for the indirect goto.
1512   if (!IndirectBranch)
1513     GetIndirectGotoBlock();
1514 
1515   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1516 
1517   // Make sure the indirect branch includes all of the address-taken blocks.
1518   IndirectBranch->addDestination(BB);
1519   return llvm::BlockAddress::get(CurFn, BB);
1520 }
1521 
1522 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1523   // If we already made the indirect branch for indirect goto, return its block.
1524   if (IndirectBranch) return IndirectBranch->getParent();
1525 
1526   CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1527 
1528   // Create the PHI node that indirect gotos will add entries to.
1529   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1530                                               "indirect.goto.dest");
1531 
1532   // Create the indirect branch instruction.
1533   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1534   return IndirectBranch->getParent();
1535 }
1536 
1537 /// Computes the length of an array in elements, as well as the base
1538 /// element type and a properly-typed first element pointer.
1539 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1540                                               QualType &baseType,
1541                                               Address &addr) {
1542   const ArrayType *arrayType = origArrayType;
1543 
1544   // If it's a VLA, we have to load the stored size.  Note that
1545   // this is the size of the VLA in bytes, not its size in elements.
1546   llvm::Value *numVLAElements = nullptr;
1547   if (isa<VariableArrayType>(arrayType)) {
1548     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1549 
1550     // Walk into all VLAs.  This doesn't require changes to addr,
1551     // which has type T* where T is the first non-VLA element type.
1552     do {
1553       QualType elementType = arrayType->getElementType();
1554       arrayType = getContext().getAsArrayType(elementType);
1555 
1556       // If we only have VLA components, 'addr' requires no adjustment.
1557       if (!arrayType) {
1558         baseType = elementType;
1559         return numVLAElements;
1560       }
1561     } while (isa<VariableArrayType>(arrayType));
1562 
1563     // We get out here only if we find a constant array type
1564     // inside the VLA.
1565   }
1566 
1567   // We have some number of constant-length arrays, so addr should
1568   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1569   // down to the first element of addr.
1570   SmallVector<llvm::Value*, 8> gepIndices;
1571 
1572   // GEP down to the array type.
1573   llvm::ConstantInt *zero = Builder.getInt32(0);
1574   gepIndices.push_back(zero);
1575 
1576   uint64_t countFromCLAs = 1;
1577   QualType eltType;
1578 
1579   llvm::ArrayType *llvmArrayType =
1580     dyn_cast<llvm::ArrayType>(addr.getElementType());
1581   while (llvmArrayType) {
1582     assert(isa<ConstantArrayType>(arrayType));
1583     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1584              == llvmArrayType->getNumElements());
1585 
1586     gepIndices.push_back(zero);
1587     countFromCLAs *= llvmArrayType->getNumElements();
1588     eltType = arrayType->getElementType();
1589 
1590     llvmArrayType =
1591       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1592     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1593     assert((!llvmArrayType || arrayType) &&
1594            "LLVM and Clang types are out-of-synch");
1595   }
1596 
1597   if (arrayType) {
1598     // From this point onwards, the Clang array type has been emitted
1599     // as some other type (probably a packed struct). Compute the array
1600     // size, and just emit the 'begin' expression as a bitcast.
1601     while (arrayType) {
1602       countFromCLAs *=
1603           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1604       eltType = arrayType->getElementType();
1605       arrayType = getContext().getAsArrayType(eltType);
1606     }
1607 
1608     llvm::Type *baseType = ConvertType(eltType);
1609     addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1610   } else {
1611     // Create the actual GEP.
1612     addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1613                                              gepIndices, "array.begin"),
1614                    addr.getAlignment());
1615   }
1616 
1617   baseType = eltType;
1618 
1619   llvm::Value *numElements
1620     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1621 
1622   // If we had any VLA dimensions, factor them in.
1623   if (numVLAElements)
1624     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1625 
1626   return numElements;
1627 }
1628 
1629 std::pair<llvm::Value*, QualType>
1630 CodeGenFunction::getVLASize(QualType type) {
1631   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1632   assert(vla && "type was not a variable array type!");
1633   return getVLASize(vla);
1634 }
1635 
1636 std::pair<llvm::Value*, QualType>
1637 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1638   // The number of elements so far; always size_t.
1639   llvm::Value *numElements = nullptr;
1640 
1641   QualType elementType;
1642   do {
1643     elementType = type->getElementType();
1644     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1645     assert(vlaSize && "no size for VLA!");
1646     assert(vlaSize->getType() == SizeTy);
1647 
1648     if (!numElements) {
1649       numElements = vlaSize;
1650     } else {
1651       // It's undefined behavior if this wraps around, so mark it that way.
1652       // FIXME: Teach -fsanitize=undefined to trap this.
1653       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1654     }
1655   } while ((type = getContext().getAsVariableArrayType(elementType)));
1656 
1657   return std::pair<llvm::Value*,QualType>(numElements, elementType);
1658 }
1659 
1660 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1661   assert(type->isVariablyModifiedType() &&
1662          "Must pass variably modified type to EmitVLASizes!");
1663 
1664   EnsureInsertPoint();
1665 
1666   // We're going to walk down into the type and look for VLA
1667   // expressions.
1668   do {
1669     assert(type->isVariablyModifiedType());
1670 
1671     const Type *ty = type.getTypePtr();
1672     switch (ty->getTypeClass()) {
1673 
1674 #define TYPE(Class, Base)
1675 #define ABSTRACT_TYPE(Class, Base)
1676 #define NON_CANONICAL_TYPE(Class, Base)
1677 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1678 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1679 #include "clang/AST/TypeNodes.def"
1680       llvm_unreachable("unexpected dependent type!");
1681 
1682     // These types are never variably-modified.
1683     case Type::Builtin:
1684     case Type::Complex:
1685     case Type::Vector:
1686     case Type::ExtVector:
1687     case Type::Record:
1688     case Type::Enum:
1689     case Type::Elaborated:
1690     case Type::TemplateSpecialization:
1691     case Type::ObjCObject:
1692     case Type::ObjCInterface:
1693     case Type::ObjCObjectPointer:
1694       llvm_unreachable("type class is never variably-modified!");
1695 
1696     case Type::Adjusted:
1697       type = cast<AdjustedType>(ty)->getAdjustedType();
1698       break;
1699 
1700     case Type::Decayed:
1701       type = cast<DecayedType>(ty)->getPointeeType();
1702       break;
1703 
1704     case Type::Pointer:
1705       type = cast<PointerType>(ty)->getPointeeType();
1706       break;
1707 
1708     case Type::BlockPointer:
1709       type = cast<BlockPointerType>(ty)->getPointeeType();
1710       break;
1711 
1712     case Type::LValueReference:
1713     case Type::RValueReference:
1714       type = cast<ReferenceType>(ty)->getPointeeType();
1715       break;
1716 
1717     case Type::MemberPointer:
1718       type = cast<MemberPointerType>(ty)->getPointeeType();
1719       break;
1720 
1721     case Type::ConstantArray:
1722     case Type::IncompleteArray:
1723       // Losing element qualification here is fine.
1724       type = cast<ArrayType>(ty)->getElementType();
1725       break;
1726 
1727     case Type::VariableArray: {
1728       // Losing element qualification here is fine.
1729       const VariableArrayType *vat = cast<VariableArrayType>(ty);
1730 
1731       // Unknown size indication requires no size computation.
1732       // Otherwise, evaluate and record it.
1733       if (const Expr *size = vat->getSizeExpr()) {
1734         // It's possible that we might have emitted this already,
1735         // e.g. with a typedef and a pointer to it.
1736         llvm::Value *&entry = VLASizeMap[size];
1737         if (!entry) {
1738           llvm::Value *Size = EmitScalarExpr(size);
1739 
1740           // C11 6.7.6.2p5:
1741           //   If the size is an expression that is not an integer constant
1742           //   expression [...] each time it is evaluated it shall have a value
1743           //   greater than zero.
1744           if (SanOpts.has(SanitizerKind::VLABound) &&
1745               size->getType()->isSignedIntegerType()) {
1746             SanitizerScope SanScope(this);
1747             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1748             llvm::Constant *StaticArgs[] = {
1749               EmitCheckSourceLocation(size->getLocStart()),
1750               EmitCheckTypeDescriptor(size->getType())
1751             };
1752             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1753                                      SanitizerKind::VLABound),
1754                       "vla_bound_not_positive", StaticArgs, Size);
1755           }
1756 
1757           // Always zexting here would be wrong if it weren't
1758           // undefined behavior to have a negative bound.
1759           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1760         }
1761       }
1762       type = vat->getElementType();
1763       break;
1764     }
1765 
1766     case Type::FunctionProto:
1767     case Type::FunctionNoProto:
1768       type = cast<FunctionType>(ty)->getReturnType();
1769       break;
1770 
1771     case Type::Paren:
1772     case Type::TypeOf:
1773     case Type::UnaryTransform:
1774     case Type::Attributed:
1775     case Type::SubstTemplateTypeParm:
1776     case Type::PackExpansion:
1777       // Keep walking after single level desugaring.
1778       type = type.getSingleStepDesugaredType(getContext());
1779       break;
1780 
1781     case Type::Typedef:
1782     case Type::Decltype:
1783     case Type::Auto:
1784       // Stop walking: nothing to do.
1785       return;
1786 
1787     case Type::TypeOfExpr:
1788       // Stop walking: emit typeof expression.
1789       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1790       return;
1791 
1792     case Type::Atomic:
1793       type = cast<AtomicType>(ty)->getValueType();
1794       break;
1795 
1796     case Type::Pipe:
1797       type = cast<PipeType>(ty)->getElementType();
1798       break;
1799     }
1800   } while (type->isVariablyModifiedType());
1801 }
1802 
1803 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1804   if (getContext().getBuiltinVaListType()->isArrayType())
1805     return EmitPointerWithAlignment(E);
1806   return EmitLValue(E).getAddress();
1807 }
1808 
1809 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1810   return EmitLValue(E).getAddress();
1811 }
1812 
1813 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1814                                               llvm::Constant *Init) {
1815   assert (Init && "Invalid DeclRefExpr initializer!");
1816   if (CGDebugInfo *Dbg = getDebugInfo())
1817     if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
1818       Dbg->EmitGlobalVariable(E->getDecl(), Init);
1819 }
1820 
1821 CodeGenFunction::PeepholeProtection
1822 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1823   // At the moment, the only aggressive peephole we do in IR gen
1824   // is trunc(zext) folding, but if we add more, we can easily
1825   // extend this protection.
1826 
1827   if (!rvalue.isScalar()) return PeepholeProtection();
1828   llvm::Value *value = rvalue.getScalarVal();
1829   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1830 
1831   // Just make an extra bitcast.
1832   assert(HaveInsertPoint());
1833   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1834                                                   Builder.GetInsertBlock());
1835 
1836   PeepholeProtection protection;
1837   protection.Inst = inst;
1838   return protection;
1839 }
1840 
1841 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1842   if (!protection.Inst) return;
1843 
1844   // In theory, we could try to duplicate the peepholes now, but whatever.
1845   protection.Inst->eraseFromParent();
1846 }
1847 
1848 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1849                                                  llvm::Value *AnnotatedVal,
1850                                                  StringRef AnnotationStr,
1851                                                  SourceLocation Location) {
1852   llvm::Value *Args[4] = {
1853     AnnotatedVal,
1854     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1855     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1856     CGM.EmitAnnotationLineNo(Location)
1857   };
1858   return Builder.CreateCall(AnnotationFn, Args);
1859 }
1860 
1861 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1862   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1863   // FIXME We create a new bitcast for every annotation because that's what
1864   // llvm-gcc was doing.
1865   for (const auto *I : D->specific_attrs<AnnotateAttr>())
1866     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1867                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1868                        I->getAnnotation(), D->getLocation());
1869 }
1870 
1871 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1872                                               Address Addr) {
1873   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1874   llvm::Value *V = Addr.getPointer();
1875   llvm::Type *VTy = V->getType();
1876   llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1877                                     CGM.Int8PtrTy);
1878 
1879   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1880     // FIXME Always emit the cast inst so we can differentiate between
1881     // annotation on the first field of a struct and annotation on the struct
1882     // itself.
1883     if (VTy != CGM.Int8PtrTy)
1884       V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1885     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1886     V = Builder.CreateBitCast(V, VTy);
1887   }
1888 
1889   return Address(V, Addr.getAlignment());
1890 }
1891 
1892 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
1893 
1894 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
1895     : CGF(CGF) {
1896   assert(!CGF->IsSanitizerScope);
1897   CGF->IsSanitizerScope = true;
1898 }
1899 
1900 CodeGenFunction::SanitizerScope::~SanitizerScope() {
1901   CGF->IsSanitizerScope = false;
1902 }
1903 
1904 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
1905                                    const llvm::Twine &Name,
1906                                    llvm::BasicBlock *BB,
1907                                    llvm::BasicBlock::iterator InsertPt) const {
1908   LoopStack.InsertHelper(I);
1909   if (IsSanitizerScope)
1910     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
1911 }
1912 
1913 void CGBuilderInserter::InsertHelper(
1914     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1915     llvm::BasicBlock::iterator InsertPt) const {
1916   llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
1917   if (CGF)
1918     CGF->InsertHelper(I, Name, BB, InsertPt);
1919 }
1920 
1921 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
1922                                 CodeGenModule &CGM, const FunctionDecl *FD,
1923                                 std::string &FirstMissing) {
1924   // If there aren't any required features listed then go ahead and return.
1925   if (ReqFeatures.empty())
1926     return false;
1927 
1928   // Now build up the set of caller features and verify that all the required
1929   // features are there.
1930   llvm::StringMap<bool> CallerFeatureMap;
1931   CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
1932 
1933   // If we have at least one of the features in the feature list return
1934   // true, otherwise return false.
1935   return std::all_of(
1936       ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
1937         SmallVector<StringRef, 1> OrFeatures;
1938         Feature.split(OrFeatures, "|");
1939         return std::any_of(OrFeatures.begin(), OrFeatures.end(),
1940                            [&](StringRef Feature) {
1941                              if (!CallerFeatureMap.lookup(Feature)) {
1942                                FirstMissing = Feature.str();
1943                                return false;
1944                              }
1945                              return true;
1946                            });
1947       });
1948 }
1949 
1950 // Emits an error if we don't have a valid set of target features for the
1951 // called function.
1952 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
1953                                           const FunctionDecl *TargetDecl) {
1954   // Early exit if this is an indirect call.
1955   if (!TargetDecl)
1956     return;
1957 
1958   // Get the current enclosing function if it exists. If it doesn't
1959   // we can't check the target features anyhow.
1960   const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
1961   if (!FD)
1962     return;
1963 
1964   // Grab the required features for the call. For a builtin this is listed in
1965   // the td file with the default cpu, for an always_inline function this is any
1966   // listed cpu and any listed features.
1967   unsigned BuiltinID = TargetDecl->getBuiltinID();
1968   std::string MissingFeature;
1969   if (BuiltinID) {
1970     SmallVector<StringRef, 1> ReqFeatures;
1971     const char *FeatureList =
1972         CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1973     // Return if the builtin doesn't have any required features.
1974     if (!FeatureList || StringRef(FeatureList) == "")
1975       return;
1976     StringRef(FeatureList).split(ReqFeatures, ",");
1977     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1978       CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
1979           << TargetDecl->getDeclName()
1980           << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1981 
1982   } else if (TargetDecl->hasAttr<TargetAttr>()) {
1983     // Get the required features for the callee.
1984     SmallVector<StringRef, 1> ReqFeatures;
1985     llvm::StringMap<bool> CalleeFeatureMap;
1986     CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
1987     for (const auto &F : CalleeFeatureMap) {
1988       // Only positive features are "required".
1989       if (F.getValue())
1990         ReqFeatures.push_back(F.getKey());
1991     }
1992     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1993       CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
1994           << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
1995   }
1996 }
1997 
1998 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
1999   if (!CGM.getCodeGenOpts().SanitizeStats)
2000     return;
2001 
2002   llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2003   IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2004   CGM.getSanStats().create(IRB, SSK);
2005 }
2006