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 "CGCUDARuntime.h"
16 #include "CGCXXABI.h"
17 #include "CGDebugInfo.h"
18 #include "CGOpenMPRuntime.h"
19 #include "CodeGenModule.h"
20 #include "CodeGenPGO.h"
21 #include "TargetInfo.h"
22 #include "clang/AST/ASTContext.h"
23 #include "clang/AST/Decl.h"
24 #include "clang/AST/DeclCXX.h"
25 #include "clang/AST/StmtCXX.h"
26 #include "clang/Basic/TargetInfo.h"
27 #include "clang/CodeGen/CGFunctionInfo.h"
28 #include "clang/Frontend/CodeGenOptions.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/MDBuilder.h"
32 #include "llvm/IR/Operator.h"
33 using namespace clang;
34 using namespace CodeGen;
35 
36 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
37     : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
38       Builder(cgm.getModule().getContext(), llvm::ConstantFolder(),
39               CGBuilderInserterTy(this)),
40       CurFn(nullptr), CapturedStmtInfo(nullptr),
41       SanOpts(CGM.getLangOpts().Sanitize), IsSanitizerScope(false),
42       CurFuncIsThunk(false), AutoreleaseResult(false), SawAsmBlock(false),
43       BlockInfo(nullptr), BlockPointer(nullptr),
44       LambdaThisCaptureField(nullptr), NormalCleanupDest(nullptr),
45       NextCleanupDestIndex(1), FirstBlockInfo(nullptr), EHResumeBlock(nullptr),
46       ExceptionSlot(nullptr), EHSelectorSlot(nullptr),
47       AbnormalTerminationSlot(nullptr), SEHPointersDecl(nullptr),
48       DebugInfo(CGM.getModuleDebugInfo()), DisableDebugInfo(false),
49       DidCallStackSave(false), IndirectBranch(nullptr), PGO(cgm),
50       SwitchInsn(nullptr), SwitchWeights(nullptr), CaseRangeBlock(nullptr),
51       UnreachableBlock(nullptr), NumReturnExprs(0), NumSimpleReturnExprs(0),
52       CXXABIThisDecl(nullptr), CXXABIThisValue(nullptr), CXXThisValue(nullptr),
53       CXXDefaultInitExprThis(nullptr), CXXStructorImplicitParamDecl(nullptr),
54       CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
55       CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
56       TerminateHandler(nullptr), TrapBB(nullptr) {
57   if (!suppressNewContext)
58     CGM.getCXXABI().getMangleContext().startNewFunction();
59 
60   llvm::FastMathFlags FMF;
61   if (CGM.getLangOpts().FastMath)
62     FMF.setUnsafeAlgebra();
63   if (CGM.getLangOpts().FiniteMathOnly) {
64     FMF.setNoNaNs();
65     FMF.setNoInfs();
66   }
67   if (CGM.getCodeGenOpts().NoNaNsFPMath) {
68     FMF.setNoNaNs();
69   }
70   if (CGM.getCodeGenOpts().NoSignedZeros) {
71     FMF.setNoSignedZeros();
72   }
73   Builder.SetFastMathFlags(FMF);
74 }
75 
76 CodeGenFunction::~CodeGenFunction() {
77   assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
78 
79   // If there are any unclaimed block infos, go ahead and destroy them
80   // now.  This can happen if IR-gen gets clever and skips evaluating
81   // something.
82   if (FirstBlockInfo)
83     destroyBlockInfos(FirstBlockInfo);
84 
85   if (getLangOpts().OpenMP) {
86     CGM.getOpenMPRuntime().functionFinished(*this);
87   }
88 }
89 
90 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
91   CharUnits Alignment;
92   if (CGM.getCXXABI().isTypeInfoCalculable(T)) {
93     Alignment = getContext().getTypeAlignInChars(T);
94     unsigned MaxAlign = getContext().getLangOpts().MaxTypeAlign;
95     if (MaxAlign && Alignment.getQuantity() > MaxAlign &&
96         !getContext().isAlignmentRequired(T))
97       Alignment = CharUnits::fromQuantity(MaxAlign);
98   }
99   return LValue::MakeAddr(V, T, Alignment, getContext(), CGM.getTBAAInfo(T));
100 }
101 
102 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
103   return CGM.getTypes().ConvertTypeForMem(T);
104 }
105 
106 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
107   return CGM.getTypes().ConvertType(T);
108 }
109 
110 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
111   type = type.getCanonicalType();
112   while (true) {
113     switch (type->getTypeClass()) {
114 #define TYPE(name, parent)
115 #define ABSTRACT_TYPE(name, parent)
116 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
117 #define DEPENDENT_TYPE(name, parent) case Type::name:
118 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
119 #include "clang/AST/TypeNodes.def"
120       llvm_unreachable("non-canonical or dependent type in IR-generation");
121 
122     case Type::Auto:
123       llvm_unreachable("undeduced auto type in IR-generation");
124 
125     // Various scalar types.
126     case Type::Builtin:
127     case Type::Pointer:
128     case Type::BlockPointer:
129     case Type::LValueReference:
130     case Type::RValueReference:
131     case Type::MemberPointer:
132     case Type::Vector:
133     case Type::ExtVector:
134     case Type::FunctionProto:
135     case Type::FunctionNoProto:
136     case Type::Enum:
137     case Type::ObjCObjectPointer:
138       return TEK_Scalar;
139 
140     // Complexes.
141     case Type::Complex:
142       return TEK_Complex;
143 
144     // Arrays, records, and Objective-C objects.
145     case Type::ConstantArray:
146     case Type::IncompleteArray:
147     case Type::VariableArray:
148     case Type::Record:
149     case Type::ObjCObject:
150     case Type::ObjCInterface:
151       return TEK_Aggregate;
152 
153     // We operate on atomic values according to their underlying type.
154     case Type::Atomic:
155       type = cast<AtomicType>(type)->getValueType();
156       continue;
157     }
158     llvm_unreachable("unknown type kind!");
159   }
160 }
161 
162 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
163   // For cleanliness, we try to avoid emitting the return block for
164   // simple cases.
165   llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
166 
167   if (CurBB) {
168     assert(!CurBB->getTerminator() && "Unexpected terminated block.");
169 
170     // We have a valid insert point, reuse it if it is empty or there are no
171     // explicit jumps to the return block.
172     if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
173       ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
174       delete ReturnBlock.getBlock();
175     } else
176       EmitBlock(ReturnBlock.getBlock());
177     return llvm::DebugLoc();
178   }
179 
180   // Otherwise, if the return block is the target of a single direct
181   // branch then we can just put the code in that block instead. This
182   // cleans up functions which started with a unified return block.
183   if (ReturnBlock.getBlock()->hasOneUse()) {
184     llvm::BranchInst *BI =
185       dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
186     if (BI && BI->isUnconditional() &&
187         BI->getSuccessor(0) == ReturnBlock.getBlock()) {
188       // Record/return the DebugLoc of the simple 'return' expression to be used
189       // later by the actual 'ret' instruction.
190       llvm::DebugLoc Loc = BI->getDebugLoc();
191       Builder.SetInsertPoint(BI->getParent());
192       BI->eraseFromParent();
193       delete ReturnBlock.getBlock();
194       return Loc;
195     }
196   }
197 
198   // FIXME: We are at an unreachable point, there is no reason to emit the block
199   // unless it has uses. However, we still need a place to put the debug
200   // region.end for now.
201 
202   EmitBlock(ReturnBlock.getBlock());
203   return llvm::DebugLoc();
204 }
205 
206 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
207   if (!BB) return;
208   if (!BB->use_empty())
209     return CGF.CurFn->getBasicBlockList().push_back(BB);
210   delete BB;
211 }
212 
213 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
214   assert(BreakContinueStack.empty() &&
215          "mismatched push/pop in break/continue stack!");
216 
217   bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
218     && NumSimpleReturnExprs == NumReturnExprs
219     && ReturnBlock.getBlock()->use_empty();
220   // Usually the return expression is evaluated before the cleanup
221   // code.  If the function contains only a simple return statement,
222   // such as a constant, the location before the cleanup code becomes
223   // the last useful breakpoint in the function, because the simple
224   // return expression will be evaluated after the cleanup code. To be
225   // safe, set the debug location for cleanup code to the location of
226   // the return statement.  Otherwise the cleanup code should be at the
227   // end of the function's lexical scope.
228   //
229   // If there are multiple branches to the return block, the branch
230   // instructions will get the location of the return statements and
231   // all will be fine.
232   if (CGDebugInfo *DI = getDebugInfo()) {
233     if (OnlySimpleReturnStmts)
234       DI->EmitLocation(Builder, LastStopPoint);
235     else
236       DI->EmitLocation(Builder, EndLoc);
237   }
238 
239   // Pop any cleanups that might have been associated with the
240   // parameters.  Do this in whatever block we're currently in; it's
241   // important to do this before we enter the return block or return
242   // edges will be *really* confused.
243   bool EmitRetDbgLoc = true;
244   if (EHStack.stable_begin() != PrologueCleanupDepth) {
245     // Make sure the line table doesn't jump back into the body for
246     // the ret after it's been at EndLoc.
247     EmitRetDbgLoc = false;
248 
249     if (CGDebugInfo *DI = getDebugInfo())
250       if (OnlySimpleReturnStmts)
251         DI->EmitLocation(Builder, EndLoc);
252 
253     PopCleanupBlocks(PrologueCleanupDepth);
254   }
255 
256   // Emit function epilog (to return).
257   llvm::DebugLoc Loc = EmitReturnBlock();
258 
259   if (ShouldInstrumentFunction())
260     EmitFunctionInstrumentation("__cyg_profile_func_exit");
261 
262   // Emit debug descriptor for function end.
263   if (CGDebugInfo *DI = getDebugInfo())
264     DI->EmitFunctionEnd(Builder);
265 
266   // Reset the debug location to that of the simple 'return' expression, if any
267   // rather than that of the end of the function's scope '}'.
268   ApplyDebugLocation AL(*this, Loc);
269   EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
270   EmitEndEHSpec(CurCodeDecl);
271 
272   assert(EHStack.empty() &&
273          "did not remove all scopes from cleanup stack!");
274 
275   // If someone did an indirect goto, emit the indirect goto block at the end of
276   // the function.
277   if (IndirectBranch) {
278     EmitBlock(IndirectBranch->getParent());
279     Builder.ClearInsertionPoint();
280   }
281 
282   // Remove the AllocaInsertPt instruction, which is just a convenience for us.
283   llvm::Instruction *Ptr = AllocaInsertPt;
284   AllocaInsertPt = nullptr;
285   Ptr->eraseFromParent();
286 
287   // If someone took the address of a label but never did an indirect goto, we
288   // made a zero entry PHI node, which is illegal, zap it now.
289   if (IndirectBranch) {
290     llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
291     if (PN->getNumIncomingValues() == 0) {
292       PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
293       PN->eraseFromParent();
294     }
295   }
296 
297   EmitIfUsed(*this, EHResumeBlock);
298   EmitIfUsed(*this, TerminateLandingPad);
299   EmitIfUsed(*this, TerminateHandler);
300   EmitIfUsed(*this, UnreachableBlock);
301 
302   if (CGM.getCodeGenOpts().EmitDeclMetadata)
303     EmitDeclMetadata();
304 
305   for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
306            I = DeferredReplacements.begin(),
307            E = DeferredReplacements.end();
308        I != E; ++I) {
309     I->first->replaceAllUsesWith(I->second);
310     I->first->eraseFromParent();
311   }
312 }
313 
314 /// ShouldInstrumentFunction - Return true if the current function should be
315 /// instrumented with __cyg_profile_func_* calls
316 bool CodeGenFunction::ShouldInstrumentFunction() {
317   if (!CGM.getCodeGenOpts().InstrumentFunctions)
318     return false;
319   if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
320     return false;
321   return true;
322 }
323 
324 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
325 /// instrumentation function with the current function and the call site, if
326 /// function instrumentation is enabled.
327 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
328   // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
329   llvm::PointerType *PointerTy = Int8PtrTy;
330   llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
331   llvm::FunctionType *FunctionTy =
332     llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
333 
334   llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
335   llvm::CallInst *CallSite = Builder.CreateCall(
336     CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
337     llvm::ConstantInt::get(Int32Ty, 0),
338     "callsite");
339 
340   llvm::Value *args[] = {
341     llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
342     CallSite
343   };
344 
345   EmitNounwindRuntimeCall(F, args);
346 }
347 
348 void CodeGenFunction::EmitMCountInstrumentation() {
349   llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
350 
351   llvm::Constant *MCountFn =
352     CGM.CreateRuntimeFunction(FTy, getTarget().getMCountName());
353   EmitNounwindRuntimeCall(MCountFn);
354 }
355 
356 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
357 // information in the program executable. The argument information stored
358 // includes the argument name, its type, the address and access qualifiers used.
359 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
360                                  CodeGenModule &CGM, llvm::LLVMContext &Context,
361                                  SmallVector<llvm::Metadata *, 5> &kernelMDArgs,
362                                  CGBuilderTy &Builder, ASTContext &ASTCtx) {
363   // Create MDNodes that represent the kernel arg metadata.
364   // Each MDNode is a list in the form of "key", N number of values which is
365   // the same number of values as their are kernel arguments.
366 
367   const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
368 
369   // MDNode for the kernel argument address space qualifiers.
370   SmallVector<llvm::Metadata *, 8> addressQuals;
371   addressQuals.push_back(llvm::MDString::get(Context, "kernel_arg_addr_space"));
372 
373   // MDNode for the kernel argument access qualifiers (images only).
374   SmallVector<llvm::Metadata *, 8> accessQuals;
375   accessQuals.push_back(llvm::MDString::get(Context, "kernel_arg_access_qual"));
376 
377   // MDNode for the kernel argument type names.
378   SmallVector<llvm::Metadata *, 8> argTypeNames;
379   argTypeNames.push_back(llvm::MDString::get(Context, "kernel_arg_type"));
380 
381   // MDNode for the kernel argument base type names.
382   SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
383   argBaseTypeNames.push_back(
384       llvm::MDString::get(Context, "kernel_arg_base_type"));
385 
386   // MDNode for the kernel argument type qualifiers.
387   SmallVector<llvm::Metadata *, 8> argTypeQuals;
388   argTypeQuals.push_back(llvm::MDString::get(Context, "kernel_arg_type_qual"));
389 
390   // MDNode for the kernel argument names.
391   SmallVector<llvm::Metadata *, 8> argNames;
392   argNames.push_back(llvm::MDString::get(Context, "kernel_arg_name"));
393 
394   for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
395     const ParmVarDecl *parm = FD->getParamDecl(i);
396     QualType ty = parm->getType();
397     std::string typeQuals;
398 
399     if (ty->isPointerType()) {
400       QualType pointeeTy = ty->getPointeeType();
401 
402       // Get address qualifier.
403       addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
404           ASTCtx.getTargetAddressSpace(pointeeTy.getAddressSpace()))));
405 
406       // Get argument type name.
407       std::string typeName =
408           pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
409 
410       // Turn "unsigned type" to "utype"
411       std::string::size_type pos = typeName.find("unsigned");
412       if (pointeeTy.isCanonical() && pos != std::string::npos)
413         typeName.erase(pos+1, 8);
414 
415       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
416 
417       std::string baseTypeName =
418           pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
419               Policy) +
420           "*";
421 
422       // Turn "unsigned type" to "utype"
423       pos = baseTypeName.find("unsigned");
424       if (pos != std::string::npos)
425         baseTypeName.erase(pos+1, 8);
426 
427       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
428 
429       // Get argument type qualifiers:
430       if (ty.isRestrictQualified())
431         typeQuals = "restrict";
432       if (pointeeTy.isConstQualified() ||
433           (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
434         typeQuals += typeQuals.empty() ? "const" : " const";
435       if (pointeeTy.isVolatileQualified())
436         typeQuals += typeQuals.empty() ? "volatile" : " volatile";
437     } else {
438       uint32_t AddrSpc = 0;
439       if (ty->isImageType())
440         AddrSpc =
441           CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
442 
443       addressQuals.push_back(
444           llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
445 
446       // Get argument type name.
447       std::string typeName = ty.getUnqualifiedType().getAsString(Policy);
448 
449       // Turn "unsigned type" to "utype"
450       std::string::size_type pos = typeName.find("unsigned");
451       if (ty.isCanonical() && pos != std::string::npos)
452         typeName.erase(pos+1, 8);
453 
454       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
455 
456       std::string baseTypeName =
457           ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
458 
459       // Turn "unsigned type" to "utype"
460       pos = baseTypeName.find("unsigned");
461       if (pos != std::string::npos)
462         baseTypeName.erase(pos+1, 8);
463 
464       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
465 
466       // Get argument type qualifiers:
467       if (ty.isConstQualified())
468         typeQuals = "const";
469       if (ty.isVolatileQualified())
470         typeQuals += typeQuals.empty() ? "volatile" : " volatile";
471     }
472 
473     argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
474 
475     // Get image access qualifier:
476     if (ty->isImageType()) {
477       const OpenCLImageAccessAttr *A = parm->getAttr<OpenCLImageAccessAttr>();
478       if (A && A->isWriteOnly())
479         accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
480       else
481         accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
482       // FIXME: what about read_write?
483     } else
484       accessQuals.push_back(llvm::MDString::get(Context, "none"));
485 
486     // Get argument name.
487     argNames.push_back(llvm::MDString::get(Context, parm->getName()));
488   }
489 
490   kernelMDArgs.push_back(llvm::MDNode::get(Context, addressQuals));
491   kernelMDArgs.push_back(llvm::MDNode::get(Context, accessQuals));
492   kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeNames));
493   kernelMDArgs.push_back(llvm::MDNode::get(Context, argBaseTypeNames));
494   kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeQuals));
495   if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
496     kernelMDArgs.push_back(llvm::MDNode::get(Context, argNames));
497 }
498 
499 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
500                                                llvm::Function *Fn)
501 {
502   if (!FD->hasAttr<OpenCLKernelAttr>())
503     return;
504 
505   llvm::LLVMContext &Context = getLLVMContext();
506 
507   SmallVector<llvm::Metadata *, 5> kernelMDArgs;
508   kernelMDArgs.push_back(llvm::ConstantAsMetadata::get(Fn));
509 
510   GenOpenCLArgMetadata(FD, Fn, CGM, Context, kernelMDArgs, Builder,
511                        getContext());
512 
513   if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
514     QualType hintQTy = A->getTypeHint();
515     const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
516     bool isSignedInteger =
517         hintQTy->isSignedIntegerType() ||
518         (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
519     llvm::Metadata *attrMDArgs[] = {
520         llvm::MDString::get(Context, "vec_type_hint"),
521         llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
522             CGM.getTypes().ConvertType(A->getTypeHint()))),
523         llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
524             llvm::IntegerType::get(Context, 32),
525             llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
526     kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
527   }
528 
529   if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
530     llvm::Metadata *attrMDArgs[] = {
531         llvm::MDString::get(Context, "work_group_size_hint"),
532         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
533         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
534         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
535     kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
536   }
537 
538   if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
539     llvm::Metadata *attrMDArgs[] = {
540         llvm::MDString::get(Context, "reqd_work_group_size"),
541         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
542         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
543         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
544     kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
545   }
546 
547   llvm::MDNode *kernelMDNode = llvm::MDNode::get(Context, kernelMDArgs);
548   llvm::NamedMDNode *OpenCLKernelMetadata =
549     CGM.getModule().getOrInsertNamedMetadata("opencl.kernels");
550   OpenCLKernelMetadata->addOperand(kernelMDNode);
551 }
552 
553 /// Determine whether the function F ends with a return stmt.
554 static bool endsWithReturn(const Decl* F) {
555   const Stmt *Body = nullptr;
556   if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
557     Body = FD->getBody();
558   else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
559     Body = OMD->getBody();
560 
561   if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
562     auto LastStmt = CS->body_rbegin();
563     if (LastStmt != CS->body_rend())
564       return isa<ReturnStmt>(*LastStmt);
565   }
566   return false;
567 }
568 
569 void CodeGenFunction::StartFunction(GlobalDecl GD,
570                                     QualType RetTy,
571                                     llvm::Function *Fn,
572                                     const CGFunctionInfo &FnInfo,
573                                     const FunctionArgList &Args,
574                                     SourceLocation Loc,
575                                     SourceLocation StartLoc) {
576   assert(!CurFn &&
577          "Do not use a CodeGenFunction object for more than one function");
578 
579   const Decl *D = GD.getDecl();
580 
581   DidCallStackSave = false;
582   CurCodeDecl = D;
583   CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
584   FnRetTy = RetTy;
585   CurFn = Fn;
586   CurFnInfo = &FnInfo;
587   assert(CurFn->isDeclaration() && "Function already has body?");
588 
589   if (CGM.isInSanitizerBlacklist(Fn, Loc))
590     SanOpts.clear();
591 
592   // Pass inline keyword to optimizer if it appears explicitly on any
593   // declaration. Also, in the case of -fno-inline attach NoInline
594   // attribute to all function that are not marked AlwaysInline.
595   if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
596     if (!CGM.getCodeGenOpts().NoInline) {
597       for (auto RI : FD->redecls())
598         if (RI->isInlineSpecified()) {
599           Fn->addFnAttr(llvm::Attribute::InlineHint);
600           break;
601         }
602     } else if (!FD->hasAttr<AlwaysInlineAttr>())
603       Fn->addFnAttr(llvm::Attribute::NoInline);
604   }
605 
606   if (getLangOpts().OpenCL) {
607     // Add metadata for a kernel function.
608     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
609       EmitOpenCLKernelMetadata(FD, Fn);
610   }
611 
612   // If we are checking function types, emit a function type signature as
613   // prologue data.
614   if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
615     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
616       if (llvm::Constant *PrologueSig =
617               CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
618         llvm::Constant *FTRTTIConst =
619             CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
620         llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
621         llvm::Constant *PrologueStructConst =
622             llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
623         Fn->setPrologueData(PrologueStructConst);
624       }
625     }
626   }
627 
628   llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
629 
630   // Create a marker to make it easy to insert allocas into the entryblock
631   // later.  Don't create this with the builder, because we don't want it
632   // folded.
633   llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
634   AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "", EntryBB);
635   if (Builder.isNamePreserving())
636     AllocaInsertPt->setName("allocapt");
637 
638   ReturnBlock = getJumpDestInCurrentScope("return");
639 
640   Builder.SetInsertPoint(EntryBB);
641 
642   // Emit subprogram debug descriptor.
643   if (CGDebugInfo *DI = getDebugInfo()) {
644     SmallVector<QualType, 16> ArgTypes;
645     for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
646 	 i != e; ++i) {
647       ArgTypes.push_back((*i)->getType());
648     }
649 
650     QualType FnType =
651       getContext().getFunctionType(RetTy, ArgTypes,
652                                    FunctionProtoType::ExtProtoInfo());
653     DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
654   }
655 
656   if (ShouldInstrumentFunction())
657     EmitFunctionInstrumentation("__cyg_profile_func_enter");
658 
659   if (CGM.getCodeGenOpts().InstrumentForProfiling)
660     EmitMCountInstrumentation();
661 
662   if (RetTy->isVoidType()) {
663     // Void type; nothing to return.
664     ReturnValue = nullptr;
665 
666     // Count the implicit return.
667     if (!endsWithReturn(D))
668       ++NumReturnExprs;
669   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
670              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
671     // Indirect aggregate return; emit returned value directly into sret slot.
672     // This reduces code size, and affects correctness in C++.
673     auto AI = CurFn->arg_begin();
674     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
675       ++AI;
676     ReturnValue = AI;
677   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
678              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
679     // Load the sret pointer from the argument struct and return into that.
680     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
681     llvm::Function::arg_iterator EI = CurFn->arg_end();
682     --EI;
683     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, EI, Idx);
684     ReturnValue = Builder.CreateLoad(Addr, "agg.result");
685   } else {
686     ReturnValue = CreateIRTemp(RetTy, "retval");
687 
688     // Tell the epilog emitter to autorelease the result.  We do this
689     // now so that various specialized functions can suppress it
690     // during their IR-generation.
691     if (getLangOpts().ObjCAutoRefCount &&
692         !CurFnInfo->isReturnsRetained() &&
693         RetTy->isObjCRetainableType())
694       AutoreleaseResult = true;
695   }
696 
697   EmitStartEHSpec(CurCodeDecl);
698 
699   PrologueCleanupDepth = EHStack.stable_begin();
700   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
701 
702   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
703     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
704     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
705     if (MD->getParent()->isLambda() &&
706         MD->getOverloadedOperator() == OO_Call) {
707       // We're in a lambda; figure out the captures.
708       MD->getParent()->getCaptureFields(LambdaCaptureFields,
709                                         LambdaThisCaptureField);
710       if (LambdaThisCaptureField) {
711         // If this lambda captures this, load it.
712         LValue ThisLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
713         CXXThisValue = EmitLoadOfLValue(ThisLValue,
714                                         SourceLocation()).getScalarVal();
715       }
716       for (auto *FD : MD->getParent()->fields()) {
717         if (FD->hasCapturedVLAType()) {
718           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
719                                            SourceLocation()).getScalarVal();
720           auto VAT = FD->getCapturedVLAType();
721           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
722         }
723       }
724     } else {
725       // Not in a lambda; just use 'this' from the method.
726       // FIXME: Should we generate a new load for each use of 'this'?  The
727       // fast register allocator would be happier...
728       CXXThisValue = CXXABIThisValue;
729     }
730   }
731 
732   // If any of the arguments have a variably modified type, make sure to
733   // emit the type size.
734   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
735        i != e; ++i) {
736     const VarDecl *VD = *i;
737 
738     // Dig out the type as written from ParmVarDecls; it's unclear whether
739     // the standard (C99 6.9.1p10) requires this, but we're following the
740     // precedent set by gcc.
741     QualType Ty;
742     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
743       Ty = PVD->getOriginalType();
744     else
745       Ty = VD->getType();
746 
747     if (Ty->isVariablyModifiedType())
748       EmitVariablyModifiedType(Ty);
749   }
750   // Emit a location at the end of the prologue.
751   if (CGDebugInfo *DI = getDebugInfo())
752     DI->EmitLocation(Builder, StartLoc);
753 }
754 
755 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
756                                        const Stmt *Body) {
757   RegionCounter Cnt = getPGORegionCounter(Body);
758   Cnt.beginRegion(Builder);
759   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
760     EmitCompoundStmtWithoutScope(*S);
761   else
762     EmitStmt(Body);
763 }
764 
765 /// When instrumenting to collect profile data, the counts for some blocks
766 /// such as switch cases need to not include the fall-through counts, so
767 /// emit a branch around the instrumentation code. When not instrumenting,
768 /// this just calls EmitBlock().
769 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
770                                                RegionCounter &Cnt) {
771   llvm::BasicBlock *SkipCountBB = nullptr;
772   if (HaveInsertPoint() && CGM.getCodeGenOpts().ProfileInstrGenerate) {
773     // When instrumenting for profiling, the fallthrough to certain
774     // statements needs to skip over the instrumentation code so that we
775     // get an accurate count.
776     SkipCountBB = createBasicBlock("skipcount");
777     EmitBranch(SkipCountBB);
778   }
779   EmitBlock(BB);
780   Cnt.beginRegion(Builder, /*AddIncomingFallThrough=*/true);
781   if (SkipCountBB)
782     EmitBlock(SkipCountBB);
783 }
784 
785 /// Tries to mark the given function nounwind based on the
786 /// non-existence of any throwing calls within it.  We believe this is
787 /// lightweight enough to do at -O0.
788 static void TryMarkNoThrow(llvm::Function *F) {
789   // LLVM treats 'nounwind' on a function as part of the type, so we
790   // can't do this on functions that can be overwritten.
791   if (F->mayBeOverridden()) return;
792 
793   for (llvm::Function::iterator FI = F->begin(), FE = F->end(); FI != FE; ++FI)
794     for (llvm::BasicBlock::iterator
795            BI = FI->begin(), BE = FI->end(); BI != BE; ++BI)
796       if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(&*BI)) {
797         if (!Call->doesNotThrow())
798           return;
799       } else if (isa<llvm::ResumeInst>(&*BI)) {
800         return;
801       }
802   F->setDoesNotThrow();
803 }
804 
805 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
806                                    const CGFunctionInfo &FnInfo) {
807   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
808 
809   // Check if we should generate debug info for this function.
810   if (FD->hasAttr<NoDebugAttr>())
811     DebugInfo = nullptr; // disable debug info indefinitely for this function
812 
813   FunctionArgList Args;
814   QualType ResTy = FD->getReturnType();
815 
816   CurGD = GD;
817   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
818   if (MD && MD->isInstance()) {
819     if (CGM.getCXXABI().HasThisReturn(GD))
820       ResTy = MD->getThisType(getContext());
821     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
822       ResTy = CGM.getContext().VoidPtrTy;
823     CGM.getCXXABI().buildThisParam(*this, Args);
824   }
825 
826   Args.append(FD->param_begin(), FD->param_end());
827 
828   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
829     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
830 
831   SourceRange BodyRange;
832   if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
833   CurEHLocation = BodyRange.getEnd();
834 
835   // Use the location of the start of the function to determine where
836   // the function definition is located. By default use the location
837   // of the declaration as the location for the subprogram. A function
838   // may lack a declaration in the source code if it is created by code
839   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
840   SourceLocation Loc = FD->getLocation();
841 
842   // If this is a function specialization then use the pattern body
843   // as the location for the function.
844   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
845     if (SpecDecl->hasBody(SpecDecl))
846       Loc = SpecDecl->getLocation();
847 
848   // Emit the standard function prologue.
849   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
850 
851   // Generate the body of the function.
852   PGO.checkGlobalDecl(GD);
853   PGO.assignRegionCounters(GD.getDecl(), CurFn);
854   if (isa<CXXDestructorDecl>(FD))
855     EmitDestructorBody(Args);
856   else if (isa<CXXConstructorDecl>(FD))
857     EmitConstructorBody(Args);
858   else if (getLangOpts().CUDA &&
859            !getLangOpts().CUDAIsDevice &&
860            FD->hasAttr<CUDAGlobalAttr>())
861     CGM.getCUDARuntime().EmitDeviceStubBody(*this, Args);
862   else if (isa<CXXConversionDecl>(FD) &&
863            cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
864     // The lambda conversion to block pointer is special; the semantics can't be
865     // expressed in the AST, so IRGen needs to special-case it.
866     EmitLambdaToBlockPointerBody(Args);
867   } else if (isa<CXXMethodDecl>(FD) &&
868              cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
869     // The lambda static invoker function is special, because it forwards or
870     // clones the body of the function call operator (but is actually static).
871     EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
872   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
873              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
874               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
875     // Implicit copy-assignment gets the same special treatment as implicit
876     // copy-constructors.
877     emitImplicitAssignmentOperatorBody(Args);
878   } else if (Stmt *Body = FD->getBody()) {
879     EmitFunctionBody(Args, Body);
880   } else
881     llvm_unreachable("no definition for emitted function");
882 
883   // C++11 [stmt.return]p2:
884   //   Flowing off the end of a function [...] results in undefined behavior in
885   //   a value-returning function.
886   // C11 6.9.1p12:
887   //   If the '}' that terminates a function is reached, and the value of the
888   //   function call is used by the caller, the behavior is undefined.
889   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
890       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
891     if (SanOpts.has(SanitizerKind::Return)) {
892       SanitizerScope SanScope(this);
893       llvm::Value *IsFalse = Builder.getFalse();
894       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
895                 "missing_return", EmitCheckSourceLocation(FD->getLocation()),
896                 None);
897     } else if (CGM.getCodeGenOpts().OptimizationLevel == 0)
898       Builder.CreateCall(CGM.getIntrinsic(llvm::Intrinsic::trap));
899     Builder.CreateUnreachable();
900     Builder.ClearInsertionPoint();
901   }
902 
903   // Emit the standard function epilogue.
904   FinishFunction(BodyRange.getEnd());
905 
906   // If we haven't marked the function nothrow through other means, do
907   // a quick pass now to see if we can.
908   if (!CurFn->doesNotThrow())
909     TryMarkNoThrow(CurFn);
910 }
911 
912 /// ContainsLabel - Return true if the statement contains a label in it.  If
913 /// this statement is not executed normally, it not containing a label means
914 /// that we can just remove the code.
915 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
916   // Null statement, not a label!
917   if (!S) return false;
918 
919   // If this is a label, we have to emit the code, consider something like:
920   // if (0) {  ...  foo:  bar(); }  goto foo;
921   //
922   // TODO: If anyone cared, we could track __label__'s, since we know that you
923   // can't jump to one from outside their declared region.
924   if (isa<LabelStmt>(S))
925     return true;
926 
927   // If this is a case/default statement, and we haven't seen a switch, we have
928   // to emit the code.
929   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
930     return true;
931 
932   // If this is a switch statement, we want to ignore cases below it.
933   if (isa<SwitchStmt>(S))
934     IgnoreCaseStmts = true;
935 
936   // Scan subexpressions for verboten labels.
937   for (Stmt::const_child_range I = S->children(); I; ++I)
938     if (ContainsLabel(*I, IgnoreCaseStmts))
939       return true;
940 
941   return false;
942 }
943 
944 /// containsBreak - Return true if the statement contains a break out of it.
945 /// If the statement (recursively) contains a switch or loop with a break
946 /// inside of it, this is fine.
947 bool CodeGenFunction::containsBreak(const Stmt *S) {
948   // Null statement, not a label!
949   if (!S) return false;
950 
951   // If this is a switch or loop that defines its own break scope, then we can
952   // include it and anything inside of it.
953   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
954       isa<ForStmt>(S))
955     return false;
956 
957   if (isa<BreakStmt>(S))
958     return true;
959 
960   // Scan subexpressions for verboten breaks.
961   for (Stmt::const_child_range I = S->children(); I; ++I)
962     if (containsBreak(*I))
963       return true;
964 
965   return false;
966 }
967 
968 
969 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
970 /// to a constant, or if it does but contains a label, return false.  If it
971 /// constant folds return true and set the boolean result in Result.
972 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
973                                                    bool &ResultBool) {
974   llvm::APSInt ResultInt;
975   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt))
976     return false;
977 
978   ResultBool = ResultInt.getBoolValue();
979   return true;
980 }
981 
982 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
983 /// to a constant, or if it does but contains a label, return false.  If it
984 /// constant folds return true and set the folded value.
985 bool CodeGenFunction::
986 ConstantFoldsToSimpleInteger(const Expr *Cond, llvm::APSInt &ResultInt) {
987   // FIXME: Rename and handle conversion of other evaluatable things
988   // to bool.
989   llvm::APSInt Int;
990   if (!Cond->EvaluateAsInt(Int, getContext()))
991     return false;  // Not foldable, not integer or not fully evaluatable.
992 
993   if (CodeGenFunction::ContainsLabel(Cond))
994     return false;  // Contains a label.
995 
996   ResultInt = Int;
997   return true;
998 }
999 
1000 
1001 
1002 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1003 /// statement) to the specified blocks.  Based on the condition, this might try
1004 /// to simplify the codegen of the conditional based on the branch.
1005 ///
1006 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1007                                            llvm::BasicBlock *TrueBlock,
1008                                            llvm::BasicBlock *FalseBlock,
1009                                            uint64_t TrueCount) {
1010   Cond = Cond->IgnoreParens();
1011 
1012   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1013 
1014     // Handle X && Y in a condition.
1015     if (CondBOp->getOpcode() == BO_LAnd) {
1016       RegionCounter Cnt = getPGORegionCounter(CondBOp);
1017 
1018       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1019       // folded if the case was simple enough.
1020       bool ConstantBool = false;
1021       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1022           ConstantBool) {
1023         // br(1 && X) -> br(X).
1024         Cnt.beginRegion(Builder);
1025         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1026                                     TrueCount);
1027       }
1028 
1029       // If we have "X && 1", simplify the code to use an uncond branch.
1030       // "X && 0" would have been constant folded to 0.
1031       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1032           ConstantBool) {
1033         // br(X && 1) -> br(X).
1034         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1035                                     TrueCount);
1036       }
1037 
1038       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1039       // want to jump to the FalseBlock.
1040       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1041       // The counter tells us how often we evaluate RHS, and all of TrueCount
1042       // can be propagated to that branch.
1043       uint64_t RHSCount = Cnt.getCount();
1044 
1045       ConditionalEvaluation eval(*this);
1046       {
1047         ApplyDebugLocation DL(*this, Cond);
1048         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1049         EmitBlock(LHSTrue);
1050       }
1051 
1052       // Any temporaries created here are conditional.
1053       Cnt.beginRegion(Builder);
1054       eval.begin(*this);
1055       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1056       eval.end(*this);
1057 
1058       return;
1059     }
1060 
1061     if (CondBOp->getOpcode() == BO_LOr) {
1062       RegionCounter Cnt = getPGORegionCounter(CondBOp);
1063 
1064       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1065       // folded if the case was simple enough.
1066       bool ConstantBool = false;
1067       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1068           !ConstantBool) {
1069         // br(0 || X) -> br(X).
1070         Cnt.beginRegion(Builder);
1071         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1072                                     TrueCount);
1073       }
1074 
1075       // If we have "X || 0", simplify the code to use an uncond branch.
1076       // "X || 1" would have been constant folded to 1.
1077       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1078           !ConstantBool) {
1079         // br(X || 0) -> br(X).
1080         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1081                                     TrueCount);
1082       }
1083 
1084       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1085       // want to jump to the TrueBlock.
1086       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1087       // We have the count for entry to the RHS and for the whole expression
1088       // being true, so we can divy up True count between the short circuit and
1089       // the RHS.
1090       uint64_t LHSCount = Cnt.getParentCount() - Cnt.getCount();
1091       uint64_t RHSCount = TrueCount - LHSCount;
1092 
1093       ConditionalEvaluation eval(*this);
1094       {
1095         ApplyDebugLocation DL(*this, Cond);
1096         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1097         EmitBlock(LHSFalse);
1098       }
1099 
1100       // Any temporaries created here are conditional.
1101       Cnt.beginRegion(Builder);
1102       eval.begin(*this);
1103       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1104 
1105       eval.end(*this);
1106 
1107       return;
1108     }
1109   }
1110 
1111   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1112     // br(!x, t, f) -> br(x, f, t)
1113     if (CondUOp->getOpcode() == UO_LNot) {
1114       // Negate the count.
1115       uint64_t FalseCount = PGO.getCurrentRegionCount() - TrueCount;
1116       // Negate the condition and swap the destination blocks.
1117       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1118                                   FalseCount);
1119     }
1120   }
1121 
1122   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1123     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1124     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1125     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1126 
1127     RegionCounter Cnt = getPGORegionCounter(CondOp);
1128     ConditionalEvaluation cond(*this);
1129     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock, Cnt.getCount());
1130 
1131     // When computing PGO branch weights, we only know the overall count for
1132     // the true block. This code is essentially doing tail duplication of the
1133     // naive code-gen, introducing new edges for which counts are not
1134     // available. Divide the counts proportionally between the LHS and RHS of
1135     // the conditional operator.
1136     uint64_t LHSScaledTrueCount = 0;
1137     if (TrueCount) {
1138       double LHSRatio = Cnt.getCount() / (double) Cnt.getParentCount();
1139       LHSScaledTrueCount = TrueCount * LHSRatio;
1140     }
1141 
1142     cond.begin(*this);
1143     EmitBlock(LHSBlock);
1144     Cnt.beginRegion(Builder);
1145     {
1146       ApplyDebugLocation DL(*this, Cond);
1147       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1148                            LHSScaledTrueCount);
1149     }
1150     cond.end(*this);
1151 
1152     cond.begin(*this);
1153     EmitBlock(RHSBlock);
1154     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1155                          TrueCount - LHSScaledTrueCount);
1156     cond.end(*this);
1157 
1158     return;
1159   }
1160 
1161   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1162     // Conditional operator handling can give us a throw expression as a
1163     // condition for a case like:
1164     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1165     // Fold this to:
1166     //   br(c, throw x, br(y, t, f))
1167     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1168     return;
1169   }
1170 
1171   // Create branch weights based on the number of times we get here and the
1172   // number of times the condition should be true.
1173   uint64_t CurrentCount = std::max(PGO.getCurrentRegionCount(), TrueCount);
1174   llvm::MDNode *Weights = PGO.createBranchWeights(TrueCount,
1175                                                   CurrentCount - TrueCount);
1176 
1177   // Emit the code with the fully general case.
1178   llvm::Value *CondV;
1179   {
1180     ApplyDebugLocation DL(*this, Cond);
1181     CondV = EvaluateExprAsBool(Cond);
1182   }
1183   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights);
1184 }
1185 
1186 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1187 /// specified stmt yet.
1188 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1189   CGM.ErrorUnsupported(S, Type);
1190 }
1191 
1192 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1193 /// variable-length array whose elements have a non-zero bit-pattern.
1194 ///
1195 /// \param baseType the inner-most element type of the array
1196 /// \param src - a char* pointing to the bit-pattern for a single
1197 /// base element of the array
1198 /// \param sizeInChars - the total size of the VLA, in chars
1199 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1200                                llvm::Value *dest, llvm::Value *src,
1201                                llvm::Value *sizeInChars) {
1202   std::pair<CharUnits,CharUnits> baseSizeAndAlign
1203     = CGF.getContext().getTypeInfoInChars(baseType);
1204 
1205   CGBuilderTy &Builder = CGF.Builder;
1206 
1207   llvm::Value *baseSizeInChars
1208     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSizeAndAlign.first.getQuantity());
1209 
1210   llvm::Type *i8p = Builder.getInt8PtrTy();
1211 
1212   llvm::Value *begin = Builder.CreateBitCast(dest, i8p, "vla.begin");
1213   llvm::Value *end = Builder.CreateInBoundsGEP(dest, sizeInChars, "vla.end");
1214 
1215   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1216   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1217   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1218 
1219   // Make a loop over the VLA.  C99 guarantees that the VLA element
1220   // count must be nonzero.
1221   CGF.EmitBlock(loopBB);
1222 
1223   llvm::PHINode *cur = Builder.CreatePHI(i8p, 2, "vla.cur");
1224   cur->addIncoming(begin, originBB);
1225 
1226   // memcpy the individual element bit-pattern.
1227   Builder.CreateMemCpy(cur, src, baseSizeInChars,
1228                        baseSizeAndAlign.second.getQuantity(),
1229                        /*volatile*/ false);
1230 
1231   // Go to the next element.
1232   llvm::Value *next = Builder.CreateConstInBoundsGEP1_32(Builder.getInt8Ty(),
1233                                                          cur, 1, "vla.next");
1234 
1235   // Leave if that's the end of the VLA.
1236   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1237   Builder.CreateCondBr(done, contBB, loopBB);
1238   cur->addIncoming(next, loopBB);
1239 
1240   CGF.EmitBlock(contBB);
1241 }
1242 
1243 void
1244 CodeGenFunction::EmitNullInitialization(llvm::Value *DestPtr, QualType Ty) {
1245   // Ignore empty classes in C++.
1246   if (getLangOpts().CPlusPlus) {
1247     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1248       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1249         return;
1250     }
1251   }
1252 
1253   // Cast the dest ptr to the appropriate i8 pointer type.
1254   unsigned DestAS =
1255     cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace();
1256   llvm::Type *BP = Builder.getInt8PtrTy(DestAS);
1257   if (DestPtr->getType() != BP)
1258     DestPtr = Builder.CreateBitCast(DestPtr, BP);
1259 
1260   // Get size and alignment info for this aggregate.
1261   std::pair<CharUnits, CharUnits> TypeInfo =
1262     getContext().getTypeInfoInChars(Ty);
1263   CharUnits Size = TypeInfo.first;
1264   CharUnits Align = TypeInfo.second;
1265 
1266   llvm::Value *SizeVal;
1267   const VariableArrayType *vla;
1268 
1269   // Don't bother emitting a zero-byte memset.
1270   if (Size.isZero()) {
1271     // But note that getTypeInfo returns 0 for a VLA.
1272     if (const VariableArrayType *vlaType =
1273           dyn_cast_or_null<VariableArrayType>(
1274                                           getContext().getAsArrayType(Ty))) {
1275       QualType eltType;
1276       llvm::Value *numElts;
1277       std::tie(numElts, eltType) = getVLASize(vlaType);
1278 
1279       SizeVal = numElts;
1280       CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1281       if (!eltSize.isOne())
1282         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1283       vla = vlaType;
1284     } else {
1285       return;
1286     }
1287   } else {
1288     SizeVal = CGM.getSize(Size);
1289     vla = nullptr;
1290   }
1291 
1292   // If the type contains a pointer to data member we can't memset it to zero.
1293   // Instead, create a null constant and copy it to the destination.
1294   // TODO: there are other patterns besides zero that we can usefully memset,
1295   // like -1, which happens to be the pattern used by member-pointers.
1296   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1297     // For a VLA, emit a single element, then splat that over the VLA.
1298     if (vla) Ty = getContext().getBaseElementType(vla);
1299 
1300     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1301 
1302     llvm::GlobalVariable *NullVariable =
1303       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1304                                /*isConstant=*/true,
1305                                llvm::GlobalVariable::PrivateLinkage,
1306                                NullConstant, Twine());
1307     llvm::Value *SrcPtr =
1308       Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy());
1309 
1310     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1311 
1312     // Get and call the appropriate llvm.memcpy overload.
1313     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, Align.getQuantity(), false);
1314     return;
1315   }
1316 
1317   // Otherwise, just memset the whole thing to zero.  This is legal
1318   // because in LLVM, all default initializers (other than the ones we just
1319   // handled above) are guaranteed to have a bit pattern of all zeros.
1320   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal,
1321                        Align.getQuantity(), false);
1322 }
1323 
1324 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1325   // Make sure that there is a block for the indirect goto.
1326   if (!IndirectBranch)
1327     GetIndirectGotoBlock();
1328 
1329   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1330 
1331   // Make sure the indirect branch includes all of the address-taken blocks.
1332   IndirectBranch->addDestination(BB);
1333   return llvm::BlockAddress::get(CurFn, BB);
1334 }
1335 
1336 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1337   // If we already made the indirect branch for indirect goto, return its block.
1338   if (IndirectBranch) return IndirectBranch->getParent();
1339 
1340   CGBuilderTy TmpBuilder(createBasicBlock("indirectgoto"));
1341 
1342   // Create the PHI node that indirect gotos will add entries to.
1343   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1344                                               "indirect.goto.dest");
1345 
1346   // Create the indirect branch instruction.
1347   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1348   return IndirectBranch->getParent();
1349 }
1350 
1351 /// Computes the length of an array in elements, as well as the base
1352 /// element type and a properly-typed first element pointer.
1353 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1354                                               QualType &baseType,
1355                                               llvm::Value *&addr) {
1356   const ArrayType *arrayType = origArrayType;
1357 
1358   // If it's a VLA, we have to load the stored size.  Note that
1359   // this is the size of the VLA in bytes, not its size in elements.
1360   llvm::Value *numVLAElements = nullptr;
1361   if (isa<VariableArrayType>(arrayType)) {
1362     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1363 
1364     // Walk into all VLAs.  This doesn't require changes to addr,
1365     // which has type T* where T is the first non-VLA element type.
1366     do {
1367       QualType elementType = arrayType->getElementType();
1368       arrayType = getContext().getAsArrayType(elementType);
1369 
1370       // If we only have VLA components, 'addr' requires no adjustment.
1371       if (!arrayType) {
1372         baseType = elementType;
1373         return numVLAElements;
1374       }
1375     } while (isa<VariableArrayType>(arrayType));
1376 
1377     // We get out here only if we find a constant array type
1378     // inside the VLA.
1379   }
1380 
1381   // We have some number of constant-length arrays, so addr should
1382   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1383   // down to the first element of addr.
1384   SmallVector<llvm::Value*, 8> gepIndices;
1385 
1386   // GEP down to the array type.
1387   llvm::ConstantInt *zero = Builder.getInt32(0);
1388   gepIndices.push_back(zero);
1389 
1390   uint64_t countFromCLAs = 1;
1391   QualType eltType;
1392 
1393   llvm::ArrayType *llvmArrayType =
1394     dyn_cast<llvm::ArrayType>(
1395       cast<llvm::PointerType>(addr->getType())->getElementType());
1396   while (llvmArrayType) {
1397     assert(isa<ConstantArrayType>(arrayType));
1398     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1399              == llvmArrayType->getNumElements());
1400 
1401     gepIndices.push_back(zero);
1402     countFromCLAs *= llvmArrayType->getNumElements();
1403     eltType = arrayType->getElementType();
1404 
1405     llvmArrayType =
1406       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1407     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1408     assert((!llvmArrayType || arrayType) &&
1409            "LLVM and Clang types are out-of-synch");
1410   }
1411 
1412   if (arrayType) {
1413     // From this point onwards, the Clang array type has been emitted
1414     // as some other type (probably a packed struct). Compute the array
1415     // size, and just emit the 'begin' expression as a bitcast.
1416     while (arrayType) {
1417       countFromCLAs *=
1418           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1419       eltType = arrayType->getElementType();
1420       arrayType = getContext().getAsArrayType(eltType);
1421     }
1422 
1423     unsigned AddressSpace = addr->getType()->getPointerAddressSpace();
1424     llvm::Type *BaseType = ConvertType(eltType)->getPointerTo(AddressSpace);
1425     addr = Builder.CreateBitCast(addr, BaseType, "array.begin");
1426   } else {
1427     // Create the actual GEP.
1428     addr = Builder.CreateInBoundsGEP(addr, gepIndices, "array.begin");
1429   }
1430 
1431   baseType = eltType;
1432 
1433   llvm::Value *numElements
1434     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1435 
1436   // If we had any VLA dimensions, factor them in.
1437   if (numVLAElements)
1438     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1439 
1440   return numElements;
1441 }
1442 
1443 std::pair<llvm::Value*, QualType>
1444 CodeGenFunction::getVLASize(QualType type) {
1445   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1446   assert(vla && "type was not a variable array type!");
1447   return getVLASize(vla);
1448 }
1449 
1450 std::pair<llvm::Value*, QualType>
1451 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1452   // The number of elements so far; always size_t.
1453   llvm::Value *numElements = nullptr;
1454 
1455   QualType elementType;
1456   do {
1457     elementType = type->getElementType();
1458     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1459     assert(vlaSize && "no size for VLA!");
1460     assert(vlaSize->getType() == SizeTy);
1461 
1462     if (!numElements) {
1463       numElements = vlaSize;
1464     } else {
1465       // It's undefined behavior if this wraps around, so mark it that way.
1466       // FIXME: Teach -fsanitize=undefined to trap this.
1467       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1468     }
1469   } while ((type = getContext().getAsVariableArrayType(elementType)));
1470 
1471   return std::pair<llvm::Value*,QualType>(numElements, elementType);
1472 }
1473 
1474 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1475   assert(type->isVariablyModifiedType() &&
1476          "Must pass variably modified type to EmitVLASizes!");
1477 
1478   EnsureInsertPoint();
1479 
1480   // We're going to walk down into the type and look for VLA
1481   // expressions.
1482   do {
1483     assert(type->isVariablyModifiedType());
1484 
1485     const Type *ty = type.getTypePtr();
1486     switch (ty->getTypeClass()) {
1487 
1488 #define TYPE(Class, Base)
1489 #define ABSTRACT_TYPE(Class, Base)
1490 #define NON_CANONICAL_TYPE(Class, Base)
1491 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1492 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1493 #include "clang/AST/TypeNodes.def"
1494       llvm_unreachable("unexpected dependent type!");
1495 
1496     // These types are never variably-modified.
1497     case Type::Builtin:
1498     case Type::Complex:
1499     case Type::Vector:
1500     case Type::ExtVector:
1501     case Type::Record:
1502     case Type::Enum:
1503     case Type::Elaborated:
1504     case Type::TemplateSpecialization:
1505     case Type::ObjCObject:
1506     case Type::ObjCInterface:
1507     case Type::ObjCObjectPointer:
1508       llvm_unreachable("type class is never variably-modified!");
1509 
1510     case Type::Adjusted:
1511       type = cast<AdjustedType>(ty)->getAdjustedType();
1512       break;
1513 
1514     case Type::Decayed:
1515       type = cast<DecayedType>(ty)->getPointeeType();
1516       break;
1517 
1518     case Type::Pointer:
1519       type = cast<PointerType>(ty)->getPointeeType();
1520       break;
1521 
1522     case Type::BlockPointer:
1523       type = cast<BlockPointerType>(ty)->getPointeeType();
1524       break;
1525 
1526     case Type::LValueReference:
1527     case Type::RValueReference:
1528       type = cast<ReferenceType>(ty)->getPointeeType();
1529       break;
1530 
1531     case Type::MemberPointer:
1532       type = cast<MemberPointerType>(ty)->getPointeeType();
1533       break;
1534 
1535     case Type::ConstantArray:
1536     case Type::IncompleteArray:
1537       // Losing element qualification here is fine.
1538       type = cast<ArrayType>(ty)->getElementType();
1539       break;
1540 
1541     case Type::VariableArray: {
1542       // Losing element qualification here is fine.
1543       const VariableArrayType *vat = cast<VariableArrayType>(ty);
1544 
1545       // Unknown size indication requires no size computation.
1546       // Otherwise, evaluate and record it.
1547       if (const Expr *size = vat->getSizeExpr()) {
1548         // It's possible that we might have emitted this already,
1549         // e.g. with a typedef and a pointer to it.
1550         llvm::Value *&entry = VLASizeMap[size];
1551         if (!entry) {
1552           llvm::Value *Size = EmitScalarExpr(size);
1553 
1554           // C11 6.7.6.2p5:
1555           //   If the size is an expression that is not an integer constant
1556           //   expression [...] each time it is evaluated it shall have a value
1557           //   greater than zero.
1558           if (SanOpts.has(SanitizerKind::VLABound) &&
1559               size->getType()->isSignedIntegerType()) {
1560             SanitizerScope SanScope(this);
1561             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1562             llvm::Constant *StaticArgs[] = {
1563               EmitCheckSourceLocation(size->getLocStart()),
1564               EmitCheckTypeDescriptor(size->getType())
1565             };
1566             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1567                                      SanitizerKind::VLABound),
1568                       "vla_bound_not_positive", StaticArgs, Size);
1569           }
1570 
1571           // Always zexting here would be wrong if it weren't
1572           // undefined behavior to have a negative bound.
1573           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1574         }
1575       }
1576       type = vat->getElementType();
1577       break;
1578     }
1579 
1580     case Type::FunctionProto:
1581     case Type::FunctionNoProto:
1582       type = cast<FunctionType>(ty)->getReturnType();
1583       break;
1584 
1585     case Type::Paren:
1586     case Type::TypeOf:
1587     case Type::UnaryTransform:
1588     case Type::Attributed:
1589     case Type::SubstTemplateTypeParm:
1590     case Type::PackExpansion:
1591       // Keep walking after single level desugaring.
1592       type = type.getSingleStepDesugaredType(getContext());
1593       break;
1594 
1595     case Type::Typedef:
1596     case Type::Decltype:
1597     case Type::Auto:
1598       // Stop walking: nothing to do.
1599       return;
1600 
1601     case Type::TypeOfExpr:
1602       // Stop walking: emit typeof expression.
1603       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1604       return;
1605 
1606     case Type::Atomic:
1607       type = cast<AtomicType>(ty)->getValueType();
1608       break;
1609     }
1610   } while (type->isVariablyModifiedType());
1611 }
1612 
1613 llvm::Value* CodeGenFunction::EmitVAListRef(const Expr* E) {
1614   if (getContext().getBuiltinVaListType()->isArrayType())
1615     return EmitScalarExpr(E);
1616   return EmitLValue(E).getAddress();
1617 }
1618 
1619 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1620                                               llvm::Constant *Init) {
1621   assert (Init && "Invalid DeclRefExpr initializer!");
1622   if (CGDebugInfo *Dbg = getDebugInfo())
1623     if (CGM.getCodeGenOpts().getDebugInfo() >= CodeGenOptions::LimitedDebugInfo)
1624       Dbg->EmitGlobalVariable(E->getDecl(), Init);
1625 }
1626 
1627 CodeGenFunction::PeepholeProtection
1628 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1629   // At the moment, the only aggressive peephole we do in IR gen
1630   // is trunc(zext) folding, but if we add more, we can easily
1631   // extend this protection.
1632 
1633   if (!rvalue.isScalar()) return PeepholeProtection();
1634   llvm::Value *value = rvalue.getScalarVal();
1635   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1636 
1637   // Just make an extra bitcast.
1638   assert(HaveInsertPoint());
1639   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1640                                                   Builder.GetInsertBlock());
1641 
1642   PeepholeProtection protection;
1643   protection.Inst = inst;
1644   return protection;
1645 }
1646 
1647 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1648   if (!protection.Inst) return;
1649 
1650   // In theory, we could try to duplicate the peepholes now, but whatever.
1651   protection.Inst->eraseFromParent();
1652 }
1653 
1654 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1655                                                  llvm::Value *AnnotatedVal,
1656                                                  StringRef AnnotationStr,
1657                                                  SourceLocation Location) {
1658   llvm::Value *Args[4] = {
1659     AnnotatedVal,
1660     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1661     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1662     CGM.EmitAnnotationLineNo(Location)
1663   };
1664   return Builder.CreateCall(AnnotationFn, Args);
1665 }
1666 
1667 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1668   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1669   // FIXME We create a new bitcast for every annotation because that's what
1670   // llvm-gcc was doing.
1671   for (const auto *I : D->specific_attrs<AnnotateAttr>())
1672     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1673                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1674                        I->getAnnotation(), D->getLocation());
1675 }
1676 
1677 llvm::Value *CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1678                                                    llvm::Value *V) {
1679   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1680   llvm::Type *VTy = V->getType();
1681   llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1682                                     CGM.Int8PtrTy);
1683 
1684   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1685     // FIXME Always emit the cast inst so we can differentiate between
1686     // annotation on the first field of a struct and annotation on the struct
1687     // itself.
1688     if (VTy != CGM.Int8PtrTy)
1689       V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1690     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1691     V = Builder.CreateBitCast(V, VTy);
1692   }
1693 
1694   return V;
1695 }
1696 
1697 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
1698 
1699 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
1700     : CGF(CGF) {
1701   assert(!CGF->IsSanitizerScope);
1702   CGF->IsSanitizerScope = true;
1703 }
1704 
1705 CodeGenFunction::SanitizerScope::~SanitizerScope() {
1706   CGF->IsSanitizerScope = false;
1707 }
1708 
1709 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
1710                                    const llvm::Twine &Name,
1711                                    llvm::BasicBlock *BB,
1712                                    llvm::BasicBlock::iterator InsertPt) const {
1713   LoopStack.InsertHelper(I);
1714   if (IsSanitizerScope)
1715     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
1716 }
1717 
1718 template <bool PreserveNames>
1719 void CGBuilderInserter<PreserveNames>::InsertHelper(
1720     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1721     llvm::BasicBlock::iterator InsertPt) const {
1722   llvm::IRBuilderDefaultInserter<PreserveNames>::InsertHelper(I, Name, BB,
1723                                                               InsertPt);
1724   if (CGF)
1725     CGF->InsertHelper(I, Name, BB, InsertPt);
1726 }
1727 
1728 #ifdef NDEBUG
1729 #define PreserveNames false
1730 #else
1731 #define PreserveNames true
1732 #endif
1733 template void CGBuilderInserter<PreserveNames>::InsertHelper(
1734     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1735     llvm::BasicBlock::iterator InsertPt) const;
1736 #undef PreserveNames
1737