xref: /llvm-project/clang/lib/CodeGen/CodeGenFunction.cpp (revision 57cc1a5d77d89a01508782dec8dabece351fbf1c)
1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This coordinates the per-function state used while generating code.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CodeGenFunction.h"
15 #include "CGBlocks.h"
16 #include "CGCleanup.h"
17 #include "CGCUDARuntime.h"
18 #include "CGCXXABI.h"
19 #include "CGDebugInfo.h"
20 #include "CGOpenMPRuntime.h"
21 #include "CodeGenModule.h"
22 #include "CodeGenPGO.h"
23 #include "TargetInfo.h"
24 #include "clang/AST/ASTContext.h"
25 #include "clang/AST/ASTLambda.h"
26 #include "clang/AST/Decl.h"
27 #include "clang/AST/DeclCXX.h"
28 #include "clang/AST/StmtCXX.h"
29 #include "clang/AST/StmtObjC.h"
30 #include "clang/Basic/Builtins.h"
31 #include "clang/Basic/TargetInfo.h"
32 #include "clang/CodeGen/CGFunctionInfo.h"
33 #include "clang/Frontend/CodeGenOptions.h"
34 #include "clang/Sema/SemaDiagnostic.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/Intrinsics.h"
38 #include "llvm/IR/MDBuilder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
41 using namespace clang;
42 using namespace CodeGen;
43 
44 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
45 /// markers.
46 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
47                                       const LangOptions &LangOpts) {
48   if (CGOpts.DisableLifetimeMarkers)
49     return false;
50 
51   // Disable lifetime markers in msan builds.
52   // FIXME: Remove this when msan works with lifetime markers.
53   if (LangOpts.Sanitize.has(SanitizerKind::Memory))
54     return false;
55 
56   // Asan uses markers for use-after-scope checks.
57   if (CGOpts.SanitizeAddressUseAfterScope)
58     return true;
59 
60   // For now, only in optimized builds.
61   return CGOpts.OptimizationLevel != 0;
62 }
63 
64 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
65     : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
66       Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
67               CGBuilderInserterTy(this)),
68       CurFn(nullptr), ReturnValue(Address::invalid()),
69       CapturedStmtInfo(nullptr), SanOpts(CGM.getLangOpts().Sanitize),
70       IsSanitizerScope(false), CurFuncIsThunk(false), AutoreleaseResult(false),
71       SawAsmBlock(false), IsOutlinedSEHHelper(false), BlockInfo(nullptr),
72       BlockPointer(nullptr), LambdaThisCaptureField(nullptr),
73       NormalCleanupDest(nullptr), NextCleanupDestIndex(1),
74       FirstBlockInfo(nullptr), EHResumeBlock(nullptr), ExceptionSlot(nullptr),
75       EHSelectorSlot(nullptr), DebugInfo(CGM.getModuleDebugInfo()),
76       DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
77       PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
78       CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
79       NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
80       CXXABIThisValue(nullptr), CXXThisValue(nullptr),
81       CXXStructorImplicitParamDecl(nullptr),
82       CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
83       CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
84       TerminateHandler(nullptr), TrapBB(nullptr),
85       ShouldEmitLifetimeMarkers(
86           shouldEmitLifetimeMarkers(CGM.getCodeGenOpts(), CGM.getLangOpts())) {
87   if (!suppressNewContext)
88     CGM.getCXXABI().getMangleContext().startNewFunction();
89 
90   llvm::FastMathFlags FMF;
91   if (CGM.getLangOpts().FastMath)
92     FMF.setFast();
93   if (CGM.getLangOpts().FiniteMathOnly) {
94     FMF.setNoNaNs();
95     FMF.setNoInfs();
96   }
97   if (CGM.getCodeGenOpts().NoNaNsFPMath) {
98     FMF.setNoNaNs();
99   }
100   if (CGM.getCodeGenOpts().NoSignedZeros) {
101     FMF.setNoSignedZeros();
102   }
103   if (CGM.getCodeGenOpts().ReciprocalMath) {
104     FMF.setAllowReciprocal();
105   }
106   if (CGM.getCodeGenOpts().Reassociate) {
107     FMF.setAllowReassoc();
108   }
109   Builder.setFastMathFlags(FMF);
110 }
111 
112 CodeGenFunction::~CodeGenFunction() {
113   assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
114 
115   // If there are any unclaimed block infos, go ahead and destroy them
116   // now.  This can happen if IR-gen gets clever and skips evaluating
117   // something.
118   if (FirstBlockInfo)
119     destroyBlockInfos(FirstBlockInfo);
120 
121   if (getLangOpts().OpenMP && CurFn)
122     CGM.getOpenMPRuntime().functionFinished(*this);
123 }
124 
125 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
126                                                     LValueBaseInfo *BaseInfo,
127                                                     TBAAAccessInfo *TBAAInfo) {
128   return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo, TBAAInfo,
129                                  /* forPointeeType= */ true);
130 }
131 
132 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
133                                                    LValueBaseInfo *BaseInfo,
134                                                    TBAAAccessInfo *TBAAInfo,
135                                                    bool forPointeeType) {
136   if (TBAAInfo)
137     *TBAAInfo = CGM.getTBAAAccessInfo(T);
138 
139   // Honor alignment typedef attributes even on incomplete types.
140   // We also honor them straight for C++ class types, even as pointees;
141   // there's an expressivity gap here.
142   if (auto TT = T->getAs<TypedefType>()) {
143     if (auto Align = TT->getDecl()->getMaxAlignment()) {
144       if (BaseInfo)
145         *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType);
146       return getContext().toCharUnitsFromBits(Align);
147     }
148   }
149 
150   if (BaseInfo)
151     *BaseInfo = LValueBaseInfo(AlignmentSource::Type);
152 
153   CharUnits Alignment;
154   if (T->isIncompleteType()) {
155     Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
156   } else {
157     // For C++ class pointees, we don't know whether we're pointing at a
158     // base or a complete object, so we generally need to use the
159     // non-virtual alignment.
160     const CXXRecordDecl *RD;
161     if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
162       Alignment = CGM.getClassPointerAlignment(RD);
163     } else {
164       Alignment = getContext().getTypeAlignInChars(T);
165       if (T.getQualifiers().hasUnaligned())
166         Alignment = CharUnits::One();
167     }
168 
169     // Cap to the global maximum type alignment unless the alignment
170     // was somehow explicit on the type.
171     if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
172       if (Alignment.getQuantity() > MaxAlign &&
173           !getContext().isAlignmentRequired(T))
174         Alignment = CharUnits::fromQuantity(MaxAlign);
175     }
176   }
177   return Alignment;
178 }
179 
180 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
181   LValueBaseInfo BaseInfo;
182   TBAAAccessInfo TBAAInfo;
183   CharUnits Alignment = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo);
184   return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo,
185                           TBAAInfo);
186 }
187 
188 /// Given a value of type T* that may not be to a complete object,
189 /// construct an l-value with the natural pointee alignment of T.
190 LValue
191 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
192   LValueBaseInfo BaseInfo;
193   TBAAAccessInfo TBAAInfo;
194   CharUnits Align = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo,
195                                             /* forPointeeType= */ true);
196   return MakeAddrLValue(Address(V, Align), T, BaseInfo, TBAAInfo);
197 }
198 
199 
200 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
201   return CGM.getTypes().ConvertTypeForMem(T);
202 }
203 
204 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
205   return CGM.getTypes().ConvertType(T);
206 }
207 
208 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
209   type = type.getCanonicalType();
210   while (true) {
211     switch (type->getTypeClass()) {
212 #define TYPE(name, parent)
213 #define ABSTRACT_TYPE(name, parent)
214 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
215 #define DEPENDENT_TYPE(name, parent) case Type::name:
216 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
217 #include "clang/AST/TypeNodes.def"
218       llvm_unreachable("non-canonical or dependent type in IR-generation");
219 
220     case Type::Auto:
221     case Type::DeducedTemplateSpecialization:
222       llvm_unreachable("undeduced type in IR-generation");
223 
224     // Various scalar types.
225     case Type::Builtin:
226     case Type::Pointer:
227     case Type::BlockPointer:
228     case Type::LValueReference:
229     case Type::RValueReference:
230     case Type::MemberPointer:
231     case Type::Vector:
232     case Type::ExtVector:
233     case Type::FunctionProto:
234     case Type::FunctionNoProto:
235     case Type::Enum:
236     case Type::ObjCObjectPointer:
237     case Type::Pipe:
238       return TEK_Scalar;
239 
240     // Complexes.
241     case Type::Complex:
242       return TEK_Complex;
243 
244     // Arrays, records, and Objective-C objects.
245     case Type::ConstantArray:
246     case Type::IncompleteArray:
247     case Type::VariableArray:
248     case Type::Record:
249     case Type::ObjCObject:
250     case Type::ObjCInterface:
251       return TEK_Aggregate;
252 
253     // We operate on atomic values according to their underlying type.
254     case Type::Atomic:
255       type = cast<AtomicType>(type)->getValueType();
256       continue;
257     }
258     llvm_unreachable("unknown type kind!");
259   }
260 }
261 
262 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
263   // For cleanliness, we try to avoid emitting the return block for
264   // simple cases.
265   llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
266 
267   if (CurBB) {
268     assert(!CurBB->getTerminator() && "Unexpected terminated block.");
269 
270     // We have a valid insert point, reuse it if it is empty or there are no
271     // explicit jumps to the return block.
272     if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
273       ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
274       delete ReturnBlock.getBlock();
275     } else
276       EmitBlock(ReturnBlock.getBlock());
277     return llvm::DebugLoc();
278   }
279 
280   // Otherwise, if the return block is the target of a single direct
281   // branch then we can just put the code in that block instead. This
282   // cleans up functions which started with a unified return block.
283   if (ReturnBlock.getBlock()->hasOneUse()) {
284     llvm::BranchInst *BI =
285       dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
286     if (BI && BI->isUnconditional() &&
287         BI->getSuccessor(0) == ReturnBlock.getBlock()) {
288       // Record/return the DebugLoc of the simple 'return' expression to be used
289       // later by the actual 'ret' instruction.
290       llvm::DebugLoc Loc = BI->getDebugLoc();
291       Builder.SetInsertPoint(BI->getParent());
292       BI->eraseFromParent();
293       delete ReturnBlock.getBlock();
294       return Loc;
295     }
296   }
297 
298   // FIXME: We are at an unreachable point, there is no reason to emit the block
299   // unless it has uses. However, we still need a place to put the debug
300   // region.end for now.
301 
302   EmitBlock(ReturnBlock.getBlock());
303   return llvm::DebugLoc();
304 }
305 
306 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
307   if (!BB) return;
308   if (!BB->use_empty())
309     return CGF.CurFn->getBasicBlockList().push_back(BB);
310   delete BB;
311 }
312 
313 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
314   assert(BreakContinueStack.empty() &&
315          "mismatched push/pop in break/continue stack!");
316 
317   bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
318     && NumSimpleReturnExprs == NumReturnExprs
319     && ReturnBlock.getBlock()->use_empty();
320   // Usually the return expression is evaluated before the cleanup
321   // code.  If the function contains only a simple return statement,
322   // such as a constant, the location before the cleanup code becomes
323   // the last useful breakpoint in the function, because the simple
324   // return expression will be evaluated after the cleanup code. To be
325   // safe, set the debug location for cleanup code to the location of
326   // the return statement.  Otherwise the cleanup code should be at the
327   // end of the function's lexical scope.
328   //
329   // If there are multiple branches to the return block, the branch
330   // instructions will get the location of the return statements and
331   // all will be fine.
332   if (CGDebugInfo *DI = getDebugInfo()) {
333     if (OnlySimpleReturnStmts)
334       DI->EmitLocation(Builder, LastStopPoint);
335     else
336       DI->EmitLocation(Builder, EndLoc);
337   }
338 
339   // Pop any cleanups that might have been associated with the
340   // parameters.  Do this in whatever block we're currently in; it's
341   // important to do this before we enter the return block or return
342   // edges will be *really* confused.
343   bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
344   bool HasOnlyLifetimeMarkers =
345       HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
346   bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
347   if (HasCleanups) {
348     // Make sure the line table doesn't jump back into the body for
349     // the ret after it's been at EndLoc.
350     if (CGDebugInfo *DI = getDebugInfo())
351       if (OnlySimpleReturnStmts)
352         DI->EmitLocation(Builder, EndLoc);
353 
354     PopCleanupBlocks(PrologueCleanupDepth);
355   }
356 
357   // Emit function epilog (to return).
358   llvm::DebugLoc Loc = EmitReturnBlock();
359 
360   if (ShouldInstrumentFunction()) {
361     if (CGM.getCodeGenOpts().InstrumentFunctions)
362       CurFn->addFnAttr("instrument-function-exit", "__cyg_profile_func_exit");
363     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
364       CurFn->addFnAttr("instrument-function-exit-inlined",
365                        "__cyg_profile_func_exit");
366   }
367 
368   // Emit debug descriptor for function end.
369   if (CGDebugInfo *DI = getDebugInfo())
370     DI->EmitFunctionEnd(Builder, CurFn);
371 
372   // Reset the debug location to that of the simple 'return' expression, if any
373   // rather than that of the end of the function's scope '}'.
374   ApplyDebugLocation AL(*this, Loc);
375   EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
376   EmitEndEHSpec(CurCodeDecl);
377 
378   assert(EHStack.empty() &&
379          "did not remove all scopes from cleanup stack!");
380 
381   // If someone did an indirect goto, emit the indirect goto block at the end of
382   // the function.
383   if (IndirectBranch) {
384     EmitBlock(IndirectBranch->getParent());
385     Builder.ClearInsertionPoint();
386   }
387 
388   // If some of our locals escaped, insert a call to llvm.localescape in the
389   // entry block.
390   if (!EscapedLocals.empty()) {
391     // Invert the map from local to index into a simple vector. There should be
392     // no holes.
393     SmallVector<llvm::Value *, 4> EscapeArgs;
394     EscapeArgs.resize(EscapedLocals.size());
395     for (auto &Pair : EscapedLocals)
396       EscapeArgs[Pair.second] = Pair.first;
397     llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
398         &CGM.getModule(), llvm::Intrinsic::localescape);
399     CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
400   }
401 
402   // Remove the AllocaInsertPt instruction, which is just a convenience for us.
403   llvm::Instruction *Ptr = AllocaInsertPt;
404   AllocaInsertPt = nullptr;
405   Ptr->eraseFromParent();
406 
407   // If someone took the address of a label but never did an indirect goto, we
408   // made a zero entry PHI node, which is illegal, zap it now.
409   if (IndirectBranch) {
410     llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
411     if (PN->getNumIncomingValues() == 0) {
412       PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
413       PN->eraseFromParent();
414     }
415   }
416 
417   EmitIfUsed(*this, EHResumeBlock);
418   EmitIfUsed(*this, TerminateLandingPad);
419   EmitIfUsed(*this, TerminateHandler);
420   EmitIfUsed(*this, UnreachableBlock);
421 
422   for (const auto &FuncletAndParent : TerminateFunclets)
423     EmitIfUsed(*this, FuncletAndParent.second);
424 
425   if (CGM.getCodeGenOpts().EmitDeclMetadata)
426     EmitDeclMetadata();
427 
428   for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
429            I = DeferredReplacements.begin(),
430            E = DeferredReplacements.end();
431        I != E; ++I) {
432     I->first->replaceAllUsesWith(I->second);
433     I->first->eraseFromParent();
434   }
435 
436   // Eliminate CleanupDestSlot alloca by replacing it with SSA values and
437   // PHIs if the current function is a coroutine. We don't do it for all
438   // functions as it may result in slight increase in numbers of instructions
439   // if compiled with no optimizations. We do it for coroutine as the lifetime
440   // of CleanupDestSlot alloca make correct coroutine frame building very
441   // difficult.
442   if (NormalCleanupDest && isCoroutine()) {
443     llvm::DominatorTree DT(*CurFn);
444     llvm::PromoteMemToReg(NormalCleanupDest, DT);
445     NormalCleanupDest = nullptr;
446   }
447 }
448 
449 /// ShouldInstrumentFunction - Return true if the current function should be
450 /// instrumented with __cyg_profile_func_* calls
451 bool CodeGenFunction::ShouldInstrumentFunction() {
452   if (!CGM.getCodeGenOpts().InstrumentFunctions &&
453       !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining &&
454       !CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
455     return false;
456   if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
457     return false;
458   return true;
459 }
460 
461 /// ShouldXRayInstrument - Return true if the current function should be
462 /// instrumented with XRay nop sleds.
463 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
464   return CGM.getCodeGenOpts().XRayInstrumentFunctions;
465 }
466 
467 /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to
468 /// the __xray_customevent(...) builin calls, when doing XRay instrumentation.
469 bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const {
470   return CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents;
471 }
472 
473 llvm::Constant *
474 CodeGenFunction::EncodeAddrForUseInPrologue(llvm::Function *F,
475                                             llvm::Constant *Addr) {
476   // Addresses stored in prologue data can't require run-time fixups and must
477   // be PC-relative. Run-time fixups are undesirable because they necessitate
478   // writable text segments, which are unsafe. And absolute addresses are
479   // undesirable because they break PIE mode.
480 
481   // Add a layer of indirection through a private global. Taking its address
482   // won't result in a run-time fixup, even if Addr has linkonce_odr linkage.
483   auto *GV = new llvm::GlobalVariable(CGM.getModule(), Addr->getType(),
484                                       /*isConstant=*/true,
485                                       llvm::GlobalValue::PrivateLinkage, Addr);
486 
487   // Create a PC-relative address.
488   auto *GOTAsInt = llvm::ConstantExpr::getPtrToInt(GV, IntPtrTy);
489   auto *FuncAsInt = llvm::ConstantExpr::getPtrToInt(F, IntPtrTy);
490   auto *PCRelAsInt = llvm::ConstantExpr::getSub(GOTAsInt, FuncAsInt);
491   return (IntPtrTy == Int32Ty)
492              ? PCRelAsInt
493              : llvm::ConstantExpr::getTrunc(PCRelAsInt, Int32Ty);
494 }
495 
496 llvm::Value *
497 CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F,
498                                           llvm::Value *EncodedAddr) {
499   // Reconstruct the address of the global.
500   auto *PCRelAsInt = Builder.CreateSExt(EncodedAddr, IntPtrTy);
501   auto *FuncAsInt = Builder.CreatePtrToInt(F, IntPtrTy, "func_addr.int");
502   auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int");
503   auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr");
504 
505   // Load the original pointer through the global.
506   return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()),
507                             "decoded_addr");
508 }
509 
510 static void removeImageAccessQualifier(std::string& TyName) {
511   std::string ReadOnlyQual("__read_only");
512   std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
513   if (ReadOnlyPos != std::string::npos)
514     // "+ 1" for the space after access qualifier.
515     TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
516   else {
517     std::string WriteOnlyQual("__write_only");
518     std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
519     if (WriteOnlyPos != std::string::npos)
520       TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
521     else {
522       std::string ReadWriteQual("__read_write");
523       std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
524       if (ReadWritePos != std::string::npos)
525         TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
526     }
527   }
528 }
529 
530 // Returns the address space id that should be produced to the
531 // kernel_arg_addr_space metadata. This is always fixed to the ids
532 // as specified in the SPIR 2.0 specification in order to differentiate
533 // for example in clGetKernelArgInfo() implementation between the address
534 // spaces with targets without unique mapping to the OpenCL address spaces
535 // (basically all single AS CPUs).
536 static unsigned ArgInfoAddressSpace(LangAS AS) {
537   switch (AS) {
538   case LangAS::opencl_global:   return 1;
539   case LangAS::opencl_constant: return 2;
540   case LangAS::opencl_local:    return 3;
541   case LangAS::opencl_generic:  return 4; // Not in SPIR 2.0 specs.
542   default:
543     return 0; // Assume private.
544   }
545 }
546 
547 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
548 // information in the program executable. The argument information stored
549 // includes the argument name, its type, the address and access qualifiers used.
550 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
551                                  CodeGenModule &CGM, llvm::LLVMContext &Context,
552                                  CGBuilderTy &Builder, ASTContext &ASTCtx) {
553   // Create MDNodes that represent the kernel arg metadata.
554   // Each MDNode is a list in the form of "key", N number of values which is
555   // the same number of values as their are kernel arguments.
556 
557   const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
558 
559   // MDNode for the kernel argument address space qualifiers.
560   SmallVector<llvm::Metadata *, 8> addressQuals;
561 
562   // MDNode for the kernel argument access qualifiers (images only).
563   SmallVector<llvm::Metadata *, 8> accessQuals;
564 
565   // MDNode for the kernel argument type names.
566   SmallVector<llvm::Metadata *, 8> argTypeNames;
567 
568   // MDNode for the kernel argument base type names.
569   SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
570 
571   // MDNode for the kernel argument type qualifiers.
572   SmallVector<llvm::Metadata *, 8> argTypeQuals;
573 
574   // MDNode for the kernel argument names.
575   SmallVector<llvm::Metadata *, 8> argNames;
576 
577   for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
578     const ParmVarDecl *parm = FD->getParamDecl(i);
579     QualType ty = parm->getType();
580     std::string typeQuals;
581 
582     if (ty->isPointerType()) {
583       QualType pointeeTy = ty->getPointeeType();
584 
585       // Get address qualifier.
586       addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
587         ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
588 
589       // Get argument type name.
590       std::string typeName =
591           pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
592 
593       // Turn "unsigned type" to "utype"
594       std::string::size_type pos = typeName.find("unsigned");
595       if (pointeeTy.isCanonical() && pos != std::string::npos)
596         typeName.erase(pos+1, 8);
597 
598       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
599 
600       std::string baseTypeName =
601           pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
602               Policy) +
603           "*";
604 
605       // Turn "unsigned type" to "utype"
606       pos = baseTypeName.find("unsigned");
607       if (pos != std::string::npos)
608         baseTypeName.erase(pos+1, 8);
609 
610       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
611 
612       // Get argument type qualifiers:
613       if (ty.isRestrictQualified())
614         typeQuals = "restrict";
615       if (pointeeTy.isConstQualified() ||
616           (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
617         typeQuals += typeQuals.empty() ? "const" : " const";
618       if (pointeeTy.isVolatileQualified())
619         typeQuals += typeQuals.empty() ? "volatile" : " volatile";
620     } else {
621       uint32_t AddrSpc = 0;
622       bool isPipe = ty->isPipeType();
623       if (ty->isImageType() || isPipe)
624         AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
625 
626       addressQuals.push_back(
627           llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
628 
629       // Get argument type name.
630       std::string typeName;
631       if (isPipe)
632         typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
633                      .getAsString(Policy);
634       else
635         typeName = ty.getUnqualifiedType().getAsString(Policy);
636 
637       // Turn "unsigned type" to "utype"
638       std::string::size_type pos = typeName.find("unsigned");
639       if (ty.isCanonical() && pos != std::string::npos)
640         typeName.erase(pos+1, 8);
641 
642       std::string baseTypeName;
643       if (isPipe)
644         baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
645                           ->getElementType().getCanonicalType()
646                           .getAsString(Policy);
647       else
648         baseTypeName =
649           ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
650 
651       // Remove access qualifiers on images
652       // (as they are inseparable from type in clang implementation,
653       // but OpenCL spec provides a special query to get access qualifier
654       // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
655       if (ty->isImageType()) {
656         removeImageAccessQualifier(typeName);
657         removeImageAccessQualifier(baseTypeName);
658       }
659 
660       argTypeNames.push_back(llvm::MDString::get(Context, typeName));
661 
662       // Turn "unsigned type" to "utype"
663       pos = baseTypeName.find("unsigned");
664       if (pos != std::string::npos)
665         baseTypeName.erase(pos+1, 8);
666 
667       argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
668 
669       if (isPipe)
670         typeQuals = "pipe";
671     }
672 
673     argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
674 
675     // Get image and pipe access qualifier:
676     if (ty->isImageType()|| ty->isPipeType()) {
677       const Decl *PDecl = parm;
678       if (auto *TD = dyn_cast<TypedefType>(ty))
679         PDecl = TD->getDecl();
680       const OpenCLAccessAttr *A = PDecl->getAttr<OpenCLAccessAttr>();
681       if (A && A->isWriteOnly())
682         accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
683       else if (A && A->isReadWrite())
684         accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
685       else
686         accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
687     } else
688       accessQuals.push_back(llvm::MDString::get(Context, "none"));
689 
690     // Get argument name.
691     argNames.push_back(llvm::MDString::get(Context, parm->getName()));
692   }
693 
694   Fn->setMetadata("kernel_arg_addr_space",
695                   llvm::MDNode::get(Context, addressQuals));
696   Fn->setMetadata("kernel_arg_access_qual",
697                   llvm::MDNode::get(Context, accessQuals));
698   Fn->setMetadata("kernel_arg_type",
699                   llvm::MDNode::get(Context, argTypeNames));
700   Fn->setMetadata("kernel_arg_base_type",
701                   llvm::MDNode::get(Context, argBaseTypeNames));
702   Fn->setMetadata("kernel_arg_type_qual",
703                   llvm::MDNode::get(Context, argTypeQuals));
704   if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
705     Fn->setMetadata("kernel_arg_name",
706                     llvm::MDNode::get(Context, argNames));
707 }
708 
709 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
710                                                llvm::Function *Fn)
711 {
712   if (!FD->hasAttr<OpenCLKernelAttr>())
713     return;
714 
715   llvm::LLVMContext &Context = getLLVMContext();
716 
717   GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
718 
719   if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
720     QualType HintQTy = A->getTypeHint();
721     const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
722     bool IsSignedInteger =
723         HintQTy->isSignedIntegerType() ||
724         (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
725     llvm::Metadata *AttrMDArgs[] = {
726         llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
727             CGM.getTypes().ConvertType(A->getTypeHint()))),
728         llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
729             llvm::IntegerType::get(Context, 32),
730             llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
731     Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
732   }
733 
734   if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
735     llvm::Metadata *AttrMDArgs[] = {
736         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
737         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
738         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
739     Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
740   }
741 
742   if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
743     llvm::Metadata *AttrMDArgs[] = {
744         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
745         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
746         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
747     Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
748   }
749 
750   if (const OpenCLIntelReqdSubGroupSizeAttr *A =
751           FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
752     llvm::Metadata *AttrMDArgs[] = {
753         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
754     Fn->setMetadata("intel_reqd_sub_group_size",
755                     llvm::MDNode::get(Context, AttrMDArgs));
756   }
757 }
758 
759 /// Determine whether the function F ends with a return stmt.
760 static bool endsWithReturn(const Decl* F) {
761   const Stmt *Body = nullptr;
762   if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
763     Body = FD->getBody();
764   else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
765     Body = OMD->getBody();
766 
767   if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
768     auto LastStmt = CS->body_rbegin();
769     if (LastStmt != CS->body_rend())
770       return isa<ReturnStmt>(*LastStmt);
771   }
772   return false;
773 }
774 
775 static void markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
776   Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
777   Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
778 }
779 
780 static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) {
781   auto *MD = dyn_cast_or_null<CXXMethodDecl>(D);
782   if (!MD || !MD->getDeclName().getAsIdentifierInfo() ||
783       !MD->getDeclName().getAsIdentifierInfo()->isStr("allocate") ||
784       (MD->getNumParams() != 1 && MD->getNumParams() != 2))
785     return false;
786 
787   if (MD->parameters()[0]->getType().getCanonicalType() != Ctx.getSizeType())
788     return false;
789 
790   if (MD->getNumParams() == 2) {
791     auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>();
792     if (!PT || !PT->isVoidPointerType() ||
793         !PT->getPointeeType().isConstQualified())
794       return false;
795   }
796 
797   return true;
798 }
799 
800 /// Return the UBSan prologue signature for \p FD if one is available.
801 static llvm::Constant *getPrologueSignature(CodeGenModule &CGM,
802                                             const FunctionDecl *FD) {
803   if (const auto *MD = dyn_cast<CXXMethodDecl>(FD))
804     if (!MD->isStatic())
805       return nullptr;
806   return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM);
807 }
808 
809 void CodeGenFunction::StartFunction(GlobalDecl GD,
810                                     QualType RetTy,
811                                     llvm::Function *Fn,
812                                     const CGFunctionInfo &FnInfo,
813                                     const FunctionArgList &Args,
814                                     SourceLocation Loc,
815                                     SourceLocation StartLoc) {
816   assert(!CurFn &&
817          "Do not use a CodeGenFunction object for more than one function");
818 
819   const Decl *D = GD.getDecl();
820 
821   DidCallStackSave = false;
822   CurCodeDecl = D;
823   if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
824     if (FD->usesSEHTry())
825       CurSEHParent = FD;
826   CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
827   FnRetTy = RetTy;
828   CurFn = Fn;
829   CurFnInfo = &FnInfo;
830   assert(CurFn->isDeclaration() && "Function already has body?");
831 
832   // If this function has been blacklisted for any of the enabled sanitizers,
833   // disable the sanitizer for the function.
834   do {
835 #define SANITIZER(NAME, ID)                                                    \
836   if (SanOpts.empty())                                                         \
837     break;                                                                     \
838   if (SanOpts.has(SanitizerKind::ID))                                          \
839     if (CGM.isInSanitizerBlacklist(SanitizerKind::ID, Fn, Loc))                \
840       SanOpts.set(SanitizerKind::ID, false);
841 
842 #include "clang/Basic/Sanitizers.def"
843 #undef SANITIZER
844   } while (0);
845 
846   if (D) {
847     // Apply the no_sanitize* attributes to SanOpts.
848     for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
849       SanOpts.Mask &= ~Attr->getMask();
850   }
851 
852   // Apply sanitizer attributes to the function.
853   if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
854     Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
855   if (SanOpts.hasOneOf(SanitizerKind::HWAddress))
856     Fn->addFnAttr(llvm::Attribute::SanitizeHWAddress);
857   if (SanOpts.has(SanitizerKind::Thread))
858     Fn->addFnAttr(llvm::Attribute::SanitizeThread);
859   if (SanOpts.has(SanitizerKind::Memory))
860     Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
861   if (SanOpts.has(SanitizerKind::SafeStack))
862     Fn->addFnAttr(llvm::Attribute::SafeStack);
863 
864   // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
865   // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
866   if (SanOpts.has(SanitizerKind::Thread)) {
867     if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
868       IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
869       if (OMD->getMethodFamily() == OMF_dealloc ||
870           OMD->getMethodFamily() == OMF_initialize ||
871           (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
872         markAsIgnoreThreadCheckingAtRuntime(Fn);
873       }
874     } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) {
875       IdentifierInfo *II = FD->getIdentifier();
876       if (II && II->isStr("__destroy_helper_block_"))
877         markAsIgnoreThreadCheckingAtRuntime(Fn);
878     }
879   }
880 
881   // Ignore unrelated casts in STL allocate() since the allocator must cast
882   // from void* to T* before object initialization completes. Don't match on the
883   // namespace because not all allocators are in std::
884   if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
885     if (matchesStlAllocatorFn(D, getContext()))
886       SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast;
887   }
888 
889   // Apply xray attributes to the function (as a string, for now)
890   bool InstrumentXray = ShouldXRayInstrumentFunction();
891   if (D && InstrumentXray) {
892     if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
893       if (XRayAttr->alwaysXRayInstrument())
894         Fn->addFnAttr("function-instrument", "xray-always");
895       if (XRayAttr->neverXRayInstrument())
896         Fn->addFnAttr("function-instrument", "xray-never");
897       if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) {
898         Fn->addFnAttr("xray-log-args",
899                       llvm::utostr(LogArgs->getArgumentCount()));
900       }
901     } else {
902       if (!CGM.imbueXRayAttrs(Fn, Loc))
903         Fn->addFnAttr(
904             "xray-instruction-threshold",
905             llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
906     }
907   }
908 
909   // Add no-jump-tables value.
910   Fn->addFnAttr("no-jump-tables",
911                 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
912 
913   // Add profile-sample-accurate value.
914   if (CGM.getCodeGenOpts().ProfileSampleAccurate)
915     Fn->addFnAttr("profile-sample-accurate");
916 
917   if (getLangOpts().OpenCL) {
918     // Add metadata for a kernel function.
919     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
920       EmitOpenCLKernelMetadata(FD, Fn);
921   }
922 
923   // If we are checking function types, emit a function type signature as
924   // prologue data.
925   if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
926     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
927       if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
928         // Remove any (C++17) exception specifications, to allow calling e.g. a
929         // noexcept function through a non-noexcept pointer.
930         auto ProtoTy =
931           getContext().getFunctionTypeWithExceptionSpec(FD->getType(),
932                                                         EST_None);
933         llvm::Constant *FTRTTIConst =
934             CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true);
935         llvm::Constant *FTRTTIConstEncoded =
936             EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
937         llvm::Constant *PrologueStructElems[] = {PrologueSig,
938                                                  FTRTTIConstEncoded};
939         llvm::Constant *PrologueStructConst =
940             llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
941         Fn->setPrologueData(PrologueStructConst);
942       }
943     }
944   }
945 
946   // If we're checking nullability, we need to know whether we can check the
947   // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
948   if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
949     auto Nullability = FnRetTy->getNullability(getContext());
950     if (Nullability && *Nullability == NullabilityKind::NonNull) {
951       if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
952             CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
953         RetValNullabilityPrecondition =
954             llvm::ConstantInt::getTrue(getLLVMContext());
955     }
956   }
957 
958   // If we're in C++ mode and the function name is "main", it is guaranteed
959   // to be norecurse by the standard (3.6.1.3 "The function main shall not be
960   // used within a program").
961   if (getLangOpts().CPlusPlus)
962     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
963       if (FD->isMain())
964         Fn->addFnAttr(llvm::Attribute::NoRecurse);
965 
966   llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
967 
968   // Create a marker to make it easy to insert allocas into the entryblock
969   // later.  Don't create this with the builder, because we don't want it
970   // folded.
971   llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
972   AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
973 
974   ReturnBlock = getJumpDestInCurrentScope("return");
975 
976   Builder.SetInsertPoint(EntryBB);
977 
978   // If we're checking the return value, allocate space for a pointer to a
979   // precise source location of the checked return statement.
980   if (requiresReturnValueCheck()) {
981     ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
982     InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
983   }
984 
985   // Emit subprogram debug descriptor.
986   if (CGDebugInfo *DI = getDebugInfo()) {
987     // Reconstruct the type from the argument list so that implicit parameters,
988     // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
989     // convention.
990     CallingConv CC = CallingConv::CC_C;
991     if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
992       if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
993         CC = SrcFnTy->getCallConv();
994     SmallVector<QualType, 16> ArgTypes;
995     for (const VarDecl *VD : Args)
996       ArgTypes.push_back(VD->getType());
997     QualType FnType = getContext().getFunctionType(
998         RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
999     DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
1000   }
1001 
1002   if (ShouldInstrumentFunction()) {
1003     if (CGM.getCodeGenOpts().InstrumentFunctions)
1004       CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
1005     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
1006       CurFn->addFnAttr("instrument-function-entry-inlined",
1007                        "__cyg_profile_func_enter");
1008     if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
1009       CurFn->addFnAttr("instrument-function-entry-inlined",
1010                        "__cyg_profile_func_enter_bare");
1011   }
1012 
1013   // Since emitting the mcount call here impacts optimizations such as function
1014   // inlining, we just add an attribute to insert a mcount call in backend.
1015   // The attribute "counting-function" is set to mcount function name which is
1016   // architecture dependent.
1017   if (CGM.getCodeGenOpts().InstrumentForProfiling) {
1018     if (CGM.getCodeGenOpts().CallFEntry)
1019       Fn->addFnAttr("fentry-call", "true");
1020     else {
1021       if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
1022         Fn->addFnAttr("instrument-function-entry-inlined",
1023                       getTarget().getMCountName());
1024       }
1025     }
1026   }
1027 
1028   if (RetTy->isVoidType()) {
1029     // Void type; nothing to return.
1030     ReturnValue = Address::invalid();
1031 
1032     // Count the implicit return.
1033     if (!endsWithReturn(D))
1034       ++NumReturnExprs;
1035   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
1036              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1037     // Indirect aggregate return; emit returned value directly into sret slot.
1038     // This reduces code size, and affects correctness in C++.
1039     auto AI = CurFn->arg_begin();
1040     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
1041       ++AI;
1042     ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
1043   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
1044              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1045     // Load the sret pointer from the argument struct and return into that.
1046     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
1047     llvm::Function::arg_iterator EI = CurFn->arg_end();
1048     --EI;
1049     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
1050     Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
1051     ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
1052   } else {
1053     ReturnValue = CreateIRTemp(RetTy, "retval");
1054 
1055     // Tell the epilog emitter to autorelease the result.  We do this
1056     // now so that various specialized functions can suppress it
1057     // during their IR-generation.
1058     if (getLangOpts().ObjCAutoRefCount &&
1059         !CurFnInfo->isReturnsRetained() &&
1060         RetTy->isObjCRetainableType())
1061       AutoreleaseResult = true;
1062   }
1063 
1064   EmitStartEHSpec(CurCodeDecl);
1065 
1066   PrologueCleanupDepth = EHStack.stable_begin();
1067   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
1068 
1069   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
1070     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
1071     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
1072     if (MD->getParent()->isLambda() &&
1073         MD->getOverloadedOperator() == OO_Call) {
1074       // We're in a lambda; figure out the captures.
1075       MD->getParent()->getCaptureFields(LambdaCaptureFields,
1076                                         LambdaThisCaptureField);
1077       if (LambdaThisCaptureField) {
1078         // If the lambda captures the object referred to by '*this' - either by
1079         // value or by reference, make sure CXXThisValue points to the correct
1080         // object.
1081 
1082         // Get the lvalue for the field (which is a copy of the enclosing object
1083         // or contains the address of the enclosing object).
1084         LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
1085         if (!LambdaThisCaptureField->getType()->isPointerType()) {
1086           // If the enclosing object was captured by value, just use its address.
1087           CXXThisValue = ThisFieldLValue.getAddress().getPointer();
1088         } else {
1089           // Load the lvalue pointed to by the field, since '*this' was captured
1090           // by reference.
1091           CXXThisValue =
1092               EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
1093         }
1094       }
1095       for (auto *FD : MD->getParent()->fields()) {
1096         if (FD->hasCapturedVLAType()) {
1097           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
1098                                            SourceLocation()).getScalarVal();
1099           auto VAT = FD->getCapturedVLAType();
1100           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
1101         }
1102       }
1103     } else {
1104       // Not in a lambda; just use 'this' from the method.
1105       // FIXME: Should we generate a new load for each use of 'this'?  The
1106       // fast register allocator would be happier...
1107       CXXThisValue = CXXABIThisValue;
1108     }
1109 
1110     // Check the 'this' pointer once per function, if it's available.
1111     if (CXXABIThisValue) {
1112       SanitizerSet SkippedChecks;
1113       SkippedChecks.set(SanitizerKind::ObjectSize, true);
1114       QualType ThisTy = MD->getThisType(getContext());
1115 
1116       // If this is the call operator of a lambda with no capture-default, it
1117       // may have a static invoker function, which may call this operator with
1118       // a null 'this' pointer.
1119       if (isLambdaCallOperator(MD) &&
1120           cast<CXXRecordDecl>(MD->getParent())->getLambdaCaptureDefault() ==
1121               LCD_None)
1122         SkippedChecks.set(SanitizerKind::Null, true);
1123 
1124       EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1125                                                 : TCK_MemberCall,
1126                     Loc, CXXABIThisValue, ThisTy,
1127                     getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1128                     SkippedChecks);
1129     }
1130   }
1131 
1132   // If any of the arguments have a variably modified type, make sure to
1133   // emit the type size.
1134   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1135        i != e; ++i) {
1136     const VarDecl *VD = *i;
1137 
1138     // Dig out the type as written from ParmVarDecls; it's unclear whether
1139     // the standard (C99 6.9.1p10) requires this, but we're following the
1140     // precedent set by gcc.
1141     QualType Ty;
1142     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1143       Ty = PVD->getOriginalType();
1144     else
1145       Ty = VD->getType();
1146 
1147     if (Ty->isVariablyModifiedType())
1148       EmitVariablyModifiedType(Ty);
1149   }
1150   // Emit a location at the end of the prologue.
1151   if (CGDebugInfo *DI = getDebugInfo())
1152     DI->EmitLocation(Builder, StartLoc);
1153 }
1154 
1155 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1156                                        const Stmt *Body) {
1157   incrementProfileCounter(Body);
1158   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1159     EmitCompoundStmtWithoutScope(*S);
1160   else
1161     EmitStmt(Body);
1162 }
1163 
1164 /// When instrumenting to collect profile data, the counts for some blocks
1165 /// such as switch cases need to not include the fall-through counts, so
1166 /// emit a branch around the instrumentation code. When not instrumenting,
1167 /// this just calls EmitBlock().
1168 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1169                                                const Stmt *S) {
1170   llvm::BasicBlock *SkipCountBB = nullptr;
1171   if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1172     // When instrumenting for profiling, the fallthrough to certain
1173     // statements needs to skip over the instrumentation code so that we
1174     // get an accurate count.
1175     SkipCountBB = createBasicBlock("skipcount");
1176     EmitBranch(SkipCountBB);
1177   }
1178   EmitBlock(BB);
1179   uint64_t CurrentCount = getCurrentProfileCount();
1180   incrementProfileCounter(S);
1181   setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1182   if (SkipCountBB)
1183     EmitBlock(SkipCountBB);
1184 }
1185 
1186 /// Tries to mark the given function nounwind based on the
1187 /// non-existence of any throwing calls within it.  We believe this is
1188 /// lightweight enough to do at -O0.
1189 static void TryMarkNoThrow(llvm::Function *F) {
1190   // LLVM treats 'nounwind' on a function as part of the type, so we
1191   // can't do this on functions that can be overwritten.
1192   if (F->isInterposable()) return;
1193 
1194   for (llvm::BasicBlock &BB : *F)
1195     for (llvm::Instruction &I : BB)
1196       if (I.mayThrow())
1197         return;
1198 
1199   F->setDoesNotThrow();
1200 }
1201 
1202 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1203                                                FunctionArgList &Args) {
1204   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1205   QualType ResTy = FD->getReturnType();
1206 
1207   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1208   if (MD && MD->isInstance()) {
1209     if (CGM.getCXXABI().HasThisReturn(GD))
1210       ResTy = MD->getThisType(getContext());
1211     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1212       ResTy = CGM.getContext().VoidPtrTy;
1213     CGM.getCXXABI().buildThisParam(*this, Args);
1214   }
1215 
1216   // The base version of an inheriting constructor whose constructed base is a
1217   // virtual base is not passed any arguments (because it doesn't actually call
1218   // the inherited constructor).
1219   bool PassedParams = true;
1220   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1221     if (auto Inherited = CD->getInheritedConstructor())
1222       PassedParams =
1223           getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1224 
1225   if (PassedParams) {
1226     for (auto *Param : FD->parameters()) {
1227       Args.push_back(Param);
1228       if (!Param->hasAttr<PassObjectSizeAttr>())
1229         continue;
1230 
1231       auto *Implicit = ImplicitParamDecl::Create(
1232           getContext(), Param->getDeclContext(), Param->getLocation(),
1233           /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1234       SizeArguments[Param] = Implicit;
1235       Args.push_back(Implicit);
1236     }
1237   }
1238 
1239   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1240     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1241 
1242   return ResTy;
1243 }
1244 
1245 static bool
1246 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1247                                              const ASTContext &Context) {
1248   QualType T = FD->getReturnType();
1249   // Avoid the optimization for functions that return a record type with a
1250   // trivial destructor or another trivially copyable type.
1251   if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1252     if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1253       return !ClassDecl->hasTrivialDestructor();
1254   }
1255   return !T.isTriviallyCopyableType(Context);
1256 }
1257 
1258 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1259                                    const CGFunctionInfo &FnInfo) {
1260   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1261   CurGD = GD;
1262 
1263   FunctionArgList Args;
1264   QualType ResTy = BuildFunctionArgList(GD, Args);
1265 
1266   // Check if we should generate debug info for this function.
1267   if (FD->hasAttr<NoDebugAttr>())
1268     DebugInfo = nullptr; // disable debug info indefinitely for this function
1269 
1270   // The function might not have a body if we're generating thunks for a
1271   // function declaration.
1272   SourceRange BodyRange;
1273   if (Stmt *Body = FD->getBody())
1274     BodyRange = Body->getSourceRange();
1275   else
1276     BodyRange = FD->getLocation();
1277   CurEHLocation = BodyRange.getEnd();
1278 
1279   // Use the location of the start of the function to determine where
1280   // the function definition is located. By default use the location
1281   // of the declaration as the location for the subprogram. A function
1282   // may lack a declaration in the source code if it is created by code
1283   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1284   SourceLocation Loc = FD->getLocation();
1285 
1286   // If this is a function specialization then use the pattern body
1287   // as the location for the function.
1288   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1289     if (SpecDecl->hasBody(SpecDecl))
1290       Loc = SpecDecl->getLocation();
1291 
1292   Stmt *Body = FD->getBody();
1293 
1294   // Initialize helper which will detect jumps which can cause invalid lifetime
1295   // markers.
1296   if (Body && ShouldEmitLifetimeMarkers)
1297     Bypasses.Init(Body);
1298 
1299   // Emit the standard function prologue.
1300   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1301 
1302   // Generate the body of the function.
1303   PGO.assignRegionCounters(GD, CurFn);
1304   if (isa<CXXDestructorDecl>(FD))
1305     EmitDestructorBody(Args);
1306   else if (isa<CXXConstructorDecl>(FD))
1307     EmitConstructorBody(Args);
1308   else if (getLangOpts().CUDA &&
1309            !getLangOpts().CUDAIsDevice &&
1310            FD->hasAttr<CUDAGlobalAttr>())
1311     CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1312   else if (isa<CXXMethodDecl>(FD) &&
1313            cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1314     // The lambda static invoker function is special, because it forwards or
1315     // clones the body of the function call operator (but is actually static).
1316     EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1317   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1318              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1319               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1320     // Implicit copy-assignment gets the same special treatment as implicit
1321     // copy-constructors.
1322     emitImplicitAssignmentOperatorBody(Args);
1323   } else if (Body) {
1324     EmitFunctionBody(Args, Body);
1325   } else
1326     llvm_unreachable("no definition for emitted function");
1327 
1328   // C++11 [stmt.return]p2:
1329   //   Flowing off the end of a function [...] results in undefined behavior in
1330   //   a value-returning function.
1331   // C11 6.9.1p12:
1332   //   If the '}' that terminates a function is reached, and the value of the
1333   //   function call is used by the caller, the behavior is undefined.
1334   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1335       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1336     bool ShouldEmitUnreachable =
1337         CGM.getCodeGenOpts().StrictReturn ||
1338         shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1339     if (SanOpts.has(SanitizerKind::Return)) {
1340       SanitizerScope SanScope(this);
1341       llvm::Value *IsFalse = Builder.getFalse();
1342       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1343                 SanitizerHandler::MissingReturn,
1344                 EmitCheckSourceLocation(FD->getLocation()), None);
1345     } else if (ShouldEmitUnreachable) {
1346       if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1347         EmitTrapCall(llvm::Intrinsic::trap);
1348     }
1349     if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1350       Builder.CreateUnreachable();
1351       Builder.ClearInsertionPoint();
1352     }
1353   }
1354 
1355   // Emit the standard function epilogue.
1356   FinishFunction(BodyRange.getEnd());
1357 
1358   // If we haven't marked the function nothrow through other means, do
1359   // a quick pass now to see if we can.
1360   if (!CurFn->doesNotThrow())
1361     TryMarkNoThrow(CurFn);
1362 }
1363 
1364 /// ContainsLabel - Return true if the statement contains a label in it.  If
1365 /// this statement is not executed normally, it not containing a label means
1366 /// that we can just remove the code.
1367 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1368   // Null statement, not a label!
1369   if (!S) return false;
1370 
1371   // If this is a label, we have to emit the code, consider something like:
1372   // if (0) {  ...  foo:  bar(); }  goto foo;
1373   //
1374   // TODO: If anyone cared, we could track __label__'s, since we know that you
1375   // can't jump to one from outside their declared region.
1376   if (isa<LabelStmt>(S))
1377     return true;
1378 
1379   // If this is a case/default statement, and we haven't seen a switch, we have
1380   // to emit the code.
1381   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1382     return true;
1383 
1384   // If this is a switch statement, we want to ignore cases below it.
1385   if (isa<SwitchStmt>(S))
1386     IgnoreCaseStmts = true;
1387 
1388   // Scan subexpressions for verboten labels.
1389   for (const Stmt *SubStmt : S->children())
1390     if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1391       return true;
1392 
1393   return false;
1394 }
1395 
1396 /// containsBreak - Return true if the statement contains a break out of it.
1397 /// If the statement (recursively) contains a switch or loop with a break
1398 /// inside of it, this is fine.
1399 bool CodeGenFunction::containsBreak(const Stmt *S) {
1400   // Null statement, not a label!
1401   if (!S) return false;
1402 
1403   // If this is a switch or loop that defines its own break scope, then we can
1404   // include it and anything inside of it.
1405   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1406       isa<ForStmt>(S))
1407     return false;
1408 
1409   if (isa<BreakStmt>(S))
1410     return true;
1411 
1412   // Scan subexpressions for verboten breaks.
1413   for (const Stmt *SubStmt : S->children())
1414     if (containsBreak(SubStmt))
1415       return true;
1416 
1417   return false;
1418 }
1419 
1420 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1421   if (!S) return false;
1422 
1423   // Some statement kinds add a scope and thus never add a decl to the current
1424   // scope. Note, this list is longer than the list of statements that might
1425   // have an unscoped decl nested within them, but this way is conservatively
1426   // correct even if more statement kinds are added.
1427   if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1428       isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1429       isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1430       isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1431     return false;
1432 
1433   if (isa<DeclStmt>(S))
1434     return true;
1435 
1436   for (const Stmt *SubStmt : S->children())
1437     if (mightAddDeclToScope(SubStmt))
1438       return true;
1439 
1440   return false;
1441 }
1442 
1443 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1444 /// to a constant, or if it does but contains a label, return false.  If it
1445 /// constant folds return true and set the boolean result in Result.
1446 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1447                                                    bool &ResultBool,
1448                                                    bool AllowLabels) {
1449   llvm::APSInt ResultInt;
1450   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1451     return false;
1452 
1453   ResultBool = ResultInt.getBoolValue();
1454   return true;
1455 }
1456 
1457 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1458 /// to a constant, or if it does but contains a label, return false.  If it
1459 /// constant folds return true and set the folded value.
1460 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1461                                                    llvm::APSInt &ResultInt,
1462                                                    bool AllowLabels) {
1463   // FIXME: Rename and handle conversion of other evaluatable things
1464   // to bool.
1465   llvm::APSInt Int;
1466   if (!Cond->EvaluateAsInt(Int, getContext()))
1467     return false;  // Not foldable, not integer or not fully evaluatable.
1468 
1469   if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1470     return false;  // Contains a label.
1471 
1472   ResultInt = Int;
1473   return true;
1474 }
1475 
1476 
1477 
1478 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1479 /// statement) to the specified blocks.  Based on the condition, this might try
1480 /// to simplify the codegen of the conditional based on the branch.
1481 ///
1482 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1483                                            llvm::BasicBlock *TrueBlock,
1484                                            llvm::BasicBlock *FalseBlock,
1485                                            uint64_t TrueCount) {
1486   Cond = Cond->IgnoreParens();
1487 
1488   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1489 
1490     // Handle X && Y in a condition.
1491     if (CondBOp->getOpcode() == BO_LAnd) {
1492       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1493       // folded if the case was simple enough.
1494       bool ConstantBool = false;
1495       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1496           ConstantBool) {
1497         // br(1 && X) -> br(X).
1498         incrementProfileCounter(CondBOp);
1499         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1500                                     TrueCount);
1501       }
1502 
1503       // If we have "X && 1", simplify the code to use an uncond branch.
1504       // "X && 0" would have been constant folded to 0.
1505       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1506           ConstantBool) {
1507         // br(X && 1) -> br(X).
1508         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1509                                     TrueCount);
1510       }
1511 
1512       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1513       // want to jump to the FalseBlock.
1514       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1515       // The counter tells us how often we evaluate RHS, and all of TrueCount
1516       // can be propagated to that branch.
1517       uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1518 
1519       ConditionalEvaluation eval(*this);
1520       {
1521         ApplyDebugLocation DL(*this, Cond);
1522         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1523         EmitBlock(LHSTrue);
1524       }
1525 
1526       incrementProfileCounter(CondBOp);
1527       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1528 
1529       // Any temporaries created here are conditional.
1530       eval.begin(*this);
1531       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1532       eval.end(*this);
1533 
1534       return;
1535     }
1536 
1537     if (CondBOp->getOpcode() == BO_LOr) {
1538       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1539       // folded if the case was simple enough.
1540       bool ConstantBool = false;
1541       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1542           !ConstantBool) {
1543         // br(0 || X) -> br(X).
1544         incrementProfileCounter(CondBOp);
1545         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1546                                     TrueCount);
1547       }
1548 
1549       // If we have "X || 0", simplify the code to use an uncond branch.
1550       // "X || 1" would have been constant folded to 1.
1551       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1552           !ConstantBool) {
1553         // br(X || 0) -> br(X).
1554         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1555                                     TrueCount);
1556       }
1557 
1558       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1559       // want to jump to the TrueBlock.
1560       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1561       // We have the count for entry to the RHS and for the whole expression
1562       // being true, so we can divy up True count between the short circuit and
1563       // the RHS.
1564       uint64_t LHSCount =
1565           getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1566       uint64_t RHSCount = TrueCount - LHSCount;
1567 
1568       ConditionalEvaluation eval(*this);
1569       {
1570         ApplyDebugLocation DL(*this, Cond);
1571         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1572         EmitBlock(LHSFalse);
1573       }
1574 
1575       incrementProfileCounter(CondBOp);
1576       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1577 
1578       // Any temporaries created here are conditional.
1579       eval.begin(*this);
1580       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1581 
1582       eval.end(*this);
1583 
1584       return;
1585     }
1586   }
1587 
1588   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1589     // br(!x, t, f) -> br(x, f, t)
1590     if (CondUOp->getOpcode() == UO_LNot) {
1591       // Negate the count.
1592       uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1593       // Negate the condition and swap the destination blocks.
1594       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1595                                   FalseCount);
1596     }
1597   }
1598 
1599   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1600     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1601     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1602     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1603 
1604     ConditionalEvaluation cond(*this);
1605     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1606                          getProfileCount(CondOp));
1607 
1608     // When computing PGO branch weights, we only know the overall count for
1609     // the true block. This code is essentially doing tail duplication of the
1610     // naive code-gen, introducing new edges for which counts are not
1611     // available. Divide the counts proportionally between the LHS and RHS of
1612     // the conditional operator.
1613     uint64_t LHSScaledTrueCount = 0;
1614     if (TrueCount) {
1615       double LHSRatio =
1616           getProfileCount(CondOp) / (double)getCurrentProfileCount();
1617       LHSScaledTrueCount = TrueCount * LHSRatio;
1618     }
1619 
1620     cond.begin(*this);
1621     EmitBlock(LHSBlock);
1622     incrementProfileCounter(CondOp);
1623     {
1624       ApplyDebugLocation DL(*this, Cond);
1625       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1626                            LHSScaledTrueCount);
1627     }
1628     cond.end(*this);
1629 
1630     cond.begin(*this);
1631     EmitBlock(RHSBlock);
1632     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1633                          TrueCount - LHSScaledTrueCount);
1634     cond.end(*this);
1635 
1636     return;
1637   }
1638 
1639   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1640     // Conditional operator handling can give us a throw expression as a
1641     // condition for a case like:
1642     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1643     // Fold this to:
1644     //   br(c, throw x, br(y, t, f))
1645     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1646     return;
1647   }
1648 
1649   // If the branch has a condition wrapped by __builtin_unpredictable,
1650   // create metadata that specifies that the branch is unpredictable.
1651   // Don't bother if not optimizing because that metadata would not be used.
1652   llvm::MDNode *Unpredictable = nullptr;
1653   auto *Call = dyn_cast<CallExpr>(Cond);
1654   if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1655     auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1656     if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1657       llvm::MDBuilder MDHelper(getLLVMContext());
1658       Unpredictable = MDHelper.createUnpredictable();
1659     }
1660   }
1661 
1662   // Create branch weights based on the number of times we get here and the
1663   // number of times the condition should be true.
1664   uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1665   llvm::MDNode *Weights =
1666       createProfileWeights(TrueCount, CurrentCount - TrueCount);
1667 
1668   // Emit the code with the fully general case.
1669   llvm::Value *CondV;
1670   {
1671     ApplyDebugLocation DL(*this, Cond);
1672     CondV = EvaluateExprAsBool(Cond);
1673   }
1674   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1675 }
1676 
1677 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1678 /// specified stmt yet.
1679 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1680   CGM.ErrorUnsupported(S, Type);
1681 }
1682 
1683 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1684 /// variable-length array whose elements have a non-zero bit-pattern.
1685 ///
1686 /// \param baseType the inner-most element type of the array
1687 /// \param src - a char* pointing to the bit-pattern for a single
1688 /// base element of the array
1689 /// \param sizeInChars - the total size of the VLA, in chars
1690 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1691                                Address dest, Address src,
1692                                llvm::Value *sizeInChars) {
1693   CGBuilderTy &Builder = CGF.Builder;
1694 
1695   CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1696   llvm::Value *baseSizeInChars
1697     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1698 
1699   Address begin =
1700     Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1701   llvm::Value *end =
1702     Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1703 
1704   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1705   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1706   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1707 
1708   // Make a loop over the VLA.  C99 guarantees that the VLA element
1709   // count must be nonzero.
1710   CGF.EmitBlock(loopBB);
1711 
1712   llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1713   cur->addIncoming(begin.getPointer(), originBB);
1714 
1715   CharUnits curAlign =
1716     dest.getAlignment().alignmentOfArrayElement(baseSize);
1717 
1718   // memcpy the individual element bit-pattern.
1719   Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1720                        /*volatile*/ false);
1721 
1722   // Go to the next element.
1723   llvm::Value *next =
1724     Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1725 
1726   // Leave if that's the end of the VLA.
1727   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1728   Builder.CreateCondBr(done, contBB, loopBB);
1729   cur->addIncoming(next, loopBB);
1730 
1731   CGF.EmitBlock(contBB);
1732 }
1733 
1734 void
1735 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1736   // Ignore empty classes in C++.
1737   if (getLangOpts().CPlusPlus) {
1738     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1739       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1740         return;
1741     }
1742   }
1743 
1744   // Cast the dest ptr to the appropriate i8 pointer type.
1745   if (DestPtr.getElementType() != Int8Ty)
1746     DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1747 
1748   // Get size and alignment info for this aggregate.
1749   CharUnits size = getContext().getTypeSizeInChars(Ty);
1750 
1751   llvm::Value *SizeVal;
1752   const VariableArrayType *vla;
1753 
1754   // Don't bother emitting a zero-byte memset.
1755   if (size.isZero()) {
1756     // But note that getTypeInfo returns 0 for a VLA.
1757     if (const VariableArrayType *vlaType =
1758           dyn_cast_or_null<VariableArrayType>(
1759                                           getContext().getAsArrayType(Ty))) {
1760       QualType eltType;
1761       llvm::Value *numElts;
1762       std::tie(numElts, eltType) = getVLASize(vlaType);
1763 
1764       SizeVal = numElts;
1765       CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1766       if (!eltSize.isOne())
1767         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1768       vla = vlaType;
1769     } else {
1770       return;
1771     }
1772   } else {
1773     SizeVal = CGM.getSize(size);
1774     vla = nullptr;
1775   }
1776 
1777   // If the type contains a pointer to data member we can't memset it to zero.
1778   // Instead, create a null constant and copy it to the destination.
1779   // TODO: there are other patterns besides zero that we can usefully memset,
1780   // like -1, which happens to be the pattern used by member-pointers.
1781   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1782     // For a VLA, emit a single element, then splat that over the VLA.
1783     if (vla) Ty = getContext().getBaseElementType(vla);
1784 
1785     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1786 
1787     llvm::GlobalVariable *NullVariable =
1788       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1789                                /*isConstant=*/true,
1790                                llvm::GlobalVariable::PrivateLinkage,
1791                                NullConstant, Twine());
1792     CharUnits NullAlign = DestPtr.getAlignment();
1793     NullVariable->setAlignment(NullAlign.getQuantity());
1794     Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1795                    NullAlign);
1796 
1797     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1798 
1799     // Get and call the appropriate llvm.memcpy overload.
1800     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1801     return;
1802   }
1803 
1804   // Otherwise, just memset the whole thing to zero.  This is legal
1805   // because in LLVM, all default initializers (other than the ones we just
1806   // handled above) are guaranteed to have a bit pattern of all zeros.
1807   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1808 }
1809 
1810 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1811   // Make sure that there is a block for the indirect goto.
1812   if (!IndirectBranch)
1813     GetIndirectGotoBlock();
1814 
1815   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1816 
1817   // Make sure the indirect branch includes all of the address-taken blocks.
1818   IndirectBranch->addDestination(BB);
1819   return llvm::BlockAddress::get(CurFn, BB);
1820 }
1821 
1822 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1823   // If we already made the indirect branch for indirect goto, return its block.
1824   if (IndirectBranch) return IndirectBranch->getParent();
1825 
1826   CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1827 
1828   // Create the PHI node that indirect gotos will add entries to.
1829   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1830                                               "indirect.goto.dest");
1831 
1832   // Create the indirect branch instruction.
1833   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1834   return IndirectBranch->getParent();
1835 }
1836 
1837 /// Computes the length of an array in elements, as well as the base
1838 /// element type and a properly-typed first element pointer.
1839 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1840                                               QualType &baseType,
1841                                               Address &addr) {
1842   const ArrayType *arrayType = origArrayType;
1843 
1844   // If it's a VLA, we have to load the stored size.  Note that
1845   // this is the size of the VLA in bytes, not its size in elements.
1846   llvm::Value *numVLAElements = nullptr;
1847   if (isa<VariableArrayType>(arrayType)) {
1848     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1849 
1850     // Walk into all VLAs.  This doesn't require changes to addr,
1851     // which has type T* where T is the first non-VLA element type.
1852     do {
1853       QualType elementType = arrayType->getElementType();
1854       arrayType = getContext().getAsArrayType(elementType);
1855 
1856       // If we only have VLA components, 'addr' requires no adjustment.
1857       if (!arrayType) {
1858         baseType = elementType;
1859         return numVLAElements;
1860       }
1861     } while (isa<VariableArrayType>(arrayType));
1862 
1863     // We get out here only if we find a constant array type
1864     // inside the VLA.
1865   }
1866 
1867   // We have some number of constant-length arrays, so addr should
1868   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1869   // down to the first element of addr.
1870   SmallVector<llvm::Value*, 8> gepIndices;
1871 
1872   // GEP down to the array type.
1873   llvm::ConstantInt *zero = Builder.getInt32(0);
1874   gepIndices.push_back(zero);
1875 
1876   uint64_t countFromCLAs = 1;
1877   QualType eltType;
1878 
1879   llvm::ArrayType *llvmArrayType =
1880     dyn_cast<llvm::ArrayType>(addr.getElementType());
1881   while (llvmArrayType) {
1882     assert(isa<ConstantArrayType>(arrayType));
1883     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1884              == llvmArrayType->getNumElements());
1885 
1886     gepIndices.push_back(zero);
1887     countFromCLAs *= llvmArrayType->getNumElements();
1888     eltType = arrayType->getElementType();
1889 
1890     llvmArrayType =
1891       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1892     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1893     assert((!llvmArrayType || arrayType) &&
1894            "LLVM and Clang types are out-of-synch");
1895   }
1896 
1897   if (arrayType) {
1898     // From this point onwards, the Clang array type has been emitted
1899     // as some other type (probably a packed struct). Compute the array
1900     // size, and just emit the 'begin' expression as a bitcast.
1901     while (arrayType) {
1902       countFromCLAs *=
1903           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1904       eltType = arrayType->getElementType();
1905       arrayType = getContext().getAsArrayType(eltType);
1906     }
1907 
1908     llvm::Type *baseType = ConvertType(eltType);
1909     addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1910   } else {
1911     // Create the actual GEP.
1912     addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1913                                              gepIndices, "array.begin"),
1914                    addr.getAlignment());
1915   }
1916 
1917   baseType = eltType;
1918 
1919   llvm::Value *numElements
1920     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1921 
1922   // If we had any VLA dimensions, factor them in.
1923   if (numVLAElements)
1924     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1925 
1926   return numElements;
1927 }
1928 
1929 std::pair<llvm::Value*, QualType>
1930 CodeGenFunction::getVLASize(QualType type) {
1931   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1932   assert(vla && "type was not a variable array type!");
1933   return getVLASize(vla);
1934 }
1935 
1936 std::pair<llvm::Value*, QualType>
1937 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1938   // The number of elements so far; always size_t.
1939   llvm::Value *numElements = nullptr;
1940 
1941   QualType elementType;
1942   do {
1943     elementType = type->getElementType();
1944     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1945     assert(vlaSize && "no size for VLA!");
1946     assert(vlaSize->getType() == SizeTy);
1947 
1948     if (!numElements) {
1949       numElements = vlaSize;
1950     } else {
1951       // It's undefined behavior if this wraps around, so mark it that way.
1952       // FIXME: Teach -fsanitize=undefined to trap this.
1953       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1954     }
1955   } while ((type = getContext().getAsVariableArrayType(elementType)));
1956 
1957   return std::pair<llvm::Value*,QualType>(numElements, elementType);
1958 }
1959 
1960 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1961   assert(type->isVariablyModifiedType() &&
1962          "Must pass variably modified type to EmitVLASizes!");
1963 
1964   EnsureInsertPoint();
1965 
1966   // We're going to walk down into the type and look for VLA
1967   // expressions.
1968   do {
1969     assert(type->isVariablyModifiedType());
1970 
1971     const Type *ty = type.getTypePtr();
1972     switch (ty->getTypeClass()) {
1973 
1974 #define TYPE(Class, Base)
1975 #define ABSTRACT_TYPE(Class, Base)
1976 #define NON_CANONICAL_TYPE(Class, Base)
1977 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1978 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1979 #include "clang/AST/TypeNodes.def"
1980       llvm_unreachable("unexpected dependent type!");
1981 
1982     // These types are never variably-modified.
1983     case Type::Builtin:
1984     case Type::Complex:
1985     case Type::Vector:
1986     case Type::ExtVector:
1987     case Type::Record:
1988     case Type::Enum:
1989     case Type::Elaborated:
1990     case Type::TemplateSpecialization:
1991     case Type::ObjCTypeParam:
1992     case Type::ObjCObject:
1993     case Type::ObjCInterface:
1994     case Type::ObjCObjectPointer:
1995       llvm_unreachable("type class is never variably-modified!");
1996 
1997     case Type::Adjusted:
1998       type = cast<AdjustedType>(ty)->getAdjustedType();
1999       break;
2000 
2001     case Type::Decayed:
2002       type = cast<DecayedType>(ty)->getPointeeType();
2003       break;
2004 
2005     case Type::Pointer:
2006       type = cast<PointerType>(ty)->getPointeeType();
2007       break;
2008 
2009     case Type::BlockPointer:
2010       type = cast<BlockPointerType>(ty)->getPointeeType();
2011       break;
2012 
2013     case Type::LValueReference:
2014     case Type::RValueReference:
2015       type = cast<ReferenceType>(ty)->getPointeeType();
2016       break;
2017 
2018     case Type::MemberPointer:
2019       type = cast<MemberPointerType>(ty)->getPointeeType();
2020       break;
2021 
2022     case Type::ConstantArray:
2023     case Type::IncompleteArray:
2024       // Losing element qualification here is fine.
2025       type = cast<ArrayType>(ty)->getElementType();
2026       break;
2027 
2028     case Type::VariableArray: {
2029       // Losing element qualification here is fine.
2030       const VariableArrayType *vat = cast<VariableArrayType>(ty);
2031 
2032       // Unknown size indication requires no size computation.
2033       // Otherwise, evaluate and record it.
2034       if (const Expr *size = vat->getSizeExpr()) {
2035         // It's possible that we might have emitted this already,
2036         // e.g. with a typedef and a pointer to it.
2037         llvm::Value *&entry = VLASizeMap[size];
2038         if (!entry) {
2039           llvm::Value *Size = EmitScalarExpr(size);
2040 
2041           // C11 6.7.6.2p5:
2042           //   If the size is an expression that is not an integer constant
2043           //   expression [...] each time it is evaluated it shall have a value
2044           //   greater than zero.
2045           if (SanOpts.has(SanitizerKind::VLABound) &&
2046               size->getType()->isSignedIntegerType()) {
2047             SanitizerScope SanScope(this);
2048             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
2049             llvm::Constant *StaticArgs[] = {
2050               EmitCheckSourceLocation(size->getLocStart()),
2051               EmitCheckTypeDescriptor(size->getType())
2052             };
2053             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
2054                                      SanitizerKind::VLABound),
2055                       SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
2056           }
2057 
2058           // Always zexting here would be wrong if it weren't
2059           // undefined behavior to have a negative bound.
2060           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
2061         }
2062       }
2063       type = vat->getElementType();
2064       break;
2065     }
2066 
2067     case Type::FunctionProto:
2068     case Type::FunctionNoProto:
2069       type = cast<FunctionType>(ty)->getReturnType();
2070       break;
2071 
2072     case Type::Paren:
2073     case Type::TypeOf:
2074     case Type::UnaryTransform:
2075     case Type::Attributed:
2076     case Type::SubstTemplateTypeParm:
2077     case Type::PackExpansion:
2078       // Keep walking after single level desugaring.
2079       type = type.getSingleStepDesugaredType(getContext());
2080       break;
2081 
2082     case Type::Typedef:
2083     case Type::Decltype:
2084     case Type::Auto:
2085     case Type::DeducedTemplateSpecialization:
2086       // Stop walking: nothing to do.
2087       return;
2088 
2089     case Type::TypeOfExpr:
2090       // Stop walking: emit typeof expression.
2091       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
2092       return;
2093 
2094     case Type::Atomic:
2095       type = cast<AtomicType>(ty)->getValueType();
2096       break;
2097 
2098     case Type::Pipe:
2099       type = cast<PipeType>(ty)->getElementType();
2100       break;
2101     }
2102   } while (type->isVariablyModifiedType());
2103 }
2104 
2105 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
2106   if (getContext().getBuiltinVaListType()->isArrayType())
2107     return EmitPointerWithAlignment(E);
2108   return EmitLValue(E).getAddress();
2109 }
2110 
2111 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
2112   return EmitLValue(E).getAddress();
2113 }
2114 
2115 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2116                                               const APValue &Init) {
2117   assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
2118   if (CGDebugInfo *Dbg = getDebugInfo())
2119     if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2120       Dbg->EmitGlobalVariable(E->getDecl(), Init);
2121 }
2122 
2123 CodeGenFunction::PeepholeProtection
2124 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2125   // At the moment, the only aggressive peephole we do in IR gen
2126   // is trunc(zext) folding, but if we add more, we can easily
2127   // extend this protection.
2128 
2129   if (!rvalue.isScalar()) return PeepholeProtection();
2130   llvm::Value *value = rvalue.getScalarVal();
2131   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2132 
2133   // Just make an extra bitcast.
2134   assert(HaveInsertPoint());
2135   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2136                                                   Builder.GetInsertBlock());
2137 
2138   PeepholeProtection protection;
2139   protection.Inst = inst;
2140   return protection;
2141 }
2142 
2143 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2144   if (!protection.Inst) return;
2145 
2146   // In theory, we could try to duplicate the peepholes now, but whatever.
2147   protection.Inst->eraseFromParent();
2148 }
2149 
2150 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2151                                                  llvm::Value *AnnotatedVal,
2152                                                  StringRef AnnotationStr,
2153                                                  SourceLocation Location) {
2154   llvm::Value *Args[4] = {
2155     AnnotatedVal,
2156     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2157     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2158     CGM.EmitAnnotationLineNo(Location)
2159   };
2160   return Builder.CreateCall(AnnotationFn, Args);
2161 }
2162 
2163 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2164   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2165   // FIXME We create a new bitcast for every annotation because that's what
2166   // llvm-gcc was doing.
2167   for (const auto *I : D->specific_attrs<AnnotateAttr>())
2168     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2169                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2170                        I->getAnnotation(), D->getLocation());
2171 }
2172 
2173 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2174                                               Address Addr) {
2175   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2176   llvm::Value *V = Addr.getPointer();
2177   llvm::Type *VTy = V->getType();
2178   llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2179                                     CGM.Int8PtrTy);
2180 
2181   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2182     // FIXME Always emit the cast inst so we can differentiate between
2183     // annotation on the first field of a struct and annotation on the struct
2184     // itself.
2185     if (VTy != CGM.Int8PtrTy)
2186       V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2187     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2188     V = Builder.CreateBitCast(V, VTy);
2189   }
2190 
2191   return Address(V, Addr.getAlignment());
2192 }
2193 
2194 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2195 
2196 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2197     : CGF(CGF) {
2198   assert(!CGF->IsSanitizerScope);
2199   CGF->IsSanitizerScope = true;
2200 }
2201 
2202 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2203   CGF->IsSanitizerScope = false;
2204 }
2205 
2206 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2207                                    const llvm::Twine &Name,
2208                                    llvm::BasicBlock *BB,
2209                                    llvm::BasicBlock::iterator InsertPt) const {
2210   LoopStack.InsertHelper(I);
2211   if (IsSanitizerScope)
2212     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2213 }
2214 
2215 void CGBuilderInserter::InsertHelper(
2216     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2217     llvm::BasicBlock::iterator InsertPt) const {
2218   llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2219   if (CGF)
2220     CGF->InsertHelper(I, Name, BB, InsertPt);
2221 }
2222 
2223 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2224                                 CodeGenModule &CGM, const FunctionDecl *FD,
2225                                 std::string &FirstMissing) {
2226   // If there aren't any required features listed then go ahead and return.
2227   if (ReqFeatures.empty())
2228     return false;
2229 
2230   // Now build up the set of caller features and verify that all the required
2231   // features are there.
2232   llvm::StringMap<bool> CallerFeatureMap;
2233   CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2234 
2235   // If we have at least one of the features in the feature list return
2236   // true, otherwise return false.
2237   return std::all_of(
2238       ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2239         SmallVector<StringRef, 1> OrFeatures;
2240         Feature.split(OrFeatures, "|");
2241         return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2242                            [&](StringRef Feature) {
2243                              if (!CallerFeatureMap.lookup(Feature)) {
2244                                FirstMissing = Feature.str();
2245                                return false;
2246                              }
2247                              return true;
2248                            });
2249       });
2250 }
2251 
2252 // Emits an error if we don't have a valid set of target features for the
2253 // called function.
2254 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2255                                           const FunctionDecl *TargetDecl) {
2256   // Early exit if this is an indirect call.
2257   if (!TargetDecl)
2258     return;
2259 
2260   // Get the current enclosing function if it exists. If it doesn't
2261   // we can't check the target features anyhow.
2262   const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2263   if (!FD)
2264     return;
2265 
2266   // Grab the required features for the call. For a builtin this is listed in
2267   // the td file with the default cpu, for an always_inline function this is any
2268   // listed cpu and any listed features.
2269   unsigned BuiltinID = TargetDecl->getBuiltinID();
2270   std::string MissingFeature;
2271   if (BuiltinID) {
2272     SmallVector<StringRef, 1> ReqFeatures;
2273     const char *FeatureList =
2274         CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2275     // Return if the builtin doesn't have any required features.
2276     if (!FeatureList || StringRef(FeatureList) == "")
2277       return;
2278     StringRef(FeatureList).split(ReqFeatures, ",");
2279     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2280       CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2281           << TargetDecl->getDeclName()
2282           << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2283 
2284   } else if (TargetDecl->hasAttr<TargetAttr>()) {
2285     // Get the required features for the callee.
2286     SmallVector<StringRef, 1> ReqFeatures;
2287     llvm::StringMap<bool> CalleeFeatureMap;
2288     CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2289     for (const auto &F : CalleeFeatureMap) {
2290       // Only positive features are "required".
2291       if (F.getValue())
2292         ReqFeatures.push_back(F.getKey());
2293     }
2294     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2295       CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2296           << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2297   }
2298 }
2299 
2300 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2301   if (!CGM.getCodeGenOpts().SanitizeStats)
2302     return;
2303 
2304   llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2305   IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2306   CGM.getSanStats().create(IRB, SSK);
2307 }
2308 
2309 llvm::Value *
2310 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
2311   llvm::Value *TrueCondition = nullptr;
2312   if (!RO.ParsedAttribute.Architecture.empty())
2313     TrueCondition = EmitX86CpuIs(RO.ParsedAttribute.Architecture);
2314 
2315   if (!RO.ParsedAttribute.Features.empty()) {
2316     SmallVector<StringRef, 8> FeatureList;
2317     llvm::for_each(RO.ParsedAttribute.Features,
2318                    [&FeatureList](const std::string &Feature) {
2319                      FeatureList.push_back(StringRef{Feature}.substr(1));
2320                    });
2321     llvm::Value *FeatureCmp = EmitX86CpuSupports(FeatureList);
2322     TrueCondition = TrueCondition ? Builder.CreateAnd(TrueCondition, FeatureCmp)
2323                                   : FeatureCmp;
2324   }
2325   return TrueCondition;
2326 }
2327 
2328 void CodeGenFunction::EmitMultiVersionResolver(
2329     llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
2330   assert((getContext().getTargetInfo().getTriple().getArch() ==
2331               llvm::Triple::x86 ||
2332           getContext().getTargetInfo().getTriple().getArch() ==
2333               llvm::Triple::x86_64) &&
2334          "Only implemented for x86 targets");
2335 
2336   // Main function's basic block.
2337   llvm::BasicBlock *CurBlock = createBasicBlock("entry", Resolver);
2338   Builder.SetInsertPoint(CurBlock);
2339   EmitX86CpuInit();
2340 
2341   llvm::Function *DefaultFunc = nullptr;
2342   for (const MultiVersionResolverOption &RO : Options) {
2343     Builder.SetInsertPoint(CurBlock);
2344     llvm::Value *TrueCondition = FormResolverCondition(RO);
2345 
2346     if (!TrueCondition) {
2347       DefaultFunc = RO.Function;
2348     } else {
2349       llvm::BasicBlock *RetBlock = createBasicBlock("ro_ret", Resolver);
2350       llvm::IRBuilder<> RetBuilder(RetBlock);
2351       RetBuilder.CreateRet(RO.Function);
2352       CurBlock = createBasicBlock("ro_else", Resolver);
2353       Builder.CreateCondBr(TrueCondition, RetBlock, CurBlock);
2354     }
2355   }
2356 
2357   assert(DefaultFunc && "No default version?");
2358   // Emit return from the 'else-ist' block.
2359   Builder.SetInsertPoint(CurBlock);
2360   Builder.CreateRet(DefaultFunc);
2361 }
2362 
2363 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2364   if (CGDebugInfo *DI = getDebugInfo())
2365     return DI->SourceLocToDebugLoc(Location);
2366 
2367   return llvm::DebugLoc();
2368 }
2369