1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This coordinates the per-function state used while generating code.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "CodeGenFunction.h"
14 #include "CGBlocks.h"
15 #include "CGCleanup.h"
16 #include "CGCUDARuntime.h"
17 #include "CGCXXABI.h"
18 #include "CGDebugInfo.h"
19 #include "CGOpenMPRuntime.h"
20 #include "CodeGenModule.h"
21 #include "CodeGenPGO.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/ASTContext.h"
24 #include "clang/AST/ASTLambda.h"
25 #include "clang/AST/Decl.h"
26 #include "clang/AST/DeclCXX.h"
27 #include "clang/AST/StmtCXX.h"
28 #include "clang/AST/StmtObjC.h"
29 #include "clang/Basic/Builtins.h"
30 #include "clang/Basic/CodeGenOptions.h"
31 #include "clang/Basic/TargetInfo.h"
32 #include "clang/CodeGen/CGFunctionInfo.h"
33 #include "clang/Frontend/FrontendDiagnostic.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/Intrinsics.h"
37 #include "llvm/IR/MDBuilder.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
40 using namespace clang;
41 using namespace CodeGen;
42 
43 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
44 /// markers.
45 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
46                                       const LangOptions &LangOpts) {
47   if (CGOpts.DisableLifetimeMarkers)
48     return false;
49 
50   // Disable lifetime markers in msan builds.
51   // FIXME: Remove this when msan works with lifetime markers.
52   if (LangOpts.Sanitize.has(SanitizerKind::Memory))
53     return false;
54 
55   // Asan uses markers for use-after-scope checks.
56   if (CGOpts.SanitizeAddressUseAfterScope)
57     return true;
58 
59   // For now, only in optimized builds.
60   return CGOpts.OptimizationLevel != 0;
61 }
62 
63 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
64     : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
65       Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
66               CGBuilderInserterTy(this)),
67       SanOpts(CGM.getLangOpts().Sanitize), DebugInfo(CGM.getModuleDebugInfo()),
68       PGO(cgm), ShouldEmitLifetimeMarkers(shouldEmitLifetimeMarkers(
69                     CGM.getCodeGenOpts(), CGM.getLangOpts())) {
70   if (!suppressNewContext)
71     CGM.getCXXABI().getMangleContext().startNewFunction();
72 
73   llvm::FastMathFlags FMF;
74   if (CGM.getLangOpts().FastMath)
75     FMF.setFast();
76   if (CGM.getLangOpts().FiniteMathOnly) {
77     FMF.setNoNaNs();
78     FMF.setNoInfs();
79   }
80   if (CGM.getCodeGenOpts().NoNaNsFPMath) {
81     FMF.setNoNaNs();
82   }
83   if (CGM.getCodeGenOpts().NoSignedZeros) {
84     FMF.setNoSignedZeros();
85   }
86   if (CGM.getCodeGenOpts().ReciprocalMath) {
87     FMF.setAllowReciprocal();
88   }
89   if (CGM.getCodeGenOpts().Reassociate) {
90     FMF.setAllowReassoc();
91   }
92   Builder.setFastMathFlags(FMF);
93 }
94 
95 CodeGenFunction::~CodeGenFunction() {
96   assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
97 
98   // If there are any unclaimed block infos, go ahead and destroy them
99   // now.  This can happen if IR-gen gets clever and skips evaluating
100   // something.
101   if (FirstBlockInfo)
102     destroyBlockInfos(FirstBlockInfo);
103 
104   if (getLangOpts().OpenMP && CurFn)
105     CGM.getOpenMPRuntime().functionFinished(*this);
106 }
107 
108 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
109                                                     LValueBaseInfo *BaseInfo,
110                                                     TBAAAccessInfo *TBAAInfo) {
111   return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo, TBAAInfo,
112                                  /* forPointeeType= */ true);
113 }
114 
115 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
116                                                    LValueBaseInfo *BaseInfo,
117                                                    TBAAAccessInfo *TBAAInfo,
118                                                    bool forPointeeType) {
119   if (TBAAInfo)
120     *TBAAInfo = CGM.getTBAAAccessInfo(T);
121 
122   // Honor alignment typedef attributes even on incomplete types.
123   // We also honor them straight for C++ class types, even as pointees;
124   // there's an expressivity gap here.
125   if (auto TT = T->getAs<TypedefType>()) {
126     if (auto Align = TT->getDecl()->getMaxAlignment()) {
127       if (BaseInfo)
128         *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType);
129       return getContext().toCharUnitsFromBits(Align);
130     }
131   }
132 
133   if (BaseInfo)
134     *BaseInfo = LValueBaseInfo(AlignmentSource::Type);
135 
136   CharUnits Alignment;
137   if (T->isIncompleteType()) {
138     Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
139   } else {
140     // For C++ class pointees, we don't know whether we're pointing at a
141     // base or a complete object, so we generally need to use the
142     // non-virtual alignment.
143     const CXXRecordDecl *RD;
144     if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
145       Alignment = CGM.getClassPointerAlignment(RD);
146     } else {
147       Alignment = getContext().getTypeAlignInChars(T);
148       if (T.getQualifiers().hasUnaligned())
149         Alignment = CharUnits::One();
150     }
151 
152     // Cap to the global maximum type alignment unless the alignment
153     // was somehow explicit on the type.
154     if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
155       if (Alignment.getQuantity() > MaxAlign &&
156           !getContext().isAlignmentRequired(T))
157         Alignment = CharUnits::fromQuantity(MaxAlign);
158     }
159   }
160   return Alignment;
161 }
162 
163 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
164   LValueBaseInfo BaseInfo;
165   TBAAAccessInfo TBAAInfo;
166   CharUnits Alignment = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo);
167   return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo,
168                           TBAAInfo);
169 }
170 
171 /// Given a value of type T* that may not be to a complete object,
172 /// construct an l-value with the natural pointee alignment of T.
173 LValue
174 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
175   LValueBaseInfo BaseInfo;
176   TBAAAccessInfo TBAAInfo;
177   CharUnits Align = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo,
178                                             /* forPointeeType= */ true);
179   return MakeAddrLValue(Address(V, Align), T, BaseInfo, TBAAInfo);
180 }
181 
182 
183 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
184   return CGM.getTypes().ConvertTypeForMem(T);
185 }
186 
187 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
188   return CGM.getTypes().ConvertType(T);
189 }
190 
191 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
192   type = type.getCanonicalType();
193   while (true) {
194     switch (type->getTypeClass()) {
195 #define TYPE(name, parent)
196 #define ABSTRACT_TYPE(name, parent)
197 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
198 #define DEPENDENT_TYPE(name, parent) case Type::name:
199 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
200 #include "clang/AST/TypeNodes.def"
201       llvm_unreachable("non-canonical or dependent type in IR-generation");
202 
203     case Type::Auto:
204     case Type::DeducedTemplateSpecialization:
205       llvm_unreachable("undeduced type in IR-generation");
206 
207     // Various scalar types.
208     case Type::Builtin:
209     case Type::Pointer:
210     case Type::BlockPointer:
211     case Type::LValueReference:
212     case Type::RValueReference:
213     case Type::MemberPointer:
214     case Type::Vector:
215     case Type::ExtVector:
216     case Type::FunctionProto:
217     case Type::FunctionNoProto:
218     case Type::Enum:
219     case Type::ObjCObjectPointer:
220     case Type::Pipe:
221       return TEK_Scalar;
222 
223     // Complexes.
224     case Type::Complex:
225       return TEK_Complex;
226 
227     // Arrays, records, and Objective-C objects.
228     case Type::ConstantArray:
229     case Type::IncompleteArray:
230     case Type::VariableArray:
231     case Type::Record:
232     case Type::ObjCObject:
233     case Type::ObjCInterface:
234       return TEK_Aggregate;
235 
236     // We operate on atomic values according to their underlying type.
237     case Type::Atomic:
238       type = cast<AtomicType>(type)->getValueType();
239       continue;
240     }
241     llvm_unreachable("unknown type kind!");
242   }
243 }
244 
245 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
246   // For cleanliness, we try to avoid emitting the return block for
247   // simple cases.
248   llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
249 
250   if (CurBB) {
251     assert(!CurBB->getTerminator() && "Unexpected terminated block.");
252 
253     // We have a valid insert point, reuse it if it is empty or there are no
254     // explicit jumps to the return block.
255     if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
256       ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
257       delete ReturnBlock.getBlock();
258       ReturnBlock = JumpDest();
259     } else
260       EmitBlock(ReturnBlock.getBlock());
261     return llvm::DebugLoc();
262   }
263 
264   // Otherwise, if the return block is the target of a single direct
265   // branch then we can just put the code in that block instead. This
266   // cleans up functions which started with a unified return block.
267   if (ReturnBlock.getBlock()->hasOneUse()) {
268     llvm::BranchInst *BI =
269       dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
270     if (BI && BI->isUnconditional() &&
271         BI->getSuccessor(0) == ReturnBlock.getBlock()) {
272       // Record/return the DebugLoc of the simple 'return' expression to be used
273       // later by the actual 'ret' instruction.
274       llvm::DebugLoc Loc = BI->getDebugLoc();
275       Builder.SetInsertPoint(BI->getParent());
276       BI->eraseFromParent();
277       delete ReturnBlock.getBlock();
278       ReturnBlock = JumpDest();
279       return Loc;
280     }
281   }
282 
283   // FIXME: We are at an unreachable point, there is no reason to emit the block
284   // unless it has uses. However, we still need a place to put the debug
285   // region.end for now.
286 
287   EmitBlock(ReturnBlock.getBlock());
288   return llvm::DebugLoc();
289 }
290 
291 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
292   if (!BB) return;
293   if (!BB->use_empty())
294     return CGF.CurFn->getBasicBlockList().push_back(BB);
295   delete BB;
296 }
297 
298 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
299   assert(BreakContinueStack.empty() &&
300          "mismatched push/pop in break/continue stack!");
301 
302   bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
303     && NumSimpleReturnExprs == NumReturnExprs
304     && ReturnBlock.getBlock()->use_empty();
305   // Usually the return expression is evaluated before the cleanup
306   // code.  If the function contains only a simple return statement,
307   // such as a constant, the location before the cleanup code becomes
308   // the last useful breakpoint in the function, because the simple
309   // return expression will be evaluated after the cleanup code. To be
310   // safe, set the debug location for cleanup code to the location of
311   // the return statement.  Otherwise the cleanup code should be at the
312   // end of the function's lexical scope.
313   //
314   // If there are multiple branches to the return block, the branch
315   // instructions will get the location of the return statements and
316   // all will be fine.
317   if (CGDebugInfo *DI = getDebugInfo()) {
318     if (OnlySimpleReturnStmts)
319       DI->EmitLocation(Builder, LastStopPoint);
320     else
321       DI->EmitLocation(Builder, EndLoc);
322   }
323 
324   // Pop any cleanups that might have been associated with the
325   // parameters.  Do this in whatever block we're currently in; it's
326   // important to do this before we enter the return block or return
327   // edges will be *really* confused.
328   bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
329   bool HasOnlyLifetimeMarkers =
330       HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
331   bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
332   if (HasCleanups) {
333     // Make sure the line table doesn't jump back into the body for
334     // the ret after it's been at EndLoc.
335     if (CGDebugInfo *DI = getDebugInfo())
336       if (OnlySimpleReturnStmts)
337         DI->EmitLocation(Builder, EndLoc);
338 
339     PopCleanupBlocks(PrologueCleanupDepth);
340   }
341 
342   // Emit function epilog (to return).
343   llvm::DebugLoc Loc = EmitReturnBlock();
344 
345   if (ShouldInstrumentFunction()) {
346     if (CGM.getCodeGenOpts().InstrumentFunctions)
347       CurFn->addFnAttr("instrument-function-exit", "__cyg_profile_func_exit");
348     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
349       CurFn->addFnAttr("instrument-function-exit-inlined",
350                        "__cyg_profile_func_exit");
351   }
352 
353   // Emit debug descriptor for function end.
354   if (CGDebugInfo *DI = getDebugInfo())
355     DI->EmitFunctionEnd(Builder, CurFn);
356 
357   // Reset the debug location to that of the simple 'return' expression, if any
358   // rather than that of the end of the function's scope '}'.
359   ApplyDebugLocation AL(*this, Loc);
360   EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
361   EmitEndEHSpec(CurCodeDecl);
362 
363   assert(EHStack.empty() &&
364          "did not remove all scopes from cleanup stack!");
365 
366   // If someone did an indirect goto, emit the indirect goto block at the end of
367   // the function.
368   if (IndirectBranch) {
369     EmitBlock(IndirectBranch->getParent());
370     Builder.ClearInsertionPoint();
371   }
372 
373   // If some of our locals escaped, insert a call to llvm.localescape in the
374   // entry block.
375   if (!EscapedLocals.empty()) {
376     // Invert the map from local to index into a simple vector. There should be
377     // no holes.
378     SmallVector<llvm::Value *, 4> EscapeArgs;
379     EscapeArgs.resize(EscapedLocals.size());
380     for (auto &Pair : EscapedLocals)
381       EscapeArgs[Pair.second] = Pair.first;
382     llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
383         &CGM.getModule(), llvm::Intrinsic::localescape);
384     CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
385   }
386 
387   // Remove the AllocaInsertPt instruction, which is just a convenience for us.
388   llvm::Instruction *Ptr = AllocaInsertPt;
389   AllocaInsertPt = nullptr;
390   Ptr->eraseFromParent();
391 
392   // If someone took the address of a label but never did an indirect goto, we
393   // made a zero entry PHI node, which is illegal, zap it now.
394   if (IndirectBranch) {
395     llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
396     if (PN->getNumIncomingValues() == 0) {
397       PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
398       PN->eraseFromParent();
399     }
400   }
401 
402   EmitIfUsed(*this, EHResumeBlock);
403   EmitIfUsed(*this, TerminateLandingPad);
404   EmitIfUsed(*this, TerminateHandler);
405   EmitIfUsed(*this, UnreachableBlock);
406 
407   for (const auto &FuncletAndParent : TerminateFunclets)
408     EmitIfUsed(*this, FuncletAndParent.second);
409 
410   if (CGM.getCodeGenOpts().EmitDeclMetadata)
411     EmitDeclMetadata();
412 
413   for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
414            I = DeferredReplacements.begin(),
415            E = DeferredReplacements.end();
416        I != E; ++I) {
417     I->first->replaceAllUsesWith(I->second);
418     I->first->eraseFromParent();
419   }
420 
421   // Eliminate CleanupDestSlot alloca by replacing it with SSA values and
422   // PHIs if the current function is a coroutine. We don't do it for all
423   // functions as it may result in slight increase in numbers of instructions
424   // if compiled with no optimizations. We do it for coroutine as the lifetime
425   // of CleanupDestSlot alloca make correct coroutine frame building very
426   // difficult.
427   if (NormalCleanupDest.isValid() && isCoroutine()) {
428     llvm::DominatorTree DT(*CurFn);
429     llvm::PromoteMemToReg(
430         cast<llvm::AllocaInst>(NormalCleanupDest.getPointer()), DT);
431     NormalCleanupDest = Address::invalid();
432   }
433 
434   // Scan function arguments for vector width.
435   for (llvm::Argument &A : CurFn->args())
436     if (auto *VT = dyn_cast<llvm::VectorType>(A.getType()))
437       LargestVectorWidth = std::max(LargestVectorWidth,
438                                     VT->getPrimitiveSizeInBits());
439 
440   // Update vector width based on return type.
441   if (auto *VT = dyn_cast<llvm::VectorType>(CurFn->getReturnType()))
442     LargestVectorWidth = std::max(LargestVectorWidth,
443                                   VT->getPrimitiveSizeInBits());
444 
445   // Add the required-vector-width attribute. This contains the max width from:
446   // 1. min-vector-width attribute used in the source program.
447   // 2. Any builtins used that have a vector width specified.
448   // 3. Values passed in and out of inline assembly.
449   // 4. Width of vector arguments and return types for this function.
450   // 5. Width of vector aguments and return types for functions called by this
451   //    function.
452   CurFn->addFnAttr("min-legal-vector-width", llvm::utostr(LargestVectorWidth));
453 
454   // If we generated an unreachable return block, delete it now.
455   if (ReturnBlock.isValid() && ReturnBlock.getBlock()->use_empty()) {
456     Builder.ClearInsertionPoint();
457     ReturnBlock.getBlock()->eraseFromParent();
458   }
459   if (ReturnValue.isValid()) {
460     auto *RetAlloca = dyn_cast<llvm::AllocaInst>(ReturnValue.getPointer());
461     if (RetAlloca && RetAlloca->use_empty()) {
462       RetAlloca->eraseFromParent();
463       ReturnValue = Address::invalid();
464     }
465   }
466 }
467 
468 /// ShouldInstrumentFunction - Return true if the current function should be
469 /// instrumented with __cyg_profile_func_* calls
470 bool CodeGenFunction::ShouldInstrumentFunction() {
471   if (!CGM.getCodeGenOpts().InstrumentFunctions &&
472       !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining &&
473       !CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
474     return false;
475   if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
476     return false;
477   return true;
478 }
479 
480 /// ShouldXRayInstrument - Return true if the current function should be
481 /// instrumented with XRay nop sleds.
482 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
483   return CGM.getCodeGenOpts().XRayInstrumentFunctions;
484 }
485 
486 /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to
487 /// the __xray_customevent(...) builtin calls, when doing XRay instrumentation.
488 bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const {
489   return CGM.getCodeGenOpts().XRayInstrumentFunctions &&
490          (CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents ||
491           CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask ==
492               XRayInstrKind::Custom);
493 }
494 
495 bool CodeGenFunction::AlwaysEmitXRayTypedEvents() const {
496   return CGM.getCodeGenOpts().XRayInstrumentFunctions &&
497          (CGM.getCodeGenOpts().XRayAlwaysEmitTypedEvents ||
498           CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask ==
499               XRayInstrKind::Typed);
500 }
501 
502 llvm::Constant *
503 CodeGenFunction::EncodeAddrForUseInPrologue(llvm::Function *F,
504                                             llvm::Constant *Addr) {
505   // Addresses stored in prologue data can't require run-time fixups and must
506   // be PC-relative. Run-time fixups are undesirable because they necessitate
507   // writable text segments, which are unsafe. And absolute addresses are
508   // undesirable because they break PIE mode.
509 
510   // Add a layer of indirection through a private global. Taking its address
511   // won't result in a run-time fixup, even if Addr has linkonce_odr linkage.
512   auto *GV = new llvm::GlobalVariable(CGM.getModule(), Addr->getType(),
513                                       /*isConstant=*/true,
514                                       llvm::GlobalValue::PrivateLinkage, Addr);
515 
516   // Create a PC-relative address.
517   auto *GOTAsInt = llvm::ConstantExpr::getPtrToInt(GV, IntPtrTy);
518   auto *FuncAsInt = llvm::ConstantExpr::getPtrToInt(F, IntPtrTy);
519   auto *PCRelAsInt = llvm::ConstantExpr::getSub(GOTAsInt, FuncAsInt);
520   return (IntPtrTy == Int32Ty)
521              ? PCRelAsInt
522              : llvm::ConstantExpr::getTrunc(PCRelAsInt, Int32Ty);
523 }
524 
525 llvm::Value *
526 CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F,
527                                           llvm::Value *EncodedAddr) {
528   // Reconstruct the address of the global.
529   auto *PCRelAsInt = Builder.CreateSExt(EncodedAddr, IntPtrTy);
530   auto *FuncAsInt = Builder.CreatePtrToInt(F, IntPtrTy, "func_addr.int");
531   auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int");
532   auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr");
533 
534   // Load the original pointer through the global.
535   return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()),
536                             "decoded_addr");
537 }
538 
539 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
540                                                llvm::Function *Fn)
541 {
542   if (!FD->hasAttr<OpenCLKernelAttr>())
543     return;
544 
545   llvm::LLVMContext &Context = getLLVMContext();
546 
547   CGM.GenOpenCLArgMetadata(Fn, FD, this);
548 
549   if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
550     QualType HintQTy = A->getTypeHint();
551     const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
552     bool IsSignedInteger =
553         HintQTy->isSignedIntegerType() ||
554         (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
555     llvm::Metadata *AttrMDArgs[] = {
556         llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
557             CGM.getTypes().ConvertType(A->getTypeHint()))),
558         llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
559             llvm::IntegerType::get(Context, 32),
560             llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
561     Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
562   }
563 
564   if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
565     llvm::Metadata *AttrMDArgs[] = {
566         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
567         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
568         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
569     Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
570   }
571 
572   if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
573     llvm::Metadata *AttrMDArgs[] = {
574         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
575         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
576         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
577     Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
578   }
579 
580   if (const OpenCLIntelReqdSubGroupSizeAttr *A =
581           FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
582     llvm::Metadata *AttrMDArgs[] = {
583         llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
584     Fn->setMetadata("intel_reqd_sub_group_size",
585                     llvm::MDNode::get(Context, AttrMDArgs));
586   }
587 }
588 
589 /// Determine whether the function F ends with a return stmt.
590 static bool endsWithReturn(const Decl* F) {
591   const Stmt *Body = nullptr;
592   if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
593     Body = FD->getBody();
594   else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
595     Body = OMD->getBody();
596 
597   if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
598     auto LastStmt = CS->body_rbegin();
599     if (LastStmt != CS->body_rend())
600       return isa<ReturnStmt>(*LastStmt);
601   }
602   return false;
603 }
604 
605 void CodeGenFunction::markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
606   if (SanOpts.has(SanitizerKind::Thread)) {
607     Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
608     Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
609   }
610 }
611 
612 static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) {
613   auto *MD = dyn_cast_or_null<CXXMethodDecl>(D);
614   if (!MD || !MD->getDeclName().getAsIdentifierInfo() ||
615       !MD->getDeclName().getAsIdentifierInfo()->isStr("allocate") ||
616       (MD->getNumParams() != 1 && MD->getNumParams() != 2))
617     return false;
618 
619   if (MD->parameters()[0]->getType().getCanonicalType() != Ctx.getSizeType())
620     return false;
621 
622   if (MD->getNumParams() == 2) {
623     auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>();
624     if (!PT || !PT->isVoidPointerType() ||
625         !PT->getPointeeType().isConstQualified())
626       return false;
627   }
628 
629   return true;
630 }
631 
632 /// Return the UBSan prologue signature for \p FD if one is available.
633 static llvm::Constant *getPrologueSignature(CodeGenModule &CGM,
634                                             const FunctionDecl *FD) {
635   if (const auto *MD = dyn_cast<CXXMethodDecl>(FD))
636     if (!MD->isStatic())
637       return nullptr;
638   return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM);
639 }
640 
641 void CodeGenFunction::StartFunction(GlobalDecl GD,
642                                     QualType RetTy,
643                                     llvm::Function *Fn,
644                                     const CGFunctionInfo &FnInfo,
645                                     const FunctionArgList &Args,
646                                     SourceLocation Loc,
647                                     SourceLocation StartLoc) {
648   assert(!CurFn &&
649          "Do not use a CodeGenFunction object for more than one function");
650 
651   const Decl *D = GD.getDecl();
652 
653   DidCallStackSave = false;
654   CurCodeDecl = D;
655   if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
656     if (FD->usesSEHTry())
657       CurSEHParent = FD;
658   CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
659   FnRetTy = RetTy;
660   CurFn = Fn;
661   CurFnInfo = &FnInfo;
662   assert(CurFn->isDeclaration() && "Function already has body?");
663 
664   // If this function has been blacklisted for any of the enabled sanitizers,
665   // disable the sanitizer for the function.
666   do {
667 #define SANITIZER(NAME, ID)                                                    \
668   if (SanOpts.empty())                                                         \
669     break;                                                                     \
670   if (SanOpts.has(SanitizerKind::ID))                                          \
671     if (CGM.isInSanitizerBlacklist(SanitizerKind::ID, Fn, Loc))                \
672       SanOpts.set(SanitizerKind::ID, false);
673 
674 #include "clang/Basic/Sanitizers.def"
675 #undef SANITIZER
676   } while (0);
677 
678   if (D) {
679     // Apply the no_sanitize* attributes to SanOpts.
680     for (auto Attr : D->specific_attrs<NoSanitizeAttr>()) {
681       SanitizerMask mask = Attr->getMask();
682       SanOpts.Mask &= ~mask;
683       if (mask & SanitizerKind::Address)
684         SanOpts.set(SanitizerKind::KernelAddress, false);
685       if (mask & SanitizerKind::KernelAddress)
686         SanOpts.set(SanitizerKind::Address, false);
687       if (mask & SanitizerKind::HWAddress)
688         SanOpts.set(SanitizerKind::KernelHWAddress, false);
689       if (mask & SanitizerKind::KernelHWAddress)
690         SanOpts.set(SanitizerKind::HWAddress, false);
691     }
692   }
693 
694   // Apply sanitizer attributes to the function.
695   if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
696     Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
697   if (SanOpts.hasOneOf(SanitizerKind::HWAddress | SanitizerKind::KernelHWAddress))
698     Fn->addFnAttr(llvm::Attribute::SanitizeHWAddress);
699   if (SanOpts.has(SanitizerKind::MemTag))
700     Fn->addFnAttr(llvm::Attribute::SanitizeMemTag);
701   if (SanOpts.has(SanitizerKind::Thread))
702     Fn->addFnAttr(llvm::Attribute::SanitizeThread);
703   if (SanOpts.hasOneOf(SanitizerKind::Memory | SanitizerKind::KernelMemory))
704     Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
705   if (SanOpts.has(SanitizerKind::SafeStack))
706     Fn->addFnAttr(llvm::Attribute::SafeStack);
707   if (SanOpts.has(SanitizerKind::ShadowCallStack))
708     Fn->addFnAttr(llvm::Attribute::ShadowCallStack);
709 
710   // Apply fuzzing attribute to the function.
711   if (SanOpts.hasOneOf(SanitizerKind::Fuzzer | SanitizerKind::FuzzerNoLink))
712     Fn->addFnAttr(llvm::Attribute::OptForFuzzing);
713 
714   // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
715   // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
716   if (SanOpts.has(SanitizerKind::Thread)) {
717     if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
718       IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
719       if (OMD->getMethodFamily() == OMF_dealloc ||
720           OMD->getMethodFamily() == OMF_initialize ||
721           (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
722         markAsIgnoreThreadCheckingAtRuntime(Fn);
723       }
724     }
725   }
726 
727   // Ignore unrelated casts in STL allocate() since the allocator must cast
728   // from void* to T* before object initialization completes. Don't match on the
729   // namespace because not all allocators are in std::
730   if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
731     if (matchesStlAllocatorFn(D, getContext()))
732       SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast;
733   }
734 
735   // Ignore null checks in coroutine functions since the coroutines passes
736   // are not aware of how to move the extra UBSan instructions across the split
737   // coroutine boundaries.
738   if (D && SanOpts.has(SanitizerKind::Null))
739     if (const auto *FD = dyn_cast<FunctionDecl>(D))
740       if (FD->getBody() &&
741           FD->getBody()->getStmtClass() == Stmt::CoroutineBodyStmtClass)
742         SanOpts.Mask &= ~SanitizerKind::Null;
743 
744   // Apply xray attributes to the function (as a string, for now)
745   if (D) {
746     if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
747       if (CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
748               XRayInstrKind::Function)) {
749         if (XRayAttr->alwaysXRayInstrument() && ShouldXRayInstrumentFunction())
750           Fn->addFnAttr("function-instrument", "xray-always");
751         if (XRayAttr->neverXRayInstrument())
752           Fn->addFnAttr("function-instrument", "xray-never");
753         if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>())
754           if (ShouldXRayInstrumentFunction())
755             Fn->addFnAttr("xray-log-args",
756                           llvm::utostr(LogArgs->getArgumentCount()));
757       }
758     } else {
759       if (ShouldXRayInstrumentFunction() && !CGM.imbueXRayAttrs(Fn, Loc))
760         Fn->addFnAttr(
761             "xray-instruction-threshold",
762             llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
763     }
764   }
765 
766   // Add no-jump-tables value.
767   Fn->addFnAttr("no-jump-tables",
768                 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
769 
770   // Add profile-sample-accurate value.
771   if (CGM.getCodeGenOpts().ProfileSampleAccurate)
772     Fn->addFnAttr("profile-sample-accurate");
773 
774   if (D && D->hasAttr<CFICanonicalJumpTableAttr>())
775     Fn->addFnAttr("cfi-canonical-jump-table");
776 
777   if (getLangOpts().OpenCL) {
778     // Add metadata for a kernel function.
779     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
780       EmitOpenCLKernelMetadata(FD, Fn);
781   }
782 
783   // If we are checking function types, emit a function type signature as
784   // prologue data.
785   if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
786     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
787       if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
788         // Remove any (C++17) exception specifications, to allow calling e.g. a
789         // noexcept function through a non-noexcept pointer.
790         auto ProtoTy =
791           getContext().getFunctionTypeWithExceptionSpec(FD->getType(),
792                                                         EST_None);
793         llvm::Constant *FTRTTIConst =
794             CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true);
795         llvm::Constant *FTRTTIConstEncoded =
796             EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
797         llvm::Constant *PrologueStructElems[] = {PrologueSig,
798                                                  FTRTTIConstEncoded};
799         llvm::Constant *PrologueStructConst =
800             llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
801         Fn->setPrologueData(PrologueStructConst);
802       }
803     }
804   }
805 
806   // If we're checking nullability, we need to know whether we can check the
807   // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
808   if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
809     auto Nullability = FnRetTy->getNullability(getContext());
810     if (Nullability && *Nullability == NullabilityKind::NonNull) {
811       if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
812             CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
813         RetValNullabilityPrecondition =
814             llvm::ConstantInt::getTrue(getLLVMContext());
815     }
816   }
817 
818   // If we're in C++ mode and the function name is "main", it is guaranteed
819   // to be norecurse by the standard (3.6.1.3 "The function main shall not be
820   // used within a program").
821   if (getLangOpts().CPlusPlus)
822     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
823       if (FD->isMain())
824         Fn->addFnAttr(llvm::Attribute::NoRecurse);
825 
826   // If a custom alignment is used, force realigning to this alignment on
827   // any main function which certainly will need it.
828   if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
829     if ((FD->isMain() || FD->isMSVCRTEntryPoint()) &&
830         CGM.getCodeGenOpts().StackAlignment)
831       Fn->addFnAttr("stackrealign");
832 
833   llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
834 
835   // Create a marker to make it easy to insert allocas into the entryblock
836   // later.  Don't create this with the builder, because we don't want it
837   // folded.
838   llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
839   AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
840 
841   ReturnBlock = getJumpDestInCurrentScope("return");
842 
843   Builder.SetInsertPoint(EntryBB);
844 
845   // If we're checking the return value, allocate space for a pointer to a
846   // precise source location of the checked return statement.
847   if (requiresReturnValueCheck()) {
848     ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
849     InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
850   }
851 
852   // Emit subprogram debug descriptor.
853   if (CGDebugInfo *DI = getDebugInfo()) {
854     // Reconstruct the type from the argument list so that implicit parameters,
855     // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
856     // convention.
857     CallingConv CC = CallingConv::CC_C;
858     if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
859       if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
860         CC = SrcFnTy->getCallConv();
861     SmallVector<QualType, 16> ArgTypes;
862     for (const VarDecl *VD : Args)
863       ArgTypes.push_back(VD->getType());
864     QualType FnType = getContext().getFunctionType(
865         RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
866     DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, CurFuncIsThunk,
867                           Builder);
868   }
869 
870   if (ShouldInstrumentFunction()) {
871     if (CGM.getCodeGenOpts().InstrumentFunctions)
872       CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
873     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
874       CurFn->addFnAttr("instrument-function-entry-inlined",
875                        "__cyg_profile_func_enter");
876     if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
877       CurFn->addFnAttr("instrument-function-entry-inlined",
878                        "__cyg_profile_func_enter_bare");
879   }
880 
881   // Since emitting the mcount call here impacts optimizations such as function
882   // inlining, we just add an attribute to insert a mcount call in backend.
883   // The attribute "counting-function" is set to mcount function name which is
884   // architecture dependent.
885   if (CGM.getCodeGenOpts().InstrumentForProfiling) {
886     // Calls to fentry/mcount should not be generated if function has
887     // the no_instrument_function attribute.
888     if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
889       if (CGM.getCodeGenOpts().CallFEntry)
890         Fn->addFnAttr("fentry-call", "true");
891       else {
892         Fn->addFnAttr("instrument-function-entry-inlined",
893                       getTarget().getMCountName());
894       }
895     }
896   }
897 
898   if (RetTy->isVoidType()) {
899     // Void type; nothing to return.
900     ReturnValue = Address::invalid();
901 
902     // Count the implicit return.
903     if (!endsWithReturn(D))
904       ++NumReturnExprs;
905   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect) {
906     // Indirect return; emit returned value directly into sret slot.
907     // This reduces code size, and affects correctness in C++.
908     auto AI = CurFn->arg_begin();
909     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
910       ++AI;
911     ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
912     if (!CurFnInfo->getReturnInfo().getIndirectByVal()) {
913       ReturnValuePointer =
914           CreateDefaultAlignTempAlloca(Int8PtrTy, "result.ptr");
915       Builder.CreateStore(Builder.CreatePointerBitCastOrAddrSpaceCast(
916                               ReturnValue.getPointer(), Int8PtrTy),
917                           ReturnValuePointer);
918     }
919   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
920              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
921     // Load the sret pointer from the argument struct and return into that.
922     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
923     llvm::Function::arg_iterator EI = CurFn->arg_end();
924     --EI;
925     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
926     ReturnValuePointer = Address(Addr, getPointerAlign());
927     Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
928     ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
929   } else {
930     ReturnValue = CreateIRTemp(RetTy, "retval");
931 
932     // Tell the epilog emitter to autorelease the result.  We do this
933     // now so that various specialized functions can suppress it
934     // during their IR-generation.
935     if (getLangOpts().ObjCAutoRefCount &&
936         !CurFnInfo->isReturnsRetained() &&
937         RetTy->isObjCRetainableType())
938       AutoreleaseResult = true;
939   }
940 
941   EmitStartEHSpec(CurCodeDecl);
942 
943   PrologueCleanupDepth = EHStack.stable_begin();
944 
945   // Emit OpenMP specific initialization of the device functions.
946   if (getLangOpts().OpenMP && CurCodeDecl)
947     CGM.getOpenMPRuntime().emitFunctionProlog(*this, CurCodeDecl);
948 
949   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
950 
951   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
952     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
953     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
954     if (MD->getParent()->isLambda() &&
955         MD->getOverloadedOperator() == OO_Call) {
956       // We're in a lambda; figure out the captures.
957       MD->getParent()->getCaptureFields(LambdaCaptureFields,
958                                         LambdaThisCaptureField);
959       if (LambdaThisCaptureField) {
960         // If the lambda captures the object referred to by '*this' - either by
961         // value or by reference, make sure CXXThisValue points to the correct
962         // object.
963 
964         // Get the lvalue for the field (which is a copy of the enclosing object
965         // or contains the address of the enclosing object).
966         LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
967         if (!LambdaThisCaptureField->getType()->isPointerType()) {
968           // If the enclosing object was captured by value, just use its address.
969           CXXThisValue = ThisFieldLValue.getAddress().getPointer();
970         } else {
971           // Load the lvalue pointed to by the field, since '*this' was captured
972           // by reference.
973           CXXThisValue =
974               EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
975         }
976       }
977       for (auto *FD : MD->getParent()->fields()) {
978         if (FD->hasCapturedVLAType()) {
979           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
980                                            SourceLocation()).getScalarVal();
981           auto VAT = FD->getCapturedVLAType();
982           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
983         }
984       }
985     } else {
986       // Not in a lambda; just use 'this' from the method.
987       // FIXME: Should we generate a new load for each use of 'this'?  The
988       // fast register allocator would be happier...
989       CXXThisValue = CXXABIThisValue;
990     }
991 
992     // Check the 'this' pointer once per function, if it's available.
993     if (CXXABIThisValue) {
994       SanitizerSet SkippedChecks;
995       SkippedChecks.set(SanitizerKind::ObjectSize, true);
996       QualType ThisTy = MD->getThisType();
997 
998       // If this is the call operator of a lambda with no capture-default, it
999       // may have a static invoker function, which may call this operator with
1000       // a null 'this' pointer.
1001       if (isLambdaCallOperator(MD) &&
1002           MD->getParent()->getLambdaCaptureDefault() == LCD_None)
1003         SkippedChecks.set(SanitizerKind::Null, true);
1004 
1005       EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1006                                                 : TCK_MemberCall,
1007                     Loc, CXXABIThisValue, ThisTy,
1008                     getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1009                     SkippedChecks);
1010     }
1011   }
1012 
1013   // If any of the arguments have a variably modified type, make sure to
1014   // emit the type size.
1015   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1016        i != e; ++i) {
1017     const VarDecl *VD = *i;
1018 
1019     // Dig out the type as written from ParmVarDecls; it's unclear whether
1020     // the standard (C99 6.9.1p10) requires this, but we're following the
1021     // precedent set by gcc.
1022     QualType Ty;
1023     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1024       Ty = PVD->getOriginalType();
1025     else
1026       Ty = VD->getType();
1027 
1028     if (Ty->isVariablyModifiedType())
1029       EmitVariablyModifiedType(Ty);
1030   }
1031   // Emit a location at the end of the prologue.
1032   if (CGDebugInfo *DI = getDebugInfo())
1033     DI->EmitLocation(Builder, StartLoc);
1034 
1035   // TODO: Do we need to handle this in two places like we do with
1036   // target-features/target-cpu?
1037   if (CurFuncDecl)
1038     if (const auto *VecWidth = CurFuncDecl->getAttr<MinVectorWidthAttr>())
1039       LargestVectorWidth = VecWidth->getVectorWidth();
1040 }
1041 
1042 void CodeGenFunction::EmitFunctionBody(const Stmt *Body) {
1043   incrementProfileCounter(Body);
1044   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1045     EmitCompoundStmtWithoutScope(*S);
1046   else
1047     EmitStmt(Body);
1048 }
1049 
1050 /// When instrumenting to collect profile data, the counts for some blocks
1051 /// such as switch cases need to not include the fall-through counts, so
1052 /// emit a branch around the instrumentation code. When not instrumenting,
1053 /// this just calls EmitBlock().
1054 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1055                                                const Stmt *S) {
1056   llvm::BasicBlock *SkipCountBB = nullptr;
1057   if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1058     // When instrumenting for profiling, the fallthrough to certain
1059     // statements needs to skip over the instrumentation code so that we
1060     // get an accurate count.
1061     SkipCountBB = createBasicBlock("skipcount");
1062     EmitBranch(SkipCountBB);
1063   }
1064   EmitBlock(BB);
1065   uint64_t CurrentCount = getCurrentProfileCount();
1066   incrementProfileCounter(S);
1067   setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1068   if (SkipCountBB)
1069     EmitBlock(SkipCountBB);
1070 }
1071 
1072 /// Tries to mark the given function nounwind based on the
1073 /// non-existence of any throwing calls within it.  We believe this is
1074 /// lightweight enough to do at -O0.
1075 static void TryMarkNoThrow(llvm::Function *F) {
1076   // LLVM treats 'nounwind' on a function as part of the type, so we
1077   // can't do this on functions that can be overwritten.
1078   if (F->isInterposable()) return;
1079 
1080   for (llvm::BasicBlock &BB : *F)
1081     for (llvm::Instruction &I : BB)
1082       if (I.mayThrow())
1083         return;
1084 
1085   F->setDoesNotThrow();
1086 }
1087 
1088 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1089                                                FunctionArgList &Args) {
1090   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1091   QualType ResTy = FD->getReturnType();
1092 
1093   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1094   if (MD && MD->isInstance()) {
1095     if (CGM.getCXXABI().HasThisReturn(GD))
1096       ResTy = MD->getThisType();
1097     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1098       ResTy = CGM.getContext().VoidPtrTy;
1099     CGM.getCXXABI().buildThisParam(*this, Args);
1100   }
1101 
1102   // The base version of an inheriting constructor whose constructed base is a
1103   // virtual base is not passed any arguments (because it doesn't actually call
1104   // the inherited constructor).
1105   bool PassedParams = true;
1106   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1107     if (auto Inherited = CD->getInheritedConstructor())
1108       PassedParams =
1109           getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1110 
1111   if (PassedParams) {
1112     for (auto *Param : FD->parameters()) {
1113       Args.push_back(Param);
1114       if (!Param->hasAttr<PassObjectSizeAttr>())
1115         continue;
1116 
1117       auto *Implicit = ImplicitParamDecl::Create(
1118           getContext(), Param->getDeclContext(), Param->getLocation(),
1119           /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1120       SizeArguments[Param] = Implicit;
1121       Args.push_back(Implicit);
1122     }
1123   }
1124 
1125   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1126     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1127 
1128   return ResTy;
1129 }
1130 
1131 static bool
1132 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1133                                              const ASTContext &Context) {
1134   QualType T = FD->getReturnType();
1135   // Avoid the optimization for functions that return a record type with a
1136   // trivial destructor or another trivially copyable type.
1137   if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1138     if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1139       return !ClassDecl->hasTrivialDestructor();
1140   }
1141   return !T.isTriviallyCopyableType(Context);
1142 }
1143 
1144 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1145                                    const CGFunctionInfo &FnInfo) {
1146   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1147   CurGD = GD;
1148 
1149   FunctionArgList Args;
1150   QualType ResTy = BuildFunctionArgList(GD, Args);
1151 
1152   // Check if we should generate debug info for this function.
1153   if (FD->hasAttr<NoDebugAttr>())
1154     DebugInfo = nullptr; // disable debug info indefinitely for this function
1155 
1156   // The function might not have a body if we're generating thunks for a
1157   // function declaration.
1158   SourceRange BodyRange;
1159   if (Stmt *Body = FD->getBody())
1160     BodyRange = Body->getSourceRange();
1161   else
1162     BodyRange = FD->getLocation();
1163   CurEHLocation = BodyRange.getEnd();
1164 
1165   // Use the location of the start of the function to determine where
1166   // the function definition is located. By default use the location
1167   // of the declaration as the location for the subprogram. A function
1168   // may lack a declaration in the source code if it is created by code
1169   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1170   SourceLocation Loc = FD->getLocation();
1171 
1172   // If this is a function specialization then use the pattern body
1173   // as the location for the function.
1174   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1175     if (SpecDecl->hasBody(SpecDecl))
1176       Loc = SpecDecl->getLocation();
1177 
1178   Stmt *Body = FD->getBody();
1179 
1180   // Initialize helper which will detect jumps which can cause invalid lifetime
1181   // markers.
1182   if (Body && ShouldEmitLifetimeMarkers)
1183     Bypasses.Init(Body);
1184 
1185   // Emit the standard function prologue.
1186   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1187 
1188   // Generate the body of the function.
1189   PGO.assignRegionCounters(GD, CurFn);
1190   if (isa<CXXDestructorDecl>(FD))
1191     EmitDestructorBody(Args);
1192   else if (isa<CXXConstructorDecl>(FD))
1193     EmitConstructorBody(Args);
1194   else if (getLangOpts().CUDA &&
1195            !getLangOpts().CUDAIsDevice &&
1196            FD->hasAttr<CUDAGlobalAttr>())
1197     CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1198   else if (isa<CXXMethodDecl>(FD) &&
1199            cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1200     // The lambda static invoker function is special, because it forwards or
1201     // clones the body of the function call operator (but is actually static).
1202     EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1203   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1204              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1205               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1206     // Implicit copy-assignment gets the same special treatment as implicit
1207     // copy-constructors.
1208     emitImplicitAssignmentOperatorBody(Args);
1209   } else if (Body) {
1210     EmitFunctionBody(Body);
1211   } else
1212     llvm_unreachable("no definition for emitted function");
1213 
1214   // C++11 [stmt.return]p2:
1215   //   Flowing off the end of a function [...] results in undefined behavior in
1216   //   a value-returning function.
1217   // C11 6.9.1p12:
1218   //   If the '}' that terminates a function is reached, and the value of the
1219   //   function call is used by the caller, the behavior is undefined.
1220   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1221       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1222     bool ShouldEmitUnreachable =
1223         CGM.getCodeGenOpts().StrictReturn ||
1224         shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1225     if (SanOpts.has(SanitizerKind::Return)) {
1226       SanitizerScope SanScope(this);
1227       llvm::Value *IsFalse = Builder.getFalse();
1228       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1229                 SanitizerHandler::MissingReturn,
1230                 EmitCheckSourceLocation(FD->getLocation()), None);
1231     } else if (ShouldEmitUnreachable) {
1232       if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1233         EmitTrapCall(llvm::Intrinsic::trap);
1234     }
1235     if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1236       Builder.CreateUnreachable();
1237       Builder.ClearInsertionPoint();
1238     }
1239   }
1240 
1241   // Emit the standard function epilogue.
1242   FinishFunction(BodyRange.getEnd());
1243 
1244   // If we haven't marked the function nothrow through other means, do
1245   // a quick pass now to see if we can.
1246   if (!CurFn->doesNotThrow())
1247     TryMarkNoThrow(CurFn);
1248 }
1249 
1250 /// ContainsLabel - Return true if the statement contains a label in it.  If
1251 /// this statement is not executed normally, it not containing a label means
1252 /// that we can just remove the code.
1253 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1254   // Null statement, not a label!
1255   if (!S) return false;
1256 
1257   // If this is a label, we have to emit the code, consider something like:
1258   // if (0) {  ...  foo:  bar(); }  goto foo;
1259   //
1260   // TODO: If anyone cared, we could track __label__'s, since we know that you
1261   // can't jump to one from outside their declared region.
1262   if (isa<LabelStmt>(S))
1263     return true;
1264 
1265   // If this is a case/default statement, and we haven't seen a switch, we have
1266   // to emit the code.
1267   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1268     return true;
1269 
1270   // If this is a switch statement, we want to ignore cases below it.
1271   if (isa<SwitchStmt>(S))
1272     IgnoreCaseStmts = true;
1273 
1274   // Scan subexpressions for verboten labels.
1275   for (const Stmt *SubStmt : S->children())
1276     if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1277       return true;
1278 
1279   return false;
1280 }
1281 
1282 /// containsBreak - Return true if the statement contains a break out of it.
1283 /// If the statement (recursively) contains a switch or loop with a break
1284 /// inside of it, this is fine.
1285 bool CodeGenFunction::containsBreak(const Stmt *S) {
1286   // Null statement, not a label!
1287   if (!S) return false;
1288 
1289   // If this is a switch or loop that defines its own break scope, then we can
1290   // include it and anything inside of it.
1291   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1292       isa<ForStmt>(S))
1293     return false;
1294 
1295   if (isa<BreakStmt>(S))
1296     return true;
1297 
1298   // Scan subexpressions for verboten breaks.
1299   for (const Stmt *SubStmt : S->children())
1300     if (containsBreak(SubStmt))
1301       return true;
1302 
1303   return false;
1304 }
1305 
1306 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1307   if (!S) return false;
1308 
1309   // Some statement kinds add a scope and thus never add a decl to the current
1310   // scope. Note, this list is longer than the list of statements that might
1311   // have an unscoped decl nested within them, but this way is conservatively
1312   // correct even if more statement kinds are added.
1313   if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1314       isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1315       isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1316       isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1317     return false;
1318 
1319   if (isa<DeclStmt>(S))
1320     return true;
1321 
1322   for (const Stmt *SubStmt : S->children())
1323     if (mightAddDeclToScope(SubStmt))
1324       return true;
1325 
1326   return false;
1327 }
1328 
1329 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1330 /// to a constant, or if it does but contains a label, return false.  If it
1331 /// constant folds return true and set the boolean result in Result.
1332 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1333                                                    bool &ResultBool,
1334                                                    bool AllowLabels) {
1335   llvm::APSInt ResultInt;
1336   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1337     return false;
1338 
1339   ResultBool = ResultInt.getBoolValue();
1340   return true;
1341 }
1342 
1343 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1344 /// to a constant, or if it does but contains a label, return false.  If it
1345 /// constant folds return true and set the folded value.
1346 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1347                                                    llvm::APSInt &ResultInt,
1348                                                    bool AllowLabels) {
1349   // FIXME: Rename and handle conversion of other evaluatable things
1350   // to bool.
1351   Expr::EvalResult Result;
1352   if (!Cond->EvaluateAsInt(Result, getContext()))
1353     return false;  // Not foldable, not integer or not fully evaluatable.
1354 
1355   llvm::APSInt Int = Result.Val.getInt();
1356   if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1357     return false;  // Contains a label.
1358 
1359   ResultInt = Int;
1360   return true;
1361 }
1362 
1363 
1364 
1365 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1366 /// statement) to the specified blocks.  Based on the condition, this might try
1367 /// to simplify the codegen of the conditional based on the branch.
1368 ///
1369 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1370                                            llvm::BasicBlock *TrueBlock,
1371                                            llvm::BasicBlock *FalseBlock,
1372                                            uint64_t TrueCount) {
1373   Cond = Cond->IgnoreParens();
1374 
1375   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1376 
1377     // Handle X && Y in a condition.
1378     if (CondBOp->getOpcode() == BO_LAnd) {
1379       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1380       // folded if the case was simple enough.
1381       bool ConstantBool = false;
1382       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1383           ConstantBool) {
1384         // br(1 && X) -> br(X).
1385         incrementProfileCounter(CondBOp);
1386         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1387                                     TrueCount);
1388       }
1389 
1390       // If we have "X && 1", simplify the code to use an uncond branch.
1391       // "X && 0" would have been constant folded to 0.
1392       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1393           ConstantBool) {
1394         // br(X && 1) -> br(X).
1395         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1396                                     TrueCount);
1397       }
1398 
1399       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1400       // want to jump to the FalseBlock.
1401       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1402       // The counter tells us how often we evaluate RHS, and all of TrueCount
1403       // can be propagated to that branch.
1404       uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1405 
1406       ConditionalEvaluation eval(*this);
1407       {
1408         ApplyDebugLocation DL(*this, Cond);
1409         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1410         EmitBlock(LHSTrue);
1411       }
1412 
1413       incrementProfileCounter(CondBOp);
1414       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1415 
1416       // Any temporaries created here are conditional.
1417       eval.begin(*this);
1418       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1419       eval.end(*this);
1420 
1421       return;
1422     }
1423 
1424     if (CondBOp->getOpcode() == BO_LOr) {
1425       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1426       // folded if the case was simple enough.
1427       bool ConstantBool = false;
1428       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1429           !ConstantBool) {
1430         // br(0 || X) -> br(X).
1431         incrementProfileCounter(CondBOp);
1432         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1433                                     TrueCount);
1434       }
1435 
1436       // If we have "X || 0", simplify the code to use an uncond branch.
1437       // "X || 1" would have been constant folded to 1.
1438       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1439           !ConstantBool) {
1440         // br(X || 0) -> br(X).
1441         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1442                                     TrueCount);
1443       }
1444 
1445       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1446       // want to jump to the TrueBlock.
1447       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1448       // We have the count for entry to the RHS and for the whole expression
1449       // being true, so we can divy up True count between the short circuit and
1450       // the RHS.
1451       uint64_t LHSCount =
1452           getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1453       uint64_t RHSCount = TrueCount - LHSCount;
1454 
1455       ConditionalEvaluation eval(*this);
1456       {
1457         ApplyDebugLocation DL(*this, Cond);
1458         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1459         EmitBlock(LHSFalse);
1460       }
1461 
1462       incrementProfileCounter(CondBOp);
1463       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1464 
1465       // Any temporaries created here are conditional.
1466       eval.begin(*this);
1467       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1468 
1469       eval.end(*this);
1470 
1471       return;
1472     }
1473   }
1474 
1475   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1476     // br(!x, t, f) -> br(x, f, t)
1477     if (CondUOp->getOpcode() == UO_LNot) {
1478       // Negate the count.
1479       uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1480       // Negate the condition and swap the destination blocks.
1481       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1482                                   FalseCount);
1483     }
1484   }
1485 
1486   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1487     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1488     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1489     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1490 
1491     ConditionalEvaluation cond(*this);
1492     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1493                          getProfileCount(CondOp));
1494 
1495     // When computing PGO branch weights, we only know the overall count for
1496     // the true block. This code is essentially doing tail duplication of the
1497     // naive code-gen, introducing new edges for which counts are not
1498     // available. Divide the counts proportionally between the LHS and RHS of
1499     // the conditional operator.
1500     uint64_t LHSScaledTrueCount = 0;
1501     if (TrueCount) {
1502       double LHSRatio =
1503           getProfileCount(CondOp) / (double)getCurrentProfileCount();
1504       LHSScaledTrueCount = TrueCount * LHSRatio;
1505     }
1506 
1507     cond.begin(*this);
1508     EmitBlock(LHSBlock);
1509     incrementProfileCounter(CondOp);
1510     {
1511       ApplyDebugLocation DL(*this, Cond);
1512       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1513                            LHSScaledTrueCount);
1514     }
1515     cond.end(*this);
1516 
1517     cond.begin(*this);
1518     EmitBlock(RHSBlock);
1519     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1520                          TrueCount - LHSScaledTrueCount);
1521     cond.end(*this);
1522 
1523     return;
1524   }
1525 
1526   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1527     // Conditional operator handling can give us a throw expression as a
1528     // condition for a case like:
1529     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1530     // Fold this to:
1531     //   br(c, throw x, br(y, t, f))
1532     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1533     return;
1534   }
1535 
1536   // If the branch has a condition wrapped by __builtin_unpredictable,
1537   // create metadata that specifies that the branch is unpredictable.
1538   // Don't bother if not optimizing because that metadata would not be used.
1539   llvm::MDNode *Unpredictable = nullptr;
1540   auto *Call = dyn_cast<CallExpr>(Cond->IgnoreImpCasts());
1541   if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1542     auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1543     if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1544       llvm::MDBuilder MDHelper(getLLVMContext());
1545       Unpredictable = MDHelper.createUnpredictable();
1546     }
1547   }
1548 
1549   // Create branch weights based on the number of times we get here and the
1550   // number of times the condition should be true.
1551   uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1552   llvm::MDNode *Weights =
1553       createProfileWeights(TrueCount, CurrentCount - TrueCount);
1554 
1555   // Emit the code with the fully general case.
1556   llvm::Value *CondV;
1557   {
1558     ApplyDebugLocation DL(*this, Cond);
1559     CondV = EvaluateExprAsBool(Cond);
1560   }
1561   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1562 }
1563 
1564 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1565 /// specified stmt yet.
1566 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1567   CGM.ErrorUnsupported(S, Type);
1568 }
1569 
1570 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1571 /// variable-length array whose elements have a non-zero bit-pattern.
1572 ///
1573 /// \param baseType the inner-most element type of the array
1574 /// \param src - a char* pointing to the bit-pattern for a single
1575 /// base element of the array
1576 /// \param sizeInChars - the total size of the VLA, in chars
1577 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1578                                Address dest, Address src,
1579                                llvm::Value *sizeInChars) {
1580   CGBuilderTy &Builder = CGF.Builder;
1581 
1582   CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1583   llvm::Value *baseSizeInChars
1584     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1585 
1586   Address begin =
1587     Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1588   llvm::Value *end =
1589     Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1590 
1591   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1592   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1593   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1594 
1595   // Make a loop over the VLA.  C99 guarantees that the VLA element
1596   // count must be nonzero.
1597   CGF.EmitBlock(loopBB);
1598 
1599   llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1600   cur->addIncoming(begin.getPointer(), originBB);
1601 
1602   CharUnits curAlign =
1603     dest.getAlignment().alignmentOfArrayElement(baseSize);
1604 
1605   // memcpy the individual element bit-pattern.
1606   Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1607                        /*volatile*/ false);
1608 
1609   // Go to the next element.
1610   llvm::Value *next =
1611     Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1612 
1613   // Leave if that's the end of the VLA.
1614   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1615   Builder.CreateCondBr(done, contBB, loopBB);
1616   cur->addIncoming(next, loopBB);
1617 
1618   CGF.EmitBlock(contBB);
1619 }
1620 
1621 void
1622 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1623   // Ignore empty classes in C++.
1624   if (getLangOpts().CPlusPlus) {
1625     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1626       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1627         return;
1628     }
1629   }
1630 
1631   // Cast the dest ptr to the appropriate i8 pointer type.
1632   if (DestPtr.getElementType() != Int8Ty)
1633     DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1634 
1635   // Get size and alignment info for this aggregate.
1636   CharUnits size = getContext().getTypeSizeInChars(Ty);
1637 
1638   llvm::Value *SizeVal;
1639   const VariableArrayType *vla;
1640 
1641   // Don't bother emitting a zero-byte memset.
1642   if (size.isZero()) {
1643     // But note that getTypeInfo returns 0 for a VLA.
1644     if (const VariableArrayType *vlaType =
1645           dyn_cast_or_null<VariableArrayType>(
1646                                           getContext().getAsArrayType(Ty))) {
1647       auto VlaSize = getVLASize(vlaType);
1648       SizeVal = VlaSize.NumElts;
1649       CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type);
1650       if (!eltSize.isOne())
1651         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1652       vla = vlaType;
1653     } else {
1654       return;
1655     }
1656   } else {
1657     SizeVal = CGM.getSize(size);
1658     vla = nullptr;
1659   }
1660 
1661   // If the type contains a pointer to data member we can't memset it to zero.
1662   // Instead, create a null constant and copy it to the destination.
1663   // TODO: there are other patterns besides zero that we can usefully memset,
1664   // like -1, which happens to be the pattern used by member-pointers.
1665   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1666     // For a VLA, emit a single element, then splat that over the VLA.
1667     if (vla) Ty = getContext().getBaseElementType(vla);
1668 
1669     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1670 
1671     llvm::GlobalVariable *NullVariable =
1672       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1673                                /*isConstant=*/true,
1674                                llvm::GlobalVariable::PrivateLinkage,
1675                                NullConstant, Twine());
1676     CharUnits NullAlign = DestPtr.getAlignment();
1677     NullVariable->setAlignment(NullAlign.getQuantity());
1678     Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1679                    NullAlign);
1680 
1681     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1682 
1683     // Get and call the appropriate llvm.memcpy overload.
1684     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1685     return;
1686   }
1687 
1688   // Otherwise, just memset the whole thing to zero.  This is legal
1689   // because in LLVM, all default initializers (other than the ones we just
1690   // handled above) are guaranteed to have a bit pattern of all zeros.
1691   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1692 }
1693 
1694 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1695   // Make sure that there is a block for the indirect goto.
1696   if (!IndirectBranch)
1697     GetIndirectGotoBlock();
1698 
1699   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1700 
1701   // Make sure the indirect branch includes all of the address-taken blocks.
1702   IndirectBranch->addDestination(BB);
1703   return llvm::BlockAddress::get(CurFn, BB);
1704 }
1705 
1706 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1707   // If we already made the indirect branch for indirect goto, return its block.
1708   if (IndirectBranch) return IndirectBranch->getParent();
1709 
1710   CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1711 
1712   // Create the PHI node that indirect gotos will add entries to.
1713   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1714                                               "indirect.goto.dest");
1715 
1716   // Create the indirect branch instruction.
1717   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1718   return IndirectBranch->getParent();
1719 }
1720 
1721 /// Computes the length of an array in elements, as well as the base
1722 /// element type and a properly-typed first element pointer.
1723 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1724                                               QualType &baseType,
1725                                               Address &addr) {
1726   const ArrayType *arrayType = origArrayType;
1727 
1728   // If it's a VLA, we have to load the stored size.  Note that
1729   // this is the size of the VLA in bytes, not its size in elements.
1730   llvm::Value *numVLAElements = nullptr;
1731   if (isa<VariableArrayType>(arrayType)) {
1732     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts;
1733 
1734     // Walk into all VLAs.  This doesn't require changes to addr,
1735     // which has type T* where T is the first non-VLA element type.
1736     do {
1737       QualType elementType = arrayType->getElementType();
1738       arrayType = getContext().getAsArrayType(elementType);
1739 
1740       // If we only have VLA components, 'addr' requires no adjustment.
1741       if (!arrayType) {
1742         baseType = elementType;
1743         return numVLAElements;
1744       }
1745     } while (isa<VariableArrayType>(arrayType));
1746 
1747     // We get out here only if we find a constant array type
1748     // inside the VLA.
1749   }
1750 
1751   // We have some number of constant-length arrays, so addr should
1752   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1753   // down to the first element of addr.
1754   SmallVector<llvm::Value*, 8> gepIndices;
1755 
1756   // GEP down to the array type.
1757   llvm::ConstantInt *zero = Builder.getInt32(0);
1758   gepIndices.push_back(zero);
1759 
1760   uint64_t countFromCLAs = 1;
1761   QualType eltType;
1762 
1763   llvm::ArrayType *llvmArrayType =
1764     dyn_cast<llvm::ArrayType>(addr.getElementType());
1765   while (llvmArrayType) {
1766     assert(isa<ConstantArrayType>(arrayType));
1767     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1768              == llvmArrayType->getNumElements());
1769 
1770     gepIndices.push_back(zero);
1771     countFromCLAs *= llvmArrayType->getNumElements();
1772     eltType = arrayType->getElementType();
1773 
1774     llvmArrayType =
1775       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1776     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1777     assert((!llvmArrayType || arrayType) &&
1778            "LLVM and Clang types are out-of-synch");
1779   }
1780 
1781   if (arrayType) {
1782     // From this point onwards, the Clang array type has been emitted
1783     // as some other type (probably a packed struct). Compute the array
1784     // size, and just emit the 'begin' expression as a bitcast.
1785     while (arrayType) {
1786       countFromCLAs *=
1787           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1788       eltType = arrayType->getElementType();
1789       arrayType = getContext().getAsArrayType(eltType);
1790     }
1791 
1792     llvm::Type *baseType = ConvertType(eltType);
1793     addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1794   } else {
1795     // Create the actual GEP.
1796     addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1797                                              gepIndices, "array.begin"),
1798                    addr.getAlignment());
1799   }
1800 
1801   baseType = eltType;
1802 
1803   llvm::Value *numElements
1804     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1805 
1806   // If we had any VLA dimensions, factor them in.
1807   if (numVLAElements)
1808     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1809 
1810   return numElements;
1811 }
1812 
1813 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) {
1814   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1815   assert(vla && "type was not a variable array type!");
1816   return getVLASize(vla);
1817 }
1818 
1819 CodeGenFunction::VlaSizePair
1820 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1821   // The number of elements so far; always size_t.
1822   llvm::Value *numElements = nullptr;
1823 
1824   QualType elementType;
1825   do {
1826     elementType = type->getElementType();
1827     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1828     assert(vlaSize && "no size for VLA!");
1829     assert(vlaSize->getType() == SizeTy);
1830 
1831     if (!numElements) {
1832       numElements = vlaSize;
1833     } else {
1834       // It's undefined behavior if this wraps around, so mark it that way.
1835       // FIXME: Teach -fsanitize=undefined to trap this.
1836       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1837     }
1838   } while ((type = getContext().getAsVariableArrayType(elementType)));
1839 
1840   return { numElements, elementType };
1841 }
1842 
1843 CodeGenFunction::VlaSizePair
1844 CodeGenFunction::getVLAElements1D(QualType type) {
1845   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1846   assert(vla && "type was not a variable array type!");
1847   return getVLAElements1D(vla);
1848 }
1849 
1850 CodeGenFunction::VlaSizePair
1851 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) {
1852   llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()];
1853   assert(VlaSize && "no size for VLA!");
1854   assert(VlaSize->getType() == SizeTy);
1855   return { VlaSize, Vla->getElementType() };
1856 }
1857 
1858 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1859   assert(type->isVariablyModifiedType() &&
1860          "Must pass variably modified type to EmitVLASizes!");
1861 
1862   EnsureInsertPoint();
1863 
1864   // We're going to walk down into the type and look for VLA
1865   // expressions.
1866   do {
1867     assert(type->isVariablyModifiedType());
1868 
1869     const Type *ty = type.getTypePtr();
1870     switch (ty->getTypeClass()) {
1871 
1872 #define TYPE(Class, Base)
1873 #define ABSTRACT_TYPE(Class, Base)
1874 #define NON_CANONICAL_TYPE(Class, Base)
1875 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1876 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1877 #include "clang/AST/TypeNodes.def"
1878       llvm_unreachable("unexpected dependent type!");
1879 
1880     // These types are never variably-modified.
1881     case Type::Builtin:
1882     case Type::Complex:
1883     case Type::Vector:
1884     case Type::ExtVector:
1885     case Type::Record:
1886     case Type::Enum:
1887     case Type::Elaborated:
1888     case Type::TemplateSpecialization:
1889     case Type::ObjCTypeParam:
1890     case Type::ObjCObject:
1891     case Type::ObjCInterface:
1892     case Type::ObjCObjectPointer:
1893       llvm_unreachable("type class is never variably-modified!");
1894 
1895     case Type::Adjusted:
1896       type = cast<AdjustedType>(ty)->getAdjustedType();
1897       break;
1898 
1899     case Type::Decayed:
1900       type = cast<DecayedType>(ty)->getPointeeType();
1901       break;
1902 
1903     case Type::Pointer:
1904       type = cast<PointerType>(ty)->getPointeeType();
1905       break;
1906 
1907     case Type::BlockPointer:
1908       type = cast<BlockPointerType>(ty)->getPointeeType();
1909       break;
1910 
1911     case Type::LValueReference:
1912     case Type::RValueReference:
1913       type = cast<ReferenceType>(ty)->getPointeeType();
1914       break;
1915 
1916     case Type::MemberPointer:
1917       type = cast<MemberPointerType>(ty)->getPointeeType();
1918       break;
1919 
1920     case Type::ConstantArray:
1921     case Type::IncompleteArray:
1922       // Losing element qualification here is fine.
1923       type = cast<ArrayType>(ty)->getElementType();
1924       break;
1925 
1926     case Type::VariableArray: {
1927       // Losing element qualification here is fine.
1928       const VariableArrayType *vat = cast<VariableArrayType>(ty);
1929 
1930       // Unknown size indication requires no size computation.
1931       // Otherwise, evaluate and record it.
1932       if (const Expr *size = vat->getSizeExpr()) {
1933         // It's possible that we might have emitted this already,
1934         // e.g. with a typedef and a pointer to it.
1935         llvm::Value *&entry = VLASizeMap[size];
1936         if (!entry) {
1937           llvm::Value *Size = EmitScalarExpr(size);
1938 
1939           // C11 6.7.6.2p5:
1940           //   If the size is an expression that is not an integer constant
1941           //   expression [...] each time it is evaluated it shall have a value
1942           //   greater than zero.
1943           if (SanOpts.has(SanitizerKind::VLABound) &&
1944               size->getType()->isSignedIntegerType()) {
1945             SanitizerScope SanScope(this);
1946             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1947             llvm::Constant *StaticArgs[] = {
1948                 EmitCheckSourceLocation(size->getBeginLoc()),
1949                 EmitCheckTypeDescriptor(size->getType())};
1950             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1951                                      SanitizerKind::VLABound),
1952                       SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
1953           }
1954 
1955           // Always zexting here would be wrong if it weren't
1956           // undefined behavior to have a negative bound.
1957           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1958         }
1959       }
1960       type = vat->getElementType();
1961       break;
1962     }
1963 
1964     case Type::FunctionProto:
1965     case Type::FunctionNoProto:
1966       type = cast<FunctionType>(ty)->getReturnType();
1967       break;
1968 
1969     case Type::Paren:
1970     case Type::TypeOf:
1971     case Type::UnaryTransform:
1972     case Type::Attributed:
1973     case Type::SubstTemplateTypeParm:
1974     case Type::PackExpansion:
1975     case Type::MacroQualified:
1976       // Keep walking after single level desugaring.
1977       type = type.getSingleStepDesugaredType(getContext());
1978       break;
1979 
1980     case Type::Typedef:
1981     case Type::Decltype:
1982     case Type::Auto:
1983     case Type::DeducedTemplateSpecialization:
1984       // Stop walking: nothing to do.
1985       return;
1986 
1987     case Type::TypeOfExpr:
1988       // Stop walking: emit typeof expression.
1989       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1990       return;
1991 
1992     case Type::Atomic:
1993       type = cast<AtomicType>(ty)->getValueType();
1994       break;
1995 
1996     case Type::Pipe:
1997       type = cast<PipeType>(ty)->getElementType();
1998       break;
1999     }
2000   } while (type->isVariablyModifiedType());
2001 }
2002 
2003 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
2004   if (getContext().getBuiltinVaListType()->isArrayType())
2005     return EmitPointerWithAlignment(E);
2006   return EmitLValue(E).getAddress();
2007 }
2008 
2009 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
2010   return EmitLValue(E).getAddress();
2011 }
2012 
2013 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2014                                               const APValue &Init) {
2015   assert(Init.hasValue() && "Invalid DeclRefExpr initializer!");
2016   if (CGDebugInfo *Dbg = getDebugInfo())
2017     if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2018       Dbg->EmitGlobalVariable(E->getDecl(), Init);
2019 }
2020 
2021 CodeGenFunction::PeepholeProtection
2022 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2023   // At the moment, the only aggressive peephole we do in IR gen
2024   // is trunc(zext) folding, but if we add more, we can easily
2025   // extend this protection.
2026 
2027   if (!rvalue.isScalar()) return PeepholeProtection();
2028   llvm::Value *value = rvalue.getScalarVal();
2029   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2030 
2031   // Just make an extra bitcast.
2032   assert(HaveInsertPoint());
2033   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2034                                                   Builder.GetInsertBlock());
2035 
2036   PeepholeProtection protection;
2037   protection.Inst = inst;
2038   return protection;
2039 }
2040 
2041 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2042   if (!protection.Inst) return;
2043 
2044   // In theory, we could try to duplicate the peepholes now, but whatever.
2045   protection.Inst->eraseFromParent();
2046 }
2047 
2048 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2049                                               QualType Ty, SourceLocation Loc,
2050                                               SourceLocation AssumptionLoc,
2051                                               llvm::Value *Alignment,
2052                                               llvm::Value *OffsetValue) {
2053   llvm::Value *TheCheck;
2054   llvm::Instruction *Assumption = Builder.CreateAlignmentAssumption(
2055       CGM.getDataLayout(), PtrValue, Alignment, OffsetValue, &TheCheck);
2056   if (SanOpts.has(SanitizerKind::Alignment)) {
2057     EmitAlignmentAssumptionCheck(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
2058                                  OffsetValue, TheCheck, Assumption);
2059   }
2060 }
2061 
2062 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2063                                               QualType Ty, SourceLocation Loc,
2064                                               SourceLocation AssumptionLoc,
2065                                               unsigned Alignment,
2066                                               llvm::Value *OffsetValue) {
2067   llvm::Value *TheCheck;
2068   llvm::Instruction *Assumption = Builder.CreateAlignmentAssumption(
2069       CGM.getDataLayout(), PtrValue, Alignment, OffsetValue, &TheCheck);
2070   if (SanOpts.has(SanitizerKind::Alignment)) {
2071     llvm::Value *AlignmentVal = llvm::ConstantInt::get(IntPtrTy, Alignment);
2072     EmitAlignmentAssumptionCheck(PtrValue, Ty, Loc, AssumptionLoc, AlignmentVal,
2073                                  OffsetValue, TheCheck, Assumption);
2074   }
2075 }
2076 
2077 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2078                                               const Expr *E,
2079                                               SourceLocation AssumptionLoc,
2080                                               unsigned Alignment,
2081                                               llvm::Value *OffsetValue) {
2082   if (auto *CE = dyn_cast<CastExpr>(E))
2083     E = CE->getSubExprAsWritten();
2084   QualType Ty = E->getType();
2085   SourceLocation Loc = E->getExprLoc();
2086 
2087   EmitAlignmentAssumption(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
2088                           OffsetValue);
2089 }
2090 
2091 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Function *AnnotationFn,
2092                                                  llvm::Value *AnnotatedVal,
2093                                                  StringRef AnnotationStr,
2094                                                  SourceLocation Location) {
2095   llvm::Value *Args[4] = {
2096     AnnotatedVal,
2097     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2098     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2099     CGM.EmitAnnotationLineNo(Location)
2100   };
2101   return Builder.CreateCall(AnnotationFn, Args);
2102 }
2103 
2104 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2105   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2106   // FIXME We create a new bitcast for every annotation because that's what
2107   // llvm-gcc was doing.
2108   for (const auto *I : D->specific_attrs<AnnotateAttr>())
2109     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2110                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2111                        I->getAnnotation(), D->getLocation());
2112 }
2113 
2114 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2115                                               Address Addr) {
2116   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2117   llvm::Value *V = Addr.getPointer();
2118   llvm::Type *VTy = V->getType();
2119   llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2120                                     CGM.Int8PtrTy);
2121 
2122   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2123     // FIXME Always emit the cast inst so we can differentiate between
2124     // annotation on the first field of a struct and annotation on the struct
2125     // itself.
2126     if (VTy != CGM.Int8PtrTy)
2127       V = Builder.CreateBitCast(V, CGM.Int8PtrTy);
2128     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2129     V = Builder.CreateBitCast(V, VTy);
2130   }
2131 
2132   return Address(V, Addr.getAlignment());
2133 }
2134 
2135 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2136 
2137 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2138     : CGF(CGF) {
2139   assert(!CGF->IsSanitizerScope);
2140   CGF->IsSanitizerScope = true;
2141 }
2142 
2143 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2144   CGF->IsSanitizerScope = false;
2145 }
2146 
2147 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2148                                    const llvm::Twine &Name,
2149                                    llvm::BasicBlock *BB,
2150                                    llvm::BasicBlock::iterator InsertPt) const {
2151   LoopStack.InsertHelper(I);
2152   if (IsSanitizerScope)
2153     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2154 }
2155 
2156 void CGBuilderInserter::InsertHelper(
2157     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2158     llvm::BasicBlock::iterator InsertPt) const {
2159   llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2160   if (CGF)
2161     CGF->InsertHelper(I, Name, BB, InsertPt);
2162 }
2163 
2164 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2165                                 CodeGenModule &CGM, const FunctionDecl *FD,
2166                                 std::string &FirstMissing) {
2167   // If there aren't any required features listed then go ahead and return.
2168   if (ReqFeatures.empty())
2169     return false;
2170 
2171   // Now build up the set of caller features and verify that all the required
2172   // features are there.
2173   llvm::StringMap<bool> CallerFeatureMap;
2174   CGM.getFunctionFeatureMap(CallerFeatureMap, GlobalDecl().getWithDecl(FD));
2175 
2176   // If we have at least one of the features in the feature list return
2177   // true, otherwise return false.
2178   return std::all_of(
2179       ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2180         SmallVector<StringRef, 1> OrFeatures;
2181         Feature.split(OrFeatures, '|');
2182         return llvm::any_of(OrFeatures, [&](StringRef Feature) {
2183           if (!CallerFeatureMap.lookup(Feature)) {
2184             FirstMissing = Feature.str();
2185             return false;
2186           }
2187           return true;
2188         });
2189       });
2190 }
2191 
2192 // Emits an error if we don't have a valid set of target features for the
2193 // called function.
2194 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2195                                           const FunctionDecl *TargetDecl) {
2196   return checkTargetFeatures(E->getBeginLoc(), TargetDecl);
2197 }
2198 
2199 // Emits an error if we don't have a valid set of target features for the
2200 // called function.
2201 void CodeGenFunction::checkTargetFeatures(SourceLocation Loc,
2202                                           const FunctionDecl *TargetDecl) {
2203   // Early exit if this is an indirect call.
2204   if (!TargetDecl)
2205     return;
2206 
2207   // Get the current enclosing function if it exists. If it doesn't
2208   // we can't check the target features anyhow.
2209   const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2210   if (!FD)
2211     return;
2212 
2213   // Grab the required features for the call. For a builtin this is listed in
2214   // the td file with the default cpu, for an always_inline function this is any
2215   // listed cpu and any listed features.
2216   unsigned BuiltinID = TargetDecl->getBuiltinID();
2217   std::string MissingFeature;
2218   if (BuiltinID) {
2219     SmallVector<StringRef, 1> ReqFeatures;
2220     const char *FeatureList =
2221         CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2222     // Return if the builtin doesn't have any required features.
2223     if (!FeatureList || StringRef(FeatureList) == "")
2224       return;
2225     StringRef(FeatureList).split(ReqFeatures, ',');
2226     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2227       CGM.getDiags().Report(Loc, diag::err_builtin_needs_feature)
2228           << TargetDecl->getDeclName()
2229           << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2230 
2231   } else if (TargetDecl->hasAttr<TargetAttr>() ||
2232              TargetDecl->hasAttr<CPUSpecificAttr>()) {
2233     // Get the required features for the callee.
2234 
2235     const TargetAttr *TD = TargetDecl->getAttr<TargetAttr>();
2236     TargetAttr::ParsedTargetAttr ParsedAttr = CGM.filterFunctionTargetAttrs(TD);
2237 
2238     SmallVector<StringRef, 1> ReqFeatures;
2239     llvm::StringMap<bool> CalleeFeatureMap;
2240     CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2241 
2242     for (const auto &F : ParsedAttr.Features) {
2243       if (F[0] == '+' && CalleeFeatureMap.lookup(F.substr(1)))
2244         ReqFeatures.push_back(StringRef(F).substr(1));
2245     }
2246 
2247     for (const auto &F : CalleeFeatureMap) {
2248       // Only positive features are "required".
2249       if (F.getValue())
2250         ReqFeatures.push_back(F.getKey());
2251     }
2252     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2253       CGM.getDiags().Report(Loc, diag::err_function_needs_feature)
2254           << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2255   }
2256 }
2257 
2258 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2259   if (!CGM.getCodeGenOpts().SanitizeStats)
2260     return;
2261 
2262   llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2263   IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2264   CGM.getSanStats().create(IRB, SSK);
2265 }
2266 
2267 llvm::Value *
2268 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
2269   llvm::Value *Condition = nullptr;
2270 
2271   if (!RO.Conditions.Architecture.empty())
2272     Condition = EmitX86CpuIs(RO.Conditions.Architecture);
2273 
2274   if (!RO.Conditions.Features.empty()) {
2275     llvm::Value *FeatureCond = EmitX86CpuSupports(RO.Conditions.Features);
2276     Condition =
2277         Condition ? Builder.CreateAnd(Condition, FeatureCond) : FeatureCond;
2278   }
2279   return Condition;
2280 }
2281 
2282 static void CreateMultiVersionResolverReturn(CodeGenModule &CGM,
2283                                              llvm::Function *Resolver,
2284                                              CGBuilderTy &Builder,
2285                                              llvm::Function *FuncToReturn,
2286                                              bool SupportsIFunc) {
2287   if (SupportsIFunc) {
2288     Builder.CreateRet(FuncToReturn);
2289     return;
2290   }
2291 
2292   llvm::SmallVector<llvm::Value *, 10> Args;
2293   llvm::for_each(Resolver->args(),
2294                  [&](llvm::Argument &Arg) { Args.push_back(&Arg); });
2295 
2296   llvm::CallInst *Result = Builder.CreateCall(FuncToReturn, Args);
2297   Result->setTailCallKind(llvm::CallInst::TCK_MustTail);
2298 
2299   if (Resolver->getReturnType()->isVoidTy())
2300     Builder.CreateRetVoid();
2301   else
2302     Builder.CreateRet(Result);
2303 }
2304 
2305 void CodeGenFunction::EmitMultiVersionResolver(
2306     llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
2307   assert((getContext().getTargetInfo().getTriple().getArch() ==
2308               llvm::Triple::x86 ||
2309           getContext().getTargetInfo().getTriple().getArch() ==
2310               llvm::Triple::x86_64) &&
2311          "Only implemented for x86 targets");
2312 
2313   bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc();
2314 
2315   // Main function's basic block.
2316   llvm::BasicBlock *CurBlock = createBasicBlock("resolver_entry", Resolver);
2317   Builder.SetInsertPoint(CurBlock);
2318   EmitX86CpuInit();
2319 
2320   for (const MultiVersionResolverOption &RO : Options) {
2321     Builder.SetInsertPoint(CurBlock);
2322     llvm::Value *Condition = FormResolverCondition(RO);
2323 
2324     // The 'default' or 'generic' case.
2325     if (!Condition) {
2326       assert(&RO == Options.end() - 1 &&
2327              "Default or Generic case must be last");
2328       CreateMultiVersionResolverReturn(CGM, Resolver, Builder, RO.Function,
2329                                        SupportsIFunc);
2330       return;
2331     }
2332 
2333     llvm::BasicBlock *RetBlock = createBasicBlock("resolver_return", Resolver);
2334     CGBuilderTy RetBuilder(*this, RetBlock);
2335     CreateMultiVersionResolverReturn(CGM, Resolver, RetBuilder, RO.Function,
2336                                      SupportsIFunc);
2337     CurBlock = createBasicBlock("resolver_else", Resolver);
2338     Builder.CreateCondBr(Condition, RetBlock, CurBlock);
2339   }
2340 
2341   // If no generic/default, emit an unreachable.
2342   Builder.SetInsertPoint(CurBlock);
2343   llvm::CallInst *TrapCall = EmitTrapCall(llvm::Intrinsic::trap);
2344   TrapCall->setDoesNotReturn();
2345   TrapCall->setDoesNotThrow();
2346   Builder.CreateUnreachable();
2347   Builder.ClearInsertionPoint();
2348 }
2349 
2350 // Loc - where the diagnostic will point, where in the source code this
2351 //  alignment has failed.
2352 // SecondaryLoc - if present (will be present if sufficiently different from
2353 //  Loc), the diagnostic will additionally point a "Note:" to this location.
2354 //  It should be the location where the __attribute__((assume_aligned))
2355 //  was written e.g.
2356 void CodeGenFunction::EmitAlignmentAssumptionCheck(
2357     llvm::Value *Ptr, QualType Ty, SourceLocation Loc,
2358     SourceLocation SecondaryLoc, llvm::Value *Alignment,
2359     llvm::Value *OffsetValue, llvm::Value *TheCheck,
2360     llvm::Instruction *Assumption) {
2361   assert(Assumption && isa<llvm::CallInst>(Assumption) &&
2362          cast<llvm::CallInst>(Assumption)->getCalledValue() ==
2363              llvm::Intrinsic::getDeclaration(
2364                  Builder.GetInsertBlock()->getParent()->getParent(),
2365                  llvm::Intrinsic::assume) &&
2366          "Assumption should be a call to llvm.assume().");
2367   assert(&(Builder.GetInsertBlock()->back()) == Assumption &&
2368          "Assumption should be the last instruction of the basic block, "
2369          "since the basic block is still being generated.");
2370 
2371   if (!SanOpts.has(SanitizerKind::Alignment))
2372     return;
2373 
2374   // Don't check pointers to volatile data. The behavior here is implementation-
2375   // defined.
2376   if (Ty->getPointeeType().isVolatileQualified())
2377     return;
2378 
2379   // We need to temorairly remove the assumption so we can insert the
2380   // sanitizer check before it, else the check will be dropped by optimizations.
2381   Assumption->removeFromParent();
2382 
2383   {
2384     SanitizerScope SanScope(this);
2385 
2386     if (!OffsetValue)
2387       OffsetValue = Builder.getInt1(0); // no offset.
2388 
2389     llvm::Constant *StaticData[] = {EmitCheckSourceLocation(Loc),
2390                                     EmitCheckSourceLocation(SecondaryLoc),
2391                                     EmitCheckTypeDescriptor(Ty)};
2392     llvm::Value *DynamicData[] = {EmitCheckValue(Ptr),
2393                                   EmitCheckValue(Alignment),
2394                                   EmitCheckValue(OffsetValue)};
2395     EmitCheck({std::make_pair(TheCheck, SanitizerKind::Alignment)},
2396               SanitizerHandler::AlignmentAssumption, StaticData, DynamicData);
2397   }
2398 
2399   // We are now in the (new, empty) "cont" basic block.
2400   // Reintroduce the assumption.
2401   Builder.Insert(Assumption);
2402   // FIXME: Assumption still has it's original basic block as it's Parent.
2403 }
2404 
2405 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2406   if (CGDebugInfo *DI = getDebugInfo())
2407     return DI->SourceLocToDebugLoc(Location);
2408 
2409   return llvm::DebugLoc();
2410 }
2411