xref: /freebsd/contrib/llvm-project/clang/lib/CodeGen/CodeGenFunction.cpp (revision 6966ac055c3b7a39266fb982493330df7a097997)
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   // Apply xray attributes to the function (as a string, for now)
736   if (D) {
737     if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
738       if (CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
739               XRayInstrKind::Function)) {
740         if (XRayAttr->alwaysXRayInstrument() && ShouldXRayInstrumentFunction())
741           Fn->addFnAttr("function-instrument", "xray-always");
742         if (XRayAttr->neverXRayInstrument())
743           Fn->addFnAttr("function-instrument", "xray-never");
744         if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>())
745           if (ShouldXRayInstrumentFunction())
746             Fn->addFnAttr("xray-log-args",
747                           llvm::utostr(LogArgs->getArgumentCount()));
748       }
749     } else {
750       if (ShouldXRayInstrumentFunction() && !CGM.imbueXRayAttrs(Fn, Loc))
751         Fn->addFnAttr(
752             "xray-instruction-threshold",
753             llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
754     }
755   }
756 
757   // Add no-jump-tables value.
758   Fn->addFnAttr("no-jump-tables",
759                 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
760 
761   // Add profile-sample-accurate value.
762   if (CGM.getCodeGenOpts().ProfileSampleAccurate)
763     Fn->addFnAttr("profile-sample-accurate");
764 
765   if (getLangOpts().OpenCL) {
766     // Add metadata for a kernel function.
767     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
768       EmitOpenCLKernelMetadata(FD, Fn);
769   }
770 
771   // If we are checking function types, emit a function type signature as
772   // prologue data.
773   if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
774     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
775       if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
776         // Remove any (C++17) exception specifications, to allow calling e.g. a
777         // noexcept function through a non-noexcept pointer.
778         auto ProtoTy =
779           getContext().getFunctionTypeWithExceptionSpec(FD->getType(),
780                                                         EST_None);
781         llvm::Constant *FTRTTIConst =
782             CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true);
783         llvm::Constant *FTRTTIConstEncoded =
784             EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
785         llvm::Constant *PrologueStructElems[] = {PrologueSig,
786                                                  FTRTTIConstEncoded};
787         llvm::Constant *PrologueStructConst =
788             llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
789         Fn->setPrologueData(PrologueStructConst);
790       }
791     }
792   }
793 
794   // If we're checking nullability, we need to know whether we can check the
795   // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
796   if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
797     auto Nullability = FnRetTy->getNullability(getContext());
798     if (Nullability && *Nullability == NullabilityKind::NonNull) {
799       if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
800             CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
801         RetValNullabilityPrecondition =
802             llvm::ConstantInt::getTrue(getLLVMContext());
803     }
804   }
805 
806   // If we're in C++ mode and the function name is "main", it is guaranteed
807   // to be norecurse by the standard (3.6.1.3 "The function main shall not be
808   // used within a program").
809   if (getLangOpts().CPlusPlus)
810     if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
811       if (FD->isMain())
812         Fn->addFnAttr(llvm::Attribute::NoRecurse);
813 
814   // If a custom alignment is used, force realigning to this alignment on
815   // any main function which certainly will need it.
816   if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
817     if ((FD->isMain() || FD->isMSVCRTEntryPoint()) &&
818         CGM.getCodeGenOpts().StackAlignment)
819       Fn->addFnAttr("stackrealign");
820 
821   llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
822 
823   // Create a marker to make it easy to insert allocas into the entryblock
824   // later.  Don't create this with the builder, because we don't want it
825   // folded.
826   llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
827   AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
828 
829   ReturnBlock = getJumpDestInCurrentScope("return");
830 
831   Builder.SetInsertPoint(EntryBB);
832 
833   // If we're checking the return value, allocate space for a pointer to a
834   // precise source location of the checked return statement.
835   if (requiresReturnValueCheck()) {
836     ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
837     InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
838   }
839 
840   // Emit subprogram debug descriptor.
841   if (CGDebugInfo *DI = getDebugInfo()) {
842     // Reconstruct the type from the argument list so that implicit parameters,
843     // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
844     // convention.
845     CallingConv CC = CallingConv::CC_C;
846     if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
847       if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
848         CC = SrcFnTy->getCallConv();
849     SmallVector<QualType, 16> ArgTypes;
850     for (const VarDecl *VD : Args)
851       ArgTypes.push_back(VD->getType());
852     QualType FnType = getContext().getFunctionType(
853         RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
854     DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, CurFuncIsThunk,
855                           Builder);
856   }
857 
858   if (ShouldInstrumentFunction()) {
859     if (CGM.getCodeGenOpts().InstrumentFunctions)
860       CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
861     if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
862       CurFn->addFnAttr("instrument-function-entry-inlined",
863                        "__cyg_profile_func_enter");
864     if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
865       CurFn->addFnAttr("instrument-function-entry-inlined",
866                        "__cyg_profile_func_enter_bare");
867   }
868 
869   // Since emitting the mcount call here impacts optimizations such as function
870   // inlining, we just add an attribute to insert a mcount call in backend.
871   // The attribute "counting-function" is set to mcount function name which is
872   // architecture dependent.
873   if (CGM.getCodeGenOpts().InstrumentForProfiling) {
874     // Calls to fentry/mcount should not be generated if function has
875     // the no_instrument_function attribute.
876     if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
877       if (CGM.getCodeGenOpts().CallFEntry)
878         Fn->addFnAttr("fentry-call", "true");
879       else {
880         Fn->addFnAttr("instrument-function-entry-inlined",
881                       getTarget().getMCountName());
882       }
883     }
884   }
885 
886   if (RetTy->isVoidType()) {
887     // Void type; nothing to return.
888     ReturnValue = Address::invalid();
889 
890     // Count the implicit return.
891     if (!endsWithReturn(D))
892       ++NumReturnExprs;
893   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect) {
894     // Indirect return; emit returned value directly into sret slot.
895     // This reduces code size, and affects correctness in C++.
896     auto AI = CurFn->arg_begin();
897     if (CurFnInfo->getReturnInfo().isSRetAfterThis())
898       ++AI;
899     ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
900     if (!CurFnInfo->getReturnInfo().getIndirectByVal()) {
901       ReturnValuePointer =
902           CreateDefaultAlignTempAlloca(Int8PtrTy, "result.ptr");
903       Builder.CreateStore(Builder.CreatePointerBitCastOrAddrSpaceCast(
904                               ReturnValue.getPointer(), Int8PtrTy),
905                           ReturnValuePointer);
906     }
907   } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
908              !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
909     // Load the sret pointer from the argument struct and return into that.
910     unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
911     llvm::Function::arg_iterator EI = CurFn->arg_end();
912     --EI;
913     llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
914     ReturnValuePointer = Address(Addr, getPointerAlign());
915     Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
916     ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
917   } else {
918     ReturnValue = CreateIRTemp(RetTy, "retval");
919 
920     // Tell the epilog emitter to autorelease the result.  We do this
921     // now so that various specialized functions can suppress it
922     // during their IR-generation.
923     if (getLangOpts().ObjCAutoRefCount &&
924         !CurFnInfo->isReturnsRetained() &&
925         RetTy->isObjCRetainableType())
926       AutoreleaseResult = true;
927   }
928 
929   EmitStartEHSpec(CurCodeDecl);
930 
931   PrologueCleanupDepth = EHStack.stable_begin();
932 
933   // Emit OpenMP specific initialization of the device functions.
934   if (getLangOpts().OpenMP && CurCodeDecl)
935     CGM.getOpenMPRuntime().emitFunctionProlog(*this, CurCodeDecl);
936 
937   EmitFunctionProlog(*CurFnInfo, CurFn, Args);
938 
939   if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
940     CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
941     const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
942     if (MD->getParent()->isLambda() &&
943         MD->getOverloadedOperator() == OO_Call) {
944       // We're in a lambda; figure out the captures.
945       MD->getParent()->getCaptureFields(LambdaCaptureFields,
946                                         LambdaThisCaptureField);
947       if (LambdaThisCaptureField) {
948         // If the lambda captures the object referred to by '*this' - either by
949         // value or by reference, make sure CXXThisValue points to the correct
950         // object.
951 
952         // Get the lvalue for the field (which is a copy of the enclosing object
953         // or contains the address of the enclosing object).
954         LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
955         if (!LambdaThisCaptureField->getType()->isPointerType()) {
956           // If the enclosing object was captured by value, just use its address.
957           CXXThisValue = ThisFieldLValue.getAddress().getPointer();
958         } else {
959           // Load the lvalue pointed to by the field, since '*this' was captured
960           // by reference.
961           CXXThisValue =
962               EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
963         }
964       }
965       for (auto *FD : MD->getParent()->fields()) {
966         if (FD->hasCapturedVLAType()) {
967           auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
968                                            SourceLocation()).getScalarVal();
969           auto VAT = FD->getCapturedVLAType();
970           VLASizeMap[VAT->getSizeExpr()] = ExprArg;
971         }
972       }
973     } else {
974       // Not in a lambda; just use 'this' from the method.
975       // FIXME: Should we generate a new load for each use of 'this'?  The
976       // fast register allocator would be happier...
977       CXXThisValue = CXXABIThisValue;
978     }
979 
980     // Check the 'this' pointer once per function, if it's available.
981     if (CXXABIThisValue) {
982       SanitizerSet SkippedChecks;
983       SkippedChecks.set(SanitizerKind::ObjectSize, true);
984       QualType ThisTy = MD->getThisType();
985 
986       // If this is the call operator of a lambda with no capture-default, it
987       // may have a static invoker function, which may call this operator with
988       // a null 'this' pointer.
989       if (isLambdaCallOperator(MD) &&
990           MD->getParent()->getLambdaCaptureDefault() == LCD_None)
991         SkippedChecks.set(SanitizerKind::Null, true);
992 
993       EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
994                                                 : TCK_MemberCall,
995                     Loc, CXXABIThisValue, ThisTy,
996                     getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
997                     SkippedChecks);
998     }
999   }
1000 
1001   // If any of the arguments have a variably modified type, make sure to
1002   // emit the type size.
1003   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1004        i != e; ++i) {
1005     const VarDecl *VD = *i;
1006 
1007     // Dig out the type as written from ParmVarDecls; it's unclear whether
1008     // the standard (C99 6.9.1p10) requires this, but we're following the
1009     // precedent set by gcc.
1010     QualType Ty;
1011     if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1012       Ty = PVD->getOriginalType();
1013     else
1014       Ty = VD->getType();
1015 
1016     if (Ty->isVariablyModifiedType())
1017       EmitVariablyModifiedType(Ty);
1018   }
1019   // Emit a location at the end of the prologue.
1020   if (CGDebugInfo *DI = getDebugInfo())
1021     DI->EmitLocation(Builder, StartLoc);
1022 
1023   // TODO: Do we need to handle this in two places like we do with
1024   // target-features/target-cpu?
1025   if (CurFuncDecl)
1026     if (const auto *VecWidth = CurFuncDecl->getAttr<MinVectorWidthAttr>())
1027       LargestVectorWidth = VecWidth->getVectorWidth();
1028 }
1029 
1030 void CodeGenFunction::EmitFunctionBody(const Stmt *Body) {
1031   incrementProfileCounter(Body);
1032   if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1033     EmitCompoundStmtWithoutScope(*S);
1034   else
1035     EmitStmt(Body);
1036 }
1037 
1038 /// When instrumenting to collect profile data, the counts for some blocks
1039 /// such as switch cases need to not include the fall-through counts, so
1040 /// emit a branch around the instrumentation code. When not instrumenting,
1041 /// this just calls EmitBlock().
1042 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1043                                                const Stmt *S) {
1044   llvm::BasicBlock *SkipCountBB = nullptr;
1045   if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1046     // When instrumenting for profiling, the fallthrough to certain
1047     // statements needs to skip over the instrumentation code so that we
1048     // get an accurate count.
1049     SkipCountBB = createBasicBlock("skipcount");
1050     EmitBranch(SkipCountBB);
1051   }
1052   EmitBlock(BB);
1053   uint64_t CurrentCount = getCurrentProfileCount();
1054   incrementProfileCounter(S);
1055   setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1056   if (SkipCountBB)
1057     EmitBlock(SkipCountBB);
1058 }
1059 
1060 /// Tries to mark the given function nounwind based on the
1061 /// non-existence of any throwing calls within it.  We believe this is
1062 /// lightweight enough to do at -O0.
1063 static void TryMarkNoThrow(llvm::Function *F) {
1064   // LLVM treats 'nounwind' on a function as part of the type, so we
1065   // can't do this on functions that can be overwritten.
1066   if (F->isInterposable()) return;
1067 
1068   for (llvm::BasicBlock &BB : *F)
1069     for (llvm::Instruction &I : BB)
1070       if (I.mayThrow())
1071         return;
1072 
1073   F->setDoesNotThrow();
1074 }
1075 
1076 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1077                                                FunctionArgList &Args) {
1078   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1079   QualType ResTy = FD->getReturnType();
1080 
1081   const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1082   if (MD && MD->isInstance()) {
1083     if (CGM.getCXXABI().HasThisReturn(GD))
1084       ResTy = MD->getThisType();
1085     else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1086       ResTy = CGM.getContext().VoidPtrTy;
1087     CGM.getCXXABI().buildThisParam(*this, Args);
1088   }
1089 
1090   // The base version of an inheriting constructor whose constructed base is a
1091   // virtual base is not passed any arguments (because it doesn't actually call
1092   // the inherited constructor).
1093   bool PassedParams = true;
1094   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1095     if (auto Inherited = CD->getInheritedConstructor())
1096       PassedParams =
1097           getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1098 
1099   if (PassedParams) {
1100     for (auto *Param : FD->parameters()) {
1101       Args.push_back(Param);
1102       if (!Param->hasAttr<PassObjectSizeAttr>())
1103         continue;
1104 
1105       auto *Implicit = ImplicitParamDecl::Create(
1106           getContext(), Param->getDeclContext(), Param->getLocation(),
1107           /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1108       SizeArguments[Param] = Implicit;
1109       Args.push_back(Implicit);
1110     }
1111   }
1112 
1113   if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1114     CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1115 
1116   return ResTy;
1117 }
1118 
1119 static bool
1120 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1121                                              const ASTContext &Context) {
1122   QualType T = FD->getReturnType();
1123   // Avoid the optimization for functions that return a record type with a
1124   // trivial destructor or another trivially copyable type.
1125   if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1126     if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1127       return !ClassDecl->hasTrivialDestructor();
1128   }
1129   return !T.isTriviallyCopyableType(Context);
1130 }
1131 
1132 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1133                                    const CGFunctionInfo &FnInfo) {
1134   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1135   CurGD = GD;
1136 
1137   FunctionArgList Args;
1138   QualType ResTy = BuildFunctionArgList(GD, Args);
1139 
1140   // Check if we should generate debug info for this function.
1141   if (FD->hasAttr<NoDebugAttr>())
1142     DebugInfo = nullptr; // disable debug info indefinitely for this function
1143 
1144   // The function might not have a body if we're generating thunks for a
1145   // function declaration.
1146   SourceRange BodyRange;
1147   if (Stmt *Body = FD->getBody())
1148     BodyRange = Body->getSourceRange();
1149   else
1150     BodyRange = FD->getLocation();
1151   CurEHLocation = BodyRange.getEnd();
1152 
1153   // Use the location of the start of the function to determine where
1154   // the function definition is located. By default use the location
1155   // of the declaration as the location for the subprogram. A function
1156   // may lack a declaration in the source code if it is created by code
1157   // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1158   SourceLocation Loc = FD->getLocation();
1159 
1160   // If this is a function specialization then use the pattern body
1161   // as the location for the function.
1162   if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1163     if (SpecDecl->hasBody(SpecDecl))
1164       Loc = SpecDecl->getLocation();
1165 
1166   Stmt *Body = FD->getBody();
1167 
1168   // Initialize helper which will detect jumps which can cause invalid lifetime
1169   // markers.
1170   if (Body && ShouldEmitLifetimeMarkers)
1171     Bypasses.Init(Body);
1172 
1173   // Emit the standard function prologue.
1174   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1175 
1176   // Generate the body of the function.
1177   PGO.assignRegionCounters(GD, CurFn);
1178   if (isa<CXXDestructorDecl>(FD))
1179     EmitDestructorBody(Args);
1180   else if (isa<CXXConstructorDecl>(FD))
1181     EmitConstructorBody(Args);
1182   else if (getLangOpts().CUDA &&
1183            !getLangOpts().CUDAIsDevice &&
1184            FD->hasAttr<CUDAGlobalAttr>())
1185     CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1186   else if (isa<CXXMethodDecl>(FD) &&
1187            cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1188     // The lambda static invoker function is special, because it forwards or
1189     // clones the body of the function call operator (but is actually static).
1190     EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1191   } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1192              (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1193               cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1194     // Implicit copy-assignment gets the same special treatment as implicit
1195     // copy-constructors.
1196     emitImplicitAssignmentOperatorBody(Args);
1197   } else if (Body) {
1198     EmitFunctionBody(Body);
1199   } else
1200     llvm_unreachable("no definition for emitted function");
1201 
1202   // C++11 [stmt.return]p2:
1203   //   Flowing off the end of a function [...] results in undefined behavior in
1204   //   a value-returning function.
1205   // C11 6.9.1p12:
1206   //   If the '}' that terminates a function is reached, and the value of the
1207   //   function call is used by the caller, the behavior is undefined.
1208   if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1209       !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1210     bool ShouldEmitUnreachable =
1211         CGM.getCodeGenOpts().StrictReturn ||
1212         shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1213     if (SanOpts.has(SanitizerKind::Return)) {
1214       SanitizerScope SanScope(this);
1215       llvm::Value *IsFalse = Builder.getFalse();
1216       EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1217                 SanitizerHandler::MissingReturn,
1218                 EmitCheckSourceLocation(FD->getLocation()), None);
1219     } else if (ShouldEmitUnreachable) {
1220       if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1221         EmitTrapCall(llvm::Intrinsic::trap);
1222     }
1223     if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1224       Builder.CreateUnreachable();
1225       Builder.ClearInsertionPoint();
1226     }
1227   }
1228 
1229   // Emit the standard function epilogue.
1230   FinishFunction(BodyRange.getEnd());
1231 
1232   // If we haven't marked the function nothrow through other means, do
1233   // a quick pass now to see if we can.
1234   if (!CurFn->doesNotThrow())
1235     TryMarkNoThrow(CurFn);
1236 }
1237 
1238 /// ContainsLabel - Return true if the statement contains a label in it.  If
1239 /// this statement is not executed normally, it not containing a label means
1240 /// that we can just remove the code.
1241 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1242   // Null statement, not a label!
1243   if (!S) return false;
1244 
1245   // If this is a label, we have to emit the code, consider something like:
1246   // if (0) {  ...  foo:  bar(); }  goto foo;
1247   //
1248   // TODO: If anyone cared, we could track __label__'s, since we know that you
1249   // can't jump to one from outside their declared region.
1250   if (isa<LabelStmt>(S))
1251     return true;
1252 
1253   // If this is a case/default statement, and we haven't seen a switch, we have
1254   // to emit the code.
1255   if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1256     return true;
1257 
1258   // If this is a switch statement, we want to ignore cases below it.
1259   if (isa<SwitchStmt>(S))
1260     IgnoreCaseStmts = true;
1261 
1262   // Scan subexpressions for verboten labels.
1263   for (const Stmt *SubStmt : S->children())
1264     if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1265       return true;
1266 
1267   return false;
1268 }
1269 
1270 /// containsBreak - Return true if the statement contains a break out of it.
1271 /// If the statement (recursively) contains a switch or loop with a break
1272 /// inside of it, this is fine.
1273 bool CodeGenFunction::containsBreak(const Stmt *S) {
1274   // Null statement, not a label!
1275   if (!S) return false;
1276 
1277   // If this is a switch or loop that defines its own break scope, then we can
1278   // include it and anything inside of it.
1279   if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1280       isa<ForStmt>(S))
1281     return false;
1282 
1283   if (isa<BreakStmt>(S))
1284     return true;
1285 
1286   // Scan subexpressions for verboten breaks.
1287   for (const Stmt *SubStmt : S->children())
1288     if (containsBreak(SubStmt))
1289       return true;
1290 
1291   return false;
1292 }
1293 
1294 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1295   if (!S) return false;
1296 
1297   // Some statement kinds add a scope and thus never add a decl to the current
1298   // scope. Note, this list is longer than the list of statements that might
1299   // have an unscoped decl nested within them, but this way is conservatively
1300   // correct even if more statement kinds are added.
1301   if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1302       isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1303       isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1304       isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1305     return false;
1306 
1307   if (isa<DeclStmt>(S))
1308     return true;
1309 
1310   for (const Stmt *SubStmt : S->children())
1311     if (mightAddDeclToScope(SubStmt))
1312       return true;
1313 
1314   return false;
1315 }
1316 
1317 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1318 /// to a constant, or if it does but contains a label, return false.  If it
1319 /// constant folds return true and set the boolean result in Result.
1320 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1321                                                    bool &ResultBool,
1322                                                    bool AllowLabels) {
1323   llvm::APSInt ResultInt;
1324   if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1325     return false;
1326 
1327   ResultBool = ResultInt.getBoolValue();
1328   return true;
1329 }
1330 
1331 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1332 /// to a constant, or if it does but contains a label, return false.  If it
1333 /// constant folds return true and set the folded value.
1334 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1335                                                    llvm::APSInt &ResultInt,
1336                                                    bool AllowLabels) {
1337   // FIXME: Rename and handle conversion of other evaluatable things
1338   // to bool.
1339   Expr::EvalResult Result;
1340   if (!Cond->EvaluateAsInt(Result, getContext()))
1341     return false;  // Not foldable, not integer or not fully evaluatable.
1342 
1343   llvm::APSInt Int = Result.Val.getInt();
1344   if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1345     return false;  // Contains a label.
1346 
1347   ResultInt = Int;
1348   return true;
1349 }
1350 
1351 
1352 
1353 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1354 /// statement) to the specified blocks.  Based on the condition, this might try
1355 /// to simplify the codegen of the conditional based on the branch.
1356 ///
1357 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1358                                            llvm::BasicBlock *TrueBlock,
1359                                            llvm::BasicBlock *FalseBlock,
1360                                            uint64_t TrueCount) {
1361   Cond = Cond->IgnoreParens();
1362 
1363   if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1364 
1365     // Handle X && Y in a condition.
1366     if (CondBOp->getOpcode() == BO_LAnd) {
1367       // If we have "1 && X", simplify the code.  "0 && X" would have constant
1368       // folded if the case was simple enough.
1369       bool ConstantBool = false;
1370       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1371           ConstantBool) {
1372         // br(1 && X) -> br(X).
1373         incrementProfileCounter(CondBOp);
1374         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1375                                     TrueCount);
1376       }
1377 
1378       // If we have "X && 1", simplify the code to use an uncond branch.
1379       // "X && 0" would have been constant folded to 0.
1380       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1381           ConstantBool) {
1382         // br(X && 1) -> br(X).
1383         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1384                                     TrueCount);
1385       }
1386 
1387       // Emit the LHS as a conditional.  If the LHS conditional is false, we
1388       // want to jump to the FalseBlock.
1389       llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1390       // The counter tells us how often we evaluate RHS, and all of TrueCount
1391       // can be propagated to that branch.
1392       uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1393 
1394       ConditionalEvaluation eval(*this);
1395       {
1396         ApplyDebugLocation DL(*this, Cond);
1397         EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1398         EmitBlock(LHSTrue);
1399       }
1400 
1401       incrementProfileCounter(CondBOp);
1402       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1403 
1404       // Any temporaries created here are conditional.
1405       eval.begin(*this);
1406       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1407       eval.end(*this);
1408 
1409       return;
1410     }
1411 
1412     if (CondBOp->getOpcode() == BO_LOr) {
1413       // If we have "0 || X", simplify the code.  "1 || X" would have constant
1414       // folded if the case was simple enough.
1415       bool ConstantBool = false;
1416       if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1417           !ConstantBool) {
1418         // br(0 || X) -> br(X).
1419         incrementProfileCounter(CondBOp);
1420         return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1421                                     TrueCount);
1422       }
1423 
1424       // If we have "X || 0", simplify the code to use an uncond branch.
1425       // "X || 1" would have been constant folded to 1.
1426       if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1427           !ConstantBool) {
1428         // br(X || 0) -> br(X).
1429         return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1430                                     TrueCount);
1431       }
1432 
1433       // Emit the LHS as a conditional.  If the LHS conditional is true, we
1434       // want to jump to the TrueBlock.
1435       llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1436       // We have the count for entry to the RHS and for the whole expression
1437       // being true, so we can divy up True count between the short circuit and
1438       // the RHS.
1439       uint64_t LHSCount =
1440           getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1441       uint64_t RHSCount = TrueCount - LHSCount;
1442 
1443       ConditionalEvaluation eval(*this);
1444       {
1445         ApplyDebugLocation DL(*this, Cond);
1446         EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1447         EmitBlock(LHSFalse);
1448       }
1449 
1450       incrementProfileCounter(CondBOp);
1451       setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1452 
1453       // Any temporaries created here are conditional.
1454       eval.begin(*this);
1455       EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1456 
1457       eval.end(*this);
1458 
1459       return;
1460     }
1461   }
1462 
1463   if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1464     // br(!x, t, f) -> br(x, f, t)
1465     if (CondUOp->getOpcode() == UO_LNot) {
1466       // Negate the count.
1467       uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1468       // Negate the condition and swap the destination blocks.
1469       return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1470                                   FalseCount);
1471     }
1472   }
1473 
1474   if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1475     // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1476     llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1477     llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1478 
1479     ConditionalEvaluation cond(*this);
1480     EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1481                          getProfileCount(CondOp));
1482 
1483     // When computing PGO branch weights, we only know the overall count for
1484     // the true block. This code is essentially doing tail duplication of the
1485     // naive code-gen, introducing new edges for which counts are not
1486     // available. Divide the counts proportionally between the LHS and RHS of
1487     // the conditional operator.
1488     uint64_t LHSScaledTrueCount = 0;
1489     if (TrueCount) {
1490       double LHSRatio =
1491           getProfileCount(CondOp) / (double)getCurrentProfileCount();
1492       LHSScaledTrueCount = TrueCount * LHSRatio;
1493     }
1494 
1495     cond.begin(*this);
1496     EmitBlock(LHSBlock);
1497     incrementProfileCounter(CondOp);
1498     {
1499       ApplyDebugLocation DL(*this, Cond);
1500       EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1501                            LHSScaledTrueCount);
1502     }
1503     cond.end(*this);
1504 
1505     cond.begin(*this);
1506     EmitBlock(RHSBlock);
1507     EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1508                          TrueCount - LHSScaledTrueCount);
1509     cond.end(*this);
1510 
1511     return;
1512   }
1513 
1514   if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1515     // Conditional operator handling can give us a throw expression as a
1516     // condition for a case like:
1517     //   br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1518     // Fold this to:
1519     //   br(c, throw x, br(y, t, f))
1520     EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1521     return;
1522   }
1523 
1524   // If the branch has a condition wrapped by __builtin_unpredictable,
1525   // create metadata that specifies that the branch is unpredictable.
1526   // Don't bother if not optimizing because that metadata would not be used.
1527   llvm::MDNode *Unpredictable = nullptr;
1528   auto *Call = dyn_cast<CallExpr>(Cond->IgnoreImpCasts());
1529   if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1530     auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1531     if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1532       llvm::MDBuilder MDHelper(getLLVMContext());
1533       Unpredictable = MDHelper.createUnpredictable();
1534     }
1535   }
1536 
1537   // Create branch weights based on the number of times we get here and the
1538   // number of times the condition should be true.
1539   uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1540   llvm::MDNode *Weights =
1541       createProfileWeights(TrueCount, CurrentCount - TrueCount);
1542 
1543   // Emit the code with the fully general case.
1544   llvm::Value *CondV;
1545   {
1546     ApplyDebugLocation DL(*this, Cond);
1547     CondV = EvaluateExprAsBool(Cond);
1548   }
1549   Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1550 }
1551 
1552 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1553 /// specified stmt yet.
1554 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1555   CGM.ErrorUnsupported(S, Type);
1556 }
1557 
1558 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1559 /// variable-length array whose elements have a non-zero bit-pattern.
1560 ///
1561 /// \param baseType the inner-most element type of the array
1562 /// \param src - a char* pointing to the bit-pattern for a single
1563 /// base element of the array
1564 /// \param sizeInChars - the total size of the VLA, in chars
1565 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1566                                Address dest, Address src,
1567                                llvm::Value *sizeInChars) {
1568   CGBuilderTy &Builder = CGF.Builder;
1569 
1570   CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1571   llvm::Value *baseSizeInChars
1572     = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1573 
1574   Address begin =
1575     Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1576   llvm::Value *end =
1577     Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1578 
1579   llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1580   llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1581   llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1582 
1583   // Make a loop over the VLA.  C99 guarantees that the VLA element
1584   // count must be nonzero.
1585   CGF.EmitBlock(loopBB);
1586 
1587   llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1588   cur->addIncoming(begin.getPointer(), originBB);
1589 
1590   CharUnits curAlign =
1591     dest.getAlignment().alignmentOfArrayElement(baseSize);
1592 
1593   // memcpy the individual element bit-pattern.
1594   Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1595                        /*volatile*/ false);
1596 
1597   // Go to the next element.
1598   llvm::Value *next =
1599     Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1600 
1601   // Leave if that's the end of the VLA.
1602   llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1603   Builder.CreateCondBr(done, contBB, loopBB);
1604   cur->addIncoming(next, loopBB);
1605 
1606   CGF.EmitBlock(contBB);
1607 }
1608 
1609 void
1610 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1611   // Ignore empty classes in C++.
1612   if (getLangOpts().CPlusPlus) {
1613     if (const RecordType *RT = Ty->getAs<RecordType>()) {
1614       if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1615         return;
1616     }
1617   }
1618 
1619   // Cast the dest ptr to the appropriate i8 pointer type.
1620   if (DestPtr.getElementType() != Int8Ty)
1621     DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1622 
1623   // Get size and alignment info for this aggregate.
1624   CharUnits size = getContext().getTypeSizeInChars(Ty);
1625 
1626   llvm::Value *SizeVal;
1627   const VariableArrayType *vla;
1628 
1629   // Don't bother emitting a zero-byte memset.
1630   if (size.isZero()) {
1631     // But note that getTypeInfo returns 0 for a VLA.
1632     if (const VariableArrayType *vlaType =
1633           dyn_cast_or_null<VariableArrayType>(
1634                                           getContext().getAsArrayType(Ty))) {
1635       auto VlaSize = getVLASize(vlaType);
1636       SizeVal = VlaSize.NumElts;
1637       CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type);
1638       if (!eltSize.isOne())
1639         SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1640       vla = vlaType;
1641     } else {
1642       return;
1643     }
1644   } else {
1645     SizeVal = CGM.getSize(size);
1646     vla = nullptr;
1647   }
1648 
1649   // If the type contains a pointer to data member we can't memset it to zero.
1650   // Instead, create a null constant and copy it to the destination.
1651   // TODO: there are other patterns besides zero that we can usefully memset,
1652   // like -1, which happens to be the pattern used by member-pointers.
1653   if (!CGM.getTypes().isZeroInitializable(Ty)) {
1654     // For a VLA, emit a single element, then splat that over the VLA.
1655     if (vla) Ty = getContext().getBaseElementType(vla);
1656 
1657     llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1658 
1659     llvm::GlobalVariable *NullVariable =
1660       new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1661                                /*isConstant=*/true,
1662                                llvm::GlobalVariable::PrivateLinkage,
1663                                NullConstant, Twine());
1664     CharUnits NullAlign = DestPtr.getAlignment();
1665     NullVariable->setAlignment(NullAlign.getQuantity());
1666     Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1667                    NullAlign);
1668 
1669     if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1670 
1671     // Get and call the appropriate llvm.memcpy overload.
1672     Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1673     return;
1674   }
1675 
1676   // Otherwise, just memset the whole thing to zero.  This is legal
1677   // because in LLVM, all default initializers (other than the ones we just
1678   // handled above) are guaranteed to have a bit pattern of all zeros.
1679   Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1680 }
1681 
1682 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1683   // Make sure that there is a block for the indirect goto.
1684   if (!IndirectBranch)
1685     GetIndirectGotoBlock();
1686 
1687   llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1688 
1689   // Make sure the indirect branch includes all of the address-taken blocks.
1690   IndirectBranch->addDestination(BB);
1691   return llvm::BlockAddress::get(CurFn, BB);
1692 }
1693 
1694 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1695   // If we already made the indirect branch for indirect goto, return its block.
1696   if (IndirectBranch) return IndirectBranch->getParent();
1697 
1698   CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1699 
1700   // Create the PHI node that indirect gotos will add entries to.
1701   llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1702                                               "indirect.goto.dest");
1703 
1704   // Create the indirect branch instruction.
1705   IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1706   return IndirectBranch->getParent();
1707 }
1708 
1709 /// Computes the length of an array in elements, as well as the base
1710 /// element type and a properly-typed first element pointer.
1711 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1712                                               QualType &baseType,
1713                                               Address &addr) {
1714   const ArrayType *arrayType = origArrayType;
1715 
1716   // If it's a VLA, we have to load the stored size.  Note that
1717   // this is the size of the VLA in bytes, not its size in elements.
1718   llvm::Value *numVLAElements = nullptr;
1719   if (isa<VariableArrayType>(arrayType)) {
1720     numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts;
1721 
1722     // Walk into all VLAs.  This doesn't require changes to addr,
1723     // which has type T* where T is the first non-VLA element type.
1724     do {
1725       QualType elementType = arrayType->getElementType();
1726       arrayType = getContext().getAsArrayType(elementType);
1727 
1728       // If we only have VLA components, 'addr' requires no adjustment.
1729       if (!arrayType) {
1730         baseType = elementType;
1731         return numVLAElements;
1732       }
1733     } while (isa<VariableArrayType>(arrayType));
1734 
1735     // We get out here only if we find a constant array type
1736     // inside the VLA.
1737   }
1738 
1739   // We have some number of constant-length arrays, so addr should
1740   // have LLVM type [M x [N x [...]]]*.  Build a GEP that walks
1741   // down to the first element of addr.
1742   SmallVector<llvm::Value*, 8> gepIndices;
1743 
1744   // GEP down to the array type.
1745   llvm::ConstantInt *zero = Builder.getInt32(0);
1746   gepIndices.push_back(zero);
1747 
1748   uint64_t countFromCLAs = 1;
1749   QualType eltType;
1750 
1751   llvm::ArrayType *llvmArrayType =
1752     dyn_cast<llvm::ArrayType>(addr.getElementType());
1753   while (llvmArrayType) {
1754     assert(isa<ConstantArrayType>(arrayType));
1755     assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1756              == llvmArrayType->getNumElements());
1757 
1758     gepIndices.push_back(zero);
1759     countFromCLAs *= llvmArrayType->getNumElements();
1760     eltType = arrayType->getElementType();
1761 
1762     llvmArrayType =
1763       dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1764     arrayType = getContext().getAsArrayType(arrayType->getElementType());
1765     assert((!llvmArrayType || arrayType) &&
1766            "LLVM and Clang types are out-of-synch");
1767   }
1768 
1769   if (arrayType) {
1770     // From this point onwards, the Clang array type has been emitted
1771     // as some other type (probably a packed struct). Compute the array
1772     // size, and just emit the 'begin' expression as a bitcast.
1773     while (arrayType) {
1774       countFromCLAs *=
1775           cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1776       eltType = arrayType->getElementType();
1777       arrayType = getContext().getAsArrayType(eltType);
1778     }
1779 
1780     llvm::Type *baseType = ConvertType(eltType);
1781     addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1782   } else {
1783     // Create the actual GEP.
1784     addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1785                                              gepIndices, "array.begin"),
1786                    addr.getAlignment());
1787   }
1788 
1789   baseType = eltType;
1790 
1791   llvm::Value *numElements
1792     = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1793 
1794   // If we had any VLA dimensions, factor them in.
1795   if (numVLAElements)
1796     numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1797 
1798   return numElements;
1799 }
1800 
1801 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) {
1802   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1803   assert(vla && "type was not a variable array type!");
1804   return getVLASize(vla);
1805 }
1806 
1807 CodeGenFunction::VlaSizePair
1808 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1809   // The number of elements so far; always size_t.
1810   llvm::Value *numElements = nullptr;
1811 
1812   QualType elementType;
1813   do {
1814     elementType = type->getElementType();
1815     llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1816     assert(vlaSize && "no size for VLA!");
1817     assert(vlaSize->getType() == SizeTy);
1818 
1819     if (!numElements) {
1820       numElements = vlaSize;
1821     } else {
1822       // It's undefined behavior if this wraps around, so mark it that way.
1823       // FIXME: Teach -fsanitize=undefined to trap this.
1824       numElements = Builder.CreateNUWMul(numElements, vlaSize);
1825     }
1826   } while ((type = getContext().getAsVariableArrayType(elementType)));
1827 
1828   return { numElements, elementType };
1829 }
1830 
1831 CodeGenFunction::VlaSizePair
1832 CodeGenFunction::getVLAElements1D(QualType type) {
1833   const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1834   assert(vla && "type was not a variable array type!");
1835   return getVLAElements1D(vla);
1836 }
1837 
1838 CodeGenFunction::VlaSizePair
1839 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) {
1840   llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()];
1841   assert(VlaSize && "no size for VLA!");
1842   assert(VlaSize->getType() == SizeTy);
1843   return { VlaSize, Vla->getElementType() };
1844 }
1845 
1846 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1847   assert(type->isVariablyModifiedType() &&
1848          "Must pass variably modified type to EmitVLASizes!");
1849 
1850   EnsureInsertPoint();
1851 
1852   // We're going to walk down into the type and look for VLA
1853   // expressions.
1854   do {
1855     assert(type->isVariablyModifiedType());
1856 
1857     const Type *ty = type.getTypePtr();
1858     switch (ty->getTypeClass()) {
1859 
1860 #define TYPE(Class, Base)
1861 #define ABSTRACT_TYPE(Class, Base)
1862 #define NON_CANONICAL_TYPE(Class, Base)
1863 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1864 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1865 #include "clang/AST/TypeNodes.def"
1866       llvm_unreachable("unexpected dependent type!");
1867 
1868     // These types are never variably-modified.
1869     case Type::Builtin:
1870     case Type::Complex:
1871     case Type::Vector:
1872     case Type::ExtVector:
1873     case Type::Record:
1874     case Type::Enum:
1875     case Type::Elaborated:
1876     case Type::TemplateSpecialization:
1877     case Type::ObjCTypeParam:
1878     case Type::ObjCObject:
1879     case Type::ObjCInterface:
1880     case Type::ObjCObjectPointer:
1881       llvm_unreachable("type class is never variably-modified!");
1882 
1883     case Type::Adjusted:
1884       type = cast<AdjustedType>(ty)->getAdjustedType();
1885       break;
1886 
1887     case Type::Decayed:
1888       type = cast<DecayedType>(ty)->getPointeeType();
1889       break;
1890 
1891     case Type::Pointer:
1892       type = cast<PointerType>(ty)->getPointeeType();
1893       break;
1894 
1895     case Type::BlockPointer:
1896       type = cast<BlockPointerType>(ty)->getPointeeType();
1897       break;
1898 
1899     case Type::LValueReference:
1900     case Type::RValueReference:
1901       type = cast<ReferenceType>(ty)->getPointeeType();
1902       break;
1903 
1904     case Type::MemberPointer:
1905       type = cast<MemberPointerType>(ty)->getPointeeType();
1906       break;
1907 
1908     case Type::ConstantArray:
1909     case Type::IncompleteArray:
1910       // Losing element qualification here is fine.
1911       type = cast<ArrayType>(ty)->getElementType();
1912       break;
1913 
1914     case Type::VariableArray: {
1915       // Losing element qualification here is fine.
1916       const VariableArrayType *vat = cast<VariableArrayType>(ty);
1917 
1918       // Unknown size indication requires no size computation.
1919       // Otherwise, evaluate and record it.
1920       if (const Expr *size = vat->getSizeExpr()) {
1921         // It's possible that we might have emitted this already,
1922         // e.g. with a typedef and a pointer to it.
1923         llvm::Value *&entry = VLASizeMap[size];
1924         if (!entry) {
1925           llvm::Value *Size = EmitScalarExpr(size);
1926 
1927           // C11 6.7.6.2p5:
1928           //   If the size is an expression that is not an integer constant
1929           //   expression [...] each time it is evaluated it shall have a value
1930           //   greater than zero.
1931           if (SanOpts.has(SanitizerKind::VLABound) &&
1932               size->getType()->isSignedIntegerType()) {
1933             SanitizerScope SanScope(this);
1934             llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1935             llvm::Constant *StaticArgs[] = {
1936                 EmitCheckSourceLocation(size->getBeginLoc()),
1937                 EmitCheckTypeDescriptor(size->getType())};
1938             EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1939                                      SanitizerKind::VLABound),
1940                       SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
1941           }
1942 
1943           // Always zexting here would be wrong if it weren't
1944           // undefined behavior to have a negative bound.
1945           entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1946         }
1947       }
1948       type = vat->getElementType();
1949       break;
1950     }
1951 
1952     case Type::FunctionProto:
1953     case Type::FunctionNoProto:
1954       type = cast<FunctionType>(ty)->getReturnType();
1955       break;
1956 
1957     case Type::Paren:
1958     case Type::TypeOf:
1959     case Type::UnaryTransform:
1960     case Type::Attributed:
1961     case Type::SubstTemplateTypeParm:
1962     case Type::PackExpansion:
1963     case Type::MacroQualified:
1964       // Keep walking after single level desugaring.
1965       type = type.getSingleStepDesugaredType(getContext());
1966       break;
1967 
1968     case Type::Typedef:
1969     case Type::Decltype:
1970     case Type::Auto:
1971     case Type::DeducedTemplateSpecialization:
1972       // Stop walking: nothing to do.
1973       return;
1974 
1975     case Type::TypeOfExpr:
1976       // Stop walking: emit typeof expression.
1977       EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1978       return;
1979 
1980     case Type::Atomic:
1981       type = cast<AtomicType>(ty)->getValueType();
1982       break;
1983 
1984     case Type::Pipe:
1985       type = cast<PipeType>(ty)->getElementType();
1986       break;
1987     }
1988   } while (type->isVariablyModifiedType());
1989 }
1990 
1991 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1992   if (getContext().getBuiltinVaListType()->isArrayType())
1993     return EmitPointerWithAlignment(E);
1994   return EmitLValue(E).getAddress();
1995 }
1996 
1997 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1998   return EmitLValue(E).getAddress();
1999 }
2000 
2001 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2002                                               const APValue &Init) {
2003   assert(Init.hasValue() && "Invalid DeclRefExpr initializer!");
2004   if (CGDebugInfo *Dbg = getDebugInfo())
2005     if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2006       Dbg->EmitGlobalVariable(E->getDecl(), Init);
2007 }
2008 
2009 CodeGenFunction::PeepholeProtection
2010 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2011   // At the moment, the only aggressive peephole we do in IR gen
2012   // is trunc(zext) folding, but if we add more, we can easily
2013   // extend this protection.
2014 
2015   if (!rvalue.isScalar()) return PeepholeProtection();
2016   llvm::Value *value = rvalue.getScalarVal();
2017   if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2018 
2019   // Just make an extra bitcast.
2020   assert(HaveInsertPoint());
2021   llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2022                                                   Builder.GetInsertBlock());
2023 
2024   PeepholeProtection protection;
2025   protection.Inst = inst;
2026   return protection;
2027 }
2028 
2029 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2030   if (!protection.Inst) return;
2031 
2032   // In theory, we could try to duplicate the peepholes now, but whatever.
2033   protection.Inst->eraseFromParent();
2034 }
2035 
2036 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2037                                               QualType Ty, SourceLocation Loc,
2038                                               SourceLocation AssumptionLoc,
2039                                               llvm::Value *Alignment,
2040                                               llvm::Value *OffsetValue) {
2041   llvm::Value *TheCheck;
2042   llvm::Instruction *Assumption = Builder.CreateAlignmentAssumption(
2043       CGM.getDataLayout(), PtrValue, Alignment, OffsetValue, &TheCheck);
2044   if (SanOpts.has(SanitizerKind::Alignment)) {
2045     EmitAlignmentAssumptionCheck(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
2046                                  OffsetValue, TheCheck, Assumption);
2047   }
2048 }
2049 
2050 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2051                                               QualType Ty, SourceLocation Loc,
2052                                               SourceLocation AssumptionLoc,
2053                                               unsigned Alignment,
2054                                               llvm::Value *OffsetValue) {
2055   llvm::Value *TheCheck;
2056   llvm::Instruction *Assumption = Builder.CreateAlignmentAssumption(
2057       CGM.getDataLayout(), PtrValue, Alignment, OffsetValue, &TheCheck);
2058   if (SanOpts.has(SanitizerKind::Alignment)) {
2059     llvm::Value *AlignmentVal = llvm::ConstantInt::get(IntPtrTy, Alignment);
2060     EmitAlignmentAssumptionCheck(PtrValue, Ty, Loc, AssumptionLoc, AlignmentVal,
2061                                  OffsetValue, TheCheck, Assumption);
2062   }
2063 }
2064 
2065 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2066                                               const Expr *E,
2067                                               SourceLocation AssumptionLoc,
2068                                               unsigned Alignment,
2069                                               llvm::Value *OffsetValue) {
2070   if (auto *CE = dyn_cast<CastExpr>(E))
2071     E = CE->getSubExprAsWritten();
2072   QualType Ty = E->getType();
2073   SourceLocation Loc = E->getExprLoc();
2074 
2075   EmitAlignmentAssumption(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
2076                           OffsetValue);
2077 }
2078 
2079 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Function *AnnotationFn,
2080                                                  llvm::Value *AnnotatedVal,
2081                                                  StringRef AnnotationStr,
2082                                                  SourceLocation Location) {
2083   llvm::Value *Args[4] = {
2084     AnnotatedVal,
2085     Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2086     Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2087     CGM.EmitAnnotationLineNo(Location)
2088   };
2089   return Builder.CreateCall(AnnotationFn, Args);
2090 }
2091 
2092 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2093   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2094   // FIXME We create a new bitcast for every annotation because that's what
2095   // llvm-gcc was doing.
2096   for (const auto *I : D->specific_attrs<AnnotateAttr>())
2097     EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2098                        Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2099                        I->getAnnotation(), D->getLocation());
2100 }
2101 
2102 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2103                                               Address Addr) {
2104   assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2105   llvm::Value *V = Addr.getPointer();
2106   llvm::Type *VTy = V->getType();
2107   llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2108                                     CGM.Int8PtrTy);
2109 
2110   for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2111     // FIXME Always emit the cast inst so we can differentiate between
2112     // annotation on the first field of a struct and annotation on the struct
2113     // itself.
2114     if (VTy != CGM.Int8PtrTy)
2115       V = Builder.CreateBitCast(V, CGM.Int8PtrTy);
2116     V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2117     V = Builder.CreateBitCast(V, VTy);
2118   }
2119 
2120   return Address(V, Addr.getAlignment());
2121 }
2122 
2123 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2124 
2125 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2126     : CGF(CGF) {
2127   assert(!CGF->IsSanitizerScope);
2128   CGF->IsSanitizerScope = true;
2129 }
2130 
2131 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2132   CGF->IsSanitizerScope = false;
2133 }
2134 
2135 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2136                                    const llvm::Twine &Name,
2137                                    llvm::BasicBlock *BB,
2138                                    llvm::BasicBlock::iterator InsertPt) const {
2139   LoopStack.InsertHelper(I);
2140   if (IsSanitizerScope)
2141     CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2142 }
2143 
2144 void CGBuilderInserter::InsertHelper(
2145     llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2146     llvm::BasicBlock::iterator InsertPt) const {
2147   llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2148   if (CGF)
2149     CGF->InsertHelper(I, Name, BB, InsertPt);
2150 }
2151 
2152 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2153                                 CodeGenModule &CGM, const FunctionDecl *FD,
2154                                 std::string &FirstMissing) {
2155   // If there aren't any required features listed then go ahead and return.
2156   if (ReqFeatures.empty())
2157     return false;
2158 
2159   // Now build up the set of caller features and verify that all the required
2160   // features are there.
2161   llvm::StringMap<bool> CallerFeatureMap;
2162   CGM.getFunctionFeatureMap(CallerFeatureMap, GlobalDecl().getWithDecl(FD));
2163 
2164   // If we have at least one of the features in the feature list return
2165   // true, otherwise return false.
2166   return std::all_of(
2167       ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2168         SmallVector<StringRef, 1> OrFeatures;
2169         Feature.split(OrFeatures, '|');
2170         return llvm::any_of(OrFeatures, [&](StringRef Feature) {
2171           if (!CallerFeatureMap.lookup(Feature)) {
2172             FirstMissing = Feature.str();
2173             return false;
2174           }
2175           return true;
2176         });
2177       });
2178 }
2179 
2180 // Emits an error if we don't have a valid set of target features for the
2181 // called function.
2182 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2183                                           const FunctionDecl *TargetDecl) {
2184   return checkTargetFeatures(E->getBeginLoc(), TargetDecl);
2185 }
2186 
2187 // Emits an error if we don't have a valid set of target features for the
2188 // called function.
2189 void CodeGenFunction::checkTargetFeatures(SourceLocation Loc,
2190                                           const FunctionDecl *TargetDecl) {
2191   // Early exit if this is an indirect call.
2192   if (!TargetDecl)
2193     return;
2194 
2195   // Get the current enclosing function if it exists. If it doesn't
2196   // we can't check the target features anyhow.
2197   const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2198   if (!FD)
2199     return;
2200 
2201   // Grab the required features for the call. For a builtin this is listed in
2202   // the td file with the default cpu, for an always_inline function this is any
2203   // listed cpu and any listed features.
2204   unsigned BuiltinID = TargetDecl->getBuiltinID();
2205   std::string MissingFeature;
2206   if (BuiltinID) {
2207     SmallVector<StringRef, 1> ReqFeatures;
2208     const char *FeatureList =
2209         CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2210     // Return if the builtin doesn't have any required features.
2211     if (!FeatureList || StringRef(FeatureList) == "")
2212       return;
2213     StringRef(FeatureList).split(ReqFeatures, ',');
2214     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2215       CGM.getDiags().Report(Loc, diag::err_builtin_needs_feature)
2216           << TargetDecl->getDeclName()
2217           << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2218 
2219   } else if (TargetDecl->hasAttr<TargetAttr>() ||
2220              TargetDecl->hasAttr<CPUSpecificAttr>()) {
2221     // Get the required features for the callee.
2222 
2223     const TargetAttr *TD = TargetDecl->getAttr<TargetAttr>();
2224     TargetAttr::ParsedTargetAttr ParsedAttr = CGM.filterFunctionTargetAttrs(TD);
2225 
2226     SmallVector<StringRef, 1> ReqFeatures;
2227     llvm::StringMap<bool> CalleeFeatureMap;
2228     CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2229 
2230     for (const auto &F : ParsedAttr.Features) {
2231       if (F[0] == '+' && CalleeFeatureMap.lookup(F.substr(1)))
2232         ReqFeatures.push_back(StringRef(F).substr(1));
2233     }
2234 
2235     for (const auto &F : CalleeFeatureMap) {
2236       // Only positive features are "required".
2237       if (F.getValue())
2238         ReqFeatures.push_back(F.getKey());
2239     }
2240     if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2241       CGM.getDiags().Report(Loc, diag::err_function_needs_feature)
2242           << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2243   }
2244 }
2245 
2246 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2247   if (!CGM.getCodeGenOpts().SanitizeStats)
2248     return;
2249 
2250   llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2251   IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2252   CGM.getSanStats().create(IRB, SSK);
2253 }
2254 
2255 llvm::Value *
2256 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
2257   llvm::Value *Condition = nullptr;
2258 
2259   if (!RO.Conditions.Architecture.empty())
2260     Condition = EmitX86CpuIs(RO.Conditions.Architecture);
2261 
2262   if (!RO.Conditions.Features.empty()) {
2263     llvm::Value *FeatureCond = EmitX86CpuSupports(RO.Conditions.Features);
2264     Condition =
2265         Condition ? Builder.CreateAnd(Condition, FeatureCond) : FeatureCond;
2266   }
2267   return Condition;
2268 }
2269 
2270 static void CreateMultiVersionResolverReturn(CodeGenModule &CGM,
2271                                              llvm::Function *Resolver,
2272                                              CGBuilderTy &Builder,
2273                                              llvm::Function *FuncToReturn,
2274                                              bool SupportsIFunc) {
2275   if (SupportsIFunc) {
2276     Builder.CreateRet(FuncToReturn);
2277     return;
2278   }
2279 
2280   llvm::SmallVector<llvm::Value *, 10> Args;
2281   llvm::for_each(Resolver->args(),
2282                  [&](llvm::Argument &Arg) { Args.push_back(&Arg); });
2283 
2284   llvm::CallInst *Result = Builder.CreateCall(FuncToReturn, Args);
2285   Result->setTailCallKind(llvm::CallInst::TCK_MustTail);
2286 
2287   if (Resolver->getReturnType()->isVoidTy())
2288     Builder.CreateRetVoid();
2289   else
2290     Builder.CreateRet(Result);
2291 }
2292 
2293 void CodeGenFunction::EmitMultiVersionResolver(
2294     llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
2295   assert((getContext().getTargetInfo().getTriple().getArch() ==
2296               llvm::Triple::x86 ||
2297           getContext().getTargetInfo().getTriple().getArch() ==
2298               llvm::Triple::x86_64) &&
2299          "Only implemented for x86 targets");
2300 
2301   bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc();
2302 
2303   // Main function's basic block.
2304   llvm::BasicBlock *CurBlock = createBasicBlock("resolver_entry", Resolver);
2305   Builder.SetInsertPoint(CurBlock);
2306   EmitX86CpuInit();
2307 
2308   for (const MultiVersionResolverOption &RO : Options) {
2309     Builder.SetInsertPoint(CurBlock);
2310     llvm::Value *Condition = FormResolverCondition(RO);
2311 
2312     // The 'default' or 'generic' case.
2313     if (!Condition) {
2314       assert(&RO == Options.end() - 1 &&
2315              "Default or Generic case must be last");
2316       CreateMultiVersionResolverReturn(CGM, Resolver, Builder, RO.Function,
2317                                        SupportsIFunc);
2318       return;
2319     }
2320 
2321     llvm::BasicBlock *RetBlock = createBasicBlock("resolver_return", Resolver);
2322     CGBuilderTy RetBuilder(*this, RetBlock);
2323     CreateMultiVersionResolverReturn(CGM, Resolver, RetBuilder, RO.Function,
2324                                      SupportsIFunc);
2325     CurBlock = createBasicBlock("resolver_else", Resolver);
2326     Builder.CreateCondBr(Condition, RetBlock, CurBlock);
2327   }
2328 
2329   // If no generic/default, emit an unreachable.
2330   Builder.SetInsertPoint(CurBlock);
2331   llvm::CallInst *TrapCall = EmitTrapCall(llvm::Intrinsic::trap);
2332   TrapCall->setDoesNotReturn();
2333   TrapCall->setDoesNotThrow();
2334   Builder.CreateUnreachable();
2335   Builder.ClearInsertionPoint();
2336 }
2337 
2338 // Loc - where the diagnostic will point, where in the source code this
2339 //  alignment has failed.
2340 // SecondaryLoc - if present (will be present if sufficiently different from
2341 //  Loc), the diagnostic will additionally point a "Note:" to this location.
2342 //  It should be the location where the __attribute__((assume_aligned))
2343 //  was written e.g.
2344 void CodeGenFunction::EmitAlignmentAssumptionCheck(
2345     llvm::Value *Ptr, QualType Ty, SourceLocation Loc,
2346     SourceLocation SecondaryLoc, llvm::Value *Alignment,
2347     llvm::Value *OffsetValue, llvm::Value *TheCheck,
2348     llvm::Instruction *Assumption) {
2349   assert(Assumption && isa<llvm::CallInst>(Assumption) &&
2350          cast<llvm::CallInst>(Assumption)->getCalledValue() ==
2351              llvm::Intrinsic::getDeclaration(
2352                  Builder.GetInsertBlock()->getParent()->getParent(),
2353                  llvm::Intrinsic::assume) &&
2354          "Assumption should be a call to llvm.assume().");
2355   assert(&(Builder.GetInsertBlock()->back()) == Assumption &&
2356          "Assumption should be the last instruction of the basic block, "
2357          "since the basic block is still being generated.");
2358 
2359   if (!SanOpts.has(SanitizerKind::Alignment))
2360     return;
2361 
2362   // Don't check pointers to volatile data. The behavior here is implementation-
2363   // defined.
2364   if (Ty->getPointeeType().isVolatileQualified())
2365     return;
2366 
2367   // We need to temorairly remove the assumption so we can insert the
2368   // sanitizer check before it, else the check will be dropped by optimizations.
2369   Assumption->removeFromParent();
2370 
2371   {
2372     SanitizerScope SanScope(this);
2373 
2374     if (!OffsetValue)
2375       OffsetValue = Builder.getInt1(0); // no offset.
2376 
2377     llvm::Constant *StaticData[] = {EmitCheckSourceLocation(Loc),
2378                                     EmitCheckSourceLocation(SecondaryLoc),
2379                                     EmitCheckTypeDescriptor(Ty)};
2380     llvm::Value *DynamicData[] = {EmitCheckValue(Ptr),
2381                                   EmitCheckValue(Alignment),
2382                                   EmitCheckValue(OffsetValue)};
2383     EmitCheck({std::make_pair(TheCheck, SanitizerKind::Alignment)},
2384               SanitizerHandler::AlignmentAssumption, StaticData, DynamicData);
2385   }
2386 
2387   // We are now in the (new, empty) "cont" basic block.
2388   // Reintroduce the assumption.
2389   Builder.Insert(Assumption);
2390   // FIXME: Assumption still has it's original basic block as it's Parent.
2391 }
2392 
2393 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2394   if (CGDebugInfo *DI = getDebugInfo())
2395     return DI->SourceLocToDebugLoc(Location);
2396 
2397   return llvm::DebugLoc();
2398 }
2399