1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This coordinates the per-function state used while generating code.
12 //===----------------------------------------------------------------------===//
14 #include "CodeGenFunction.h"
16 #include "CGCleanup.h"
17 #include "CGCUDARuntime.h"
19 #include "CGDebugInfo.h"
20 #include "CGOpenMPRuntime.h"
21 #include "CodeGenModule.h"
22 #include "CodeGenPGO.h"
23 #include "TargetInfo.h"
24 #include "clang/AST/ASTContext.h"
25 #include "clang/AST/Decl.h"
26 #include "clang/AST/DeclCXX.h"
27 #include "clang/AST/StmtCXX.h"
28 #include "clang/Basic/Builtins.h"
29 #include "clang/Basic/TargetInfo.h"
30 #include "clang/CodeGen/CGFunctionInfo.h"
31 #include "clang/Frontend/CodeGenOptions.h"
32 #include "clang/Sema/SemaDiagnostic.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Operator.h"
37 using namespace clang;
38 using namespace CodeGen;
40 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
41 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
42 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
43 CGBuilderInserterTy(this)),
44 CurFn(nullptr), ReturnValue(Address::invalid()),
45 CapturedStmtInfo(nullptr),
46 SanOpts(CGM.getLangOpts().Sanitize), IsSanitizerScope(false),
47 CurFuncIsThunk(false), AutoreleaseResult(false), SawAsmBlock(false),
48 IsOutlinedSEHHelper(false),
49 BlockInfo(nullptr), BlockPointer(nullptr),
50 LambdaThisCaptureField(nullptr), NormalCleanupDest(nullptr),
51 NextCleanupDestIndex(1), FirstBlockInfo(nullptr), EHResumeBlock(nullptr),
52 ExceptionSlot(nullptr), EHSelectorSlot(nullptr),
53 DebugInfo(CGM.getModuleDebugInfo()),
54 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
55 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
56 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
57 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
58 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
59 CXXStructorImplicitParamDecl(nullptr),
60 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
61 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
62 TerminateHandler(nullptr), TrapBB(nullptr) {
63 if (!suppressNewContext)
64 CGM.getCXXABI().getMangleContext().startNewFunction();
66 llvm::FastMathFlags FMF;
67 if (CGM.getLangOpts().FastMath)
68 FMF.setUnsafeAlgebra();
69 if (CGM.getLangOpts().FiniteMathOnly) {
73 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
76 if (CGM.getCodeGenOpts().NoSignedZeros) {
77 FMF.setNoSignedZeros();
79 if (CGM.getCodeGenOpts().ReciprocalMath) {
80 FMF.setAllowReciprocal();
82 Builder.setFastMathFlags(FMF);
85 CodeGenFunction::~CodeGenFunction() {
86 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
88 // If there are any unclaimed block infos, go ahead and destroy them
89 // now. This can happen if IR-gen gets clever and skips evaluating
92 destroyBlockInfos(FirstBlockInfo);
94 if (getLangOpts().OpenMP) {
95 CGM.getOpenMPRuntime().functionFinished(*this);
99 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
100 AlignmentSource *Source) {
101 return getNaturalTypeAlignment(T->getPointeeType(), Source,
102 /*forPointee*/ true);
105 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
106 AlignmentSource *Source,
107 bool forPointeeType) {
108 // Honor alignment typedef attributes even on incomplete types.
109 // We also honor them straight for C++ class types, even as pointees;
110 // there's an expressivity gap here.
111 if (auto TT = T->getAs<TypedefType>()) {
112 if (auto Align = TT->getDecl()->getMaxAlignment()) {
113 if (Source) *Source = AlignmentSource::AttributedType;
114 return getContext().toCharUnitsFromBits(Align);
118 if (Source) *Source = AlignmentSource::Type;
121 if (T->isIncompleteType()) {
122 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
124 // For C++ class pointees, we don't know whether we're pointing at a
125 // base or a complete object, so we generally need to use the
126 // non-virtual alignment.
127 const CXXRecordDecl *RD;
128 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
129 Alignment = CGM.getClassPointerAlignment(RD);
131 Alignment = getContext().getTypeAlignInChars(T);
134 // Cap to the global maximum type alignment unless the alignment
135 // was somehow explicit on the type.
136 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
137 if (Alignment.getQuantity() > MaxAlign &&
138 !getContext().isAlignmentRequired(T))
139 Alignment = CharUnits::fromQuantity(MaxAlign);
145 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
146 AlignmentSource AlignSource;
147 CharUnits Alignment = getNaturalTypeAlignment(T, &AlignSource);
148 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), AlignSource,
152 /// Given a value of type T* that may not be to a complete object,
153 /// construct an l-value with the natural pointee alignment of T.
155 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
156 AlignmentSource AlignSource;
157 CharUnits Align = getNaturalTypeAlignment(T, &AlignSource, /*pointee*/ true);
158 return MakeAddrLValue(Address(V, Align), T, AlignSource);
162 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
163 return CGM.getTypes().ConvertTypeForMem(T);
166 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
167 return CGM.getTypes().ConvertType(T);
170 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
171 type = type.getCanonicalType();
173 switch (type->getTypeClass()) {
174 #define TYPE(name, parent)
175 #define ABSTRACT_TYPE(name, parent)
176 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
177 #define DEPENDENT_TYPE(name, parent) case Type::name:
178 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
179 #include "clang/AST/TypeNodes.def"
180 llvm_unreachable("non-canonical or dependent type in IR-generation");
183 llvm_unreachable("undeduced auto type in IR-generation");
185 // Various scalar types.
188 case Type::BlockPointer:
189 case Type::LValueReference:
190 case Type::RValueReference:
191 case Type::MemberPointer:
193 case Type::ExtVector:
194 case Type::FunctionProto:
195 case Type::FunctionNoProto:
197 case Type::ObjCObjectPointer:
205 // Arrays, records, and Objective-C objects.
206 case Type::ConstantArray:
207 case Type::IncompleteArray:
208 case Type::VariableArray:
210 case Type::ObjCObject:
211 case Type::ObjCInterface:
212 return TEK_Aggregate;
214 // We operate on atomic values according to their underlying type.
216 type = cast<AtomicType>(type)->getValueType();
219 llvm_unreachable("unknown type kind!");
223 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
224 // For cleanliness, we try to avoid emitting the return block for
226 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
229 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
231 // We have a valid insert point, reuse it if it is empty or there are no
232 // explicit jumps to the return block.
233 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
234 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
235 delete ReturnBlock.getBlock();
237 EmitBlock(ReturnBlock.getBlock());
238 return llvm::DebugLoc();
241 // Otherwise, if the return block is the target of a single direct
242 // branch then we can just put the code in that block instead. This
243 // cleans up functions which started with a unified return block.
244 if (ReturnBlock.getBlock()->hasOneUse()) {
245 llvm::BranchInst *BI =
246 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
247 if (BI && BI->isUnconditional() &&
248 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
249 // Record/return the DebugLoc of the simple 'return' expression to be used
250 // later by the actual 'ret' instruction.
251 llvm::DebugLoc Loc = BI->getDebugLoc();
252 Builder.SetInsertPoint(BI->getParent());
253 BI->eraseFromParent();
254 delete ReturnBlock.getBlock();
259 // FIXME: We are at an unreachable point, there is no reason to emit the block
260 // unless it has uses. However, we still need a place to put the debug
261 // region.end for now.
263 EmitBlock(ReturnBlock.getBlock());
264 return llvm::DebugLoc();
267 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
269 if (!BB->use_empty())
270 return CGF.CurFn->getBasicBlockList().push_back(BB);
274 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
275 assert(BreakContinueStack.empty() &&
276 "mismatched push/pop in break/continue stack!");
278 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
279 && NumSimpleReturnExprs == NumReturnExprs
280 && ReturnBlock.getBlock()->use_empty();
281 // Usually the return expression is evaluated before the cleanup
282 // code. If the function contains only a simple return statement,
283 // such as a constant, the location before the cleanup code becomes
284 // the last useful breakpoint in the function, because the simple
285 // return expression will be evaluated after the cleanup code. To be
286 // safe, set the debug location for cleanup code to the location of
287 // the return statement. Otherwise the cleanup code should be at the
288 // end of the function's lexical scope.
290 // If there are multiple branches to the return block, the branch
291 // instructions will get the location of the return statements and
293 if (CGDebugInfo *DI = getDebugInfo()) {
294 if (OnlySimpleReturnStmts)
295 DI->EmitLocation(Builder, LastStopPoint);
297 DI->EmitLocation(Builder, EndLoc);
300 // Pop any cleanups that might have been associated with the
301 // parameters. Do this in whatever block we're currently in; it's
302 // important to do this before we enter the return block or return
303 // edges will be *really* confused.
304 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
305 bool HasOnlyLifetimeMarkers =
306 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
307 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
309 // Make sure the line table doesn't jump back into the body for
310 // the ret after it's been at EndLoc.
311 if (CGDebugInfo *DI = getDebugInfo())
312 if (OnlySimpleReturnStmts)
313 DI->EmitLocation(Builder, EndLoc);
315 PopCleanupBlocks(PrologueCleanupDepth);
318 // Emit function epilog (to return).
319 llvm::DebugLoc Loc = EmitReturnBlock();
321 if (ShouldInstrumentFunction())
322 EmitFunctionInstrumentation("__cyg_profile_func_exit");
324 // Emit debug descriptor for function end.
325 if (CGDebugInfo *DI = getDebugInfo())
326 DI->EmitFunctionEnd(Builder);
328 // Reset the debug location to that of the simple 'return' expression, if any
329 // rather than that of the end of the function's scope '}'.
330 ApplyDebugLocation AL(*this, Loc);
331 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
332 EmitEndEHSpec(CurCodeDecl);
334 assert(EHStack.empty() &&
335 "did not remove all scopes from cleanup stack!");
337 // If someone did an indirect goto, emit the indirect goto block at the end of
339 if (IndirectBranch) {
340 EmitBlock(IndirectBranch->getParent());
341 Builder.ClearInsertionPoint();
344 // If some of our locals escaped, insert a call to llvm.localescape in the
346 if (!EscapedLocals.empty()) {
347 // Invert the map from local to index into a simple vector. There should be
349 SmallVector<llvm::Value *, 4> EscapeArgs;
350 EscapeArgs.resize(EscapedLocals.size());
351 for (auto &Pair : EscapedLocals)
352 EscapeArgs[Pair.second] = Pair.first;
353 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
354 &CGM.getModule(), llvm::Intrinsic::localescape);
355 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
358 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
359 llvm::Instruction *Ptr = AllocaInsertPt;
360 AllocaInsertPt = nullptr;
361 Ptr->eraseFromParent();
363 // If someone took the address of a label but never did an indirect goto, we
364 // made a zero entry PHI node, which is illegal, zap it now.
365 if (IndirectBranch) {
366 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
367 if (PN->getNumIncomingValues() == 0) {
368 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
369 PN->eraseFromParent();
373 EmitIfUsed(*this, EHResumeBlock);
374 EmitIfUsed(*this, TerminateLandingPad);
375 EmitIfUsed(*this, TerminateHandler);
376 EmitIfUsed(*this, UnreachableBlock);
378 if (CGM.getCodeGenOpts().EmitDeclMetadata)
381 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
382 I = DeferredReplacements.begin(),
383 E = DeferredReplacements.end();
385 I->first->replaceAllUsesWith(I->second);
386 I->first->eraseFromParent();
390 /// ShouldInstrumentFunction - Return true if the current function should be
391 /// instrumented with __cyg_profile_func_* calls
392 bool CodeGenFunction::ShouldInstrumentFunction() {
393 if (!CGM.getCodeGenOpts().InstrumentFunctions)
395 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
400 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
401 /// instrumentation function with the current function and the call site, if
402 /// function instrumentation is enabled.
403 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
404 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
405 llvm::PointerType *PointerTy = Int8PtrTy;
406 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
407 llvm::FunctionType *FunctionTy =
408 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
410 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
411 llvm::CallInst *CallSite = Builder.CreateCall(
412 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
413 llvm::ConstantInt::get(Int32Ty, 0),
416 llvm::Value *args[] = {
417 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
421 EmitNounwindRuntimeCall(F, args);
424 void CodeGenFunction::EmitMCountInstrumentation() {
425 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
427 llvm::Constant *MCountFn =
428 CGM.CreateRuntimeFunction(FTy, getTarget().getMCountName());
429 EmitNounwindRuntimeCall(MCountFn);
432 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
433 // information in the program executable. The argument information stored
434 // includes the argument name, its type, the address and access qualifiers used.
435 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
436 CodeGenModule &CGM, llvm::LLVMContext &Context,
437 SmallVector<llvm::Metadata *, 5> &kernelMDArgs,
438 CGBuilderTy &Builder, ASTContext &ASTCtx) {
439 // Create MDNodes that represent the kernel arg metadata.
440 // Each MDNode is a list in the form of "key", N number of values which is
441 // the same number of values as their are kernel arguments.
443 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
445 // MDNode for the kernel argument address space qualifiers.
446 SmallVector<llvm::Metadata *, 8> addressQuals;
447 addressQuals.push_back(llvm::MDString::get(Context, "kernel_arg_addr_space"));
449 // MDNode for the kernel argument access qualifiers (images only).
450 SmallVector<llvm::Metadata *, 8> accessQuals;
451 accessQuals.push_back(llvm::MDString::get(Context, "kernel_arg_access_qual"));
453 // MDNode for the kernel argument type names.
454 SmallVector<llvm::Metadata *, 8> argTypeNames;
455 argTypeNames.push_back(llvm::MDString::get(Context, "kernel_arg_type"));
457 // MDNode for the kernel argument base type names.
458 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
459 argBaseTypeNames.push_back(
460 llvm::MDString::get(Context, "kernel_arg_base_type"));
462 // MDNode for the kernel argument type qualifiers.
463 SmallVector<llvm::Metadata *, 8> argTypeQuals;
464 argTypeQuals.push_back(llvm::MDString::get(Context, "kernel_arg_type_qual"));
466 // MDNode for the kernel argument names.
467 SmallVector<llvm::Metadata *, 8> argNames;
468 argNames.push_back(llvm::MDString::get(Context, "kernel_arg_name"));
470 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
471 const ParmVarDecl *parm = FD->getParamDecl(i);
472 QualType ty = parm->getType();
473 std::string typeQuals;
475 if (ty->isPointerType()) {
476 QualType pointeeTy = ty->getPointeeType();
478 // Get address qualifier.
479 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
480 ASTCtx.getTargetAddressSpace(pointeeTy.getAddressSpace()))));
482 // Get argument type name.
483 std::string typeName =
484 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
486 // Turn "unsigned type" to "utype"
487 std::string::size_type pos = typeName.find("unsigned");
488 if (pointeeTy.isCanonical() && pos != std::string::npos)
489 typeName.erase(pos+1, 8);
491 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
493 std::string baseTypeName =
494 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
498 // Turn "unsigned type" to "utype"
499 pos = baseTypeName.find("unsigned");
500 if (pos != std::string::npos)
501 baseTypeName.erase(pos+1, 8);
503 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
505 // Get argument type qualifiers:
506 if (ty.isRestrictQualified())
507 typeQuals = "restrict";
508 if (pointeeTy.isConstQualified() ||
509 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
510 typeQuals += typeQuals.empty() ? "const" : " const";
511 if (pointeeTy.isVolatileQualified())
512 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
514 uint32_t AddrSpc = 0;
515 bool isPipe = ty->isPipeType();
516 if (ty->isImageType() || isPipe)
518 CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
520 addressQuals.push_back(
521 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
523 // Get argument type name.
524 std::string typeName;
526 typeName = cast<PipeType>(ty)->getElementType().getAsString(Policy);
528 typeName = ty.getUnqualifiedType().getAsString(Policy);
530 // Turn "unsigned type" to "utype"
531 std::string::size_type pos = typeName.find("unsigned");
532 if (ty.isCanonical() && pos != std::string::npos)
533 typeName.erase(pos+1, 8);
535 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
537 std::string baseTypeName;
540 cast<PipeType>(ty)->getElementType().getCanonicalType().getAsString(Policy);
543 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
545 // Turn "unsigned type" to "utype"
546 pos = baseTypeName.find("unsigned");
547 if (pos != std::string::npos)
548 baseTypeName.erase(pos+1, 8);
550 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
552 // Get argument type qualifiers:
553 if (ty.isConstQualified())
555 if (ty.isVolatileQualified())
556 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
561 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
563 // Get image and pipe access qualifier:
564 // FIXME: now image and pipe share the same access qualifier maybe we can
565 // refine it to OpenCL access qualifier and also handle write_read
566 if (ty->isImageType()|| ty->isPipeType()) {
567 const OpenCLImageAccessAttr *A = parm->getAttr<OpenCLImageAccessAttr>();
568 if (A && A->isWriteOnly())
569 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
571 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
572 // FIXME: what about read_write?
574 accessQuals.push_back(llvm::MDString::get(Context, "none"));
576 // Get argument name.
577 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
580 kernelMDArgs.push_back(llvm::MDNode::get(Context, addressQuals));
581 kernelMDArgs.push_back(llvm::MDNode::get(Context, accessQuals));
582 kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeNames));
583 kernelMDArgs.push_back(llvm::MDNode::get(Context, argBaseTypeNames));
584 kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeQuals));
585 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
586 kernelMDArgs.push_back(llvm::MDNode::get(Context, argNames));
589 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
592 if (!FD->hasAttr<OpenCLKernelAttr>())
595 llvm::LLVMContext &Context = getLLVMContext();
597 SmallVector<llvm::Metadata *, 5> kernelMDArgs;
598 kernelMDArgs.push_back(llvm::ConstantAsMetadata::get(Fn));
600 GenOpenCLArgMetadata(FD, Fn, CGM, Context, kernelMDArgs, Builder,
603 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
604 QualType hintQTy = A->getTypeHint();
605 const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
606 bool isSignedInteger =
607 hintQTy->isSignedIntegerType() ||
608 (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
609 llvm::Metadata *attrMDArgs[] = {
610 llvm::MDString::get(Context, "vec_type_hint"),
611 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
612 CGM.getTypes().ConvertType(A->getTypeHint()))),
613 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
614 llvm::IntegerType::get(Context, 32),
615 llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
616 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
619 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
620 llvm::Metadata *attrMDArgs[] = {
621 llvm::MDString::get(Context, "work_group_size_hint"),
622 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
623 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
624 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
625 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
628 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
629 llvm::Metadata *attrMDArgs[] = {
630 llvm::MDString::get(Context, "reqd_work_group_size"),
631 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
632 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
633 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
634 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
637 llvm::MDNode *kernelMDNode = llvm::MDNode::get(Context, kernelMDArgs);
638 llvm::NamedMDNode *OpenCLKernelMetadata =
639 CGM.getModule().getOrInsertNamedMetadata("opencl.kernels");
640 OpenCLKernelMetadata->addOperand(kernelMDNode);
643 /// Determine whether the function F ends with a return stmt.
644 static bool endsWithReturn(const Decl* F) {
645 const Stmt *Body = nullptr;
646 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
647 Body = FD->getBody();
648 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
649 Body = OMD->getBody();
651 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
652 auto LastStmt = CS->body_rbegin();
653 if (LastStmt != CS->body_rend())
654 return isa<ReturnStmt>(*LastStmt);
659 void CodeGenFunction::StartFunction(GlobalDecl GD,
662 const CGFunctionInfo &FnInfo,
663 const FunctionArgList &Args,
665 SourceLocation StartLoc) {
667 "Do not use a CodeGenFunction object for more than one function");
669 const Decl *D = GD.getDecl();
671 DidCallStackSave = false;
673 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
677 assert(CurFn->isDeclaration() && "Function already has body?");
679 if (CGM.isInSanitizerBlacklist(Fn, Loc))
683 // Apply the no_sanitize* attributes to SanOpts.
684 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
685 SanOpts.Mask &= ~Attr->getMask();
688 // Apply sanitizer attributes to the function.
689 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
690 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
691 if (SanOpts.has(SanitizerKind::Thread))
692 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
693 if (SanOpts.has(SanitizerKind::Memory))
694 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
695 if (SanOpts.has(SanitizerKind::SafeStack))
696 Fn->addFnAttr(llvm::Attribute::SafeStack);
698 // Pass inline keyword to optimizer if it appears explicitly on any
699 // declaration. Also, in the case of -fno-inline attach NoInline
700 // attribute to all function that are not marked AlwaysInline.
701 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
702 if (!CGM.getCodeGenOpts().NoInline) {
703 for (auto RI : FD->redecls())
704 if (RI->isInlineSpecified()) {
705 Fn->addFnAttr(llvm::Attribute::InlineHint);
708 } else if (!FD->hasAttr<AlwaysInlineAttr>())
709 Fn->addFnAttr(llvm::Attribute::NoInline);
712 if (getLangOpts().OpenCL) {
713 // Add metadata for a kernel function.
714 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
715 EmitOpenCLKernelMetadata(FD, Fn);
718 // If we are checking function types, emit a function type signature as
720 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
721 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
722 if (llvm::Constant *PrologueSig =
723 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
724 llvm::Constant *FTRTTIConst =
725 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
726 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
727 llvm::Constant *PrologueStructConst =
728 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
729 Fn->setPrologueData(PrologueStructConst);
734 // If we're in C++ mode and the function name is "main", it is guaranteed
735 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
736 // used within a program").
737 if (getLangOpts().CPlusPlus)
738 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
740 Fn->addFnAttr(llvm::Attribute::NoRecurse);
742 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
744 // Create a marker to make it easy to insert allocas into the entryblock
745 // later. Don't create this with the builder, because we don't want it
747 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
748 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "", EntryBB);
749 if (Builder.isNamePreserving())
750 AllocaInsertPt->setName("allocapt");
752 ReturnBlock = getJumpDestInCurrentScope("return");
754 Builder.SetInsertPoint(EntryBB);
756 // Emit subprogram debug descriptor.
757 if (CGDebugInfo *DI = getDebugInfo()) {
758 SmallVector<QualType, 16> ArgTypes;
759 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
761 ArgTypes.push_back((*i)->getType());
765 getContext().getFunctionType(RetTy, ArgTypes,
766 FunctionProtoType::ExtProtoInfo());
767 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
770 if (ShouldInstrumentFunction())
771 EmitFunctionInstrumentation("__cyg_profile_func_enter");
773 if (CGM.getCodeGenOpts().InstrumentForProfiling)
774 EmitMCountInstrumentation();
776 if (RetTy->isVoidType()) {
777 // Void type; nothing to return.
778 ReturnValue = Address::invalid();
780 // Count the implicit return.
781 if (!endsWithReturn(D))
783 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
784 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
785 // Indirect aggregate return; emit returned value directly into sret slot.
786 // This reduces code size, and affects correctness in C++.
787 auto AI = CurFn->arg_begin();
788 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
790 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
791 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
792 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
793 // Load the sret pointer from the argument struct and return into that.
794 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
795 llvm::Function::arg_iterator EI = CurFn->arg_end();
797 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
798 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
799 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
801 ReturnValue = CreateIRTemp(RetTy, "retval");
803 // Tell the epilog emitter to autorelease the result. We do this
804 // now so that various specialized functions can suppress it
805 // during their IR-generation.
806 if (getLangOpts().ObjCAutoRefCount &&
807 !CurFnInfo->isReturnsRetained() &&
808 RetTy->isObjCRetainableType())
809 AutoreleaseResult = true;
812 EmitStartEHSpec(CurCodeDecl);
814 PrologueCleanupDepth = EHStack.stable_begin();
815 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
817 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
818 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
819 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
820 if (MD->getParent()->isLambda() &&
821 MD->getOverloadedOperator() == OO_Call) {
822 // We're in a lambda; figure out the captures.
823 MD->getParent()->getCaptureFields(LambdaCaptureFields,
824 LambdaThisCaptureField);
825 if (LambdaThisCaptureField) {
826 // If this lambda captures this, load it.
827 LValue ThisLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
828 CXXThisValue = EmitLoadOfLValue(ThisLValue,
829 SourceLocation()).getScalarVal();
831 for (auto *FD : MD->getParent()->fields()) {
832 if (FD->hasCapturedVLAType()) {
833 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
834 SourceLocation()).getScalarVal();
835 auto VAT = FD->getCapturedVLAType();
836 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
840 // Not in a lambda; just use 'this' from the method.
841 // FIXME: Should we generate a new load for each use of 'this'? The
842 // fast register allocator would be happier...
843 CXXThisValue = CXXABIThisValue;
847 // If any of the arguments have a variably modified type, make sure to
848 // emit the type size.
849 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
851 const VarDecl *VD = *i;
853 // Dig out the type as written from ParmVarDecls; it's unclear whether
854 // the standard (C99 6.9.1p10) requires this, but we're following the
855 // precedent set by gcc.
857 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
858 Ty = PVD->getOriginalType();
862 if (Ty->isVariablyModifiedType())
863 EmitVariablyModifiedType(Ty);
865 // Emit a location at the end of the prologue.
866 if (CGDebugInfo *DI = getDebugInfo())
867 DI->EmitLocation(Builder, StartLoc);
870 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
872 incrementProfileCounter(Body);
873 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
874 EmitCompoundStmtWithoutScope(*S);
879 /// When instrumenting to collect profile data, the counts for some blocks
880 /// such as switch cases need to not include the fall-through counts, so
881 /// emit a branch around the instrumentation code. When not instrumenting,
882 /// this just calls EmitBlock().
883 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
885 llvm::BasicBlock *SkipCountBB = nullptr;
886 if (HaveInsertPoint() && CGM.getCodeGenOpts().ProfileInstrGenerate) {
887 // When instrumenting for profiling, the fallthrough to certain
888 // statements needs to skip over the instrumentation code so that we
889 // get an accurate count.
890 SkipCountBB = createBasicBlock("skipcount");
891 EmitBranch(SkipCountBB);
894 uint64_t CurrentCount = getCurrentProfileCount();
895 incrementProfileCounter(S);
896 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
898 EmitBlock(SkipCountBB);
901 /// Tries to mark the given function nounwind based on the
902 /// non-existence of any throwing calls within it. We believe this is
903 /// lightweight enough to do at -O0.
904 static void TryMarkNoThrow(llvm::Function *F) {
905 // LLVM treats 'nounwind' on a function as part of the type, so we
906 // can't do this on functions that can be overwritten.
907 if (F->mayBeOverridden()) return;
909 for (llvm::BasicBlock &BB : *F)
910 for (llvm::Instruction &I : BB)
914 F->setDoesNotThrow();
917 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
918 const CGFunctionInfo &FnInfo) {
919 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
921 // Check if we should generate debug info for this function.
922 if (FD->hasAttr<NoDebugAttr>())
923 DebugInfo = nullptr; // disable debug info indefinitely for this function
925 FunctionArgList Args;
926 QualType ResTy = FD->getReturnType();
929 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
930 if (MD && MD->isInstance()) {
931 if (CGM.getCXXABI().HasThisReturn(GD))
932 ResTy = MD->getThisType(getContext());
933 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
934 ResTy = CGM.getContext().VoidPtrTy;
935 CGM.getCXXABI().buildThisParam(*this, Args);
938 for (auto *Param : FD->params()) {
939 Args.push_back(Param);
940 if (!Param->hasAttr<PassObjectSizeAttr>())
943 IdentifierInfo *NoID = nullptr;
944 auto *Implicit = ImplicitParamDecl::Create(
945 getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
946 getContext().getSizeType());
947 SizeArguments[Param] = Implicit;
948 Args.push_back(Implicit);
951 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
952 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
954 SourceRange BodyRange;
955 if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
956 CurEHLocation = BodyRange.getEnd();
958 // Use the location of the start of the function to determine where
959 // the function definition is located. By default use the location
960 // of the declaration as the location for the subprogram. A function
961 // may lack a declaration in the source code if it is created by code
962 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
963 SourceLocation Loc = FD->getLocation();
965 // If this is a function specialization then use the pattern body
966 // as the location for the function.
967 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
968 if (SpecDecl->hasBody(SpecDecl))
969 Loc = SpecDecl->getLocation();
971 // Emit the standard function prologue.
972 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
974 // Generate the body of the function.
975 PGO.assignRegionCounters(GD, CurFn);
976 if (isa<CXXDestructorDecl>(FD))
977 EmitDestructorBody(Args);
978 else if (isa<CXXConstructorDecl>(FD))
979 EmitConstructorBody(Args);
980 else if (getLangOpts().CUDA &&
981 !getLangOpts().CUDAIsDevice &&
982 FD->hasAttr<CUDAGlobalAttr>())
983 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
984 else if (isa<CXXConversionDecl>(FD) &&
985 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
986 // The lambda conversion to block pointer is special; the semantics can't be
987 // expressed in the AST, so IRGen needs to special-case it.
988 EmitLambdaToBlockPointerBody(Args);
989 } else if (isa<CXXMethodDecl>(FD) &&
990 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
991 // The lambda static invoker function is special, because it forwards or
992 // clones the body of the function call operator (but is actually static).
993 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
994 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
995 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
996 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
997 // Implicit copy-assignment gets the same special treatment as implicit
998 // copy-constructors.
999 emitImplicitAssignmentOperatorBody(Args);
1000 } else if (Stmt *Body = FD->getBody()) {
1001 EmitFunctionBody(Args, Body);
1003 llvm_unreachable("no definition for emitted function");
1005 // C++11 [stmt.return]p2:
1006 // Flowing off the end of a function [...] results in undefined behavior in
1007 // a value-returning function.
1009 // If the '}' that terminates a function is reached, and the value of the
1010 // function call is used by the caller, the behavior is undefined.
1011 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1012 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1013 if (SanOpts.has(SanitizerKind::Return)) {
1014 SanitizerScope SanScope(this);
1015 llvm::Value *IsFalse = Builder.getFalse();
1016 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1017 "missing_return", EmitCheckSourceLocation(FD->getLocation()),
1019 } else if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
1020 EmitTrapCall(llvm::Intrinsic::trap);
1022 Builder.CreateUnreachable();
1023 Builder.ClearInsertionPoint();
1026 // Emit the standard function epilogue.
1027 FinishFunction(BodyRange.getEnd());
1029 // If we haven't marked the function nothrow through other means, do
1030 // a quick pass now to see if we can.
1031 if (!CurFn->doesNotThrow())
1032 TryMarkNoThrow(CurFn);
1035 /// ContainsLabel - Return true if the statement contains a label in it. If
1036 /// this statement is not executed normally, it not containing a label means
1037 /// that we can just remove the code.
1038 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1039 // Null statement, not a label!
1040 if (!S) return false;
1042 // If this is a label, we have to emit the code, consider something like:
1043 // if (0) { ... foo: bar(); } goto foo;
1045 // TODO: If anyone cared, we could track __label__'s, since we know that you
1046 // can't jump to one from outside their declared region.
1047 if (isa<LabelStmt>(S))
1050 // If this is a case/default statement, and we haven't seen a switch, we have
1051 // to emit the code.
1052 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1055 // If this is a switch statement, we want to ignore cases below it.
1056 if (isa<SwitchStmt>(S))
1057 IgnoreCaseStmts = true;
1059 // Scan subexpressions for verboten labels.
1060 for (const Stmt *SubStmt : S->children())
1061 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1067 /// containsBreak - Return true if the statement contains a break out of it.
1068 /// If the statement (recursively) contains a switch or loop with a break
1069 /// inside of it, this is fine.
1070 bool CodeGenFunction::containsBreak(const Stmt *S) {
1071 // Null statement, not a label!
1072 if (!S) return false;
1074 // If this is a switch or loop that defines its own break scope, then we can
1075 // include it and anything inside of it.
1076 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1080 if (isa<BreakStmt>(S))
1083 // Scan subexpressions for verboten breaks.
1084 for (const Stmt *SubStmt : S->children())
1085 if (containsBreak(SubStmt))
1092 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1093 /// to a constant, or if it does but contains a label, return false. If it
1094 /// constant folds return true and set the boolean result in Result.
1095 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1097 llvm::APSInt ResultInt;
1098 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt))
1101 ResultBool = ResultInt.getBoolValue();
1105 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1106 /// to a constant, or if it does but contains a label, return false. If it
1107 /// constant folds return true and set the folded value.
1108 bool CodeGenFunction::
1109 ConstantFoldsToSimpleInteger(const Expr *Cond, llvm::APSInt &ResultInt) {
1110 // FIXME: Rename and handle conversion of other evaluatable things
1113 if (!Cond->EvaluateAsInt(Int, getContext()))
1114 return false; // Not foldable, not integer or not fully evaluatable.
1116 if (CodeGenFunction::ContainsLabel(Cond))
1117 return false; // Contains a label.
1125 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1126 /// statement) to the specified blocks. Based on the condition, this might try
1127 /// to simplify the codegen of the conditional based on the branch.
1129 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1130 llvm::BasicBlock *TrueBlock,
1131 llvm::BasicBlock *FalseBlock,
1132 uint64_t TrueCount) {
1133 Cond = Cond->IgnoreParens();
1135 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1137 // Handle X && Y in a condition.
1138 if (CondBOp->getOpcode() == BO_LAnd) {
1139 // If we have "1 && X", simplify the code. "0 && X" would have constant
1140 // folded if the case was simple enough.
1141 bool ConstantBool = false;
1142 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1144 // br(1 && X) -> br(X).
1145 incrementProfileCounter(CondBOp);
1146 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1150 // If we have "X && 1", simplify the code to use an uncond branch.
1151 // "X && 0" would have been constant folded to 0.
1152 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1154 // br(X && 1) -> br(X).
1155 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1159 // Emit the LHS as a conditional. If the LHS conditional is false, we
1160 // want to jump to the FalseBlock.
1161 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1162 // The counter tells us how often we evaluate RHS, and all of TrueCount
1163 // can be propagated to that branch.
1164 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1166 ConditionalEvaluation eval(*this);
1168 ApplyDebugLocation DL(*this, Cond);
1169 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1173 incrementProfileCounter(CondBOp);
1174 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1176 // Any temporaries created here are conditional.
1178 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1184 if (CondBOp->getOpcode() == BO_LOr) {
1185 // If we have "0 || X", simplify the code. "1 || X" would have constant
1186 // folded if the case was simple enough.
1187 bool ConstantBool = false;
1188 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1190 // br(0 || X) -> br(X).
1191 incrementProfileCounter(CondBOp);
1192 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1196 // If we have "X || 0", simplify the code to use an uncond branch.
1197 // "X || 1" would have been constant folded to 1.
1198 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1200 // br(X || 0) -> br(X).
1201 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1205 // Emit the LHS as a conditional. If the LHS conditional is true, we
1206 // want to jump to the TrueBlock.
1207 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1208 // We have the count for entry to the RHS and for the whole expression
1209 // being true, so we can divy up True count between the short circuit and
1212 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1213 uint64_t RHSCount = TrueCount - LHSCount;
1215 ConditionalEvaluation eval(*this);
1217 ApplyDebugLocation DL(*this, Cond);
1218 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1219 EmitBlock(LHSFalse);
1222 incrementProfileCounter(CondBOp);
1223 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1225 // Any temporaries created here are conditional.
1227 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1235 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1236 // br(!x, t, f) -> br(x, f, t)
1237 if (CondUOp->getOpcode() == UO_LNot) {
1238 // Negate the count.
1239 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1240 // Negate the condition and swap the destination blocks.
1241 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1246 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1247 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1248 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1249 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1251 ConditionalEvaluation cond(*this);
1252 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1253 getProfileCount(CondOp));
1255 // When computing PGO branch weights, we only know the overall count for
1256 // the true block. This code is essentially doing tail duplication of the
1257 // naive code-gen, introducing new edges for which counts are not
1258 // available. Divide the counts proportionally between the LHS and RHS of
1259 // the conditional operator.
1260 uint64_t LHSScaledTrueCount = 0;
1263 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1264 LHSScaledTrueCount = TrueCount * LHSRatio;
1268 EmitBlock(LHSBlock);
1269 incrementProfileCounter(CondOp);
1271 ApplyDebugLocation DL(*this, Cond);
1272 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1273 LHSScaledTrueCount);
1278 EmitBlock(RHSBlock);
1279 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1280 TrueCount - LHSScaledTrueCount);
1286 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1287 // Conditional operator handling can give us a throw expression as a
1288 // condition for a case like:
1289 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1291 // br(c, throw x, br(y, t, f))
1292 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1296 // If the branch has a condition wrapped by __builtin_unpredictable,
1297 // create metadata that specifies that the branch is unpredictable.
1298 // Don't bother if not optimizing because that metadata would not be used.
1299 llvm::MDNode *Unpredictable = nullptr;
1300 if (CGM.getCodeGenOpts().OptimizationLevel != 0) {
1301 if (const CallExpr *Call = dyn_cast<CallExpr>(Cond)) {
1302 const Decl *TargetDecl = Call->getCalleeDecl();
1303 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl)) {
1304 if (FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1305 llvm::MDBuilder MDHelper(getLLVMContext());
1306 Unpredictable = MDHelper.createUnpredictable();
1312 // Create branch weights based on the number of times we get here and the
1313 // number of times the condition should be true.
1314 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1315 llvm::MDNode *Weights =
1316 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1318 // Emit the code with the fully general case.
1321 ApplyDebugLocation DL(*this, Cond);
1322 CondV = EvaluateExprAsBool(Cond);
1324 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1327 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1328 /// specified stmt yet.
1329 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1330 CGM.ErrorUnsupported(S, Type);
1333 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1334 /// variable-length array whose elements have a non-zero bit-pattern.
1336 /// \param baseType the inner-most element type of the array
1337 /// \param src - a char* pointing to the bit-pattern for a single
1338 /// base element of the array
1339 /// \param sizeInChars - the total size of the VLA, in chars
1340 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1341 Address dest, Address src,
1342 llvm::Value *sizeInChars) {
1343 CGBuilderTy &Builder = CGF.Builder;
1345 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1346 llvm::Value *baseSizeInChars
1347 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1350 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1352 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1354 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1355 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1356 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1358 // Make a loop over the VLA. C99 guarantees that the VLA element
1359 // count must be nonzero.
1360 CGF.EmitBlock(loopBB);
1362 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1363 cur->addIncoming(begin.getPointer(), originBB);
1365 CharUnits curAlign =
1366 dest.getAlignment().alignmentOfArrayElement(baseSize);
1368 // memcpy the individual element bit-pattern.
1369 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1370 /*volatile*/ false);
1372 // Go to the next element.
1374 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1376 // Leave if that's the end of the VLA.
1377 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1378 Builder.CreateCondBr(done, contBB, loopBB);
1379 cur->addIncoming(next, loopBB);
1381 CGF.EmitBlock(contBB);
1385 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1386 // Ignore empty classes in C++.
1387 if (getLangOpts().CPlusPlus) {
1388 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1389 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1394 // Cast the dest ptr to the appropriate i8 pointer type.
1395 if (DestPtr.getElementType() != Int8Ty)
1396 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1398 // Get size and alignment info for this aggregate.
1399 CharUnits size = getContext().getTypeSizeInChars(Ty);
1401 llvm::Value *SizeVal;
1402 const VariableArrayType *vla;
1404 // Don't bother emitting a zero-byte memset.
1405 if (size.isZero()) {
1406 // But note that getTypeInfo returns 0 for a VLA.
1407 if (const VariableArrayType *vlaType =
1408 dyn_cast_or_null<VariableArrayType>(
1409 getContext().getAsArrayType(Ty))) {
1411 llvm::Value *numElts;
1412 std::tie(numElts, eltType) = getVLASize(vlaType);
1415 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1416 if (!eltSize.isOne())
1417 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1423 SizeVal = CGM.getSize(size);
1427 // If the type contains a pointer to data member we can't memset it to zero.
1428 // Instead, create a null constant and copy it to the destination.
1429 // TODO: there are other patterns besides zero that we can usefully memset,
1430 // like -1, which happens to be the pattern used by member-pointers.
1431 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1432 // For a VLA, emit a single element, then splat that over the VLA.
1433 if (vla) Ty = getContext().getBaseElementType(vla);
1435 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1437 llvm::GlobalVariable *NullVariable =
1438 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1439 /*isConstant=*/true,
1440 llvm::GlobalVariable::PrivateLinkage,
1441 NullConstant, Twine());
1442 CharUnits NullAlign = DestPtr.getAlignment();
1443 NullVariable->setAlignment(NullAlign.getQuantity());
1444 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1447 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1449 // Get and call the appropriate llvm.memcpy overload.
1450 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1454 // Otherwise, just memset the whole thing to zero. This is legal
1455 // because in LLVM, all default initializers (other than the ones we just
1456 // handled above) are guaranteed to have a bit pattern of all zeros.
1457 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1460 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1461 // Make sure that there is a block for the indirect goto.
1462 if (!IndirectBranch)
1463 GetIndirectGotoBlock();
1465 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1467 // Make sure the indirect branch includes all of the address-taken blocks.
1468 IndirectBranch->addDestination(BB);
1469 return llvm::BlockAddress::get(CurFn, BB);
1472 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1473 // If we already made the indirect branch for indirect goto, return its block.
1474 if (IndirectBranch) return IndirectBranch->getParent();
1476 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1478 // Create the PHI node that indirect gotos will add entries to.
1479 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1480 "indirect.goto.dest");
1482 // Create the indirect branch instruction.
1483 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1484 return IndirectBranch->getParent();
1487 /// Computes the length of an array in elements, as well as the base
1488 /// element type and a properly-typed first element pointer.
1489 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1492 const ArrayType *arrayType = origArrayType;
1494 // If it's a VLA, we have to load the stored size. Note that
1495 // this is the size of the VLA in bytes, not its size in elements.
1496 llvm::Value *numVLAElements = nullptr;
1497 if (isa<VariableArrayType>(arrayType)) {
1498 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1500 // Walk into all VLAs. This doesn't require changes to addr,
1501 // which has type T* where T is the first non-VLA element type.
1503 QualType elementType = arrayType->getElementType();
1504 arrayType = getContext().getAsArrayType(elementType);
1506 // If we only have VLA components, 'addr' requires no adjustment.
1508 baseType = elementType;
1509 return numVLAElements;
1511 } while (isa<VariableArrayType>(arrayType));
1513 // We get out here only if we find a constant array type
1517 // We have some number of constant-length arrays, so addr should
1518 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1519 // down to the first element of addr.
1520 SmallVector<llvm::Value*, 8> gepIndices;
1522 // GEP down to the array type.
1523 llvm::ConstantInt *zero = Builder.getInt32(0);
1524 gepIndices.push_back(zero);
1526 uint64_t countFromCLAs = 1;
1529 llvm::ArrayType *llvmArrayType =
1530 dyn_cast<llvm::ArrayType>(addr.getElementType());
1531 while (llvmArrayType) {
1532 assert(isa<ConstantArrayType>(arrayType));
1533 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1534 == llvmArrayType->getNumElements());
1536 gepIndices.push_back(zero);
1537 countFromCLAs *= llvmArrayType->getNumElements();
1538 eltType = arrayType->getElementType();
1541 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1542 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1543 assert((!llvmArrayType || arrayType) &&
1544 "LLVM and Clang types are out-of-synch");
1548 // From this point onwards, the Clang array type has been emitted
1549 // as some other type (probably a packed struct). Compute the array
1550 // size, and just emit the 'begin' expression as a bitcast.
1553 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1554 eltType = arrayType->getElementType();
1555 arrayType = getContext().getAsArrayType(eltType);
1558 llvm::Type *baseType = ConvertType(eltType);
1559 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1561 // Create the actual GEP.
1562 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1563 gepIndices, "array.begin"),
1564 addr.getAlignment());
1569 llvm::Value *numElements
1570 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1572 // If we had any VLA dimensions, factor them in.
1574 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1579 std::pair<llvm::Value*, QualType>
1580 CodeGenFunction::getVLASize(QualType type) {
1581 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1582 assert(vla && "type was not a variable array type!");
1583 return getVLASize(vla);
1586 std::pair<llvm::Value*, QualType>
1587 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1588 // The number of elements so far; always size_t.
1589 llvm::Value *numElements = nullptr;
1591 QualType elementType;
1593 elementType = type->getElementType();
1594 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1595 assert(vlaSize && "no size for VLA!");
1596 assert(vlaSize->getType() == SizeTy);
1599 numElements = vlaSize;
1601 // It's undefined behavior if this wraps around, so mark it that way.
1602 // FIXME: Teach -fsanitize=undefined to trap this.
1603 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1605 } while ((type = getContext().getAsVariableArrayType(elementType)));
1607 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1610 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1611 assert(type->isVariablyModifiedType() &&
1612 "Must pass variably modified type to EmitVLASizes!");
1614 EnsureInsertPoint();
1616 // We're going to walk down into the type and look for VLA
1619 assert(type->isVariablyModifiedType());
1621 const Type *ty = type.getTypePtr();
1622 switch (ty->getTypeClass()) {
1624 #define TYPE(Class, Base)
1625 #define ABSTRACT_TYPE(Class, Base)
1626 #define NON_CANONICAL_TYPE(Class, Base)
1627 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1628 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1629 #include "clang/AST/TypeNodes.def"
1630 llvm_unreachable("unexpected dependent type!");
1632 // These types are never variably-modified.
1636 case Type::ExtVector:
1639 case Type::Elaborated:
1640 case Type::TemplateSpecialization:
1641 case Type::ObjCObject:
1642 case Type::ObjCInterface:
1643 case Type::ObjCObjectPointer:
1644 llvm_unreachable("type class is never variably-modified!");
1646 case Type::Adjusted:
1647 type = cast<AdjustedType>(ty)->getAdjustedType();
1651 type = cast<DecayedType>(ty)->getPointeeType();
1655 type = cast<PointerType>(ty)->getPointeeType();
1658 case Type::BlockPointer:
1659 type = cast<BlockPointerType>(ty)->getPointeeType();
1662 case Type::LValueReference:
1663 case Type::RValueReference:
1664 type = cast<ReferenceType>(ty)->getPointeeType();
1667 case Type::MemberPointer:
1668 type = cast<MemberPointerType>(ty)->getPointeeType();
1671 case Type::ConstantArray:
1672 case Type::IncompleteArray:
1673 // Losing element qualification here is fine.
1674 type = cast<ArrayType>(ty)->getElementType();
1677 case Type::VariableArray: {
1678 // Losing element qualification here is fine.
1679 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1681 // Unknown size indication requires no size computation.
1682 // Otherwise, evaluate and record it.
1683 if (const Expr *size = vat->getSizeExpr()) {
1684 // It's possible that we might have emitted this already,
1685 // e.g. with a typedef and a pointer to it.
1686 llvm::Value *&entry = VLASizeMap[size];
1688 llvm::Value *Size = EmitScalarExpr(size);
1691 // If the size is an expression that is not an integer constant
1692 // expression [...] each time it is evaluated it shall have a value
1693 // greater than zero.
1694 if (SanOpts.has(SanitizerKind::VLABound) &&
1695 size->getType()->isSignedIntegerType()) {
1696 SanitizerScope SanScope(this);
1697 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1698 llvm::Constant *StaticArgs[] = {
1699 EmitCheckSourceLocation(size->getLocStart()),
1700 EmitCheckTypeDescriptor(size->getType())
1702 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1703 SanitizerKind::VLABound),
1704 "vla_bound_not_positive", StaticArgs, Size);
1707 // Always zexting here would be wrong if it weren't
1708 // undefined behavior to have a negative bound.
1709 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1712 type = vat->getElementType();
1716 case Type::FunctionProto:
1717 case Type::FunctionNoProto:
1718 type = cast<FunctionType>(ty)->getReturnType();
1723 case Type::UnaryTransform:
1724 case Type::Attributed:
1725 case Type::SubstTemplateTypeParm:
1726 case Type::PackExpansion:
1727 // Keep walking after single level desugaring.
1728 type = type.getSingleStepDesugaredType(getContext());
1732 case Type::Decltype:
1734 // Stop walking: nothing to do.
1737 case Type::TypeOfExpr:
1738 // Stop walking: emit typeof expression.
1739 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1743 type = cast<AtomicType>(ty)->getValueType();
1747 type = cast<PipeType>(ty)->getElementType();
1750 } while (type->isVariablyModifiedType());
1753 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1754 if (getContext().getBuiltinVaListType()->isArrayType())
1755 return EmitPointerWithAlignment(E);
1756 return EmitLValue(E).getAddress();
1759 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1760 return EmitLValue(E).getAddress();
1763 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1764 llvm::Constant *Init) {
1765 assert (Init && "Invalid DeclRefExpr initializer!");
1766 if (CGDebugInfo *Dbg = getDebugInfo())
1767 if (CGM.getCodeGenOpts().getDebugInfo() >= CodeGenOptions::LimitedDebugInfo)
1768 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1771 CodeGenFunction::PeepholeProtection
1772 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1773 // At the moment, the only aggressive peephole we do in IR gen
1774 // is trunc(zext) folding, but if we add more, we can easily
1775 // extend this protection.
1777 if (!rvalue.isScalar()) return PeepholeProtection();
1778 llvm::Value *value = rvalue.getScalarVal();
1779 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1781 // Just make an extra bitcast.
1782 assert(HaveInsertPoint());
1783 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1784 Builder.GetInsertBlock());
1786 PeepholeProtection protection;
1787 protection.Inst = inst;
1791 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1792 if (!protection.Inst) return;
1794 // In theory, we could try to duplicate the peepholes now, but whatever.
1795 protection.Inst->eraseFromParent();
1798 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1799 llvm::Value *AnnotatedVal,
1800 StringRef AnnotationStr,
1801 SourceLocation Location) {
1802 llvm::Value *Args[4] = {
1804 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1805 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1806 CGM.EmitAnnotationLineNo(Location)
1808 return Builder.CreateCall(AnnotationFn, Args);
1811 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1812 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1813 // FIXME We create a new bitcast for every annotation because that's what
1814 // llvm-gcc was doing.
1815 for (const auto *I : D->specific_attrs<AnnotateAttr>())
1816 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1817 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1818 I->getAnnotation(), D->getLocation());
1821 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1823 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1824 llvm::Value *V = Addr.getPointer();
1825 llvm::Type *VTy = V->getType();
1826 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1829 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1830 // FIXME Always emit the cast inst so we can differentiate between
1831 // annotation on the first field of a struct and annotation on the struct
1833 if (VTy != CGM.Int8PtrTy)
1834 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1835 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1836 V = Builder.CreateBitCast(V, VTy);
1839 return Address(V, Addr.getAlignment());
1842 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
1844 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
1846 assert(!CGF->IsSanitizerScope);
1847 CGF->IsSanitizerScope = true;
1850 CodeGenFunction::SanitizerScope::~SanitizerScope() {
1851 CGF->IsSanitizerScope = false;
1854 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
1855 const llvm::Twine &Name,
1856 llvm::BasicBlock *BB,
1857 llvm::BasicBlock::iterator InsertPt) const {
1858 LoopStack.InsertHelper(I);
1859 if (IsSanitizerScope)
1860 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
1863 template <bool PreserveNames>
1864 void CGBuilderInserter<PreserveNames>::InsertHelper(
1865 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1866 llvm::BasicBlock::iterator InsertPt) const {
1867 llvm::IRBuilderDefaultInserter<PreserveNames>::InsertHelper(I, Name, BB,
1870 CGF->InsertHelper(I, Name, BB, InsertPt);
1874 #define PreserveNames false
1876 #define PreserveNames true
1878 template void CGBuilderInserter<PreserveNames>::InsertHelper(
1879 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1880 llvm::BasicBlock::iterator InsertPt) const;
1881 #undef PreserveNames
1883 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
1884 CodeGenModule &CGM, const FunctionDecl *FD,
1885 std::string &FirstMissing) {
1886 // If there aren't any required features listed then go ahead and return.
1887 if (ReqFeatures.empty())
1890 // Now build up the set of caller features and verify that all the required
1891 // features are there.
1892 llvm::StringMap<bool> CallerFeatureMap;
1893 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
1895 // If we have at least one of the features in the feature list return
1896 // true, otherwise return false.
1898 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
1899 SmallVector<StringRef, 1> OrFeatures;
1900 Feature.split(OrFeatures, "|");
1901 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
1902 [&](StringRef Feature) {
1903 if (!CallerFeatureMap.lookup(Feature)) {
1904 FirstMissing = Feature.str();
1912 // Emits an error if we don't have a valid set of target features for the
1914 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
1915 const FunctionDecl *TargetDecl) {
1916 // Early exit if this is an indirect call.
1920 // Get the current enclosing function if it exists. If it doesn't
1921 // we can't check the target features anyhow.
1922 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
1926 // Grab the required features for the call. For a builtin this is listed in
1927 // the td file with the default cpu, for an always_inline function this is any
1928 // listed cpu and any listed features.
1929 unsigned BuiltinID = TargetDecl->getBuiltinID();
1930 std::string MissingFeature;
1932 SmallVector<StringRef, 1> ReqFeatures;
1933 const char *FeatureList =
1934 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1935 // Return if the builtin doesn't have any required features.
1936 if (!FeatureList || StringRef(FeatureList) == "")
1938 StringRef(FeatureList).split(ReqFeatures, ",");
1939 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1940 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
1941 << TargetDecl->getDeclName()
1942 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1944 } else if (TargetDecl->hasAttr<TargetAttr>()) {
1945 // Get the required features for the callee.
1946 SmallVector<StringRef, 1> ReqFeatures;
1947 llvm::StringMap<bool> CalleeFeatureMap;
1948 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
1949 for (const auto &F : CalleeFeatureMap) {
1950 // Only positive features are "required".
1952 ReqFeatures.push_back(F.getKey());
1954 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1955 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
1956 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;