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/AST/StmtObjC.h"
29 #include "clang/Basic/Builtins.h"
30 #include "clang/Basic/TargetInfo.h"
31 #include "clang/CodeGen/CGFunctionInfo.h"
32 #include "clang/Frontend/CodeGenOptions.h"
33 #include "clang/Sema/SemaDiagnostic.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/Intrinsics.h"
36 #include "llvm/IR/MDBuilder.h"
37 #include "llvm/IR/Operator.h"
38 using namespace clang;
39 using namespace CodeGen;
41 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
42 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
43 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
44 CGBuilderInserterTy(this)),
45 CurFn(nullptr), ReturnValue(Address::invalid()),
46 CapturedStmtInfo(nullptr),
47 SanOpts(CGM.getLangOpts().Sanitize), IsSanitizerScope(false),
48 CurFuncIsThunk(false), AutoreleaseResult(false), SawAsmBlock(false),
49 IsOutlinedSEHHelper(false),
50 BlockInfo(nullptr), BlockPointer(nullptr),
51 LambdaThisCaptureField(nullptr), NormalCleanupDest(nullptr),
52 NextCleanupDestIndex(1), FirstBlockInfo(nullptr), EHResumeBlock(nullptr),
53 ExceptionSlot(nullptr), EHSelectorSlot(nullptr),
54 DebugInfo(CGM.getModuleDebugInfo()),
55 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
56 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
57 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
58 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
59 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
60 CXXStructorImplicitParamDecl(nullptr),
61 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
62 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
63 TerminateHandler(nullptr), TrapBB(nullptr) {
64 if (!suppressNewContext)
65 CGM.getCXXABI().getMangleContext().startNewFunction();
67 llvm::FastMathFlags FMF;
68 if (CGM.getLangOpts().FastMath)
69 FMF.setUnsafeAlgebra();
70 if (CGM.getLangOpts().FiniteMathOnly) {
74 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
77 if (CGM.getCodeGenOpts().NoSignedZeros) {
78 FMF.setNoSignedZeros();
80 if (CGM.getCodeGenOpts().ReciprocalMath) {
81 FMF.setAllowReciprocal();
83 Builder.setFastMathFlags(FMF);
86 CodeGenFunction::~CodeGenFunction() {
87 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
89 // If there are any unclaimed block infos, go ahead and destroy them
90 // now. This can happen if IR-gen gets clever and skips evaluating
93 destroyBlockInfos(FirstBlockInfo);
95 if (getLangOpts().OpenMP) {
96 CGM.getOpenMPRuntime().functionFinished(*this);
100 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
101 AlignmentSource *Source) {
102 return getNaturalTypeAlignment(T->getPointeeType(), Source,
103 /*forPointee*/ true);
106 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
107 AlignmentSource *Source,
108 bool forPointeeType) {
109 // Honor alignment typedef attributes even on incomplete types.
110 // We also honor them straight for C++ class types, even as pointees;
111 // there's an expressivity gap here.
112 if (auto TT = T->getAs<TypedefType>()) {
113 if (auto Align = TT->getDecl()->getMaxAlignment()) {
114 if (Source) *Source = AlignmentSource::AttributedType;
115 return getContext().toCharUnitsFromBits(Align);
119 if (Source) *Source = AlignmentSource::Type;
122 if (T->isIncompleteType()) {
123 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
125 // For C++ class pointees, we don't know whether we're pointing at a
126 // base or a complete object, so we generally need to use the
127 // non-virtual alignment.
128 const CXXRecordDecl *RD;
129 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
130 Alignment = CGM.getClassPointerAlignment(RD);
132 Alignment = getContext().getTypeAlignInChars(T);
135 // Cap to the global maximum type alignment unless the alignment
136 // was somehow explicit on the type.
137 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
138 if (Alignment.getQuantity() > MaxAlign &&
139 !getContext().isAlignmentRequired(T))
140 Alignment = CharUnits::fromQuantity(MaxAlign);
146 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
147 AlignmentSource AlignSource;
148 CharUnits Alignment = getNaturalTypeAlignment(T, &AlignSource);
149 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), AlignSource,
153 /// Given a value of type T* that may not be to a complete object,
154 /// construct an l-value with the natural pointee alignment of T.
156 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
157 AlignmentSource AlignSource;
158 CharUnits Align = getNaturalTypeAlignment(T, &AlignSource, /*pointee*/ true);
159 return MakeAddrLValue(Address(V, Align), T, AlignSource);
163 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
164 return CGM.getTypes().ConvertTypeForMem(T);
167 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
168 return CGM.getTypes().ConvertType(T);
171 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
172 type = type.getCanonicalType();
174 switch (type->getTypeClass()) {
175 #define TYPE(name, parent)
176 #define ABSTRACT_TYPE(name, parent)
177 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
178 #define DEPENDENT_TYPE(name, parent) case Type::name:
179 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
180 #include "clang/AST/TypeNodes.def"
181 llvm_unreachable("non-canonical or dependent type in IR-generation");
184 llvm_unreachable("undeduced auto type in IR-generation");
186 // Various scalar types.
189 case Type::BlockPointer:
190 case Type::LValueReference:
191 case Type::RValueReference:
192 case Type::MemberPointer:
194 case Type::ExtVector:
195 case Type::FunctionProto:
196 case Type::FunctionNoProto:
198 case Type::ObjCObjectPointer:
206 // Arrays, records, and Objective-C objects.
207 case Type::ConstantArray:
208 case Type::IncompleteArray:
209 case Type::VariableArray:
211 case Type::ObjCObject:
212 case Type::ObjCInterface:
213 return TEK_Aggregate;
215 // We operate on atomic values according to their underlying type.
217 type = cast<AtomicType>(type)->getValueType();
220 llvm_unreachable("unknown type kind!");
224 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
225 // For cleanliness, we try to avoid emitting the return block for
227 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
230 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
232 // We have a valid insert point, reuse it if it is empty or there are no
233 // explicit jumps to the return block.
234 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
235 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
236 delete ReturnBlock.getBlock();
238 EmitBlock(ReturnBlock.getBlock());
239 return llvm::DebugLoc();
242 // Otherwise, if the return block is the target of a single direct
243 // branch then we can just put the code in that block instead. This
244 // cleans up functions which started with a unified return block.
245 if (ReturnBlock.getBlock()->hasOneUse()) {
246 llvm::BranchInst *BI =
247 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
248 if (BI && BI->isUnconditional() &&
249 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
250 // Record/return the DebugLoc of the simple 'return' expression to be used
251 // later by the actual 'ret' instruction.
252 llvm::DebugLoc Loc = BI->getDebugLoc();
253 Builder.SetInsertPoint(BI->getParent());
254 BI->eraseFromParent();
255 delete ReturnBlock.getBlock();
260 // FIXME: We are at an unreachable point, there is no reason to emit the block
261 // unless it has uses. However, we still need a place to put the debug
262 // region.end for now.
264 EmitBlock(ReturnBlock.getBlock());
265 return llvm::DebugLoc();
268 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
270 if (!BB->use_empty())
271 return CGF.CurFn->getBasicBlockList().push_back(BB);
275 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
276 assert(BreakContinueStack.empty() &&
277 "mismatched push/pop in break/continue stack!");
279 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
280 && NumSimpleReturnExprs == NumReturnExprs
281 && ReturnBlock.getBlock()->use_empty();
282 // Usually the return expression is evaluated before the cleanup
283 // code. If the function contains only a simple return statement,
284 // such as a constant, the location before the cleanup code becomes
285 // the last useful breakpoint in the function, because the simple
286 // return expression will be evaluated after the cleanup code. To be
287 // safe, set the debug location for cleanup code to the location of
288 // the return statement. Otherwise the cleanup code should be at the
289 // end of the function's lexical scope.
291 // If there are multiple branches to the return block, the branch
292 // instructions will get the location of the return statements and
294 if (CGDebugInfo *DI = getDebugInfo()) {
295 if (OnlySimpleReturnStmts)
296 DI->EmitLocation(Builder, LastStopPoint);
298 DI->EmitLocation(Builder, EndLoc);
301 // Pop any cleanups that might have been associated with the
302 // parameters. Do this in whatever block we're currently in; it's
303 // important to do this before we enter the return block or return
304 // edges will be *really* confused.
305 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
306 bool HasOnlyLifetimeMarkers =
307 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
308 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
310 // Make sure the line table doesn't jump back into the body for
311 // the ret after it's been at EndLoc.
312 if (CGDebugInfo *DI = getDebugInfo())
313 if (OnlySimpleReturnStmts)
314 DI->EmitLocation(Builder, EndLoc);
316 PopCleanupBlocks(PrologueCleanupDepth);
319 // Emit function epilog (to return).
320 llvm::DebugLoc Loc = EmitReturnBlock();
322 if (ShouldInstrumentFunction())
323 EmitFunctionInstrumentation("__cyg_profile_func_exit");
325 // Emit debug descriptor for function end.
326 if (CGDebugInfo *DI = getDebugInfo())
327 DI->EmitFunctionEnd(Builder);
329 // Reset the debug location to that of the simple 'return' expression, if any
330 // rather than that of the end of the function's scope '}'.
331 ApplyDebugLocation AL(*this, Loc);
332 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
333 EmitEndEHSpec(CurCodeDecl);
335 assert(EHStack.empty() &&
336 "did not remove all scopes from cleanup stack!");
338 // If someone did an indirect goto, emit the indirect goto block at the end of
340 if (IndirectBranch) {
341 EmitBlock(IndirectBranch->getParent());
342 Builder.ClearInsertionPoint();
345 // If some of our locals escaped, insert a call to llvm.localescape in the
347 if (!EscapedLocals.empty()) {
348 // Invert the map from local to index into a simple vector. There should be
350 SmallVector<llvm::Value *, 4> EscapeArgs;
351 EscapeArgs.resize(EscapedLocals.size());
352 for (auto &Pair : EscapedLocals)
353 EscapeArgs[Pair.second] = Pair.first;
354 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
355 &CGM.getModule(), llvm::Intrinsic::localescape);
356 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
359 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
360 llvm::Instruction *Ptr = AllocaInsertPt;
361 AllocaInsertPt = nullptr;
362 Ptr->eraseFromParent();
364 // If someone took the address of a label but never did an indirect goto, we
365 // made a zero entry PHI node, which is illegal, zap it now.
366 if (IndirectBranch) {
367 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
368 if (PN->getNumIncomingValues() == 0) {
369 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
370 PN->eraseFromParent();
374 EmitIfUsed(*this, EHResumeBlock);
375 EmitIfUsed(*this, TerminateLandingPad);
376 EmitIfUsed(*this, TerminateHandler);
377 EmitIfUsed(*this, UnreachableBlock);
379 if (CGM.getCodeGenOpts().EmitDeclMetadata)
382 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
383 I = DeferredReplacements.begin(),
384 E = DeferredReplacements.end();
386 I->first->replaceAllUsesWith(I->second);
387 I->first->eraseFromParent();
391 /// ShouldInstrumentFunction - Return true if the current function should be
392 /// instrumented with __cyg_profile_func_* calls
393 bool CodeGenFunction::ShouldInstrumentFunction() {
394 if (!CGM.getCodeGenOpts().InstrumentFunctions)
396 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
401 /// ShouldXRayInstrument - Return true if the current function should be
402 /// instrumented with XRay nop sleds.
403 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
404 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
407 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
408 /// instrumentation function with the current function and the call site, if
409 /// function instrumentation is enabled.
410 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
411 auto NL = ApplyDebugLocation::CreateArtificial(*this);
412 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
413 llvm::PointerType *PointerTy = Int8PtrTy;
414 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
415 llvm::FunctionType *FunctionTy =
416 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
418 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
419 llvm::CallInst *CallSite = Builder.CreateCall(
420 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
421 llvm::ConstantInt::get(Int32Ty, 0),
424 llvm::Value *args[] = {
425 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
429 EmitNounwindRuntimeCall(F, args);
432 void CodeGenFunction::EmitMCountInstrumentation() {
433 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
435 llvm::Constant *MCountFn =
436 CGM.CreateRuntimeFunction(FTy, getTarget().getMCountName());
437 EmitNounwindRuntimeCall(MCountFn);
440 // Returns the address space id that should be produced to the
441 // kernel_arg_addr_space metadata. This is always fixed to the ids
442 // as specified in the SPIR 2.0 specification in order to differentiate
443 // for example in clGetKernelArgInfo() implementation between the address
444 // spaces with targets without unique mapping to the OpenCL address spaces
445 // (basically all single AS CPUs).
446 static unsigned ArgInfoAddressSpace(unsigned LangAS) {
448 case LangAS::opencl_global: return 1;
449 case LangAS::opencl_constant: return 2;
450 case LangAS::opencl_local: return 3;
451 case LangAS::opencl_generic: return 4; // Not in SPIR 2.0 specs.
453 return 0; // Assume private.
457 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
458 // information in the program executable. The argument information stored
459 // includes the argument name, its type, the address and access qualifiers used.
460 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
461 CodeGenModule &CGM, llvm::LLVMContext &Context,
462 CGBuilderTy &Builder, ASTContext &ASTCtx) {
463 // Create MDNodes that represent the kernel arg metadata.
464 // Each MDNode is a list in the form of "key", N number of values which is
465 // the same number of values as their are kernel arguments.
467 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
469 // MDNode for the kernel argument address space qualifiers.
470 SmallVector<llvm::Metadata *, 8> addressQuals;
472 // MDNode for the kernel argument access qualifiers (images only).
473 SmallVector<llvm::Metadata *, 8> accessQuals;
475 // MDNode for the kernel argument type names.
476 SmallVector<llvm::Metadata *, 8> argTypeNames;
478 // MDNode for the kernel argument base type names.
479 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
481 // MDNode for the kernel argument type qualifiers.
482 SmallVector<llvm::Metadata *, 8> argTypeQuals;
484 // MDNode for the kernel argument names.
485 SmallVector<llvm::Metadata *, 8> argNames;
487 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
488 const ParmVarDecl *parm = FD->getParamDecl(i);
489 QualType ty = parm->getType();
490 std::string typeQuals;
492 if (ty->isPointerType()) {
493 QualType pointeeTy = ty->getPointeeType();
495 // Get address qualifier.
496 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
497 ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
499 // Get argument type name.
500 std::string typeName =
501 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
503 // Turn "unsigned type" to "utype"
504 std::string::size_type pos = typeName.find("unsigned");
505 if (pointeeTy.isCanonical() && pos != std::string::npos)
506 typeName.erase(pos+1, 8);
508 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
510 std::string baseTypeName =
511 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
515 // Turn "unsigned type" to "utype"
516 pos = baseTypeName.find("unsigned");
517 if (pos != std::string::npos)
518 baseTypeName.erase(pos+1, 8);
520 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
522 // Get argument type qualifiers:
523 if (ty.isRestrictQualified())
524 typeQuals = "restrict";
525 if (pointeeTy.isConstQualified() ||
526 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
527 typeQuals += typeQuals.empty() ? "const" : " const";
528 if (pointeeTy.isVolatileQualified())
529 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
531 uint32_t AddrSpc = 0;
532 bool isPipe = ty->isPipeType();
533 if (ty->isImageType() || isPipe)
534 AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
536 addressQuals.push_back(
537 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
539 // Get argument type name.
540 std::string typeName;
542 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
543 .getAsString(Policy);
545 typeName = ty.getUnqualifiedType().getAsString(Policy);
547 // Turn "unsigned type" to "utype"
548 std::string::size_type pos = typeName.find("unsigned");
549 if (ty.isCanonical() && pos != std::string::npos)
550 typeName.erase(pos+1, 8);
552 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
554 std::string baseTypeName;
556 baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
557 ->getElementType().getCanonicalType()
558 .getAsString(Policy);
561 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
563 // Turn "unsigned type" to "utype"
564 pos = baseTypeName.find("unsigned");
565 if (pos != std::string::npos)
566 baseTypeName.erase(pos+1, 8);
568 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
570 // Get argument type qualifiers:
571 if (ty.isConstQualified())
573 if (ty.isVolatileQualified())
574 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
579 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
581 // Get image and pipe access qualifier:
582 if (ty->isImageType()|| ty->isPipeType()) {
583 const OpenCLAccessAttr *A = parm->getAttr<OpenCLAccessAttr>();
584 if (A && A->isWriteOnly())
585 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
586 else if (A && A->isReadWrite())
587 accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
589 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
591 accessQuals.push_back(llvm::MDString::get(Context, "none"));
593 // Get argument name.
594 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
597 Fn->setMetadata("kernel_arg_addr_space",
598 llvm::MDNode::get(Context, addressQuals));
599 Fn->setMetadata("kernel_arg_access_qual",
600 llvm::MDNode::get(Context, accessQuals));
601 Fn->setMetadata("kernel_arg_type",
602 llvm::MDNode::get(Context, argTypeNames));
603 Fn->setMetadata("kernel_arg_base_type",
604 llvm::MDNode::get(Context, argBaseTypeNames));
605 Fn->setMetadata("kernel_arg_type_qual",
606 llvm::MDNode::get(Context, argTypeQuals));
607 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
608 Fn->setMetadata("kernel_arg_name",
609 llvm::MDNode::get(Context, argNames));
612 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
615 if (!FD->hasAttr<OpenCLKernelAttr>())
618 llvm::LLVMContext &Context = getLLVMContext();
620 GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
622 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
623 QualType hintQTy = A->getTypeHint();
624 const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
625 bool isSignedInteger =
626 hintQTy->isSignedIntegerType() ||
627 (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
628 llvm::Metadata *attrMDArgs[] = {
629 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
630 CGM.getTypes().ConvertType(A->getTypeHint()))),
631 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
632 llvm::IntegerType::get(Context, 32),
633 llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
634 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, attrMDArgs));
637 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
638 llvm::Metadata *attrMDArgs[] = {
639 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
640 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
641 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
642 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, attrMDArgs));
645 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
646 llvm::Metadata *attrMDArgs[] = {
647 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
648 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
649 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
650 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, attrMDArgs));
654 /// Determine whether the function F ends with a return stmt.
655 static bool endsWithReturn(const Decl* F) {
656 const Stmt *Body = nullptr;
657 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
658 Body = FD->getBody();
659 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
660 Body = OMD->getBody();
662 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
663 auto LastStmt = CS->body_rbegin();
664 if (LastStmt != CS->body_rend())
665 return isa<ReturnStmt>(*LastStmt);
670 void CodeGenFunction::StartFunction(GlobalDecl GD,
673 const CGFunctionInfo &FnInfo,
674 const FunctionArgList &Args,
676 SourceLocation StartLoc) {
678 "Do not use a CodeGenFunction object for more than one function");
680 const Decl *D = GD.getDecl();
682 DidCallStackSave = false;
684 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
685 if (FD->usesSEHTry())
687 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
691 assert(CurFn->isDeclaration() && "Function already has body?");
693 if (CGM.isInSanitizerBlacklist(Fn, Loc))
697 // Apply the no_sanitize* attributes to SanOpts.
698 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
699 SanOpts.Mask &= ~Attr->getMask();
702 // Apply sanitizer attributes to the function.
703 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
704 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
705 if (SanOpts.has(SanitizerKind::Thread))
706 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
707 if (SanOpts.has(SanitizerKind::Memory))
708 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
709 if (SanOpts.has(SanitizerKind::SafeStack))
710 Fn->addFnAttr(llvm::Attribute::SafeStack);
712 // Apply xray attributes to the function (as a string, for now)
713 if (D && ShouldXRayInstrumentFunction()) {
714 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
715 if (XRayAttr->alwaysXRayInstrument())
716 Fn->addFnAttr("function-instrument", "xray-always");
717 if (XRayAttr->neverXRayInstrument())
718 Fn->addFnAttr("function-instrument", "xray-never");
721 "xray-instruction-threshold",
722 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
726 // Pass inline keyword to optimizer if it appears explicitly on any
727 // declaration. Also, in the case of -fno-inline attach NoInline
728 // attribute to all functions that are not marked AlwaysInline, or
729 // to all functions that are not marked inline or implicitly inline
730 // in the case of -finline-hint-functions.
731 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
732 const CodeGenOptions& CodeGenOpts = CGM.getCodeGenOpts();
733 if (!CodeGenOpts.NoInline) {
734 for (auto RI : FD->redecls())
735 if (RI->isInlineSpecified()) {
736 Fn->addFnAttr(llvm::Attribute::InlineHint);
739 if (CodeGenOpts.getInlining() == CodeGenOptions::OnlyHintInlining &&
740 !FD->isInlined() && !Fn->hasFnAttribute(llvm::Attribute::InlineHint))
741 Fn->addFnAttr(llvm::Attribute::NoInline);
742 } else if (!FD->hasAttr<AlwaysInlineAttr>())
743 Fn->addFnAttr(llvm::Attribute::NoInline);
744 if (CGM.getLangOpts().OpenMP && FD->hasAttr<OMPDeclareSimdDeclAttr>())
745 CGM.getOpenMPRuntime().emitDeclareSimdFunction(FD, Fn);
748 // Add no-jump-tables value.
749 Fn->addFnAttr("no-jump-tables",
750 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
752 if (getLangOpts().OpenCL) {
753 // Add metadata for a kernel function.
754 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
755 EmitOpenCLKernelMetadata(FD, Fn);
758 // If we are checking function types, emit a function type signature as
760 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
761 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
762 if (llvm::Constant *PrologueSig =
763 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
764 llvm::Constant *FTRTTIConst =
765 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
766 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
767 llvm::Constant *PrologueStructConst =
768 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
769 Fn->setPrologueData(PrologueStructConst);
774 // If we're in C++ mode and the function name is "main", it is guaranteed
775 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
776 // used within a program").
777 if (getLangOpts().CPlusPlus)
778 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
780 Fn->addFnAttr(llvm::Attribute::NoRecurse);
782 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
784 // Create a marker to make it easy to insert allocas into the entryblock
785 // later. Don't create this with the builder, because we don't want it
787 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
788 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
790 ReturnBlock = getJumpDestInCurrentScope("return");
792 Builder.SetInsertPoint(EntryBB);
794 // Emit subprogram debug descriptor.
795 if (CGDebugInfo *DI = getDebugInfo()) {
796 // Reconstruct the type from the argument list so that implicit parameters,
797 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
799 CallingConv CC = CallingConv::CC_C;
800 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
801 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
802 CC = SrcFnTy->getCallConv();
803 SmallVector<QualType, 16> ArgTypes;
804 for (const VarDecl *VD : Args)
805 ArgTypes.push_back(VD->getType());
806 QualType FnType = getContext().getFunctionType(
807 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
808 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
811 if (ShouldInstrumentFunction())
812 EmitFunctionInstrumentation("__cyg_profile_func_enter");
814 if (CGM.getCodeGenOpts().InstrumentForProfiling)
815 EmitMCountInstrumentation();
817 if (RetTy->isVoidType()) {
818 // Void type; nothing to return.
819 ReturnValue = Address::invalid();
821 // Count the implicit return.
822 if (!endsWithReturn(D))
824 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
825 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
826 // Indirect aggregate return; emit returned value directly into sret slot.
827 // This reduces code size, and affects correctness in C++.
828 auto AI = CurFn->arg_begin();
829 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
831 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
832 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
833 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
834 // Load the sret pointer from the argument struct and return into that.
835 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
836 llvm::Function::arg_iterator EI = CurFn->arg_end();
838 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
839 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
840 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
842 ReturnValue = CreateIRTemp(RetTy, "retval");
844 // Tell the epilog emitter to autorelease the result. We do this
845 // now so that various specialized functions can suppress it
846 // during their IR-generation.
847 if (getLangOpts().ObjCAutoRefCount &&
848 !CurFnInfo->isReturnsRetained() &&
849 RetTy->isObjCRetainableType())
850 AutoreleaseResult = true;
853 EmitStartEHSpec(CurCodeDecl);
855 PrologueCleanupDepth = EHStack.stable_begin();
856 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
858 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
859 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
860 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
861 if (MD->getParent()->isLambda() &&
862 MD->getOverloadedOperator() == OO_Call) {
863 // We're in a lambda; figure out the captures.
864 MD->getParent()->getCaptureFields(LambdaCaptureFields,
865 LambdaThisCaptureField);
866 if (LambdaThisCaptureField) {
867 // If the lambda captures the object referred to by '*this' - either by
868 // value or by reference, make sure CXXThisValue points to the correct
871 // Get the lvalue for the field (which is a copy of the enclosing object
872 // or contains the address of the enclosing object).
873 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
874 if (!LambdaThisCaptureField->getType()->isPointerType()) {
875 // If the enclosing object was captured by value, just use its address.
876 CXXThisValue = ThisFieldLValue.getAddress().getPointer();
878 // Load the lvalue pointed to by the field, since '*this' was captured
881 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
884 for (auto *FD : MD->getParent()->fields()) {
885 if (FD->hasCapturedVLAType()) {
886 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
887 SourceLocation()).getScalarVal();
888 auto VAT = FD->getCapturedVLAType();
889 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
893 // Not in a lambda; just use 'this' from the method.
894 // FIXME: Should we generate a new load for each use of 'this'? The
895 // fast register allocator would be happier...
896 CXXThisValue = CXXABIThisValue;
900 // If any of the arguments have a variably modified type, make sure to
901 // emit the type size.
902 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
904 const VarDecl *VD = *i;
906 // Dig out the type as written from ParmVarDecls; it's unclear whether
907 // the standard (C99 6.9.1p10) requires this, but we're following the
908 // precedent set by gcc.
910 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
911 Ty = PVD->getOriginalType();
915 if (Ty->isVariablyModifiedType())
916 EmitVariablyModifiedType(Ty);
918 // Emit a location at the end of the prologue.
919 if (CGDebugInfo *DI = getDebugInfo())
920 DI->EmitLocation(Builder, StartLoc);
923 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
925 incrementProfileCounter(Body);
926 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
927 EmitCompoundStmtWithoutScope(*S);
932 /// When instrumenting to collect profile data, the counts for some blocks
933 /// such as switch cases need to not include the fall-through counts, so
934 /// emit a branch around the instrumentation code. When not instrumenting,
935 /// this just calls EmitBlock().
936 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
938 llvm::BasicBlock *SkipCountBB = nullptr;
939 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
940 // When instrumenting for profiling, the fallthrough to certain
941 // statements needs to skip over the instrumentation code so that we
942 // get an accurate count.
943 SkipCountBB = createBasicBlock("skipcount");
944 EmitBranch(SkipCountBB);
947 uint64_t CurrentCount = getCurrentProfileCount();
948 incrementProfileCounter(S);
949 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
951 EmitBlock(SkipCountBB);
954 /// Tries to mark the given function nounwind based on the
955 /// non-existence of any throwing calls within it. We believe this is
956 /// lightweight enough to do at -O0.
957 static void TryMarkNoThrow(llvm::Function *F) {
958 // LLVM treats 'nounwind' on a function as part of the type, so we
959 // can't do this on functions that can be overwritten.
960 if (F->isInterposable()) return;
962 for (llvm::BasicBlock &BB : *F)
963 for (llvm::Instruction &I : BB)
967 F->setDoesNotThrow();
970 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
971 FunctionArgList &Args) {
972 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
973 QualType ResTy = FD->getReturnType();
975 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
976 if (MD && MD->isInstance()) {
977 if (CGM.getCXXABI().HasThisReturn(GD))
978 ResTy = MD->getThisType(getContext());
979 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
980 ResTy = CGM.getContext().VoidPtrTy;
981 CGM.getCXXABI().buildThisParam(*this, Args);
984 // The base version of an inheriting constructor whose constructed base is a
985 // virtual base is not passed any arguments (because it doesn't actually call
986 // the inherited constructor).
987 bool PassedParams = true;
988 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
989 if (auto Inherited = CD->getInheritedConstructor())
991 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
994 for (auto *Param : FD->parameters()) {
995 Args.push_back(Param);
996 if (!Param->hasAttr<PassObjectSizeAttr>())
999 IdentifierInfo *NoID = nullptr;
1000 auto *Implicit = ImplicitParamDecl::Create(
1001 getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
1002 getContext().getSizeType());
1003 SizeArguments[Param] = Implicit;
1004 Args.push_back(Implicit);
1008 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1009 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1014 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1015 const CGFunctionInfo &FnInfo) {
1016 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1019 FunctionArgList Args;
1020 QualType ResTy = BuildFunctionArgList(GD, Args);
1022 // Check if we should generate debug info for this function.
1023 if (FD->hasAttr<NoDebugAttr>())
1024 DebugInfo = nullptr; // disable debug info indefinitely for this function
1026 SourceRange BodyRange;
1027 if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
1028 CurEHLocation = BodyRange.getEnd();
1030 // Use the location of the start of the function to determine where
1031 // the function definition is located. By default use the location
1032 // of the declaration as the location for the subprogram. A function
1033 // may lack a declaration in the source code if it is created by code
1034 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1035 SourceLocation Loc = FD->getLocation();
1037 // If this is a function specialization then use the pattern body
1038 // as the location for the function.
1039 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1040 if (SpecDecl->hasBody(SpecDecl))
1041 Loc = SpecDecl->getLocation();
1043 // Emit the standard function prologue.
1044 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1046 // Generate the body of the function.
1047 PGO.assignRegionCounters(GD, CurFn);
1048 if (isa<CXXDestructorDecl>(FD))
1049 EmitDestructorBody(Args);
1050 else if (isa<CXXConstructorDecl>(FD))
1051 EmitConstructorBody(Args);
1052 else if (getLangOpts().CUDA &&
1053 !getLangOpts().CUDAIsDevice &&
1054 FD->hasAttr<CUDAGlobalAttr>())
1055 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1056 else if (isa<CXXConversionDecl>(FD) &&
1057 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
1058 // The lambda conversion to block pointer is special; the semantics can't be
1059 // expressed in the AST, so IRGen needs to special-case it.
1060 EmitLambdaToBlockPointerBody(Args);
1061 } else if (isa<CXXMethodDecl>(FD) &&
1062 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1063 // The lambda static invoker function is special, because it forwards or
1064 // clones the body of the function call operator (but is actually static).
1065 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
1066 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1067 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1068 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1069 // Implicit copy-assignment gets the same special treatment as implicit
1070 // copy-constructors.
1071 emitImplicitAssignmentOperatorBody(Args);
1072 } else if (Stmt *Body = FD->getBody()) {
1073 EmitFunctionBody(Args, Body);
1075 llvm_unreachable("no definition for emitted function");
1077 // C++11 [stmt.return]p2:
1078 // Flowing off the end of a function [...] results in undefined behavior in
1079 // a value-returning function.
1081 // If the '}' that terminates a function is reached, and the value of the
1082 // function call is used by the caller, the behavior is undefined.
1083 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1084 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1085 if (SanOpts.has(SanitizerKind::Return)) {
1086 SanitizerScope SanScope(this);
1087 llvm::Value *IsFalse = Builder.getFalse();
1088 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1089 "missing_return", EmitCheckSourceLocation(FD->getLocation()),
1091 } else if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
1092 EmitTrapCall(llvm::Intrinsic::trap);
1094 Builder.CreateUnreachable();
1095 Builder.ClearInsertionPoint();
1098 // Emit the standard function epilogue.
1099 FinishFunction(BodyRange.getEnd());
1101 // If we haven't marked the function nothrow through other means, do
1102 // a quick pass now to see if we can.
1103 if (!CurFn->doesNotThrow())
1104 TryMarkNoThrow(CurFn);
1107 /// ContainsLabel - Return true if the statement contains a label in it. If
1108 /// this statement is not executed normally, it not containing a label means
1109 /// that we can just remove the code.
1110 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1111 // Null statement, not a label!
1112 if (!S) return false;
1114 // If this is a label, we have to emit the code, consider something like:
1115 // if (0) { ... foo: bar(); } goto foo;
1117 // TODO: If anyone cared, we could track __label__'s, since we know that you
1118 // can't jump to one from outside their declared region.
1119 if (isa<LabelStmt>(S))
1122 // If this is a case/default statement, and we haven't seen a switch, we have
1123 // to emit the code.
1124 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1127 // If this is a switch statement, we want to ignore cases below it.
1128 if (isa<SwitchStmt>(S))
1129 IgnoreCaseStmts = true;
1131 // Scan subexpressions for verboten labels.
1132 for (const Stmt *SubStmt : S->children())
1133 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1139 /// containsBreak - Return true if the statement contains a break out of it.
1140 /// If the statement (recursively) contains a switch or loop with a break
1141 /// inside of it, this is fine.
1142 bool CodeGenFunction::containsBreak(const Stmt *S) {
1143 // Null statement, not a label!
1144 if (!S) return false;
1146 // If this is a switch or loop that defines its own break scope, then we can
1147 // include it and anything inside of it.
1148 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1152 if (isa<BreakStmt>(S))
1155 // Scan subexpressions for verboten breaks.
1156 for (const Stmt *SubStmt : S->children())
1157 if (containsBreak(SubStmt))
1163 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1164 if (!S) return false;
1166 // Some statement kinds add a scope and thus never add a decl to the current
1167 // scope. Note, this list is longer than the list of statements that might
1168 // have an unscoped decl nested within them, but this way is conservatively
1169 // correct even if more statement kinds are added.
1170 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1171 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1172 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1173 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1176 if (isa<DeclStmt>(S))
1179 for (const Stmt *SubStmt : S->children())
1180 if (mightAddDeclToScope(SubStmt))
1186 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1187 /// to a constant, or if it does but contains a label, return false. If it
1188 /// constant folds return true and set the boolean result in Result.
1189 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1192 llvm::APSInt ResultInt;
1193 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1196 ResultBool = ResultInt.getBoolValue();
1200 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1201 /// to a constant, or if it does but contains a label, return false. If it
1202 /// constant folds return true and set the folded value.
1203 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1204 llvm::APSInt &ResultInt,
1206 // FIXME: Rename and handle conversion of other evaluatable things
1209 if (!Cond->EvaluateAsInt(Int, getContext()))
1210 return false; // Not foldable, not integer or not fully evaluatable.
1212 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1213 return false; // Contains a label.
1221 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1222 /// statement) to the specified blocks. Based on the condition, this might try
1223 /// to simplify the codegen of the conditional based on the branch.
1225 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1226 llvm::BasicBlock *TrueBlock,
1227 llvm::BasicBlock *FalseBlock,
1228 uint64_t TrueCount) {
1229 Cond = Cond->IgnoreParens();
1231 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1233 // Handle X && Y in a condition.
1234 if (CondBOp->getOpcode() == BO_LAnd) {
1235 // If we have "1 && X", simplify the code. "0 && X" would have constant
1236 // folded if the case was simple enough.
1237 bool ConstantBool = false;
1238 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1240 // br(1 && X) -> br(X).
1241 incrementProfileCounter(CondBOp);
1242 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1246 // If we have "X && 1", simplify the code to use an uncond branch.
1247 // "X && 0" would have been constant folded to 0.
1248 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1250 // br(X && 1) -> br(X).
1251 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1255 // Emit the LHS as a conditional. If the LHS conditional is false, we
1256 // want to jump to the FalseBlock.
1257 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1258 // The counter tells us how often we evaluate RHS, and all of TrueCount
1259 // can be propagated to that branch.
1260 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1262 ConditionalEvaluation eval(*this);
1264 ApplyDebugLocation DL(*this, Cond);
1265 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1269 incrementProfileCounter(CondBOp);
1270 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1272 // Any temporaries created here are conditional.
1274 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1280 if (CondBOp->getOpcode() == BO_LOr) {
1281 // If we have "0 || X", simplify the code. "1 || X" would have constant
1282 // folded if the case was simple enough.
1283 bool ConstantBool = false;
1284 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1286 // br(0 || X) -> br(X).
1287 incrementProfileCounter(CondBOp);
1288 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1292 // If we have "X || 0", simplify the code to use an uncond branch.
1293 // "X || 1" would have been constant folded to 1.
1294 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1296 // br(X || 0) -> br(X).
1297 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1301 // Emit the LHS as a conditional. If the LHS conditional is true, we
1302 // want to jump to the TrueBlock.
1303 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1304 // We have the count for entry to the RHS and for the whole expression
1305 // being true, so we can divy up True count between the short circuit and
1308 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1309 uint64_t RHSCount = TrueCount - LHSCount;
1311 ConditionalEvaluation eval(*this);
1313 ApplyDebugLocation DL(*this, Cond);
1314 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1315 EmitBlock(LHSFalse);
1318 incrementProfileCounter(CondBOp);
1319 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1321 // Any temporaries created here are conditional.
1323 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1331 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1332 // br(!x, t, f) -> br(x, f, t)
1333 if (CondUOp->getOpcode() == UO_LNot) {
1334 // Negate the count.
1335 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1336 // Negate the condition and swap the destination blocks.
1337 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1342 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1343 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1344 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1345 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1347 ConditionalEvaluation cond(*this);
1348 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1349 getProfileCount(CondOp));
1351 // When computing PGO branch weights, we only know the overall count for
1352 // the true block. This code is essentially doing tail duplication of the
1353 // naive code-gen, introducing new edges for which counts are not
1354 // available. Divide the counts proportionally between the LHS and RHS of
1355 // the conditional operator.
1356 uint64_t LHSScaledTrueCount = 0;
1359 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1360 LHSScaledTrueCount = TrueCount * LHSRatio;
1364 EmitBlock(LHSBlock);
1365 incrementProfileCounter(CondOp);
1367 ApplyDebugLocation DL(*this, Cond);
1368 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1369 LHSScaledTrueCount);
1374 EmitBlock(RHSBlock);
1375 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1376 TrueCount - LHSScaledTrueCount);
1382 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1383 // Conditional operator handling can give us a throw expression as a
1384 // condition for a case like:
1385 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1387 // br(c, throw x, br(y, t, f))
1388 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1392 // If the branch has a condition wrapped by __builtin_unpredictable,
1393 // create metadata that specifies that the branch is unpredictable.
1394 // Don't bother if not optimizing because that metadata would not be used.
1395 llvm::MDNode *Unpredictable = nullptr;
1396 auto *Call = dyn_cast<CallExpr>(Cond);
1397 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1398 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1399 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1400 llvm::MDBuilder MDHelper(getLLVMContext());
1401 Unpredictable = MDHelper.createUnpredictable();
1405 // Create branch weights based on the number of times we get here and the
1406 // number of times the condition should be true.
1407 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1408 llvm::MDNode *Weights =
1409 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1411 // Emit the code with the fully general case.
1414 ApplyDebugLocation DL(*this, Cond);
1415 CondV = EvaluateExprAsBool(Cond);
1417 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1420 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1421 /// specified stmt yet.
1422 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1423 CGM.ErrorUnsupported(S, Type);
1426 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1427 /// variable-length array whose elements have a non-zero bit-pattern.
1429 /// \param baseType the inner-most element type of the array
1430 /// \param src - a char* pointing to the bit-pattern for a single
1431 /// base element of the array
1432 /// \param sizeInChars - the total size of the VLA, in chars
1433 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1434 Address dest, Address src,
1435 llvm::Value *sizeInChars) {
1436 CGBuilderTy &Builder = CGF.Builder;
1438 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1439 llvm::Value *baseSizeInChars
1440 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1443 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1445 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1447 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1448 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1449 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1451 // Make a loop over the VLA. C99 guarantees that the VLA element
1452 // count must be nonzero.
1453 CGF.EmitBlock(loopBB);
1455 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1456 cur->addIncoming(begin.getPointer(), originBB);
1458 CharUnits curAlign =
1459 dest.getAlignment().alignmentOfArrayElement(baseSize);
1461 // memcpy the individual element bit-pattern.
1462 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1463 /*volatile*/ false);
1465 // Go to the next element.
1467 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1469 // Leave if that's the end of the VLA.
1470 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1471 Builder.CreateCondBr(done, contBB, loopBB);
1472 cur->addIncoming(next, loopBB);
1474 CGF.EmitBlock(contBB);
1478 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1479 // Ignore empty classes in C++.
1480 if (getLangOpts().CPlusPlus) {
1481 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1482 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1487 // Cast the dest ptr to the appropriate i8 pointer type.
1488 if (DestPtr.getElementType() != Int8Ty)
1489 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1491 // Get size and alignment info for this aggregate.
1492 CharUnits size = getContext().getTypeSizeInChars(Ty);
1494 llvm::Value *SizeVal;
1495 const VariableArrayType *vla;
1497 // Don't bother emitting a zero-byte memset.
1498 if (size.isZero()) {
1499 // But note that getTypeInfo returns 0 for a VLA.
1500 if (const VariableArrayType *vlaType =
1501 dyn_cast_or_null<VariableArrayType>(
1502 getContext().getAsArrayType(Ty))) {
1504 llvm::Value *numElts;
1505 std::tie(numElts, eltType) = getVLASize(vlaType);
1508 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1509 if (!eltSize.isOne())
1510 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1516 SizeVal = CGM.getSize(size);
1520 // If the type contains a pointer to data member we can't memset it to zero.
1521 // Instead, create a null constant and copy it to the destination.
1522 // TODO: there are other patterns besides zero that we can usefully memset,
1523 // like -1, which happens to be the pattern used by member-pointers.
1524 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1525 // For a VLA, emit a single element, then splat that over the VLA.
1526 if (vla) Ty = getContext().getBaseElementType(vla);
1528 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1530 llvm::GlobalVariable *NullVariable =
1531 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1532 /*isConstant=*/true,
1533 llvm::GlobalVariable::PrivateLinkage,
1534 NullConstant, Twine());
1535 CharUnits NullAlign = DestPtr.getAlignment();
1536 NullVariable->setAlignment(NullAlign.getQuantity());
1537 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1540 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1542 // Get and call the appropriate llvm.memcpy overload.
1543 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1547 // Otherwise, just memset the whole thing to zero. This is legal
1548 // because in LLVM, all default initializers (other than the ones we just
1549 // handled above) are guaranteed to have a bit pattern of all zeros.
1550 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1553 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1554 // Make sure that there is a block for the indirect goto.
1555 if (!IndirectBranch)
1556 GetIndirectGotoBlock();
1558 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1560 // Make sure the indirect branch includes all of the address-taken blocks.
1561 IndirectBranch->addDestination(BB);
1562 return llvm::BlockAddress::get(CurFn, BB);
1565 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1566 // If we already made the indirect branch for indirect goto, return its block.
1567 if (IndirectBranch) return IndirectBranch->getParent();
1569 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1571 // Create the PHI node that indirect gotos will add entries to.
1572 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1573 "indirect.goto.dest");
1575 // Create the indirect branch instruction.
1576 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1577 return IndirectBranch->getParent();
1580 /// Computes the length of an array in elements, as well as the base
1581 /// element type and a properly-typed first element pointer.
1582 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1585 const ArrayType *arrayType = origArrayType;
1587 // If it's a VLA, we have to load the stored size. Note that
1588 // this is the size of the VLA in bytes, not its size in elements.
1589 llvm::Value *numVLAElements = nullptr;
1590 if (isa<VariableArrayType>(arrayType)) {
1591 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1593 // Walk into all VLAs. This doesn't require changes to addr,
1594 // which has type T* where T is the first non-VLA element type.
1596 QualType elementType = arrayType->getElementType();
1597 arrayType = getContext().getAsArrayType(elementType);
1599 // If we only have VLA components, 'addr' requires no adjustment.
1601 baseType = elementType;
1602 return numVLAElements;
1604 } while (isa<VariableArrayType>(arrayType));
1606 // We get out here only if we find a constant array type
1610 // We have some number of constant-length arrays, so addr should
1611 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1612 // down to the first element of addr.
1613 SmallVector<llvm::Value*, 8> gepIndices;
1615 // GEP down to the array type.
1616 llvm::ConstantInt *zero = Builder.getInt32(0);
1617 gepIndices.push_back(zero);
1619 uint64_t countFromCLAs = 1;
1622 llvm::ArrayType *llvmArrayType =
1623 dyn_cast<llvm::ArrayType>(addr.getElementType());
1624 while (llvmArrayType) {
1625 assert(isa<ConstantArrayType>(arrayType));
1626 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1627 == llvmArrayType->getNumElements());
1629 gepIndices.push_back(zero);
1630 countFromCLAs *= llvmArrayType->getNumElements();
1631 eltType = arrayType->getElementType();
1634 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1635 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1636 assert((!llvmArrayType || arrayType) &&
1637 "LLVM and Clang types are out-of-synch");
1641 // From this point onwards, the Clang array type has been emitted
1642 // as some other type (probably a packed struct). Compute the array
1643 // size, and just emit the 'begin' expression as a bitcast.
1646 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1647 eltType = arrayType->getElementType();
1648 arrayType = getContext().getAsArrayType(eltType);
1651 llvm::Type *baseType = ConvertType(eltType);
1652 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1654 // Create the actual GEP.
1655 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1656 gepIndices, "array.begin"),
1657 addr.getAlignment());
1662 llvm::Value *numElements
1663 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1665 // If we had any VLA dimensions, factor them in.
1667 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1672 std::pair<llvm::Value*, QualType>
1673 CodeGenFunction::getVLASize(QualType type) {
1674 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1675 assert(vla && "type was not a variable array type!");
1676 return getVLASize(vla);
1679 std::pair<llvm::Value*, QualType>
1680 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1681 // The number of elements so far; always size_t.
1682 llvm::Value *numElements = nullptr;
1684 QualType elementType;
1686 elementType = type->getElementType();
1687 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1688 assert(vlaSize && "no size for VLA!");
1689 assert(vlaSize->getType() == SizeTy);
1692 numElements = vlaSize;
1694 // It's undefined behavior if this wraps around, so mark it that way.
1695 // FIXME: Teach -fsanitize=undefined to trap this.
1696 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1698 } while ((type = getContext().getAsVariableArrayType(elementType)));
1700 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1703 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1704 assert(type->isVariablyModifiedType() &&
1705 "Must pass variably modified type to EmitVLASizes!");
1707 EnsureInsertPoint();
1709 // We're going to walk down into the type and look for VLA
1712 assert(type->isVariablyModifiedType());
1714 const Type *ty = type.getTypePtr();
1715 switch (ty->getTypeClass()) {
1717 #define TYPE(Class, Base)
1718 #define ABSTRACT_TYPE(Class, Base)
1719 #define NON_CANONICAL_TYPE(Class, Base)
1720 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1721 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1722 #include "clang/AST/TypeNodes.def"
1723 llvm_unreachable("unexpected dependent type!");
1725 // These types are never variably-modified.
1729 case Type::ExtVector:
1732 case Type::Elaborated:
1733 case Type::TemplateSpecialization:
1734 case Type::ObjCObject:
1735 case Type::ObjCInterface:
1736 case Type::ObjCObjectPointer:
1737 llvm_unreachable("type class is never variably-modified!");
1739 case Type::Adjusted:
1740 type = cast<AdjustedType>(ty)->getAdjustedType();
1744 type = cast<DecayedType>(ty)->getPointeeType();
1748 type = cast<PointerType>(ty)->getPointeeType();
1751 case Type::BlockPointer:
1752 type = cast<BlockPointerType>(ty)->getPointeeType();
1755 case Type::LValueReference:
1756 case Type::RValueReference:
1757 type = cast<ReferenceType>(ty)->getPointeeType();
1760 case Type::MemberPointer:
1761 type = cast<MemberPointerType>(ty)->getPointeeType();
1764 case Type::ConstantArray:
1765 case Type::IncompleteArray:
1766 // Losing element qualification here is fine.
1767 type = cast<ArrayType>(ty)->getElementType();
1770 case Type::VariableArray: {
1771 // Losing element qualification here is fine.
1772 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1774 // Unknown size indication requires no size computation.
1775 // Otherwise, evaluate and record it.
1776 if (const Expr *size = vat->getSizeExpr()) {
1777 // It's possible that we might have emitted this already,
1778 // e.g. with a typedef and a pointer to it.
1779 llvm::Value *&entry = VLASizeMap[size];
1781 llvm::Value *Size = EmitScalarExpr(size);
1784 // If the size is an expression that is not an integer constant
1785 // expression [...] each time it is evaluated it shall have a value
1786 // greater than zero.
1787 if (SanOpts.has(SanitizerKind::VLABound) &&
1788 size->getType()->isSignedIntegerType()) {
1789 SanitizerScope SanScope(this);
1790 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1791 llvm::Constant *StaticArgs[] = {
1792 EmitCheckSourceLocation(size->getLocStart()),
1793 EmitCheckTypeDescriptor(size->getType())
1795 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1796 SanitizerKind::VLABound),
1797 "vla_bound_not_positive", StaticArgs, Size);
1800 // Always zexting here would be wrong if it weren't
1801 // undefined behavior to have a negative bound.
1802 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1805 type = vat->getElementType();
1809 case Type::FunctionProto:
1810 case Type::FunctionNoProto:
1811 type = cast<FunctionType>(ty)->getReturnType();
1816 case Type::UnaryTransform:
1817 case Type::Attributed:
1818 case Type::SubstTemplateTypeParm:
1819 case Type::PackExpansion:
1820 // Keep walking after single level desugaring.
1821 type = type.getSingleStepDesugaredType(getContext());
1825 case Type::Decltype:
1827 // Stop walking: nothing to do.
1830 case Type::TypeOfExpr:
1831 // Stop walking: emit typeof expression.
1832 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1836 type = cast<AtomicType>(ty)->getValueType();
1840 type = cast<PipeType>(ty)->getElementType();
1843 } while (type->isVariablyModifiedType());
1846 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1847 if (getContext().getBuiltinVaListType()->isArrayType())
1848 return EmitPointerWithAlignment(E);
1849 return EmitLValue(E).getAddress();
1852 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1853 return EmitLValue(E).getAddress();
1856 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1857 llvm::Constant *Init) {
1858 assert (Init && "Invalid DeclRefExpr initializer!");
1859 if (CGDebugInfo *Dbg = getDebugInfo())
1860 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
1861 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1864 CodeGenFunction::PeepholeProtection
1865 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1866 // At the moment, the only aggressive peephole we do in IR gen
1867 // is trunc(zext) folding, but if we add more, we can easily
1868 // extend this protection.
1870 if (!rvalue.isScalar()) return PeepholeProtection();
1871 llvm::Value *value = rvalue.getScalarVal();
1872 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1874 // Just make an extra bitcast.
1875 assert(HaveInsertPoint());
1876 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1877 Builder.GetInsertBlock());
1879 PeepholeProtection protection;
1880 protection.Inst = inst;
1884 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1885 if (!protection.Inst) return;
1887 // In theory, we could try to duplicate the peepholes now, but whatever.
1888 protection.Inst->eraseFromParent();
1891 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1892 llvm::Value *AnnotatedVal,
1893 StringRef AnnotationStr,
1894 SourceLocation Location) {
1895 llvm::Value *Args[4] = {
1897 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1898 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1899 CGM.EmitAnnotationLineNo(Location)
1901 return Builder.CreateCall(AnnotationFn, Args);
1904 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1905 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1906 // FIXME We create a new bitcast for every annotation because that's what
1907 // llvm-gcc was doing.
1908 for (const auto *I : D->specific_attrs<AnnotateAttr>())
1909 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1910 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1911 I->getAnnotation(), D->getLocation());
1914 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1916 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1917 llvm::Value *V = Addr.getPointer();
1918 llvm::Type *VTy = V->getType();
1919 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1922 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1923 // FIXME Always emit the cast inst so we can differentiate between
1924 // annotation on the first field of a struct and annotation on the struct
1926 if (VTy != CGM.Int8PtrTy)
1927 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1928 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1929 V = Builder.CreateBitCast(V, VTy);
1932 return Address(V, Addr.getAlignment());
1935 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
1937 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
1939 assert(!CGF->IsSanitizerScope);
1940 CGF->IsSanitizerScope = true;
1943 CodeGenFunction::SanitizerScope::~SanitizerScope() {
1944 CGF->IsSanitizerScope = false;
1947 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
1948 const llvm::Twine &Name,
1949 llvm::BasicBlock *BB,
1950 llvm::BasicBlock::iterator InsertPt) const {
1951 LoopStack.InsertHelper(I);
1952 if (IsSanitizerScope)
1953 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
1956 void CGBuilderInserter::InsertHelper(
1957 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1958 llvm::BasicBlock::iterator InsertPt) const {
1959 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
1961 CGF->InsertHelper(I, Name, BB, InsertPt);
1964 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
1965 CodeGenModule &CGM, const FunctionDecl *FD,
1966 std::string &FirstMissing) {
1967 // If there aren't any required features listed then go ahead and return.
1968 if (ReqFeatures.empty())
1971 // Now build up the set of caller features and verify that all the required
1972 // features are there.
1973 llvm::StringMap<bool> CallerFeatureMap;
1974 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
1976 // If we have at least one of the features in the feature list return
1977 // true, otherwise return false.
1979 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
1980 SmallVector<StringRef, 1> OrFeatures;
1981 Feature.split(OrFeatures, "|");
1982 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
1983 [&](StringRef Feature) {
1984 if (!CallerFeatureMap.lookup(Feature)) {
1985 FirstMissing = Feature.str();
1993 // Emits an error if we don't have a valid set of target features for the
1995 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
1996 const FunctionDecl *TargetDecl) {
1997 // Early exit if this is an indirect call.
2001 // Get the current enclosing function if it exists. If it doesn't
2002 // we can't check the target features anyhow.
2003 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2007 // Grab the required features for the call. For a builtin this is listed in
2008 // the td file with the default cpu, for an always_inline function this is any
2009 // listed cpu and any listed features.
2010 unsigned BuiltinID = TargetDecl->getBuiltinID();
2011 std::string MissingFeature;
2013 SmallVector<StringRef, 1> ReqFeatures;
2014 const char *FeatureList =
2015 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2016 // Return if the builtin doesn't have any required features.
2017 if (!FeatureList || StringRef(FeatureList) == "")
2019 StringRef(FeatureList).split(ReqFeatures, ",");
2020 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2021 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2022 << TargetDecl->getDeclName()
2023 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2025 } else if (TargetDecl->hasAttr<TargetAttr>()) {
2026 // Get the required features for the callee.
2027 SmallVector<StringRef, 1> ReqFeatures;
2028 llvm::StringMap<bool> CalleeFeatureMap;
2029 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2030 for (const auto &F : CalleeFeatureMap) {
2031 // Only positive features are "required".
2033 ReqFeatures.push_back(F.getKey());
2035 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2036 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2037 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2041 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2042 if (!CGM.getCodeGenOpts().SanitizeStats)
2045 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2046 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2047 CGM.getSanStats().create(IRB, SSK);