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"
15 #include "CGCleanup.h"
16 #include "CGCUDARuntime.h"
18 #include "CGDebugInfo.h"
19 #include "CGOpenMPRuntime.h"
20 #include "CodeGenModule.h"
21 #include "CodeGenPGO.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/ASTContext.h"
24 #include "clang/AST/Decl.h"
25 #include "clang/AST/DeclCXX.h"
26 #include "clang/AST/StmtCXX.h"
27 #include "clang/Basic/TargetInfo.h"
28 #include "clang/CodeGen/CGFunctionInfo.h"
29 #include "clang/Frontend/CodeGenOptions.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/MDBuilder.h"
33 #include "llvm/IR/Operator.h"
34 using namespace clang;
35 using namespace CodeGen;
37 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
38 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
39 Builder(cgm.getModule().getContext(), llvm::ConstantFolder(),
40 CGBuilderInserterTy(this)),
41 CurFn(nullptr), CapturedStmtInfo(nullptr),
42 SanOpts(CGM.getLangOpts().Sanitize), IsSanitizerScope(false),
43 CurFuncIsThunk(false), AutoreleaseResult(false), SawAsmBlock(false),
44 IsOutlinedSEHHelper(false), BlockInfo(nullptr), BlockPointer(nullptr),
45 LambdaThisCaptureField(nullptr), NormalCleanupDest(nullptr),
46 NextCleanupDestIndex(1), FirstBlockInfo(nullptr), EHResumeBlock(nullptr),
47 ExceptionSlot(nullptr), EHSelectorSlot(nullptr),
48 DebugInfo(CGM.getModuleDebugInfo()),
49 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
50 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
51 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
52 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
53 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
54 CXXDefaultInitExprThis(nullptr), CXXStructorImplicitParamDecl(nullptr),
55 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
56 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
57 TerminateHandler(nullptr), TrapBB(nullptr) {
58 if (!suppressNewContext)
59 CGM.getCXXABI().getMangleContext().startNewFunction();
61 llvm::FastMathFlags FMF;
62 if (CGM.getLangOpts().FastMath)
63 FMF.setUnsafeAlgebra();
64 if (CGM.getLangOpts().FiniteMathOnly) {
68 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
71 if (CGM.getCodeGenOpts().NoSignedZeros) {
72 FMF.setNoSignedZeros();
74 if (CGM.getCodeGenOpts().ReciprocalMath) {
75 FMF.setAllowReciprocal();
77 Builder.SetFastMathFlags(FMF);
80 CodeGenFunction::~CodeGenFunction() {
81 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
83 // If there are any unclaimed block infos, go ahead and destroy them
84 // now. This can happen if IR-gen gets clever and skips evaluating
87 destroyBlockInfos(FirstBlockInfo);
89 if (getLangOpts().OpenMP) {
90 CGM.getOpenMPRuntime().functionFinished(*this);
94 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
96 if (CGM.getCXXABI().isTypeInfoCalculable(T)) {
97 Alignment = getContext().getTypeAlignInChars(T);
98 unsigned MaxAlign = getContext().getLangOpts().MaxTypeAlign;
99 if (MaxAlign && Alignment.getQuantity() > MaxAlign &&
100 !getContext().isAlignmentRequired(T))
101 Alignment = CharUnits::fromQuantity(MaxAlign);
103 return LValue::MakeAddr(V, T, Alignment, getContext(), CGM.getTBAAInfo(T));
106 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
107 return CGM.getTypes().ConvertTypeForMem(T);
110 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
111 return CGM.getTypes().ConvertType(T);
114 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
115 type = type.getCanonicalType();
117 switch (type->getTypeClass()) {
118 #define TYPE(name, parent)
119 #define ABSTRACT_TYPE(name, parent)
120 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
121 #define DEPENDENT_TYPE(name, parent) case Type::name:
122 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
123 #include "clang/AST/TypeNodes.def"
124 llvm_unreachable("non-canonical or dependent type in IR-generation");
127 llvm_unreachable("undeduced auto type in IR-generation");
129 // Various scalar types.
132 case Type::BlockPointer:
133 case Type::LValueReference:
134 case Type::RValueReference:
135 case Type::MemberPointer:
137 case Type::ExtVector:
138 case Type::FunctionProto:
139 case Type::FunctionNoProto:
141 case Type::ObjCObjectPointer:
148 // Arrays, records, and Objective-C objects.
149 case Type::ConstantArray:
150 case Type::IncompleteArray:
151 case Type::VariableArray:
153 case Type::ObjCObject:
154 case Type::ObjCInterface:
155 return TEK_Aggregate;
157 // We operate on atomic values according to their underlying type.
159 type = cast<AtomicType>(type)->getValueType();
162 llvm_unreachable("unknown type kind!");
166 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
167 // For cleanliness, we try to avoid emitting the return block for
169 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
172 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
174 // We have a valid insert point, reuse it if it is empty or there are no
175 // explicit jumps to the return block.
176 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
177 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
178 delete ReturnBlock.getBlock();
180 EmitBlock(ReturnBlock.getBlock());
181 return llvm::DebugLoc();
184 // Otherwise, if the return block is the target of a single direct
185 // branch then we can just put the code in that block instead. This
186 // cleans up functions which started with a unified return block.
187 if (ReturnBlock.getBlock()->hasOneUse()) {
188 llvm::BranchInst *BI =
189 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
190 if (BI && BI->isUnconditional() &&
191 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
192 // Record/return the DebugLoc of the simple 'return' expression to be used
193 // later by the actual 'ret' instruction.
194 llvm::DebugLoc Loc = BI->getDebugLoc();
195 Builder.SetInsertPoint(BI->getParent());
196 BI->eraseFromParent();
197 delete ReturnBlock.getBlock();
202 // FIXME: We are at an unreachable point, there is no reason to emit the block
203 // unless it has uses. However, we still need a place to put the debug
204 // region.end for now.
206 EmitBlock(ReturnBlock.getBlock());
207 return llvm::DebugLoc();
210 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
212 if (!BB->use_empty())
213 return CGF.CurFn->getBasicBlockList().push_back(BB);
217 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
218 assert(BreakContinueStack.empty() &&
219 "mismatched push/pop in break/continue stack!");
221 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
222 && NumSimpleReturnExprs == NumReturnExprs
223 && ReturnBlock.getBlock()->use_empty();
224 // Usually the return expression is evaluated before the cleanup
225 // code. If the function contains only a simple return statement,
226 // such as a constant, the location before the cleanup code becomes
227 // the last useful breakpoint in the function, because the simple
228 // return expression will be evaluated after the cleanup code. To be
229 // safe, set the debug location for cleanup code to the location of
230 // the return statement. Otherwise the cleanup code should be at the
231 // end of the function's lexical scope.
233 // If there are multiple branches to the return block, the branch
234 // instructions will get the location of the return statements and
236 if (CGDebugInfo *DI = getDebugInfo()) {
237 if (OnlySimpleReturnStmts)
238 DI->EmitLocation(Builder, LastStopPoint);
240 DI->EmitLocation(Builder, EndLoc);
243 // Pop any cleanups that might have been associated with the
244 // parameters. Do this in whatever block we're currently in; it's
245 // important to do this before we enter the return block or return
246 // edges will be *really* confused.
247 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
248 bool HasOnlyLifetimeMarkers =
249 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
250 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
252 // Make sure the line table doesn't jump back into the body for
253 // the ret after it's been at EndLoc.
254 if (CGDebugInfo *DI = getDebugInfo())
255 if (OnlySimpleReturnStmts)
256 DI->EmitLocation(Builder, EndLoc);
258 PopCleanupBlocks(PrologueCleanupDepth);
261 // Emit function epilog (to return).
262 llvm::DebugLoc Loc = EmitReturnBlock();
264 if (ShouldInstrumentFunction())
265 EmitFunctionInstrumentation("__cyg_profile_func_exit");
267 // Emit debug descriptor for function end.
268 if (CGDebugInfo *DI = getDebugInfo())
269 DI->EmitFunctionEnd(Builder);
271 // Reset the debug location to that of the simple 'return' expression, if any
272 // rather than that of the end of the function's scope '}'.
273 ApplyDebugLocation AL(*this, Loc);
274 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
275 EmitEndEHSpec(CurCodeDecl);
277 assert(EHStack.empty() &&
278 "did not remove all scopes from cleanup stack!");
280 // If someone did an indirect goto, emit the indirect goto block at the end of
282 if (IndirectBranch) {
283 EmitBlock(IndirectBranch->getParent());
284 Builder.ClearInsertionPoint();
287 // If some of our locals escaped, insert a call to llvm.localescape in the
289 if (!EscapedLocals.empty()) {
290 // Invert the map from local to index into a simple vector. There should be
292 SmallVector<llvm::Value *, 4> EscapeArgs;
293 EscapeArgs.resize(EscapedLocals.size());
294 for (auto &Pair : EscapedLocals)
295 EscapeArgs[Pair.second] = Pair.first;
296 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
297 &CGM.getModule(), llvm::Intrinsic::localescape);
298 CGBuilderTy(AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
301 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
302 llvm::Instruction *Ptr = AllocaInsertPt;
303 AllocaInsertPt = nullptr;
304 Ptr->eraseFromParent();
306 // If someone took the address of a label but never did an indirect goto, we
307 // made a zero entry PHI node, which is illegal, zap it now.
308 if (IndirectBranch) {
309 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
310 if (PN->getNumIncomingValues() == 0) {
311 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
312 PN->eraseFromParent();
316 EmitIfUsed(*this, EHResumeBlock);
317 EmitIfUsed(*this, TerminateLandingPad);
318 EmitIfUsed(*this, TerminateHandler);
319 EmitIfUsed(*this, UnreachableBlock);
321 if (CGM.getCodeGenOpts().EmitDeclMetadata)
324 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
325 I = DeferredReplacements.begin(),
326 E = DeferredReplacements.end();
328 I->first->replaceAllUsesWith(I->second);
329 I->first->eraseFromParent();
333 /// ShouldInstrumentFunction - Return true if the current function should be
334 /// instrumented with __cyg_profile_func_* calls
335 bool CodeGenFunction::ShouldInstrumentFunction() {
336 if (!CGM.getCodeGenOpts().InstrumentFunctions)
338 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
343 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
344 /// instrumentation function with the current function and the call site, if
345 /// function instrumentation is enabled.
346 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
347 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
348 llvm::PointerType *PointerTy = Int8PtrTy;
349 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
350 llvm::FunctionType *FunctionTy =
351 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
353 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
354 llvm::CallInst *CallSite = Builder.CreateCall(
355 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
356 llvm::ConstantInt::get(Int32Ty, 0),
359 llvm::Value *args[] = {
360 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
364 EmitNounwindRuntimeCall(F, args);
367 void CodeGenFunction::EmitMCountInstrumentation() {
368 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
370 llvm::Constant *MCountFn =
371 CGM.CreateRuntimeFunction(FTy, getTarget().getMCountName());
372 EmitNounwindRuntimeCall(MCountFn);
375 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
376 // information in the program executable. The argument information stored
377 // includes the argument name, its type, the address and access qualifiers used.
378 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
379 CodeGenModule &CGM, llvm::LLVMContext &Context,
380 SmallVector<llvm::Metadata *, 5> &kernelMDArgs,
381 CGBuilderTy &Builder, ASTContext &ASTCtx) {
382 // Create MDNodes that represent the kernel arg metadata.
383 // Each MDNode is a list in the form of "key", N number of values which is
384 // the same number of values as their are kernel arguments.
386 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
388 // MDNode for the kernel argument address space qualifiers.
389 SmallVector<llvm::Metadata *, 8> addressQuals;
390 addressQuals.push_back(llvm::MDString::get(Context, "kernel_arg_addr_space"));
392 // MDNode for the kernel argument access qualifiers (images only).
393 SmallVector<llvm::Metadata *, 8> accessQuals;
394 accessQuals.push_back(llvm::MDString::get(Context, "kernel_arg_access_qual"));
396 // MDNode for the kernel argument type names.
397 SmallVector<llvm::Metadata *, 8> argTypeNames;
398 argTypeNames.push_back(llvm::MDString::get(Context, "kernel_arg_type"));
400 // MDNode for the kernel argument base type names.
401 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
402 argBaseTypeNames.push_back(
403 llvm::MDString::get(Context, "kernel_arg_base_type"));
405 // MDNode for the kernel argument type qualifiers.
406 SmallVector<llvm::Metadata *, 8> argTypeQuals;
407 argTypeQuals.push_back(llvm::MDString::get(Context, "kernel_arg_type_qual"));
409 // MDNode for the kernel argument names.
410 SmallVector<llvm::Metadata *, 8> argNames;
411 argNames.push_back(llvm::MDString::get(Context, "kernel_arg_name"));
413 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
414 const ParmVarDecl *parm = FD->getParamDecl(i);
415 QualType ty = parm->getType();
416 std::string typeQuals;
418 if (ty->isPointerType()) {
419 QualType pointeeTy = ty->getPointeeType();
421 // Get address qualifier.
422 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
423 ASTCtx.getTargetAddressSpace(pointeeTy.getAddressSpace()))));
425 // Get argument type name.
426 std::string typeName =
427 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
429 // Turn "unsigned type" to "utype"
430 std::string::size_type pos = typeName.find("unsigned");
431 if (pointeeTy.isCanonical() && pos != std::string::npos)
432 typeName.erase(pos+1, 8);
434 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
436 std::string baseTypeName =
437 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
441 // Turn "unsigned type" to "utype"
442 pos = baseTypeName.find("unsigned");
443 if (pos != std::string::npos)
444 baseTypeName.erase(pos+1, 8);
446 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
448 // Get argument type qualifiers:
449 if (ty.isRestrictQualified())
450 typeQuals = "restrict";
451 if (pointeeTy.isConstQualified() ||
452 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
453 typeQuals += typeQuals.empty() ? "const" : " const";
454 if (pointeeTy.isVolatileQualified())
455 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
457 uint32_t AddrSpc = 0;
458 if (ty->isImageType())
460 CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
462 addressQuals.push_back(
463 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
465 // Get argument type name.
466 std::string typeName = ty.getUnqualifiedType().getAsString(Policy);
468 // Turn "unsigned type" to "utype"
469 std::string::size_type pos = typeName.find("unsigned");
470 if (ty.isCanonical() && pos != std::string::npos)
471 typeName.erase(pos+1, 8);
473 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
475 std::string baseTypeName =
476 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
478 // Turn "unsigned type" to "utype"
479 pos = baseTypeName.find("unsigned");
480 if (pos != std::string::npos)
481 baseTypeName.erase(pos+1, 8);
483 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
485 // Get argument type qualifiers:
486 if (ty.isConstQualified())
488 if (ty.isVolatileQualified())
489 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
492 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
494 // Get image access qualifier:
495 if (ty->isImageType()) {
496 const OpenCLImageAccessAttr *A = parm->getAttr<OpenCLImageAccessAttr>();
497 if (A && A->isWriteOnly())
498 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
500 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
501 // FIXME: what about read_write?
503 accessQuals.push_back(llvm::MDString::get(Context, "none"));
505 // Get argument name.
506 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
509 kernelMDArgs.push_back(llvm::MDNode::get(Context, addressQuals));
510 kernelMDArgs.push_back(llvm::MDNode::get(Context, accessQuals));
511 kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeNames));
512 kernelMDArgs.push_back(llvm::MDNode::get(Context, argBaseTypeNames));
513 kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeQuals));
514 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
515 kernelMDArgs.push_back(llvm::MDNode::get(Context, argNames));
518 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
521 if (!FD->hasAttr<OpenCLKernelAttr>())
524 llvm::LLVMContext &Context = getLLVMContext();
526 SmallVector<llvm::Metadata *, 5> kernelMDArgs;
527 kernelMDArgs.push_back(llvm::ConstantAsMetadata::get(Fn));
529 GenOpenCLArgMetadata(FD, Fn, CGM, Context, kernelMDArgs, Builder,
532 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
533 QualType hintQTy = A->getTypeHint();
534 const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
535 bool isSignedInteger =
536 hintQTy->isSignedIntegerType() ||
537 (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
538 llvm::Metadata *attrMDArgs[] = {
539 llvm::MDString::get(Context, "vec_type_hint"),
540 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
541 CGM.getTypes().ConvertType(A->getTypeHint()))),
542 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
543 llvm::IntegerType::get(Context, 32),
544 llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
545 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
548 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
549 llvm::Metadata *attrMDArgs[] = {
550 llvm::MDString::get(Context, "work_group_size_hint"),
551 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
552 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
553 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
554 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
557 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
558 llvm::Metadata *attrMDArgs[] = {
559 llvm::MDString::get(Context, "reqd_work_group_size"),
560 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
561 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
562 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
563 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
566 llvm::MDNode *kernelMDNode = llvm::MDNode::get(Context, kernelMDArgs);
567 llvm::NamedMDNode *OpenCLKernelMetadata =
568 CGM.getModule().getOrInsertNamedMetadata("opencl.kernels");
569 OpenCLKernelMetadata->addOperand(kernelMDNode);
572 /// Determine whether the function F ends with a return stmt.
573 static bool endsWithReturn(const Decl* F) {
574 const Stmt *Body = nullptr;
575 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
576 Body = FD->getBody();
577 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
578 Body = OMD->getBody();
580 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
581 auto LastStmt = CS->body_rbegin();
582 if (LastStmt != CS->body_rend())
583 return isa<ReturnStmt>(*LastStmt);
588 void CodeGenFunction::StartFunction(GlobalDecl GD,
591 const CGFunctionInfo &FnInfo,
592 const FunctionArgList &Args,
594 SourceLocation StartLoc) {
596 "Do not use a CodeGenFunction object for more than one function");
598 const Decl *D = GD.getDecl();
600 DidCallStackSave = false;
602 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
606 assert(CurFn->isDeclaration() && "Function already has body?");
608 if (CGM.isInSanitizerBlacklist(Fn, Loc))
612 // Apply the no_sanitize* attributes to SanOpts.
613 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
614 SanOpts.Mask &= ~Attr->getMask();
617 // Apply sanitizer attributes to the function.
618 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
619 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
620 if (SanOpts.has(SanitizerKind::Thread))
621 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
622 if (SanOpts.has(SanitizerKind::Memory))
623 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
624 if (SanOpts.has(SanitizerKind::SafeStack))
625 Fn->addFnAttr(llvm::Attribute::SafeStack);
627 // Pass inline keyword to optimizer if it appears explicitly on any
628 // declaration. Also, in the case of -fno-inline attach NoInline
629 // attribute to all function that are not marked AlwaysInline.
630 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
631 if (!CGM.getCodeGenOpts().NoInline) {
632 for (auto RI : FD->redecls())
633 if (RI->isInlineSpecified()) {
634 Fn->addFnAttr(llvm::Attribute::InlineHint);
637 } else if (!FD->hasAttr<AlwaysInlineAttr>())
638 Fn->addFnAttr(llvm::Attribute::NoInline);
641 if (getLangOpts().OpenCL) {
642 // Add metadata for a kernel function.
643 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
644 EmitOpenCLKernelMetadata(FD, Fn);
647 // If we are checking function types, emit a function type signature as
649 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
650 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
651 if (llvm::Constant *PrologueSig =
652 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
653 llvm::Constant *FTRTTIConst =
654 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
655 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
656 llvm::Constant *PrologueStructConst =
657 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
658 Fn->setPrologueData(PrologueStructConst);
663 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
665 // Create a marker to make it easy to insert allocas into the entryblock
666 // later. Don't create this with the builder, because we don't want it
668 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
669 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "", EntryBB);
670 if (Builder.isNamePreserving())
671 AllocaInsertPt->setName("allocapt");
673 ReturnBlock = getJumpDestInCurrentScope("return");
675 Builder.SetInsertPoint(EntryBB);
677 // Emit subprogram debug descriptor.
678 if (CGDebugInfo *DI = getDebugInfo()) {
679 SmallVector<QualType, 16> ArgTypes;
680 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
682 ArgTypes.push_back((*i)->getType());
686 getContext().getFunctionType(RetTy, ArgTypes,
687 FunctionProtoType::ExtProtoInfo());
688 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
691 if (ShouldInstrumentFunction())
692 EmitFunctionInstrumentation("__cyg_profile_func_enter");
694 if (CGM.getCodeGenOpts().InstrumentForProfiling)
695 EmitMCountInstrumentation();
697 if (RetTy->isVoidType()) {
698 // Void type; nothing to return.
699 ReturnValue = nullptr;
701 // Count the implicit return.
702 if (!endsWithReturn(D))
704 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
705 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
706 // Indirect aggregate return; emit returned value directly into sret slot.
707 // This reduces code size, and affects correctness in C++.
708 auto AI = CurFn->arg_begin();
709 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
712 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
713 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
714 // Load the sret pointer from the argument struct and return into that.
715 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
716 llvm::Function::arg_iterator EI = CurFn->arg_end();
718 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, EI, Idx);
719 ReturnValue = Builder.CreateLoad(Addr, "agg.result");
721 ReturnValue = CreateIRTemp(RetTy, "retval");
723 // Tell the epilog emitter to autorelease the result. We do this
724 // now so that various specialized functions can suppress it
725 // during their IR-generation.
726 if (getLangOpts().ObjCAutoRefCount &&
727 !CurFnInfo->isReturnsRetained() &&
728 RetTy->isObjCRetainableType())
729 AutoreleaseResult = true;
732 EmitStartEHSpec(CurCodeDecl);
734 PrologueCleanupDepth = EHStack.stable_begin();
735 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
737 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
738 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
739 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
740 if (MD->getParent()->isLambda() &&
741 MD->getOverloadedOperator() == OO_Call) {
742 // We're in a lambda; figure out the captures.
743 MD->getParent()->getCaptureFields(LambdaCaptureFields,
744 LambdaThisCaptureField);
745 if (LambdaThisCaptureField) {
746 // If this lambda captures this, load it.
747 LValue ThisLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
748 CXXThisValue = EmitLoadOfLValue(ThisLValue,
749 SourceLocation()).getScalarVal();
751 for (auto *FD : MD->getParent()->fields()) {
752 if (FD->hasCapturedVLAType()) {
753 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
754 SourceLocation()).getScalarVal();
755 auto VAT = FD->getCapturedVLAType();
756 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
760 // Not in a lambda; just use 'this' from the method.
761 // FIXME: Should we generate a new load for each use of 'this'? The
762 // fast register allocator would be happier...
763 CXXThisValue = CXXABIThisValue;
767 // If any of the arguments have a variably modified type, make sure to
768 // emit the type size.
769 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
771 const VarDecl *VD = *i;
773 // Dig out the type as written from ParmVarDecls; it's unclear whether
774 // the standard (C99 6.9.1p10) requires this, but we're following the
775 // precedent set by gcc.
777 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
778 Ty = PVD->getOriginalType();
782 if (Ty->isVariablyModifiedType())
783 EmitVariablyModifiedType(Ty);
785 // Emit a location at the end of the prologue.
786 if (CGDebugInfo *DI = getDebugInfo())
787 DI->EmitLocation(Builder, StartLoc);
790 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
792 incrementProfileCounter(Body);
793 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
794 EmitCompoundStmtWithoutScope(*S);
799 /// When instrumenting to collect profile data, the counts for some blocks
800 /// such as switch cases need to not include the fall-through counts, so
801 /// emit a branch around the instrumentation code. When not instrumenting,
802 /// this just calls EmitBlock().
803 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
805 llvm::BasicBlock *SkipCountBB = nullptr;
806 if (HaveInsertPoint() && CGM.getCodeGenOpts().ProfileInstrGenerate) {
807 // When instrumenting for profiling, the fallthrough to certain
808 // statements needs to skip over the instrumentation code so that we
809 // get an accurate count.
810 SkipCountBB = createBasicBlock("skipcount");
811 EmitBranch(SkipCountBB);
814 uint64_t CurrentCount = getCurrentProfileCount();
815 incrementProfileCounter(S);
816 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
818 EmitBlock(SkipCountBB);
821 /// Tries to mark the given function nounwind based on the
822 /// non-existence of any throwing calls within it. We believe this is
823 /// lightweight enough to do at -O0.
824 static void TryMarkNoThrow(llvm::Function *F) {
825 // LLVM treats 'nounwind' on a function as part of the type, so we
826 // can't do this on functions that can be overwritten.
827 if (F->mayBeOverridden()) return;
829 for (llvm::Function::iterator FI = F->begin(), FE = F->end(); FI != FE; ++FI)
830 for (llvm::BasicBlock::iterator
831 BI = FI->begin(), BE = FI->end(); BI != BE; ++BI)
832 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(&*BI)) {
833 if (!Call->doesNotThrow())
835 } else if (isa<llvm::ResumeInst>(&*BI)) {
838 F->setDoesNotThrow();
841 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
842 const CGFunctionInfo &FnInfo) {
843 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
845 // Check if we should generate debug info for this function.
846 if (FD->hasAttr<NoDebugAttr>())
847 DebugInfo = nullptr; // disable debug info indefinitely for this function
849 FunctionArgList Args;
850 QualType ResTy = FD->getReturnType();
853 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
854 if (MD && MD->isInstance()) {
855 if (CGM.getCXXABI().HasThisReturn(GD))
856 ResTy = MD->getThisType(getContext());
857 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
858 ResTy = CGM.getContext().VoidPtrTy;
859 CGM.getCXXABI().buildThisParam(*this, Args);
862 Args.append(FD->param_begin(), FD->param_end());
864 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
865 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
867 SourceRange BodyRange;
868 if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
869 CurEHLocation = BodyRange.getEnd();
871 // Use the location of the start of the function to determine where
872 // the function definition is located. By default use the location
873 // of the declaration as the location for the subprogram. A function
874 // may lack a declaration in the source code if it is created by code
875 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
876 SourceLocation Loc = FD->getLocation();
878 // If this is a function specialization then use the pattern body
879 // as the location for the function.
880 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
881 if (SpecDecl->hasBody(SpecDecl))
882 Loc = SpecDecl->getLocation();
884 // Emit the standard function prologue.
885 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
887 // Generate the body of the function.
888 PGO.checkGlobalDecl(GD);
889 PGO.assignRegionCounters(GD.getDecl(), CurFn);
890 if (isa<CXXDestructorDecl>(FD))
891 EmitDestructorBody(Args);
892 else if (isa<CXXConstructorDecl>(FD))
893 EmitConstructorBody(Args);
894 else if (getLangOpts().CUDA &&
895 !getLangOpts().CUDAIsDevice &&
896 FD->hasAttr<CUDAGlobalAttr>())
897 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
898 else if (isa<CXXConversionDecl>(FD) &&
899 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
900 // The lambda conversion to block pointer is special; the semantics can't be
901 // expressed in the AST, so IRGen needs to special-case it.
902 EmitLambdaToBlockPointerBody(Args);
903 } else if (isa<CXXMethodDecl>(FD) &&
904 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
905 // The lambda static invoker function is special, because it forwards or
906 // clones the body of the function call operator (but is actually static).
907 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
908 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
909 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
910 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
911 // Implicit copy-assignment gets the same special treatment as implicit
912 // copy-constructors.
913 emitImplicitAssignmentOperatorBody(Args);
914 } else if (Stmt *Body = FD->getBody()) {
915 EmitFunctionBody(Args, Body);
917 llvm_unreachable("no definition for emitted function");
919 // C++11 [stmt.return]p2:
920 // Flowing off the end of a function [...] results in undefined behavior in
921 // a value-returning function.
923 // If the '}' that terminates a function is reached, and the value of the
924 // function call is used by the caller, the behavior is undefined.
925 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
926 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
927 if (SanOpts.has(SanitizerKind::Return)) {
928 SanitizerScope SanScope(this);
929 llvm::Value *IsFalse = Builder.getFalse();
930 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
931 "missing_return", EmitCheckSourceLocation(FD->getLocation()),
933 } else if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
934 EmitTrapCall(llvm::Intrinsic::trap);
936 Builder.CreateUnreachable();
937 Builder.ClearInsertionPoint();
940 // Emit the standard function epilogue.
941 FinishFunction(BodyRange.getEnd());
943 // If we haven't marked the function nothrow through other means, do
944 // a quick pass now to see if we can.
945 if (!CurFn->doesNotThrow())
946 TryMarkNoThrow(CurFn);
949 /// ContainsLabel - Return true if the statement contains a label in it. If
950 /// this statement is not executed normally, it not containing a label means
951 /// that we can just remove the code.
952 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
953 // Null statement, not a label!
954 if (!S) return false;
956 // If this is a label, we have to emit the code, consider something like:
957 // if (0) { ... foo: bar(); } goto foo;
959 // TODO: If anyone cared, we could track __label__'s, since we know that you
960 // can't jump to one from outside their declared region.
961 if (isa<LabelStmt>(S))
964 // If this is a case/default statement, and we haven't seen a switch, we have
966 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
969 // If this is a switch statement, we want to ignore cases below it.
970 if (isa<SwitchStmt>(S))
971 IgnoreCaseStmts = true;
973 // Scan subexpressions for verboten labels.
974 for (const Stmt *SubStmt : S->children())
975 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
981 /// containsBreak - Return true if the statement contains a break out of it.
982 /// If the statement (recursively) contains a switch or loop with a break
983 /// inside of it, this is fine.
984 bool CodeGenFunction::containsBreak(const Stmt *S) {
985 // Null statement, not a label!
986 if (!S) return false;
988 // If this is a switch or loop that defines its own break scope, then we can
989 // include it and anything inside of it.
990 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
994 if (isa<BreakStmt>(S))
997 // Scan subexpressions for verboten breaks.
998 for (const Stmt *SubStmt : S->children())
999 if (containsBreak(SubStmt))
1006 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1007 /// to a constant, or if it does but contains a label, return false. If it
1008 /// constant folds return true and set the boolean result in Result.
1009 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1011 llvm::APSInt ResultInt;
1012 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt))
1015 ResultBool = ResultInt.getBoolValue();
1019 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1020 /// to a constant, or if it does but contains a label, return false. If it
1021 /// constant folds return true and set the folded value.
1022 bool CodeGenFunction::
1023 ConstantFoldsToSimpleInteger(const Expr *Cond, llvm::APSInt &ResultInt) {
1024 // FIXME: Rename and handle conversion of other evaluatable things
1027 if (!Cond->EvaluateAsInt(Int, getContext()))
1028 return false; // Not foldable, not integer or not fully evaluatable.
1030 if (CodeGenFunction::ContainsLabel(Cond))
1031 return false; // Contains a label.
1039 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1040 /// statement) to the specified blocks. Based on the condition, this might try
1041 /// to simplify the codegen of the conditional based on the branch.
1043 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1044 llvm::BasicBlock *TrueBlock,
1045 llvm::BasicBlock *FalseBlock,
1046 uint64_t TrueCount) {
1047 Cond = Cond->IgnoreParens();
1049 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1051 // Handle X && Y in a condition.
1052 if (CondBOp->getOpcode() == BO_LAnd) {
1053 // If we have "1 && X", simplify the code. "0 && X" would have constant
1054 // folded if the case was simple enough.
1055 bool ConstantBool = false;
1056 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1058 // br(1 && X) -> br(X).
1059 incrementProfileCounter(CondBOp);
1060 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1064 // If we have "X && 1", simplify the code to use an uncond branch.
1065 // "X && 0" would have been constant folded to 0.
1066 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1068 // br(X && 1) -> br(X).
1069 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1073 // Emit the LHS as a conditional. If the LHS conditional is false, we
1074 // want to jump to the FalseBlock.
1075 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1076 // The counter tells us how often we evaluate RHS, and all of TrueCount
1077 // can be propagated to that branch.
1078 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1080 ConditionalEvaluation eval(*this);
1082 ApplyDebugLocation DL(*this, Cond);
1083 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1087 incrementProfileCounter(CondBOp);
1088 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1090 // Any temporaries created here are conditional.
1092 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1098 if (CondBOp->getOpcode() == BO_LOr) {
1099 // If we have "0 || X", simplify the code. "1 || X" would have constant
1100 // folded if the case was simple enough.
1101 bool ConstantBool = false;
1102 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1104 // br(0 || X) -> br(X).
1105 incrementProfileCounter(CondBOp);
1106 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1110 // If we have "X || 0", simplify the code to use an uncond branch.
1111 // "X || 1" would have been constant folded to 1.
1112 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1114 // br(X || 0) -> br(X).
1115 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1119 // Emit the LHS as a conditional. If the LHS conditional is true, we
1120 // want to jump to the TrueBlock.
1121 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1122 // We have the count for entry to the RHS and for the whole expression
1123 // being true, so we can divy up True count between the short circuit and
1126 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1127 uint64_t RHSCount = TrueCount - LHSCount;
1129 ConditionalEvaluation eval(*this);
1131 ApplyDebugLocation DL(*this, Cond);
1132 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1133 EmitBlock(LHSFalse);
1136 incrementProfileCounter(CondBOp);
1137 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1139 // Any temporaries created here are conditional.
1141 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1149 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1150 // br(!x, t, f) -> br(x, f, t)
1151 if (CondUOp->getOpcode() == UO_LNot) {
1152 // Negate the count.
1153 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1154 // Negate the condition and swap the destination blocks.
1155 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1160 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1161 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1162 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1163 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1165 ConditionalEvaluation cond(*this);
1166 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1167 getProfileCount(CondOp));
1169 // When computing PGO branch weights, we only know the overall count for
1170 // the true block. This code is essentially doing tail duplication of the
1171 // naive code-gen, introducing new edges for which counts are not
1172 // available. Divide the counts proportionally between the LHS and RHS of
1173 // the conditional operator.
1174 uint64_t LHSScaledTrueCount = 0;
1177 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1178 LHSScaledTrueCount = TrueCount * LHSRatio;
1182 EmitBlock(LHSBlock);
1183 incrementProfileCounter(CondOp);
1185 ApplyDebugLocation DL(*this, Cond);
1186 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1187 LHSScaledTrueCount);
1192 EmitBlock(RHSBlock);
1193 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1194 TrueCount - LHSScaledTrueCount);
1200 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1201 // Conditional operator handling can give us a throw expression as a
1202 // condition for a case like:
1203 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1205 // br(c, throw x, br(y, t, f))
1206 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1210 // Create branch weights based on the number of times we get here and the
1211 // number of times the condition should be true.
1212 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1213 llvm::MDNode *Weights =
1214 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1216 // Emit the code with the fully general case.
1219 ApplyDebugLocation DL(*this, Cond);
1220 CondV = EvaluateExprAsBool(Cond);
1222 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights);
1225 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1226 /// specified stmt yet.
1227 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1228 CGM.ErrorUnsupported(S, Type);
1231 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1232 /// variable-length array whose elements have a non-zero bit-pattern.
1234 /// \param baseType the inner-most element type of the array
1235 /// \param src - a char* pointing to the bit-pattern for a single
1236 /// base element of the array
1237 /// \param sizeInChars - the total size of the VLA, in chars
1238 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1239 llvm::Value *dest, llvm::Value *src,
1240 llvm::Value *sizeInChars) {
1241 std::pair<CharUnits,CharUnits> baseSizeAndAlign
1242 = CGF.getContext().getTypeInfoInChars(baseType);
1244 CGBuilderTy &Builder = CGF.Builder;
1246 llvm::Value *baseSizeInChars
1247 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSizeAndAlign.first.getQuantity());
1249 llvm::Type *i8p = Builder.getInt8PtrTy();
1251 llvm::Value *begin = Builder.CreateBitCast(dest, i8p, "vla.begin");
1252 llvm::Value *end = Builder.CreateInBoundsGEP(dest, sizeInChars, "vla.end");
1254 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1255 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1256 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1258 // Make a loop over the VLA. C99 guarantees that the VLA element
1259 // count must be nonzero.
1260 CGF.EmitBlock(loopBB);
1262 llvm::PHINode *cur = Builder.CreatePHI(i8p, 2, "vla.cur");
1263 cur->addIncoming(begin, originBB);
1265 // memcpy the individual element bit-pattern.
1266 Builder.CreateMemCpy(cur, src, baseSizeInChars,
1267 baseSizeAndAlign.second.getQuantity(),
1268 /*volatile*/ false);
1270 // Go to the next element.
1271 llvm::Value *next = Builder.CreateConstInBoundsGEP1_32(Builder.getInt8Ty(),
1272 cur, 1, "vla.next");
1274 // Leave if that's the end of the VLA.
1275 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1276 Builder.CreateCondBr(done, contBB, loopBB);
1277 cur->addIncoming(next, loopBB);
1279 CGF.EmitBlock(contBB);
1283 CodeGenFunction::EmitNullInitialization(llvm::Value *DestPtr, QualType Ty) {
1284 // Ignore empty classes in C++.
1285 if (getLangOpts().CPlusPlus) {
1286 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1287 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1292 // Cast the dest ptr to the appropriate i8 pointer type.
1294 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace();
1295 llvm::Type *BP = Builder.getInt8PtrTy(DestAS);
1296 if (DestPtr->getType() != BP)
1297 DestPtr = Builder.CreateBitCast(DestPtr, BP);
1299 // Get size and alignment info for this aggregate.
1300 std::pair<CharUnits, CharUnits> TypeInfo =
1301 getContext().getTypeInfoInChars(Ty);
1302 CharUnits Size = TypeInfo.first;
1303 CharUnits Align = TypeInfo.second;
1305 llvm::Value *SizeVal;
1306 const VariableArrayType *vla;
1308 // Don't bother emitting a zero-byte memset.
1309 if (Size.isZero()) {
1310 // But note that getTypeInfo returns 0 for a VLA.
1311 if (const VariableArrayType *vlaType =
1312 dyn_cast_or_null<VariableArrayType>(
1313 getContext().getAsArrayType(Ty))) {
1315 llvm::Value *numElts;
1316 std::tie(numElts, eltType) = getVLASize(vlaType);
1319 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1320 if (!eltSize.isOne())
1321 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1327 SizeVal = CGM.getSize(Size);
1331 // If the type contains a pointer to data member we can't memset it to zero.
1332 // Instead, create a null constant and copy it to the destination.
1333 // TODO: there are other patterns besides zero that we can usefully memset,
1334 // like -1, which happens to be the pattern used by member-pointers.
1335 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1336 // For a VLA, emit a single element, then splat that over the VLA.
1337 if (vla) Ty = getContext().getBaseElementType(vla);
1339 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1341 llvm::GlobalVariable *NullVariable =
1342 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1343 /*isConstant=*/true,
1344 llvm::GlobalVariable::PrivateLinkage,
1345 NullConstant, Twine());
1346 llvm::Value *SrcPtr =
1347 Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy());
1349 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1351 // Get and call the appropriate llvm.memcpy overload.
1352 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, Align.getQuantity(), false);
1356 // Otherwise, just memset the whole thing to zero. This is legal
1357 // because in LLVM, all default initializers (other than the ones we just
1358 // handled above) are guaranteed to have a bit pattern of all zeros.
1359 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal,
1360 Align.getQuantity(), false);
1363 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1364 // Make sure that there is a block for the indirect goto.
1365 if (!IndirectBranch)
1366 GetIndirectGotoBlock();
1368 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1370 // Make sure the indirect branch includes all of the address-taken blocks.
1371 IndirectBranch->addDestination(BB);
1372 return llvm::BlockAddress::get(CurFn, BB);
1375 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1376 // If we already made the indirect branch for indirect goto, return its block.
1377 if (IndirectBranch) return IndirectBranch->getParent();
1379 CGBuilderTy TmpBuilder(createBasicBlock("indirectgoto"));
1381 // Create the PHI node that indirect gotos will add entries to.
1382 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1383 "indirect.goto.dest");
1385 // Create the indirect branch instruction.
1386 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1387 return IndirectBranch->getParent();
1390 /// Computes the length of an array in elements, as well as the base
1391 /// element type and a properly-typed first element pointer.
1392 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1394 llvm::Value *&addr) {
1395 const ArrayType *arrayType = origArrayType;
1397 // If it's a VLA, we have to load the stored size. Note that
1398 // this is the size of the VLA in bytes, not its size in elements.
1399 llvm::Value *numVLAElements = nullptr;
1400 if (isa<VariableArrayType>(arrayType)) {
1401 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1403 // Walk into all VLAs. This doesn't require changes to addr,
1404 // which has type T* where T is the first non-VLA element type.
1406 QualType elementType = arrayType->getElementType();
1407 arrayType = getContext().getAsArrayType(elementType);
1409 // If we only have VLA components, 'addr' requires no adjustment.
1411 baseType = elementType;
1412 return numVLAElements;
1414 } while (isa<VariableArrayType>(arrayType));
1416 // We get out here only if we find a constant array type
1420 // We have some number of constant-length arrays, so addr should
1421 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1422 // down to the first element of addr.
1423 SmallVector<llvm::Value*, 8> gepIndices;
1425 // GEP down to the array type.
1426 llvm::ConstantInt *zero = Builder.getInt32(0);
1427 gepIndices.push_back(zero);
1429 uint64_t countFromCLAs = 1;
1432 llvm::ArrayType *llvmArrayType =
1433 dyn_cast<llvm::ArrayType>(
1434 cast<llvm::PointerType>(addr->getType())->getElementType());
1435 while (llvmArrayType) {
1436 assert(isa<ConstantArrayType>(arrayType));
1437 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1438 == llvmArrayType->getNumElements());
1440 gepIndices.push_back(zero);
1441 countFromCLAs *= llvmArrayType->getNumElements();
1442 eltType = arrayType->getElementType();
1445 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1446 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1447 assert((!llvmArrayType || arrayType) &&
1448 "LLVM and Clang types are out-of-synch");
1452 // From this point onwards, the Clang array type has been emitted
1453 // as some other type (probably a packed struct). Compute the array
1454 // size, and just emit the 'begin' expression as a bitcast.
1457 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1458 eltType = arrayType->getElementType();
1459 arrayType = getContext().getAsArrayType(eltType);
1462 unsigned AddressSpace = addr->getType()->getPointerAddressSpace();
1463 llvm::Type *BaseType = ConvertType(eltType)->getPointerTo(AddressSpace);
1464 addr = Builder.CreateBitCast(addr, BaseType, "array.begin");
1466 // Create the actual GEP.
1467 addr = Builder.CreateInBoundsGEP(addr, gepIndices, "array.begin");
1472 llvm::Value *numElements
1473 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1475 // If we had any VLA dimensions, factor them in.
1477 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1482 std::pair<llvm::Value*, QualType>
1483 CodeGenFunction::getVLASize(QualType type) {
1484 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1485 assert(vla && "type was not a variable array type!");
1486 return getVLASize(vla);
1489 std::pair<llvm::Value*, QualType>
1490 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1491 // The number of elements so far; always size_t.
1492 llvm::Value *numElements = nullptr;
1494 QualType elementType;
1496 elementType = type->getElementType();
1497 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1498 assert(vlaSize && "no size for VLA!");
1499 assert(vlaSize->getType() == SizeTy);
1502 numElements = vlaSize;
1504 // It's undefined behavior if this wraps around, so mark it that way.
1505 // FIXME: Teach -fsanitize=undefined to trap this.
1506 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1508 } while ((type = getContext().getAsVariableArrayType(elementType)));
1510 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1513 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1514 assert(type->isVariablyModifiedType() &&
1515 "Must pass variably modified type to EmitVLASizes!");
1517 EnsureInsertPoint();
1519 // We're going to walk down into the type and look for VLA
1522 assert(type->isVariablyModifiedType());
1524 const Type *ty = type.getTypePtr();
1525 switch (ty->getTypeClass()) {
1527 #define TYPE(Class, Base)
1528 #define ABSTRACT_TYPE(Class, Base)
1529 #define NON_CANONICAL_TYPE(Class, Base)
1530 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1531 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1532 #include "clang/AST/TypeNodes.def"
1533 llvm_unreachable("unexpected dependent type!");
1535 // These types are never variably-modified.
1539 case Type::ExtVector:
1542 case Type::Elaborated:
1543 case Type::TemplateSpecialization:
1544 case Type::ObjCObject:
1545 case Type::ObjCInterface:
1546 case Type::ObjCObjectPointer:
1547 llvm_unreachable("type class is never variably-modified!");
1549 case Type::Adjusted:
1550 type = cast<AdjustedType>(ty)->getAdjustedType();
1554 type = cast<DecayedType>(ty)->getPointeeType();
1558 type = cast<PointerType>(ty)->getPointeeType();
1561 case Type::BlockPointer:
1562 type = cast<BlockPointerType>(ty)->getPointeeType();
1565 case Type::LValueReference:
1566 case Type::RValueReference:
1567 type = cast<ReferenceType>(ty)->getPointeeType();
1570 case Type::MemberPointer:
1571 type = cast<MemberPointerType>(ty)->getPointeeType();
1574 case Type::ConstantArray:
1575 case Type::IncompleteArray:
1576 // Losing element qualification here is fine.
1577 type = cast<ArrayType>(ty)->getElementType();
1580 case Type::VariableArray: {
1581 // Losing element qualification here is fine.
1582 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1584 // Unknown size indication requires no size computation.
1585 // Otherwise, evaluate and record it.
1586 if (const Expr *size = vat->getSizeExpr()) {
1587 // It's possible that we might have emitted this already,
1588 // e.g. with a typedef and a pointer to it.
1589 llvm::Value *&entry = VLASizeMap[size];
1591 llvm::Value *Size = EmitScalarExpr(size);
1594 // If the size is an expression that is not an integer constant
1595 // expression [...] each time it is evaluated it shall have a value
1596 // greater than zero.
1597 if (SanOpts.has(SanitizerKind::VLABound) &&
1598 size->getType()->isSignedIntegerType()) {
1599 SanitizerScope SanScope(this);
1600 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1601 llvm::Constant *StaticArgs[] = {
1602 EmitCheckSourceLocation(size->getLocStart()),
1603 EmitCheckTypeDescriptor(size->getType())
1605 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1606 SanitizerKind::VLABound),
1607 "vla_bound_not_positive", StaticArgs, Size);
1610 // Always zexting here would be wrong if it weren't
1611 // undefined behavior to have a negative bound.
1612 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1615 type = vat->getElementType();
1619 case Type::FunctionProto:
1620 case Type::FunctionNoProto:
1621 type = cast<FunctionType>(ty)->getReturnType();
1626 case Type::UnaryTransform:
1627 case Type::Attributed:
1628 case Type::SubstTemplateTypeParm:
1629 case Type::PackExpansion:
1630 // Keep walking after single level desugaring.
1631 type = type.getSingleStepDesugaredType(getContext());
1635 case Type::Decltype:
1637 // Stop walking: nothing to do.
1640 case Type::TypeOfExpr:
1641 // Stop walking: emit typeof expression.
1642 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1646 type = cast<AtomicType>(ty)->getValueType();
1649 } while (type->isVariablyModifiedType());
1652 llvm::Value* CodeGenFunction::EmitVAListRef(const Expr* E) {
1653 if (getContext().getBuiltinVaListType()->isArrayType())
1654 return EmitScalarExpr(E);
1655 return EmitLValue(E).getAddress();
1658 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1659 llvm::Constant *Init) {
1660 assert (Init && "Invalid DeclRefExpr initializer!");
1661 if (CGDebugInfo *Dbg = getDebugInfo())
1662 if (CGM.getCodeGenOpts().getDebugInfo() >= CodeGenOptions::LimitedDebugInfo)
1663 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1666 CodeGenFunction::PeepholeProtection
1667 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1668 // At the moment, the only aggressive peephole we do in IR gen
1669 // is trunc(zext) folding, but if we add more, we can easily
1670 // extend this protection.
1672 if (!rvalue.isScalar()) return PeepholeProtection();
1673 llvm::Value *value = rvalue.getScalarVal();
1674 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1676 // Just make an extra bitcast.
1677 assert(HaveInsertPoint());
1678 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1679 Builder.GetInsertBlock());
1681 PeepholeProtection protection;
1682 protection.Inst = inst;
1686 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1687 if (!protection.Inst) return;
1689 // In theory, we could try to duplicate the peepholes now, but whatever.
1690 protection.Inst->eraseFromParent();
1693 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1694 llvm::Value *AnnotatedVal,
1695 StringRef AnnotationStr,
1696 SourceLocation Location) {
1697 llvm::Value *Args[4] = {
1699 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1700 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1701 CGM.EmitAnnotationLineNo(Location)
1703 return Builder.CreateCall(AnnotationFn, Args);
1706 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1707 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1708 // FIXME We create a new bitcast for every annotation because that's what
1709 // llvm-gcc was doing.
1710 for (const auto *I : D->specific_attrs<AnnotateAttr>())
1711 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1712 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1713 I->getAnnotation(), D->getLocation());
1716 llvm::Value *CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1718 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1719 llvm::Type *VTy = V->getType();
1720 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1723 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1724 // FIXME Always emit the cast inst so we can differentiate between
1725 // annotation on the first field of a struct and annotation on the struct
1727 if (VTy != CGM.Int8PtrTy)
1728 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1729 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1730 V = Builder.CreateBitCast(V, VTy);
1736 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
1738 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
1740 assert(!CGF->IsSanitizerScope);
1741 CGF->IsSanitizerScope = true;
1744 CodeGenFunction::SanitizerScope::~SanitizerScope() {
1745 CGF->IsSanitizerScope = false;
1748 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
1749 const llvm::Twine &Name,
1750 llvm::BasicBlock *BB,
1751 llvm::BasicBlock::iterator InsertPt) const {
1752 LoopStack.InsertHelper(I);
1753 if (IsSanitizerScope)
1754 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
1757 template <bool PreserveNames>
1758 void CGBuilderInserter<PreserveNames>::InsertHelper(
1759 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1760 llvm::BasicBlock::iterator InsertPt) const {
1761 llvm::IRBuilderDefaultInserter<PreserveNames>::InsertHelper(I, Name, BB,
1764 CGF->InsertHelper(I, Name, BB, InsertPt);
1768 #define PreserveNames false
1770 #define PreserveNames true
1772 template void CGBuilderInserter<PreserveNames>::InsertHelper(
1773 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1774 llvm::BasicBlock::iterator InsertPt) const;
1775 #undef PreserveNames