1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This coordinates the per-function state used while generating code.
11 //===----------------------------------------------------------------------===//
13 #include "CodeGenFunction.h"
15 #include "CGCUDARuntime.h"
17 #include "CGCleanup.h"
18 #include "CGDebugInfo.h"
19 #include "CGOpenMPRuntime.h"
20 #include "CodeGenModule.h"
21 #include "CodeGenPGO.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/ASTContext.h"
24 #include "clang/AST/ASTLambda.h"
25 #include "clang/AST/Attr.h"
26 #include "clang/AST/Decl.h"
27 #include "clang/AST/DeclCXX.h"
28 #include "clang/AST/StmtCXX.h"
29 #include "clang/AST/StmtObjC.h"
30 #include "clang/Basic/Builtins.h"
31 #include "clang/Basic/CodeGenOptions.h"
32 #include "clang/Basic/TargetInfo.h"
33 #include "clang/CodeGen/CGFunctionInfo.h"
34 #include "clang/Frontend/FrontendDiagnostic.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/FPEnv.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Intrinsics.h"
40 #include "llvm/IR/MDBuilder.h"
41 #include "llvm/IR/Operator.h"
42 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
43 using namespace clang;
44 using namespace CodeGen;
46 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
48 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
49 const LangOptions &LangOpts) {
50 if (CGOpts.DisableLifetimeMarkers)
53 // Sanitizers may use markers.
54 if (CGOpts.SanitizeAddressUseAfterScope ||
55 LangOpts.Sanitize.has(SanitizerKind::HWAddress) ||
56 LangOpts.Sanitize.has(SanitizerKind::Memory))
59 // For now, only in optimized builds.
60 return CGOpts.OptimizationLevel != 0;
63 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
64 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
65 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
66 CGBuilderInserterTy(this)),
67 SanOpts(CGM.getLangOpts().Sanitize), DebugInfo(CGM.getModuleDebugInfo()),
68 PGO(cgm), ShouldEmitLifetimeMarkers(shouldEmitLifetimeMarkers(
69 CGM.getCodeGenOpts(), CGM.getLangOpts())) {
70 if (!suppressNewContext)
71 CGM.getCXXABI().getMangleContext().startNewFunction();
73 llvm::FastMathFlags FMF;
74 if (CGM.getLangOpts().FastMath)
76 if (CGM.getLangOpts().FiniteMathOnly) {
80 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
83 if (CGM.getCodeGenOpts().NoSignedZeros) {
84 FMF.setNoSignedZeros();
86 if (CGM.getCodeGenOpts().ReciprocalMath) {
87 FMF.setAllowReciprocal();
89 if (CGM.getCodeGenOpts().Reassociate) {
90 FMF.setAllowReassoc();
92 Builder.setFastMathFlags(FMF);
96 CodeGenFunction::~CodeGenFunction() {
97 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
99 // If there are any unclaimed block infos, go ahead and destroy them
100 // now. This can happen if IR-gen gets clever and skips evaluating
103 destroyBlockInfos(FirstBlockInfo);
105 if (getLangOpts().OpenMP && CurFn)
106 CGM.getOpenMPRuntime().functionFinished(*this);
109 // Map the LangOption for rounding mode into
110 // the corresponding enum in the IR.
111 static llvm::fp::RoundingMode ToConstrainedRoundingMD(
112 LangOptions::FPRoundingModeKind Kind) {
115 case LangOptions::FPR_ToNearest: return llvm::fp::rmToNearest;
116 case LangOptions::FPR_Downward: return llvm::fp::rmDownward;
117 case LangOptions::FPR_Upward: return llvm::fp::rmUpward;
118 case LangOptions::FPR_TowardZero: return llvm::fp::rmTowardZero;
119 case LangOptions::FPR_Dynamic: return llvm::fp::rmDynamic;
121 llvm_unreachable("Unsupported FP RoundingMode");
124 // Map the LangOption for exception behavior into
125 // the corresponding enum in the IR.
126 static llvm::fp::ExceptionBehavior ToConstrainedExceptMD(
127 LangOptions::FPExceptionModeKind Kind) {
130 case LangOptions::FPE_Ignore: return llvm::fp::ebIgnore;
131 case LangOptions::FPE_MayTrap: return llvm::fp::ebMayTrap;
132 case LangOptions::FPE_Strict: return llvm::fp::ebStrict;
134 llvm_unreachable("Unsupported FP Exception Behavior");
137 void CodeGenFunction::SetFPModel() {
138 auto fpRoundingMode = ToConstrainedRoundingMD(
139 getLangOpts().getFPRoundingMode());
140 auto fpExceptionBehavior = ToConstrainedExceptMD(
141 getLangOpts().getFPExceptionMode());
143 if (fpExceptionBehavior == llvm::fp::ebIgnore &&
144 fpRoundingMode == llvm::fp::rmToNearest)
145 // Constrained intrinsics are not used.
148 Builder.setIsFPConstrained(true);
149 Builder.setDefaultConstrainedRounding(fpRoundingMode);
150 Builder.setDefaultConstrainedExcept(fpExceptionBehavior);
154 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
155 LValueBaseInfo *BaseInfo,
156 TBAAAccessInfo *TBAAInfo) {
157 return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo, TBAAInfo,
158 /* forPointeeType= */ true);
161 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
162 LValueBaseInfo *BaseInfo,
163 TBAAAccessInfo *TBAAInfo,
164 bool forPointeeType) {
166 *TBAAInfo = CGM.getTBAAAccessInfo(T);
168 // Honor alignment typedef attributes even on incomplete types.
169 // We also honor them straight for C++ class types, even as pointees;
170 // there's an expressivity gap here.
171 if (auto TT = T->getAs<TypedefType>()) {
172 if (auto Align = TT->getDecl()->getMaxAlignment()) {
174 *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType);
175 return getContext().toCharUnitsFromBits(Align);
180 *BaseInfo = LValueBaseInfo(AlignmentSource::Type);
183 if (T->isIncompleteType()) {
184 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
186 // For C++ class pointees, we don't know whether we're pointing at a
187 // base or a complete object, so we generally need to use the
188 // non-virtual alignment.
189 const CXXRecordDecl *RD;
190 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
191 Alignment = CGM.getClassPointerAlignment(RD);
193 Alignment = getContext().getTypeAlignInChars(T);
194 if (T.getQualifiers().hasUnaligned())
195 Alignment = CharUnits::One();
198 // Cap to the global maximum type alignment unless the alignment
199 // was somehow explicit on the type.
200 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
201 if (Alignment.getQuantity() > MaxAlign &&
202 !getContext().isAlignmentRequired(T))
203 Alignment = CharUnits::fromQuantity(MaxAlign);
209 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
210 LValueBaseInfo BaseInfo;
211 TBAAAccessInfo TBAAInfo;
212 CharUnits Alignment = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo);
213 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo,
217 /// Given a value of type T* that may not be to a complete object,
218 /// construct an l-value with the natural pointee alignment of T.
220 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
221 LValueBaseInfo BaseInfo;
222 TBAAAccessInfo TBAAInfo;
223 CharUnits Align = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo,
224 /* forPointeeType= */ true);
225 return MakeAddrLValue(Address(V, Align), T, BaseInfo, TBAAInfo);
229 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
230 return CGM.getTypes().ConvertTypeForMem(T);
233 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
234 return CGM.getTypes().ConvertType(T);
237 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
238 type = type.getCanonicalType();
240 switch (type->getTypeClass()) {
241 #define TYPE(name, parent)
242 #define ABSTRACT_TYPE(name, parent)
243 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
244 #define DEPENDENT_TYPE(name, parent) case Type::name:
245 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
246 #include "clang/AST/TypeNodes.inc"
247 llvm_unreachable("non-canonical or dependent type in IR-generation");
250 case Type::DeducedTemplateSpecialization:
251 llvm_unreachable("undeduced type in IR-generation");
253 // Various scalar types.
256 case Type::BlockPointer:
257 case Type::LValueReference:
258 case Type::RValueReference:
259 case Type::MemberPointer:
261 case Type::ExtVector:
262 case Type::FunctionProto:
263 case Type::FunctionNoProto:
265 case Type::ObjCObjectPointer:
273 // Arrays, records, and Objective-C objects.
274 case Type::ConstantArray:
275 case Type::IncompleteArray:
276 case Type::VariableArray:
278 case Type::ObjCObject:
279 case Type::ObjCInterface:
280 return TEK_Aggregate;
282 // We operate on atomic values according to their underlying type.
284 type = cast<AtomicType>(type)->getValueType();
287 llvm_unreachable("unknown type kind!");
291 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
292 // For cleanliness, we try to avoid emitting the return block for
294 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
297 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
299 // We have a valid insert point, reuse it if it is empty or there are no
300 // explicit jumps to the return block.
301 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
302 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
303 delete ReturnBlock.getBlock();
304 ReturnBlock = JumpDest();
306 EmitBlock(ReturnBlock.getBlock());
307 return llvm::DebugLoc();
310 // Otherwise, if the return block is the target of a single direct
311 // branch then we can just put the code in that block instead. This
312 // cleans up functions which started with a unified return block.
313 if (ReturnBlock.getBlock()->hasOneUse()) {
314 llvm::BranchInst *BI =
315 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
316 if (BI && BI->isUnconditional() &&
317 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
318 // Record/return the DebugLoc of the simple 'return' expression to be used
319 // later by the actual 'ret' instruction.
320 llvm::DebugLoc Loc = BI->getDebugLoc();
321 Builder.SetInsertPoint(BI->getParent());
322 BI->eraseFromParent();
323 delete ReturnBlock.getBlock();
324 ReturnBlock = JumpDest();
329 // FIXME: We are at an unreachable point, there is no reason to emit the block
330 // unless it has uses. However, we still need a place to put the debug
331 // region.end for now.
333 EmitBlock(ReturnBlock.getBlock());
334 return llvm::DebugLoc();
337 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
339 if (!BB->use_empty())
340 return CGF.CurFn->getBasicBlockList().push_back(BB);
344 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
345 assert(BreakContinueStack.empty() &&
346 "mismatched push/pop in break/continue stack!");
348 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
349 && NumSimpleReturnExprs == NumReturnExprs
350 && ReturnBlock.getBlock()->use_empty();
351 // Usually the return expression is evaluated before the cleanup
352 // code. If the function contains only a simple return statement,
353 // such as a constant, the location before the cleanup code becomes
354 // the last useful breakpoint in the function, because the simple
355 // return expression will be evaluated after the cleanup code. To be
356 // safe, set the debug location for cleanup code to the location of
357 // the return statement. Otherwise the cleanup code should be at the
358 // end of the function's lexical scope.
360 // If there are multiple branches to the return block, the branch
361 // instructions will get the location of the return statements and
363 if (CGDebugInfo *DI = getDebugInfo()) {
364 if (OnlySimpleReturnStmts)
365 DI->EmitLocation(Builder, LastStopPoint);
367 DI->EmitLocation(Builder, EndLoc);
370 // Pop any cleanups that might have been associated with the
371 // parameters. Do this in whatever block we're currently in; it's
372 // important to do this before we enter the return block or return
373 // edges will be *really* confused.
374 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
375 bool HasOnlyLifetimeMarkers =
376 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
377 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
379 // Make sure the line table doesn't jump back into the body for
380 // the ret after it's been at EndLoc.
381 Optional<ApplyDebugLocation> AL;
382 if (CGDebugInfo *DI = getDebugInfo()) {
383 if (OnlySimpleReturnStmts)
384 DI->EmitLocation(Builder, EndLoc);
386 // We may not have a valid end location. Try to apply it anyway, and
387 // fall back to an artificial location if needed.
388 AL = ApplyDebugLocation::CreateDefaultArtificial(*this, EndLoc);
391 PopCleanupBlocks(PrologueCleanupDepth);
394 // Emit function epilog (to return).
395 llvm::DebugLoc Loc = EmitReturnBlock();
397 if (ShouldInstrumentFunction()) {
398 if (CGM.getCodeGenOpts().InstrumentFunctions)
399 CurFn->addFnAttr("instrument-function-exit", "__cyg_profile_func_exit");
400 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
401 CurFn->addFnAttr("instrument-function-exit-inlined",
402 "__cyg_profile_func_exit");
405 // Emit debug descriptor for function end.
406 if (CGDebugInfo *DI = getDebugInfo())
407 DI->EmitFunctionEnd(Builder, CurFn);
409 // Reset the debug location to that of the simple 'return' expression, if any
410 // rather than that of the end of the function's scope '}'.
411 ApplyDebugLocation AL(*this, Loc);
412 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
413 EmitEndEHSpec(CurCodeDecl);
415 assert(EHStack.empty() &&
416 "did not remove all scopes from cleanup stack!");
418 // If someone did an indirect goto, emit the indirect goto block at the end of
420 if (IndirectBranch) {
421 EmitBlock(IndirectBranch->getParent());
422 Builder.ClearInsertionPoint();
425 // If some of our locals escaped, insert a call to llvm.localescape in the
427 if (!EscapedLocals.empty()) {
428 // Invert the map from local to index into a simple vector. There should be
430 SmallVector<llvm::Value *, 4> EscapeArgs;
431 EscapeArgs.resize(EscapedLocals.size());
432 for (auto &Pair : EscapedLocals)
433 EscapeArgs[Pair.second] = Pair.first;
434 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
435 &CGM.getModule(), llvm::Intrinsic::localescape);
436 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
439 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
440 llvm::Instruction *Ptr = AllocaInsertPt;
441 AllocaInsertPt = nullptr;
442 Ptr->eraseFromParent();
444 // If someone took the address of a label but never did an indirect goto, we
445 // made a zero entry PHI node, which is illegal, zap it now.
446 if (IndirectBranch) {
447 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
448 if (PN->getNumIncomingValues() == 0) {
449 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
450 PN->eraseFromParent();
454 EmitIfUsed(*this, EHResumeBlock);
455 EmitIfUsed(*this, TerminateLandingPad);
456 EmitIfUsed(*this, TerminateHandler);
457 EmitIfUsed(*this, UnreachableBlock);
459 for (const auto &FuncletAndParent : TerminateFunclets)
460 EmitIfUsed(*this, FuncletAndParent.second);
462 if (CGM.getCodeGenOpts().EmitDeclMetadata)
465 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
466 I = DeferredReplacements.begin(),
467 E = DeferredReplacements.end();
469 I->first->replaceAllUsesWith(I->second);
470 I->first->eraseFromParent();
473 // Eliminate CleanupDestSlot alloca by replacing it with SSA values and
474 // PHIs if the current function is a coroutine. We don't do it for all
475 // functions as it may result in slight increase in numbers of instructions
476 // if compiled with no optimizations. We do it for coroutine as the lifetime
477 // of CleanupDestSlot alloca make correct coroutine frame building very
479 if (NormalCleanupDest.isValid() && isCoroutine()) {
480 llvm::DominatorTree DT(*CurFn);
481 llvm::PromoteMemToReg(
482 cast<llvm::AllocaInst>(NormalCleanupDest.getPointer()), DT);
483 NormalCleanupDest = Address::invalid();
486 // Scan function arguments for vector width.
487 for (llvm::Argument &A : CurFn->args())
488 if (auto *VT = dyn_cast<llvm::VectorType>(A.getType()))
489 LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
490 VT->getPrimitiveSizeInBits().getFixedSize());
492 // Update vector width based on return type.
493 if (auto *VT = dyn_cast<llvm::VectorType>(CurFn->getReturnType()))
494 LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
495 VT->getPrimitiveSizeInBits().getFixedSize());
497 // Add the required-vector-width attribute. This contains the max width from:
498 // 1. min-vector-width attribute used in the source program.
499 // 2. Any builtins used that have a vector width specified.
500 // 3. Values passed in and out of inline assembly.
501 // 4. Width of vector arguments and return types for this function.
502 // 5. Width of vector aguments and return types for functions called by this
504 CurFn->addFnAttr("min-legal-vector-width", llvm::utostr(LargestVectorWidth));
506 // If we generated an unreachable return block, delete it now.
507 if (ReturnBlock.isValid() && ReturnBlock.getBlock()->use_empty()) {
508 Builder.ClearInsertionPoint();
509 ReturnBlock.getBlock()->eraseFromParent();
511 if (ReturnValue.isValid()) {
512 auto *RetAlloca = dyn_cast<llvm::AllocaInst>(ReturnValue.getPointer());
513 if (RetAlloca && RetAlloca->use_empty()) {
514 RetAlloca->eraseFromParent();
515 ReturnValue = Address::invalid();
520 /// ShouldInstrumentFunction - Return true if the current function should be
521 /// instrumented with __cyg_profile_func_* calls
522 bool CodeGenFunction::ShouldInstrumentFunction() {
523 if (!CGM.getCodeGenOpts().InstrumentFunctions &&
524 !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining &&
525 !CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
527 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
532 /// ShouldXRayInstrument - Return true if the current function should be
533 /// instrumented with XRay nop sleds.
534 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
535 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
538 /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to
539 /// the __xray_customevent(...) builtin calls, when doing XRay instrumentation.
540 bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const {
541 return CGM.getCodeGenOpts().XRayInstrumentFunctions &&
542 (CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents ||
543 CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask ==
544 XRayInstrKind::Custom);
547 bool CodeGenFunction::AlwaysEmitXRayTypedEvents() const {
548 return CGM.getCodeGenOpts().XRayInstrumentFunctions &&
549 (CGM.getCodeGenOpts().XRayAlwaysEmitTypedEvents ||
550 CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask ==
551 XRayInstrKind::Typed);
555 CodeGenFunction::EncodeAddrForUseInPrologue(llvm::Function *F,
556 llvm::Constant *Addr) {
557 // Addresses stored in prologue data can't require run-time fixups and must
558 // be PC-relative. Run-time fixups are undesirable because they necessitate
559 // writable text segments, which are unsafe. And absolute addresses are
560 // undesirable because they break PIE mode.
562 // Add a layer of indirection through a private global. Taking its address
563 // won't result in a run-time fixup, even if Addr has linkonce_odr linkage.
564 auto *GV = new llvm::GlobalVariable(CGM.getModule(), Addr->getType(),
566 llvm::GlobalValue::PrivateLinkage, Addr);
568 // Create a PC-relative address.
569 auto *GOTAsInt = llvm::ConstantExpr::getPtrToInt(GV, IntPtrTy);
570 auto *FuncAsInt = llvm::ConstantExpr::getPtrToInt(F, IntPtrTy);
571 auto *PCRelAsInt = llvm::ConstantExpr::getSub(GOTAsInt, FuncAsInt);
572 return (IntPtrTy == Int32Ty)
574 : llvm::ConstantExpr::getTrunc(PCRelAsInt, Int32Ty);
578 CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F,
579 llvm::Value *EncodedAddr) {
580 // Reconstruct the address of the global.
581 auto *PCRelAsInt = Builder.CreateSExt(EncodedAddr, IntPtrTy);
582 auto *FuncAsInt = Builder.CreatePtrToInt(F, IntPtrTy, "func_addr.int");
583 auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int");
584 auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr");
586 // Load the original pointer through the global.
587 return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()),
591 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
594 if (!FD->hasAttr<OpenCLKernelAttr>())
597 llvm::LLVMContext &Context = getLLVMContext();
599 CGM.GenOpenCLArgMetadata(Fn, FD, this);
601 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
602 QualType HintQTy = A->getTypeHint();
603 const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
604 bool IsSignedInteger =
605 HintQTy->isSignedIntegerType() ||
606 (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
607 llvm::Metadata *AttrMDArgs[] = {
608 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
609 CGM.getTypes().ConvertType(A->getTypeHint()))),
610 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
611 llvm::IntegerType::get(Context, 32),
612 llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
613 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
616 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
617 llvm::Metadata *AttrMDArgs[] = {
618 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
619 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
620 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
621 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
624 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
625 llvm::Metadata *AttrMDArgs[] = {
626 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
627 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
628 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
629 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
632 if (const OpenCLIntelReqdSubGroupSizeAttr *A =
633 FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
634 llvm::Metadata *AttrMDArgs[] = {
635 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
636 Fn->setMetadata("intel_reqd_sub_group_size",
637 llvm::MDNode::get(Context, AttrMDArgs));
641 /// Determine whether the function F ends with a return stmt.
642 static bool endsWithReturn(const Decl* F) {
643 const Stmt *Body = nullptr;
644 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
645 Body = FD->getBody();
646 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
647 Body = OMD->getBody();
649 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
650 auto LastStmt = CS->body_rbegin();
651 if (LastStmt != CS->body_rend())
652 return isa<ReturnStmt>(*LastStmt);
657 void CodeGenFunction::markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
658 if (SanOpts.has(SanitizerKind::Thread)) {
659 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
660 Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
664 /// Check if the return value of this function requires sanitization.
665 bool CodeGenFunction::requiresReturnValueCheck() const {
666 return requiresReturnValueNullabilityCheck() ||
667 (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) && CurCodeDecl &&
668 CurCodeDecl->getAttr<ReturnsNonNullAttr>());
671 static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) {
672 auto *MD = dyn_cast_or_null<CXXMethodDecl>(D);
673 if (!MD || !MD->getDeclName().getAsIdentifierInfo() ||
674 !MD->getDeclName().getAsIdentifierInfo()->isStr("allocate") ||
675 (MD->getNumParams() != 1 && MD->getNumParams() != 2))
678 if (MD->parameters()[0]->getType().getCanonicalType() != Ctx.getSizeType())
681 if (MD->getNumParams() == 2) {
682 auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>();
683 if (!PT || !PT->isVoidPointerType() ||
684 !PT->getPointeeType().isConstQualified())
691 /// Return the UBSan prologue signature for \p FD if one is available.
692 static llvm::Constant *getPrologueSignature(CodeGenModule &CGM,
693 const FunctionDecl *FD) {
694 if (const auto *MD = dyn_cast<CXXMethodDecl>(FD))
697 return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM);
700 void CodeGenFunction::StartFunction(GlobalDecl GD, QualType RetTy,
702 const CGFunctionInfo &FnInfo,
703 const FunctionArgList &Args,
705 SourceLocation StartLoc) {
707 "Do not use a CodeGenFunction object for more than one function");
709 const Decl *D = GD.getDecl();
711 DidCallStackSave = false;
713 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
714 if (FD->usesSEHTry())
716 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
720 assert(CurFn->isDeclaration() && "Function already has body?");
722 // If this function has been blacklisted for any of the enabled sanitizers,
723 // disable the sanitizer for the function.
725 #define SANITIZER(NAME, ID) \
726 if (SanOpts.empty()) \
728 if (SanOpts.has(SanitizerKind::ID)) \
729 if (CGM.isInSanitizerBlacklist(SanitizerKind::ID, Fn, Loc)) \
730 SanOpts.set(SanitizerKind::ID, false);
732 #include "clang/Basic/Sanitizers.def"
737 // Apply the no_sanitize* attributes to SanOpts.
738 for (auto Attr : D->specific_attrs<NoSanitizeAttr>()) {
739 SanitizerMask mask = Attr->getMask();
740 SanOpts.Mask &= ~mask;
741 if (mask & SanitizerKind::Address)
742 SanOpts.set(SanitizerKind::KernelAddress, false);
743 if (mask & SanitizerKind::KernelAddress)
744 SanOpts.set(SanitizerKind::Address, false);
745 if (mask & SanitizerKind::HWAddress)
746 SanOpts.set(SanitizerKind::KernelHWAddress, false);
747 if (mask & SanitizerKind::KernelHWAddress)
748 SanOpts.set(SanitizerKind::HWAddress, false);
752 // Apply sanitizer attributes to the function.
753 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
754 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
755 if (SanOpts.hasOneOf(SanitizerKind::HWAddress | SanitizerKind::KernelHWAddress))
756 Fn->addFnAttr(llvm::Attribute::SanitizeHWAddress);
757 if (SanOpts.has(SanitizerKind::MemTag))
758 Fn->addFnAttr(llvm::Attribute::SanitizeMemTag);
759 if (SanOpts.has(SanitizerKind::Thread))
760 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
761 if (SanOpts.hasOneOf(SanitizerKind::Memory | SanitizerKind::KernelMemory))
762 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
763 if (SanOpts.has(SanitizerKind::SafeStack))
764 Fn->addFnAttr(llvm::Attribute::SafeStack);
765 if (SanOpts.has(SanitizerKind::ShadowCallStack))
766 Fn->addFnAttr(llvm::Attribute::ShadowCallStack);
768 // Apply fuzzing attribute to the function.
769 if (SanOpts.hasOneOf(SanitizerKind::Fuzzer | SanitizerKind::FuzzerNoLink))
770 Fn->addFnAttr(llvm::Attribute::OptForFuzzing);
772 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
773 // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
774 if (SanOpts.has(SanitizerKind::Thread)) {
775 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
776 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
777 if (OMD->getMethodFamily() == OMF_dealloc ||
778 OMD->getMethodFamily() == OMF_initialize ||
779 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
780 markAsIgnoreThreadCheckingAtRuntime(Fn);
785 // Ignore unrelated casts in STL allocate() since the allocator must cast
786 // from void* to T* before object initialization completes. Don't match on the
787 // namespace because not all allocators are in std::
788 if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
789 if (matchesStlAllocatorFn(D, getContext()))
790 SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast;
793 // Ignore null checks in coroutine functions since the coroutines passes
794 // are not aware of how to move the extra UBSan instructions across the split
795 // coroutine boundaries.
796 if (D && SanOpts.has(SanitizerKind::Null))
797 if (const auto *FD = dyn_cast<FunctionDecl>(D))
799 FD->getBody()->getStmtClass() == Stmt::CoroutineBodyStmtClass)
800 SanOpts.Mask &= ~SanitizerKind::Null;
803 // Apply xray attributes to the function (as a string, for now)
804 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
805 if (CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
806 XRayInstrKind::Function)) {
807 if (XRayAttr->alwaysXRayInstrument() && ShouldXRayInstrumentFunction())
808 Fn->addFnAttr("function-instrument", "xray-always");
809 if (XRayAttr->neverXRayInstrument())
810 Fn->addFnAttr("function-instrument", "xray-never");
811 if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>())
812 if (ShouldXRayInstrumentFunction())
813 Fn->addFnAttr("xray-log-args",
814 llvm::utostr(LogArgs->getArgumentCount()));
817 if (ShouldXRayInstrumentFunction() && !CGM.imbueXRayAttrs(Fn, Loc))
819 "xray-instruction-threshold",
820 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
823 unsigned Count, Offset;
824 if (const auto *Attr = D->getAttr<PatchableFunctionEntryAttr>()) {
825 Count = Attr->getCount();
826 Offset = Attr->getOffset();
828 Count = CGM.getCodeGenOpts().PatchableFunctionEntryCount;
829 Offset = CGM.getCodeGenOpts().PatchableFunctionEntryOffset;
831 if (Count && Offset <= Count) {
832 Fn->addFnAttr("patchable-function-entry", std::to_string(Count - Offset));
834 Fn->addFnAttr("patchable-function-prefix", std::to_string(Offset));
838 // Add no-jump-tables value.
839 Fn->addFnAttr("no-jump-tables",
840 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
842 // Add no-inline-line-tables value.
843 if (CGM.getCodeGenOpts().NoInlineLineTables)
844 Fn->addFnAttr("no-inline-line-tables");
846 // Add profile-sample-accurate value.
847 if (CGM.getCodeGenOpts().ProfileSampleAccurate)
848 Fn->addFnAttr("profile-sample-accurate");
850 if (D && D->hasAttr<CFICanonicalJumpTableAttr>())
851 Fn->addFnAttr("cfi-canonical-jump-table");
853 if (getLangOpts().OpenCL) {
854 // Add metadata for a kernel function.
855 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
856 EmitOpenCLKernelMetadata(FD, Fn);
859 // If we are checking function types, emit a function type signature as
861 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
862 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
863 if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
864 // Remove any (C++17) exception specifications, to allow calling e.g. a
865 // noexcept function through a non-noexcept pointer.
867 getContext().getFunctionTypeWithExceptionSpec(FD->getType(),
869 llvm::Constant *FTRTTIConst =
870 CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true);
871 llvm::Constant *FTRTTIConstEncoded =
872 EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
873 llvm::Constant *PrologueStructElems[] = {PrologueSig,
875 llvm::Constant *PrologueStructConst =
876 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
877 Fn->setPrologueData(PrologueStructConst);
882 // If we're checking nullability, we need to know whether we can check the
883 // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
884 if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
885 auto Nullability = FnRetTy->getNullability(getContext());
886 if (Nullability && *Nullability == NullabilityKind::NonNull) {
887 if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
888 CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
889 RetValNullabilityPrecondition =
890 llvm::ConstantInt::getTrue(getLLVMContext());
894 // If we're in C++ mode and the function name is "main", it is guaranteed
895 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
896 // used within a program").
897 if (getLangOpts().CPlusPlus)
898 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
900 Fn->addFnAttr(llvm::Attribute::NoRecurse);
902 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
903 if (FD->usesFPIntrin())
904 Fn->addFnAttr(llvm::Attribute::StrictFP);
906 // If a custom alignment is used, force realigning to this alignment on
907 // any main function which certainly will need it.
908 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
909 if ((FD->isMain() || FD->isMSVCRTEntryPoint()) &&
910 CGM.getCodeGenOpts().StackAlignment)
911 Fn->addFnAttr("stackrealign");
913 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
915 // Create a marker to make it easy to insert allocas into the entryblock
916 // later. Don't create this with the builder, because we don't want it
918 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
919 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
921 ReturnBlock = getJumpDestInCurrentScope("return");
923 Builder.SetInsertPoint(EntryBB);
925 // If we're checking the return value, allocate space for a pointer to a
926 // precise source location of the checked return statement.
927 if (requiresReturnValueCheck()) {
928 ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
929 InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
932 // Emit subprogram debug descriptor.
933 if (CGDebugInfo *DI = getDebugInfo()) {
934 // Reconstruct the type from the argument list so that implicit parameters,
935 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
937 CallingConv CC = CallingConv::CC_C;
938 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
939 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
940 CC = SrcFnTy->getCallConv();
941 SmallVector<QualType, 16> ArgTypes;
942 for (const VarDecl *VD : Args)
943 ArgTypes.push_back(VD->getType());
944 QualType FnType = getContext().getFunctionType(
945 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
946 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, CurFuncIsThunk,
950 if (ShouldInstrumentFunction()) {
951 if (CGM.getCodeGenOpts().InstrumentFunctions)
952 CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
953 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
954 CurFn->addFnAttr("instrument-function-entry-inlined",
955 "__cyg_profile_func_enter");
956 if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
957 CurFn->addFnAttr("instrument-function-entry-inlined",
958 "__cyg_profile_func_enter_bare");
961 // Since emitting the mcount call here impacts optimizations such as function
962 // inlining, we just add an attribute to insert a mcount call in backend.
963 // The attribute "counting-function" is set to mcount function name which is
964 // architecture dependent.
965 if (CGM.getCodeGenOpts().InstrumentForProfiling) {
966 // Calls to fentry/mcount should not be generated if function has
967 // the no_instrument_function attribute.
968 if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
969 if (CGM.getCodeGenOpts().CallFEntry)
970 Fn->addFnAttr("fentry-call", "true");
972 Fn->addFnAttr("instrument-function-entry-inlined",
973 getTarget().getMCountName());
975 if (CGM.getCodeGenOpts().MNopMCount) {
976 if (!CGM.getCodeGenOpts().CallFEntry)
977 CGM.getDiags().Report(diag::err_opt_not_valid_without_opt)
978 << "-mnop-mcount" << "-mfentry";
979 Fn->addFnAttr("mnop-mcount");
982 if (CGM.getCodeGenOpts().RecordMCount) {
983 if (!CGM.getCodeGenOpts().CallFEntry)
984 CGM.getDiags().Report(diag::err_opt_not_valid_without_opt)
985 << "-mrecord-mcount" << "-mfentry";
986 Fn->addFnAttr("mrecord-mcount");
991 if (CGM.getCodeGenOpts().PackedStack) {
992 if (getContext().getTargetInfo().getTriple().getArch() !=
993 llvm::Triple::systemz)
994 CGM.getDiags().Report(diag::err_opt_not_valid_on_target)
996 Fn->addFnAttr("packed-stack");
999 if (RetTy->isVoidType()) {
1000 // Void type; nothing to return.
1001 ReturnValue = Address::invalid();
1003 // Count the implicit return.
1004 if (!endsWithReturn(D))
1006 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect) {
1007 // Indirect return; emit returned value directly into sret slot.
1008 // This reduces code size, and affects correctness in C++.
1009 auto AI = CurFn->arg_begin();
1010 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
1012 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
1013 if (!CurFnInfo->getReturnInfo().getIndirectByVal()) {
1014 ReturnValuePointer =
1015 CreateDefaultAlignTempAlloca(Int8PtrTy, "result.ptr");
1016 Builder.CreateStore(Builder.CreatePointerBitCastOrAddrSpaceCast(
1017 ReturnValue.getPointer(), Int8PtrTy),
1018 ReturnValuePointer);
1020 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
1021 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1022 // Load the sret pointer from the argument struct and return into that.
1023 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
1024 llvm::Function::arg_iterator EI = CurFn->arg_end();
1026 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
1027 ReturnValuePointer = Address(Addr, getPointerAlign());
1028 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
1029 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
1031 ReturnValue = CreateIRTemp(RetTy, "retval");
1033 // Tell the epilog emitter to autorelease the result. We do this
1034 // now so that various specialized functions can suppress it
1035 // during their IR-generation.
1036 if (getLangOpts().ObjCAutoRefCount &&
1037 !CurFnInfo->isReturnsRetained() &&
1038 RetTy->isObjCRetainableType())
1039 AutoreleaseResult = true;
1042 EmitStartEHSpec(CurCodeDecl);
1044 PrologueCleanupDepth = EHStack.stable_begin();
1046 // Emit OpenMP specific initialization of the device functions.
1047 if (getLangOpts().OpenMP && CurCodeDecl)
1048 CGM.getOpenMPRuntime().emitFunctionProlog(*this, CurCodeDecl);
1050 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
1052 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
1053 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
1054 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
1055 if (MD->getParent()->isLambda() &&
1056 MD->getOverloadedOperator() == OO_Call) {
1057 // We're in a lambda; figure out the captures.
1058 MD->getParent()->getCaptureFields(LambdaCaptureFields,
1059 LambdaThisCaptureField);
1060 if (LambdaThisCaptureField) {
1061 // If the lambda captures the object referred to by '*this' - either by
1062 // value or by reference, make sure CXXThisValue points to the correct
1065 // Get the lvalue for the field (which is a copy of the enclosing object
1066 // or contains the address of the enclosing object).
1067 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
1068 if (!LambdaThisCaptureField->getType()->isPointerType()) {
1069 // If the enclosing object was captured by value, just use its address.
1070 CXXThisValue = ThisFieldLValue.getAddress(*this).getPointer();
1072 // Load the lvalue pointed to by the field, since '*this' was captured
1075 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
1078 for (auto *FD : MD->getParent()->fields()) {
1079 if (FD->hasCapturedVLAType()) {
1080 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
1081 SourceLocation()).getScalarVal();
1082 auto VAT = FD->getCapturedVLAType();
1083 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
1087 // Not in a lambda; just use 'this' from the method.
1088 // FIXME: Should we generate a new load for each use of 'this'? The
1089 // fast register allocator would be happier...
1090 CXXThisValue = CXXABIThisValue;
1093 // Check the 'this' pointer once per function, if it's available.
1094 if (CXXABIThisValue) {
1095 SanitizerSet SkippedChecks;
1096 SkippedChecks.set(SanitizerKind::ObjectSize, true);
1097 QualType ThisTy = MD->getThisType();
1099 // If this is the call operator of a lambda with no capture-default, it
1100 // may have a static invoker function, which may call this operator with
1101 // a null 'this' pointer.
1102 if (isLambdaCallOperator(MD) &&
1103 MD->getParent()->getLambdaCaptureDefault() == LCD_None)
1104 SkippedChecks.set(SanitizerKind::Null, true);
1106 EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1108 Loc, CXXABIThisValue, ThisTy,
1109 getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1114 // If any of the arguments have a variably modified type, make sure to
1115 // emit the type size.
1116 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1118 const VarDecl *VD = *i;
1120 // Dig out the type as written from ParmVarDecls; it's unclear whether
1121 // the standard (C99 6.9.1p10) requires this, but we're following the
1122 // precedent set by gcc.
1124 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1125 Ty = PVD->getOriginalType();
1129 if (Ty->isVariablyModifiedType())
1130 EmitVariablyModifiedType(Ty);
1132 // Emit a location at the end of the prologue.
1133 if (CGDebugInfo *DI = getDebugInfo())
1134 DI->EmitLocation(Builder, StartLoc);
1136 // TODO: Do we need to handle this in two places like we do with
1137 // target-features/target-cpu?
1139 if (const auto *VecWidth = CurFuncDecl->getAttr<MinVectorWidthAttr>())
1140 LargestVectorWidth = VecWidth->getVectorWidth();
1143 void CodeGenFunction::EmitFunctionBody(const Stmt *Body) {
1144 incrementProfileCounter(Body);
1145 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1146 EmitCompoundStmtWithoutScope(*S);
1151 /// When instrumenting to collect profile data, the counts for some blocks
1152 /// such as switch cases need to not include the fall-through counts, so
1153 /// emit a branch around the instrumentation code. When not instrumenting,
1154 /// this just calls EmitBlock().
1155 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1157 llvm::BasicBlock *SkipCountBB = nullptr;
1158 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1159 // When instrumenting for profiling, the fallthrough to certain
1160 // statements needs to skip over the instrumentation code so that we
1161 // get an accurate count.
1162 SkipCountBB = createBasicBlock("skipcount");
1163 EmitBranch(SkipCountBB);
1166 uint64_t CurrentCount = getCurrentProfileCount();
1167 incrementProfileCounter(S);
1168 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1170 EmitBlock(SkipCountBB);
1173 /// Tries to mark the given function nounwind based on the
1174 /// non-existence of any throwing calls within it. We believe this is
1175 /// lightweight enough to do at -O0.
1176 static void TryMarkNoThrow(llvm::Function *F) {
1177 // LLVM treats 'nounwind' on a function as part of the type, so we
1178 // can't do this on functions that can be overwritten.
1179 if (F->isInterposable()) return;
1181 for (llvm::BasicBlock &BB : *F)
1182 for (llvm::Instruction &I : BB)
1186 F->setDoesNotThrow();
1189 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1190 FunctionArgList &Args) {
1191 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1192 QualType ResTy = FD->getReturnType();
1194 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1195 if (MD && MD->isInstance()) {
1196 if (CGM.getCXXABI().HasThisReturn(GD))
1197 ResTy = MD->getThisType();
1198 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1199 ResTy = CGM.getContext().VoidPtrTy;
1200 CGM.getCXXABI().buildThisParam(*this, Args);
1203 // The base version of an inheriting constructor whose constructed base is a
1204 // virtual base is not passed any arguments (because it doesn't actually call
1205 // the inherited constructor).
1206 bool PassedParams = true;
1207 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1208 if (auto Inherited = CD->getInheritedConstructor())
1210 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1213 for (auto *Param : FD->parameters()) {
1214 Args.push_back(Param);
1215 if (!Param->hasAttr<PassObjectSizeAttr>())
1218 auto *Implicit = ImplicitParamDecl::Create(
1219 getContext(), Param->getDeclContext(), Param->getLocation(),
1220 /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1221 SizeArguments[Param] = Implicit;
1222 Args.push_back(Implicit);
1226 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1227 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1233 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1234 const ASTContext &Context) {
1235 QualType T = FD->getReturnType();
1236 // Avoid the optimization for functions that return a record type with a
1237 // trivial destructor or another trivially copyable type.
1238 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1239 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1240 return !ClassDecl->hasTrivialDestructor();
1242 return !T.isTriviallyCopyableType(Context);
1245 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1246 const CGFunctionInfo &FnInfo) {
1247 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1250 FunctionArgList Args;
1251 QualType ResTy = BuildFunctionArgList(GD, Args);
1253 // Check if we should generate debug info for this function.
1254 if (FD->hasAttr<NoDebugAttr>())
1255 DebugInfo = nullptr; // disable debug info indefinitely for this function
1257 // The function might not have a body if we're generating thunks for a
1258 // function declaration.
1259 SourceRange BodyRange;
1260 if (Stmt *Body = FD->getBody())
1261 BodyRange = Body->getSourceRange();
1263 BodyRange = FD->getLocation();
1264 CurEHLocation = BodyRange.getEnd();
1266 // Use the location of the start of the function to determine where
1267 // the function definition is located. By default use the location
1268 // of the declaration as the location for the subprogram. A function
1269 // may lack a declaration in the source code if it is created by code
1270 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1271 SourceLocation Loc = FD->getLocation();
1273 // If this is a function specialization then use the pattern body
1274 // as the location for the function.
1275 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1276 if (SpecDecl->hasBody(SpecDecl))
1277 Loc = SpecDecl->getLocation();
1279 Stmt *Body = FD->getBody();
1281 // Initialize helper which will detect jumps which can cause invalid lifetime
1283 if (Body && ShouldEmitLifetimeMarkers)
1284 Bypasses.Init(Body);
1286 // Emit the standard function prologue.
1287 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1289 // Generate the body of the function.
1290 PGO.assignRegionCounters(GD, CurFn);
1291 if (isa<CXXDestructorDecl>(FD))
1292 EmitDestructorBody(Args);
1293 else if (isa<CXXConstructorDecl>(FD))
1294 EmitConstructorBody(Args);
1295 else if (getLangOpts().CUDA &&
1296 !getLangOpts().CUDAIsDevice &&
1297 FD->hasAttr<CUDAGlobalAttr>())
1298 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1299 else if (isa<CXXMethodDecl>(FD) &&
1300 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1301 // The lambda static invoker function is special, because it forwards or
1302 // clones the body of the function call operator (but is actually static).
1303 EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1304 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1305 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1306 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1307 // Implicit copy-assignment gets the same special treatment as implicit
1308 // copy-constructors.
1309 emitImplicitAssignmentOperatorBody(Args);
1311 EmitFunctionBody(Body);
1313 llvm_unreachable("no definition for emitted function");
1315 // C++11 [stmt.return]p2:
1316 // Flowing off the end of a function [...] results in undefined behavior in
1317 // a value-returning function.
1319 // If the '}' that terminates a function is reached, and the value of the
1320 // function call is used by the caller, the behavior is undefined.
1321 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1322 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1323 bool ShouldEmitUnreachable =
1324 CGM.getCodeGenOpts().StrictReturn ||
1325 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1326 if (SanOpts.has(SanitizerKind::Return)) {
1327 SanitizerScope SanScope(this);
1328 llvm::Value *IsFalse = Builder.getFalse();
1329 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1330 SanitizerHandler::MissingReturn,
1331 EmitCheckSourceLocation(FD->getLocation()), None);
1332 } else if (ShouldEmitUnreachable) {
1333 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1334 EmitTrapCall(llvm::Intrinsic::trap);
1336 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1337 Builder.CreateUnreachable();
1338 Builder.ClearInsertionPoint();
1342 // Emit the standard function epilogue.
1343 FinishFunction(BodyRange.getEnd());
1345 // If we haven't marked the function nothrow through other means, do
1346 // a quick pass now to see if we can.
1347 if (!CurFn->doesNotThrow())
1348 TryMarkNoThrow(CurFn);
1351 /// ContainsLabel - Return true if the statement contains a label in it. If
1352 /// this statement is not executed normally, it not containing a label means
1353 /// that we can just remove the code.
1354 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1355 // Null statement, not a label!
1356 if (!S) return false;
1358 // If this is a label, we have to emit the code, consider something like:
1359 // if (0) { ... foo: bar(); } goto foo;
1361 // TODO: If anyone cared, we could track __label__'s, since we know that you
1362 // can't jump to one from outside their declared region.
1363 if (isa<LabelStmt>(S))
1366 // If this is a case/default statement, and we haven't seen a switch, we have
1367 // to emit the code.
1368 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1371 // If this is a switch statement, we want to ignore cases below it.
1372 if (isa<SwitchStmt>(S))
1373 IgnoreCaseStmts = true;
1375 // Scan subexpressions for verboten labels.
1376 for (const Stmt *SubStmt : S->children())
1377 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1383 /// containsBreak - Return true if the statement contains a break out of it.
1384 /// If the statement (recursively) contains a switch or loop with a break
1385 /// inside of it, this is fine.
1386 bool CodeGenFunction::containsBreak(const Stmt *S) {
1387 // Null statement, not a label!
1388 if (!S) return false;
1390 // If this is a switch or loop that defines its own break scope, then we can
1391 // include it and anything inside of it.
1392 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1396 if (isa<BreakStmt>(S))
1399 // Scan subexpressions for verboten breaks.
1400 for (const Stmt *SubStmt : S->children())
1401 if (containsBreak(SubStmt))
1407 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1408 if (!S) return false;
1410 // Some statement kinds add a scope and thus never add a decl to the current
1411 // scope. Note, this list is longer than the list of statements that might
1412 // have an unscoped decl nested within them, but this way is conservatively
1413 // correct even if more statement kinds are added.
1414 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1415 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1416 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1417 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1420 if (isa<DeclStmt>(S))
1423 for (const Stmt *SubStmt : S->children())
1424 if (mightAddDeclToScope(SubStmt))
1430 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1431 /// to a constant, or if it does but contains a label, return false. If it
1432 /// constant folds return true and set the boolean result in Result.
1433 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1436 llvm::APSInt ResultInt;
1437 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1440 ResultBool = ResultInt.getBoolValue();
1444 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1445 /// to a constant, or if it does but contains a label, return false. If it
1446 /// constant folds return true and set the folded value.
1447 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1448 llvm::APSInt &ResultInt,
1450 // FIXME: Rename and handle conversion of other evaluatable things
1452 Expr::EvalResult Result;
1453 if (!Cond->EvaluateAsInt(Result, getContext()))
1454 return false; // Not foldable, not integer or not fully evaluatable.
1456 llvm::APSInt Int = Result.Val.getInt();
1457 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1458 return false; // Contains a label.
1466 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1467 /// statement) to the specified blocks. Based on the condition, this might try
1468 /// to simplify the codegen of the conditional based on the branch.
1470 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1471 llvm::BasicBlock *TrueBlock,
1472 llvm::BasicBlock *FalseBlock,
1473 uint64_t TrueCount) {
1474 Cond = Cond->IgnoreParens();
1476 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1478 // Handle X && Y in a condition.
1479 if (CondBOp->getOpcode() == BO_LAnd) {
1480 // If we have "1 && X", simplify the code. "0 && X" would have constant
1481 // folded if the case was simple enough.
1482 bool ConstantBool = false;
1483 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1485 // br(1 && X) -> br(X).
1486 incrementProfileCounter(CondBOp);
1487 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1491 // If we have "X && 1", simplify the code to use an uncond branch.
1492 // "X && 0" would have been constant folded to 0.
1493 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1495 // br(X && 1) -> br(X).
1496 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1500 // Emit the LHS as a conditional. If the LHS conditional is false, we
1501 // want to jump to the FalseBlock.
1502 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1503 // The counter tells us how often we evaluate RHS, and all of TrueCount
1504 // can be propagated to that branch.
1505 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1507 ConditionalEvaluation eval(*this);
1509 ApplyDebugLocation DL(*this, Cond);
1510 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1514 incrementProfileCounter(CondBOp);
1515 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1517 // Any temporaries created here are conditional.
1519 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1525 if (CondBOp->getOpcode() == BO_LOr) {
1526 // If we have "0 || X", simplify the code. "1 || X" would have constant
1527 // folded if the case was simple enough.
1528 bool ConstantBool = false;
1529 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1531 // br(0 || X) -> br(X).
1532 incrementProfileCounter(CondBOp);
1533 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1537 // If we have "X || 0", simplify the code to use an uncond branch.
1538 // "X || 1" would have been constant folded to 1.
1539 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1541 // br(X || 0) -> br(X).
1542 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1546 // Emit the LHS as a conditional. If the LHS conditional is true, we
1547 // want to jump to the TrueBlock.
1548 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1549 // We have the count for entry to the RHS and for the whole expression
1550 // being true, so we can divy up True count between the short circuit and
1553 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1554 uint64_t RHSCount = TrueCount - LHSCount;
1556 ConditionalEvaluation eval(*this);
1558 ApplyDebugLocation DL(*this, Cond);
1559 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1560 EmitBlock(LHSFalse);
1563 incrementProfileCounter(CondBOp);
1564 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1566 // Any temporaries created here are conditional.
1568 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1576 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1577 // br(!x, t, f) -> br(x, f, t)
1578 if (CondUOp->getOpcode() == UO_LNot) {
1579 // Negate the count.
1580 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1581 // Negate the condition and swap the destination blocks.
1582 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1587 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1588 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1589 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1590 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1592 ConditionalEvaluation cond(*this);
1593 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1594 getProfileCount(CondOp));
1596 // When computing PGO branch weights, we only know the overall count for
1597 // the true block. This code is essentially doing tail duplication of the
1598 // naive code-gen, introducing new edges for which counts are not
1599 // available. Divide the counts proportionally between the LHS and RHS of
1600 // the conditional operator.
1601 uint64_t LHSScaledTrueCount = 0;
1604 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1605 LHSScaledTrueCount = TrueCount * LHSRatio;
1609 EmitBlock(LHSBlock);
1610 incrementProfileCounter(CondOp);
1612 ApplyDebugLocation DL(*this, Cond);
1613 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1614 LHSScaledTrueCount);
1619 EmitBlock(RHSBlock);
1620 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1621 TrueCount - LHSScaledTrueCount);
1627 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1628 // Conditional operator handling can give us a throw expression as a
1629 // condition for a case like:
1630 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1632 // br(c, throw x, br(y, t, f))
1633 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1637 // If the branch has a condition wrapped by __builtin_unpredictable,
1638 // create metadata that specifies that the branch is unpredictable.
1639 // Don't bother if not optimizing because that metadata would not be used.
1640 llvm::MDNode *Unpredictable = nullptr;
1641 auto *Call = dyn_cast<CallExpr>(Cond->IgnoreImpCasts());
1642 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1643 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1644 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1645 llvm::MDBuilder MDHelper(getLLVMContext());
1646 Unpredictable = MDHelper.createUnpredictable();
1650 // Create branch weights based on the number of times we get here and the
1651 // number of times the condition should be true.
1652 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1653 llvm::MDNode *Weights =
1654 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1656 // Emit the code with the fully general case.
1659 ApplyDebugLocation DL(*this, Cond);
1660 CondV = EvaluateExprAsBool(Cond);
1662 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1665 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1666 /// specified stmt yet.
1667 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1668 CGM.ErrorUnsupported(S, Type);
1671 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1672 /// variable-length array whose elements have a non-zero bit-pattern.
1674 /// \param baseType the inner-most element type of the array
1675 /// \param src - a char* pointing to the bit-pattern for a single
1676 /// base element of the array
1677 /// \param sizeInChars - the total size of the VLA, in chars
1678 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1679 Address dest, Address src,
1680 llvm::Value *sizeInChars) {
1681 CGBuilderTy &Builder = CGF.Builder;
1683 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1684 llvm::Value *baseSizeInChars
1685 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1688 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1690 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1692 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1693 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1694 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1696 // Make a loop over the VLA. C99 guarantees that the VLA element
1697 // count must be nonzero.
1698 CGF.EmitBlock(loopBB);
1700 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1701 cur->addIncoming(begin.getPointer(), originBB);
1703 CharUnits curAlign =
1704 dest.getAlignment().alignmentOfArrayElement(baseSize);
1706 // memcpy the individual element bit-pattern.
1707 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1708 /*volatile*/ false);
1710 // Go to the next element.
1712 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1714 // Leave if that's the end of the VLA.
1715 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1716 Builder.CreateCondBr(done, contBB, loopBB);
1717 cur->addIncoming(next, loopBB);
1719 CGF.EmitBlock(contBB);
1723 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1724 // Ignore empty classes in C++.
1725 if (getLangOpts().CPlusPlus) {
1726 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1727 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1732 // Cast the dest ptr to the appropriate i8 pointer type.
1733 if (DestPtr.getElementType() != Int8Ty)
1734 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1736 // Get size and alignment info for this aggregate.
1737 CharUnits size = getContext().getTypeSizeInChars(Ty);
1739 llvm::Value *SizeVal;
1740 const VariableArrayType *vla;
1742 // Don't bother emitting a zero-byte memset.
1743 if (size.isZero()) {
1744 // But note that getTypeInfo returns 0 for a VLA.
1745 if (const VariableArrayType *vlaType =
1746 dyn_cast_or_null<VariableArrayType>(
1747 getContext().getAsArrayType(Ty))) {
1748 auto VlaSize = getVLASize(vlaType);
1749 SizeVal = VlaSize.NumElts;
1750 CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type);
1751 if (!eltSize.isOne())
1752 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1758 SizeVal = CGM.getSize(size);
1762 // If the type contains a pointer to data member we can't memset it to zero.
1763 // Instead, create a null constant and copy it to the destination.
1764 // TODO: there are other patterns besides zero that we can usefully memset,
1765 // like -1, which happens to be the pattern used by member-pointers.
1766 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1767 // For a VLA, emit a single element, then splat that over the VLA.
1768 if (vla) Ty = getContext().getBaseElementType(vla);
1770 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1772 llvm::GlobalVariable *NullVariable =
1773 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1774 /*isConstant=*/true,
1775 llvm::GlobalVariable::PrivateLinkage,
1776 NullConstant, Twine());
1777 CharUnits NullAlign = DestPtr.getAlignment();
1778 NullVariable->setAlignment(NullAlign.getAsAlign());
1779 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1782 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1784 // Get and call the appropriate llvm.memcpy overload.
1785 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1789 // Otherwise, just memset the whole thing to zero. This is legal
1790 // because in LLVM, all default initializers (other than the ones we just
1791 // handled above) are guaranteed to have a bit pattern of all zeros.
1792 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1795 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1796 // Make sure that there is a block for the indirect goto.
1797 if (!IndirectBranch)
1798 GetIndirectGotoBlock();
1800 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1802 // Make sure the indirect branch includes all of the address-taken blocks.
1803 IndirectBranch->addDestination(BB);
1804 return llvm::BlockAddress::get(CurFn, BB);
1807 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1808 // If we already made the indirect branch for indirect goto, return its block.
1809 if (IndirectBranch) return IndirectBranch->getParent();
1811 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1813 // Create the PHI node that indirect gotos will add entries to.
1814 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1815 "indirect.goto.dest");
1817 // Create the indirect branch instruction.
1818 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1819 return IndirectBranch->getParent();
1822 /// Computes the length of an array in elements, as well as the base
1823 /// element type and a properly-typed first element pointer.
1824 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1827 const ArrayType *arrayType = origArrayType;
1829 // If it's a VLA, we have to load the stored size. Note that
1830 // this is the size of the VLA in bytes, not its size in elements.
1831 llvm::Value *numVLAElements = nullptr;
1832 if (isa<VariableArrayType>(arrayType)) {
1833 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts;
1835 // Walk into all VLAs. This doesn't require changes to addr,
1836 // which has type T* where T is the first non-VLA element type.
1838 QualType elementType = arrayType->getElementType();
1839 arrayType = getContext().getAsArrayType(elementType);
1841 // If we only have VLA components, 'addr' requires no adjustment.
1843 baseType = elementType;
1844 return numVLAElements;
1846 } while (isa<VariableArrayType>(arrayType));
1848 // We get out here only if we find a constant array type
1852 // We have some number of constant-length arrays, so addr should
1853 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1854 // down to the first element of addr.
1855 SmallVector<llvm::Value*, 8> gepIndices;
1857 // GEP down to the array type.
1858 llvm::ConstantInt *zero = Builder.getInt32(0);
1859 gepIndices.push_back(zero);
1861 uint64_t countFromCLAs = 1;
1864 llvm::ArrayType *llvmArrayType =
1865 dyn_cast<llvm::ArrayType>(addr.getElementType());
1866 while (llvmArrayType) {
1867 assert(isa<ConstantArrayType>(arrayType));
1868 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1869 == llvmArrayType->getNumElements());
1871 gepIndices.push_back(zero);
1872 countFromCLAs *= llvmArrayType->getNumElements();
1873 eltType = arrayType->getElementType();
1876 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1877 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1878 assert((!llvmArrayType || arrayType) &&
1879 "LLVM and Clang types are out-of-synch");
1883 // From this point onwards, the Clang array type has been emitted
1884 // as some other type (probably a packed struct). Compute the array
1885 // size, and just emit the 'begin' expression as a bitcast.
1888 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1889 eltType = arrayType->getElementType();
1890 arrayType = getContext().getAsArrayType(eltType);
1893 llvm::Type *baseType = ConvertType(eltType);
1894 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1896 // Create the actual GEP.
1897 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1898 gepIndices, "array.begin"),
1899 addr.getAlignment());
1904 llvm::Value *numElements
1905 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1907 // If we had any VLA dimensions, factor them in.
1909 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1914 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) {
1915 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1916 assert(vla && "type was not a variable array type!");
1917 return getVLASize(vla);
1920 CodeGenFunction::VlaSizePair
1921 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1922 // The number of elements so far; always size_t.
1923 llvm::Value *numElements = nullptr;
1925 QualType elementType;
1927 elementType = type->getElementType();
1928 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1929 assert(vlaSize && "no size for VLA!");
1930 assert(vlaSize->getType() == SizeTy);
1933 numElements = vlaSize;
1935 // It's undefined behavior if this wraps around, so mark it that way.
1936 // FIXME: Teach -fsanitize=undefined to trap this.
1937 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1939 } while ((type = getContext().getAsVariableArrayType(elementType)));
1941 return { numElements, elementType };
1944 CodeGenFunction::VlaSizePair
1945 CodeGenFunction::getVLAElements1D(QualType type) {
1946 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1947 assert(vla && "type was not a variable array type!");
1948 return getVLAElements1D(vla);
1951 CodeGenFunction::VlaSizePair
1952 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) {
1953 llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()];
1954 assert(VlaSize && "no size for VLA!");
1955 assert(VlaSize->getType() == SizeTy);
1956 return { VlaSize, Vla->getElementType() };
1959 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1960 assert(type->isVariablyModifiedType() &&
1961 "Must pass variably modified type to EmitVLASizes!");
1963 EnsureInsertPoint();
1965 // We're going to walk down into the type and look for VLA
1968 assert(type->isVariablyModifiedType());
1970 const Type *ty = type.getTypePtr();
1971 switch (ty->getTypeClass()) {
1973 #define TYPE(Class, Base)
1974 #define ABSTRACT_TYPE(Class, Base)
1975 #define NON_CANONICAL_TYPE(Class, Base)
1976 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1977 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1978 #include "clang/AST/TypeNodes.inc"
1979 llvm_unreachable("unexpected dependent type!");
1981 // These types are never variably-modified.
1985 case Type::ExtVector:
1988 case Type::Elaborated:
1989 case Type::TemplateSpecialization:
1990 case Type::ObjCTypeParam:
1991 case Type::ObjCObject:
1992 case Type::ObjCInterface:
1993 case Type::ObjCObjectPointer:
1994 llvm_unreachable("type class is never variably-modified!");
1996 case Type::Adjusted:
1997 type = cast<AdjustedType>(ty)->getAdjustedType();
2001 type = cast<DecayedType>(ty)->getPointeeType();
2005 type = cast<PointerType>(ty)->getPointeeType();
2008 case Type::BlockPointer:
2009 type = cast<BlockPointerType>(ty)->getPointeeType();
2012 case Type::LValueReference:
2013 case Type::RValueReference:
2014 type = cast<ReferenceType>(ty)->getPointeeType();
2017 case Type::MemberPointer:
2018 type = cast<MemberPointerType>(ty)->getPointeeType();
2021 case Type::ConstantArray:
2022 case Type::IncompleteArray:
2023 // Losing element qualification here is fine.
2024 type = cast<ArrayType>(ty)->getElementType();
2027 case Type::VariableArray: {
2028 // Losing element qualification here is fine.
2029 const VariableArrayType *vat = cast<VariableArrayType>(ty);
2031 // Unknown size indication requires no size computation.
2032 // Otherwise, evaluate and record it.
2033 if (const Expr *size = vat->getSizeExpr()) {
2034 // It's possible that we might have emitted this already,
2035 // e.g. with a typedef and a pointer to it.
2036 llvm::Value *&entry = VLASizeMap[size];
2038 llvm::Value *Size = EmitScalarExpr(size);
2041 // If the size is an expression that is not an integer constant
2042 // expression [...] each time it is evaluated it shall have a value
2043 // greater than zero.
2044 if (SanOpts.has(SanitizerKind::VLABound) &&
2045 size->getType()->isSignedIntegerType()) {
2046 SanitizerScope SanScope(this);
2047 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
2048 llvm::Constant *StaticArgs[] = {
2049 EmitCheckSourceLocation(size->getBeginLoc()),
2050 EmitCheckTypeDescriptor(size->getType())};
2051 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
2052 SanitizerKind::VLABound),
2053 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
2056 // Always zexting here would be wrong if it weren't
2057 // undefined behavior to have a negative bound.
2058 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
2061 type = vat->getElementType();
2065 case Type::FunctionProto:
2066 case Type::FunctionNoProto:
2067 type = cast<FunctionType>(ty)->getReturnType();
2072 case Type::UnaryTransform:
2073 case Type::Attributed:
2074 case Type::SubstTemplateTypeParm:
2075 case Type::PackExpansion:
2076 case Type::MacroQualified:
2077 // Keep walking after single level desugaring.
2078 type = type.getSingleStepDesugaredType(getContext());
2082 case Type::Decltype:
2084 case Type::DeducedTemplateSpecialization:
2085 // Stop walking: nothing to do.
2088 case Type::TypeOfExpr:
2089 // Stop walking: emit typeof expression.
2090 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
2094 type = cast<AtomicType>(ty)->getValueType();
2098 type = cast<PipeType>(ty)->getElementType();
2101 } while (type->isVariablyModifiedType());
2104 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
2105 if (getContext().getBuiltinVaListType()->isArrayType())
2106 return EmitPointerWithAlignment(E);
2107 return EmitLValue(E).getAddress(*this);
2110 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
2111 return EmitLValue(E).getAddress(*this);
2114 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2115 const APValue &Init) {
2116 assert(Init.hasValue() && "Invalid DeclRefExpr initializer!");
2117 if (CGDebugInfo *Dbg = getDebugInfo())
2118 if (CGM.getCodeGenOpts().hasReducedDebugInfo())
2119 Dbg->EmitGlobalVariable(E->getDecl(), Init);
2122 CodeGenFunction::PeepholeProtection
2123 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2124 // At the moment, the only aggressive peephole we do in IR gen
2125 // is trunc(zext) folding, but if we add more, we can easily
2126 // extend this protection.
2128 if (!rvalue.isScalar()) return PeepholeProtection();
2129 llvm::Value *value = rvalue.getScalarVal();
2130 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2132 // Just make an extra bitcast.
2133 assert(HaveInsertPoint());
2134 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2135 Builder.GetInsertBlock());
2137 PeepholeProtection protection;
2138 protection.Inst = inst;
2142 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2143 if (!protection.Inst) return;
2145 // In theory, we could try to duplicate the peepholes now, but whatever.
2146 protection.Inst->eraseFromParent();
2149 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2150 QualType Ty, SourceLocation Loc,
2151 SourceLocation AssumptionLoc,
2152 llvm::Value *Alignment,
2153 llvm::Value *OffsetValue) {
2154 llvm::Value *TheCheck;
2155 llvm::Instruction *Assumption = Builder.CreateAlignmentAssumption(
2156 CGM.getDataLayout(), PtrValue, Alignment, OffsetValue, &TheCheck);
2157 if (SanOpts.has(SanitizerKind::Alignment)) {
2158 EmitAlignmentAssumptionCheck(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
2159 OffsetValue, TheCheck, Assumption);
2163 void CodeGenFunction::EmitAlignmentAssumption(llvm::Value *PtrValue,
2165 SourceLocation AssumptionLoc,
2166 llvm::Value *Alignment,
2167 llvm::Value *OffsetValue) {
2168 if (auto *CE = dyn_cast<CastExpr>(E))
2169 E = CE->getSubExprAsWritten();
2170 QualType Ty = E->getType();
2171 SourceLocation Loc = E->getExprLoc();
2173 EmitAlignmentAssumption(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
2177 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Function *AnnotationFn,
2178 llvm::Value *AnnotatedVal,
2179 StringRef AnnotationStr,
2180 SourceLocation Location) {
2181 llvm::Value *Args[4] = {
2183 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2184 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2185 CGM.EmitAnnotationLineNo(Location)
2187 return Builder.CreateCall(AnnotationFn, Args);
2190 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2191 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2192 // FIXME We create a new bitcast for every annotation because that's what
2193 // llvm-gcc was doing.
2194 for (const auto *I : D->specific_attrs<AnnotateAttr>())
2195 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2196 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2197 I->getAnnotation(), D->getLocation());
2200 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2202 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2203 llvm::Value *V = Addr.getPointer();
2204 llvm::Type *VTy = V->getType();
2205 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2208 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2209 // FIXME Always emit the cast inst so we can differentiate between
2210 // annotation on the first field of a struct and annotation on the struct
2212 if (VTy != CGM.Int8PtrTy)
2213 V = Builder.CreateBitCast(V, CGM.Int8PtrTy);
2214 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2215 V = Builder.CreateBitCast(V, VTy);
2218 return Address(V, Addr.getAlignment());
2221 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2223 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2225 assert(!CGF->IsSanitizerScope);
2226 CGF->IsSanitizerScope = true;
2229 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2230 CGF->IsSanitizerScope = false;
2233 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2234 const llvm::Twine &Name,
2235 llvm::BasicBlock *BB,
2236 llvm::BasicBlock::iterator InsertPt) const {
2237 LoopStack.InsertHelper(I);
2238 if (IsSanitizerScope)
2239 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2242 void CGBuilderInserter::InsertHelper(
2243 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2244 llvm::BasicBlock::iterator InsertPt) const {
2245 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2247 CGF->InsertHelper(I, Name, BB, InsertPt);
2250 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2251 CodeGenModule &CGM, const FunctionDecl *FD,
2252 std::string &FirstMissing) {
2253 // If there aren't any required features listed then go ahead and return.
2254 if (ReqFeatures.empty())
2257 // Now build up the set of caller features and verify that all the required
2258 // features are there.
2259 llvm::StringMap<bool> CallerFeatureMap;
2260 CGM.getContext().getFunctionFeatureMap(CallerFeatureMap, FD);
2262 // If we have at least one of the features in the feature list return
2263 // true, otherwise return false.
2265 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2266 SmallVector<StringRef, 1> OrFeatures;
2267 Feature.split(OrFeatures, '|');
2268 return llvm::any_of(OrFeatures, [&](StringRef Feature) {
2269 if (!CallerFeatureMap.lookup(Feature)) {
2270 FirstMissing = Feature.str();
2278 // Emits an error if we don't have a valid set of target features for the
2280 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2281 const FunctionDecl *TargetDecl) {
2282 return checkTargetFeatures(E->getBeginLoc(), TargetDecl);
2285 // Emits an error if we don't have a valid set of target features for the
2287 void CodeGenFunction::checkTargetFeatures(SourceLocation Loc,
2288 const FunctionDecl *TargetDecl) {
2289 // Early exit if this is an indirect call.
2293 // Get the current enclosing function if it exists. If it doesn't
2294 // we can't check the target features anyhow.
2295 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl);
2299 // Grab the required features for the call. For a builtin this is listed in
2300 // the td file with the default cpu, for an always_inline function this is any
2301 // listed cpu and any listed features.
2302 unsigned BuiltinID = TargetDecl->getBuiltinID();
2303 std::string MissingFeature;
2305 SmallVector<StringRef, 1> ReqFeatures;
2306 const char *FeatureList =
2307 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2308 // Return if the builtin doesn't have any required features.
2309 if (!FeatureList || StringRef(FeatureList) == "")
2311 StringRef(FeatureList).split(ReqFeatures, ',');
2312 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2313 CGM.getDiags().Report(Loc, diag::err_builtin_needs_feature)
2314 << TargetDecl->getDeclName()
2315 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2317 } else if (!TargetDecl->isMultiVersion() &&
2318 TargetDecl->hasAttr<TargetAttr>()) {
2319 // Get the required features for the callee.
2321 const TargetAttr *TD = TargetDecl->getAttr<TargetAttr>();
2322 ParsedTargetAttr ParsedAttr =
2323 CGM.getContext().filterFunctionTargetAttrs(TD);
2325 SmallVector<StringRef, 1> ReqFeatures;
2326 llvm::StringMap<bool> CalleeFeatureMap;
2327 CGM.getContext().getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2329 for (const auto &F : ParsedAttr.Features) {
2330 if (F[0] == '+' && CalleeFeatureMap.lookup(F.substr(1)))
2331 ReqFeatures.push_back(StringRef(F).substr(1));
2334 for (const auto &F : CalleeFeatureMap) {
2335 // Only positive features are "required".
2337 ReqFeatures.push_back(F.getKey());
2339 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2340 CGM.getDiags().Report(Loc, diag::err_function_needs_feature)
2341 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2345 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2346 if (!CGM.getCodeGenOpts().SanitizeStats)
2349 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2350 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2351 CGM.getSanStats().create(IRB, SSK);
2355 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
2356 llvm::Value *Condition = nullptr;
2358 if (!RO.Conditions.Architecture.empty())
2359 Condition = EmitX86CpuIs(RO.Conditions.Architecture);
2361 if (!RO.Conditions.Features.empty()) {
2362 llvm::Value *FeatureCond = EmitX86CpuSupports(RO.Conditions.Features);
2364 Condition ? Builder.CreateAnd(Condition, FeatureCond) : FeatureCond;
2369 static void CreateMultiVersionResolverReturn(CodeGenModule &CGM,
2370 llvm::Function *Resolver,
2371 CGBuilderTy &Builder,
2372 llvm::Function *FuncToReturn,
2373 bool SupportsIFunc) {
2374 if (SupportsIFunc) {
2375 Builder.CreateRet(FuncToReturn);
2379 llvm::SmallVector<llvm::Value *, 10> Args;
2380 llvm::for_each(Resolver->args(),
2381 [&](llvm::Argument &Arg) { Args.push_back(&Arg); });
2383 llvm::CallInst *Result = Builder.CreateCall(FuncToReturn, Args);
2384 Result->setTailCallKind(llvm::CallInst::TCK_MustTail);
2386 if (Resolver->getReturnType()->isVoidTy())
2387 Builder.CreateRetVoid();
2389 Builder.CreateRet(Result);
2392 void CodeGenFunction::EmitMultiVersionResolver(
2393 llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
2394 assert(getContext().getTargetInfo().getTriple().isX86() &&
2395 "Only implemented for x86 targets");
2397 bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc();
2399 // Main function's basic block.
2400 llvm::BasicBlock *CurBlock = createBasicBlock("resolver_entry", Resolver);
2401 Builder.SetInsertPoint(CurBlock);
2404 for (const MultiVersionResolverOption &RO : Options) {
2405 Builder.SetInsertPoint(CurBlock);
2406 llvm::Value *Condition = FormResolverCondition(RO);
2408 // The 'default' or 'generic' case.
2410 assert(&RO == Options.end() - 1 &&
2411 "Default or Generic case must be last");
2412 CreateMultiVersionResolverReturn(CGM, Resolver, Builder, RO.Function,
2417 llvm::BasicBlock *RetBlock = createBasicBlock("resolver_return", Resolver);
2418 CGBuilderTy RetBuilder(*this, RetBlock);
2419 CreateMultiVersionResolverReturn(CGM, Resolver, RetBuilder, RO.Function,
2421 CurBlock = createBasicBlock("resolver_else", Resolver);
2422 Builder.CreateCondBr(Condition, RetBlock, CurBlock);
2425 // If no generic/default, emit an unreachable.
2426 Builder.SetInsertPoint(CurBlock);
2427 llvm::CallInst *TrapCall = EmitTrapCall(llvm::Intrinsic::trap);
2428 TrapCall->setDoesNotReturn();
2429 TrapCall->setDoesNotThrow();
2430 Builder.CreateUnreachable();
2431 Builder.ClearInsertionPoint();
2434 // Loc - where the diagnostic will point, where in the source code this
2435 // alignment has failed.
2436 // SecondaryLoc - if present (will be present if sufficiently different from
2437 // Loc), the diagnostic will additionally point a "Note:" to this location.
2438 // It should be the location where the __attribute__((assume_aligned))
2440 void CodeGenFunction::EmitAlignmentAssumptionCheck(
2441 llvm::Value *Ptr, QualType Ty, SourceLocation Loc,
2442 SourceLocation SecondaryLoc, llvm::Value *Alignment,
2443 llvm::Value *OffsetValue, llvm::Value *TheCheck,
2444 llvm::Instruction *Assumption) {
2445 assert(Assumption && isa<llvm::CallInst>(Assumption) &&
2446 cast<llvm::CallInst>(Assumption)->getCalledValue() ==
2447 llvm::Intrinsic::getDeclaration(
2448 Builder.GetInsertBlock()->getParent()->getParent(),
2449 llvm::Intrinsic::assume) &&
2450 "Assumption should be a call to llvm.assume().");
2451 assert(&(Builder.GetInsertBlock()->back()) == Assumption &&
2452 "Assumption should be the last instruction of the basic block, "
2453 "since the basic block is still being generated.");
2455 if (!SanOpts.has(SanitizerKind::Alignment))
2458 // Don't check pointers to volatile data. The behavior here is implementation-
2460 if (Ty->getPointeeType().isVolatileQualified())
2463 // We need to temorairly remove the assumption so we can insert the
2464 // sanitizer check before it, else the check will be dropped by optimizations.
2465 Assumption->removeFromParent();
2468 SanitizerScope SanScope(this);
2471 OffsetValue = Builder.getInt1(0); // no offset.
2473 llvm::Constant *StaticData[] = {EmitCheckSourceLocation(Loc),
2474 EmitCheckSourceLocation(SecondaryLoc),
2475 EmitCheckTypeDescriptor(Ty)};
2476 llvm::Value *DynamicData[] = {EmitCheckValue(Ptr),
2477 EmitCheckValue(Alignment),
2478 EmitCheckValue(OffsetValue)};
2479 EmitCheck({std::make_pair(TheCheck, SanitizerKind::Alignment)},
2480 SanitizerHandler::AlignmentAssumption, StaticData, DynamicData);
2483 // We are now in the (new, empty) "cont" basic block.
2484 // Reintroduce the assumption.
2485 Builder.Insert(Assumption);
2486 // FIXME: Assumption still has it's original basic block as it's Parent.
2489 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2490 if (CGDebugInfo *DI = getDebugInfo())
2491 return DI->SourceLocToDebugLoc(Location);
2493 return llvm::DebugLoc();