1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This coordinates the per-function state used while generating code.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "CodeGenFunction.h"
15 #include "CGBlocks.h"
16 #include "CGCleanup.h"
17 #include "CGCUDARuntime.h"
18 #include "CGCXXABI.h"
19 #include "CGDebugInfo.h"
20 #include "CGOpenMPRuntime.h"
21 #include "CodeGenModule.h"
22 #include "CodeGenPGO.h"
23 #include "TargetInfo.h"
24 #include "clang/AST/ASTContext.h"
25 #include "clang/AST/Decl.h"
26 #include "clang/AST/DeclCXX.h"
27 #include "clang/AST/StmtCXX.h"
28 #include "clang/Basic/Builtins.h"
29 #include "clang/Basic/TargetInfo.h"
30 #include "clang/CodeGen/CGFunctionInfo.h"
31 #include "clang/Frontend/CodeGenOptions.h"
32 #include "clang/Sema/SemaDiagnostic.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Operator.h"
37 using namespace clang;
38 using namespace CodeGen;
39
CodeGenFunction(CodeGenModule & cgm,bool suppressNewContext)40 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
41 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
42 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
43 CGBuilderInserterTy(this)),
44 CurFn(nullptr), ReturnValue(Address::invalid()),
45 CapturedStmtInfo(nullptr),
46 SanOpts(CGM.getLangOpts().Sanitize), IsSanitizerScope(false),
47 CurFuncIsThunk(false), AutoreleaseResult(false), SawAsmBlock(false),
48 IsOutlinedSEHHelper(false),
49 BlockInfo(nullptr), BlockPointer(nullptr),
50 LambdaThisCaptureField(nullptr), NormalCleanupDest(nullptr),
51 NextCleanupDestIndex(1), FirstBlockInfo(nullptr), EHResumeBlock(nullptr),
52 ExceptionSlot(nullptr), EHSelectorSlot(nullptr),
53 DebugInfo(CGM.getModuleDebugInfo()),
54 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
55 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
56 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
57 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
58 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
59 CXXStructorImplicitParamDecl(nullptr),
60 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
61 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
62 TerminateHandler(nullptr), TrapBB(nullptr) {
63 if (!suppressNewContext)
64 CGM.getCXXABI().getMangleContext().startNewFunction();
65
66 llvm::FastMathFlags FMF;
67 if (CGM.getLangOpts().FastMath)
68 FMF.setUnsafeAlgebra();
69 if (CGM.getLangOpts().FiniteMathOnly) {
70 FMF.setNoNaNs();
71 FMF.setNoInfs();
72 }
73 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
74 FMF.setNoNaNs();
75 }
76 if (CGM.getCodeGenOpts().NoSignedZeros) {
77 FMF.setNoSignedZeros();
78 }
79 if (CGM.getCodeGenOpts().ReciprocalMath) {
80 FMF.setAllowReciprocal();
81 }
82 Builder.setFastMathFlags(FMF);
83 }
84
~CodeGenFunction()85 CodeGenFunction::~CodeGenFunction() {
86 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
87
88 // If there are any unclaimed block infos, go ahead and destroy them
89 // now. This can happen if IR-gen gets clever and skips evaluating
90 // something.
91 if (FirstBlockInfo)
92 destroyBlockInfos(FirstBlockInfo);
93
94 if (getLangOpts().OpenMP) {
95 CGM.getOpenMPRuntime().functionFinished(*this);
96 }
97 }
98
getNaturalPointeeTypeAlignment(QualType T,AlignmentSource * Source)99 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
100 AlignmentSource *Source) {
101 return getNaturalTypeAlignment(T->getPointeeType(), Source,
102 /*forPointee*/ true);
103 }
104
getNaturalTypeAlignment(QualType T,AlignmentSource * Source,bool forPointeeType)105 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
106 AlignmentSource *Source,
107 bool forPointeeType) {
108 // Honor alignment typedef attributes even on incomplete types.
109 // We also honor them straight for C++ class types, even as pointees;
110 // there's an expressivity gap here.
111 if (auto TT = T->getAs<TypedefType>()) {
112 if (auto Align = TT->getDecl()->getMaxAlignment()) {
113 if (Source) *Source = AlignmentSource::AttributedType;
114 return getContext().toCharUnitsFromBits(Align);
115 }
116 }
117
118 if (Source) *Source = AlignmentSource::Type;
119
120 CharUnits Alignment;
121 if (T->isIncompleteType()) {
122 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
123 } else {
124 // For C++ class pointees, we don't know whether we're pointing at a
125 // base or a complete object, so we generally need to use the
126 // non-virtual alignment.
127 const CXXRecordDecl *RD;
128 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
129 Alignment = CGM.getClassPointerAlignment(RD);
130 } else {
131 Alignment = getContext().getTypeAlignInChars(T);
132 }
133
134 // Cap to the global maximum type alignment unless the alignment
135 // was somehow explicit on the type.
136 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
137 if (Alignment.getQuantity() > MaxAlign &&
138 !getContext().isAlignmentRequired(T))
139 Alignment = CharUnits::fromQuantity(MaxAlign);
140 }
141 }
142 return Alignment;
143 }
144
MakeNaturalAlignAddrLValue(llvm::Value * V,QualType T)145 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
146 AlignmentSource AlignSource;
147 CharUnits Alignment = getNaturalTypeAlignment(T, &AlignSource);
148 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), AlignSource,
149 CGM.getTBAAInfo(T));
150 }
151
152 /// Given a value of type T* that may not be to a complete object,
153 /// construct an l-value with the natural pointee alignment of T.
154 LValue
MakeNaturalAlignPointeeAddrLValue(llvm::Value * V,QualType T)155 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
156 AlignmentSource AlignSource;
157 CharUnits Align = getNaturalTypeAlignment(T, &AlignSource, /*pointee*/ true);
158 return MakeAddrLValue(Address(V, Align), T, AlignSource);
159 }
160
161
ConvertTypeForMem(QualType T)162 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
163 return CGM.getTypes().ConvertTypeForMem(T);
164 }
165
ConvertType(QualType T)166 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
167 return CGM.getTypes().ConvertType(T);
168 }
169
getEvaluationKind(QualType type)170 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
171 type = type.getCanonicalType();
172 while (true) {
173 switch (type->getTypeClass()) {
174 #define TYPE(name, parent)
175 #define ABSTRACT_TYPE(name, parent)
176 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
177 #define DEPENDENT_TYPE(name, parent) case Type::name:
178 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
179 #include "clang/AST/TypeNodes.def"
180 llvm_unreachable("non-canonical or dependent type in IR-generation");
181
182 case Type::Auto:
183 llvm_unreachable("undeduced auto type in IR-generation");
184
185 // Various scalar types.
186 case Type::Builtin:
187 case Type::Pointer:
188 case Type::BlockPointer:
189 case Type::LValueReference:
190 case Type::RValueReference:
191 case Type::MemberPointer:
192 case Type::Vector:
193 case Type::ExtVector:
194 case Type::FunctionProto:
195 case Type::FunctionNoProto:
196 case Type::Enum:
197 case Type::ObjCObjectPointer:
198 case Type::Pipe:
199 return TEK_Scalar;
200
201 // Complexes.
202 case Type::Complex:
203 return TEK_Complex;
204
205 // Arrays, records, and Objective-C objects.
206 case Type::ConstantArray:
207 case Type::IncompleteArray:
208 case Type::VariableArray:
209 case Type::Record:
210 case Type::ObjCObject:
211 case Type::ObjCInterface:
212 return TEK_Aggregate;
213
214 // We operate on atomic values according to their underlying type.
215 case Type::Atomic:
216 type = cast<AtomicType>(type)->getValueType();
217 continue;
218 }
219 llvm_unreachable("unknown type kind!");
220 }
221 }
222
EmitReturnBlock()223 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
224 // For cleanliness, we try to avoid emitting the return block for
225 // simple cases.
226 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
227
228 if (CurBB) {
229 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
230
231 // We have a valid insert point, reuse it if it is empty or there are no
232 // explicit jumps to the return block.
233 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
234 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
235 delete ReturnBlock.getBlock();
236 } else
237 EmitBlock(ReturnBlock.getBlock());
238 return llvm::DebugLoc();
239 }
240
241 // Otherwise, if the return block is the target of a single direct
242 // branch then we can just put the code in that block instead. This
243 // cleans up functions which started with a unified return block.
244 if (ReturnBlock.getBlock()->hasOneUse()) {
245 llvm::BranchInst *BI =
246 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
247 if (BI && BI->isUnconditional() &&
248 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
249 // Record/return the DebugLoc of the simple 'return' expression to be used
250 // later by the actual 'ret' instruction.
251 llvm::DebugLoc Loc = BI->getDebugLoc();
252 Builder.SetInsertPoint(BI->getParent());
253 BI->eraseFromParent();
254 delete ReturnBlock.getBlock();
255 return Loc;
256 }
257 }
258
259 // FIXME: We are at an unreachable point, there is no reason to emit the block
260 // unless it has uses. However, we still need a place to put the debug
261 // region.end for now.
262
263 EmitBlock(ReturnBlock.getBlock());
264 return llvm::DebugLoc();
265 }
266
EmitIfUsed(CodeGenFunction & CGF,llvm::BasicBlock * BB)267 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
268 if (!BB) return;
269 if (!BB->use_empty())
270 return CGF.CurFn->getBasicBlockList().push_back(BB);
271 delete BB;
272 }
273
FinishFunction(SourceLocation EndLoc)274 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
275 assert(BreakContinueStack.empty() &&
276 "mismatched push/pop in break/continue stack!");
277
278 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
279 && NumSimpleReturnExprs == NumReturnExprs
280 && ReturnBlock.getBlock()->use_empty();
281 // Usually the return expression is evaluated before the cleanup
282 // code. If the function contains only a simple return statement,
283 // such as a constant, the location before the cleanup code becomes
284 // the last useful breakpoint in the function, because the simple
285 // return expression will be evaluated after the cleanup code. To be
286 // safe, set the debug location for cleanup code to the location of
287 // the return statement. Otherwise the cleanup code should be at the
288 // end of the function's lexical scope.
289 //
290 // If there are multiple branches to the return block, the branch
291 // instructions will get the location of the return statements and
292 // all will be fine.
293 if (CGDebugInfo *DI = getDebugInfo()) {
294 if (OnlySimpleReturnStmts)
295 DI->EmitLocation(Builder, LastStopPoint);
296 else
297 DI->EmitLocation(Builder, EndLoc);
298 }
299
300 // Pop any cleanups that might have been associated with the
301 // parameters. Do this in whatever block we're currently in; it's
302 // important to do this before we enter the return block or return
303 // edges will be *really* confused.
304 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
305 bool HasOnlyLifetimeMarkers =
306 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
307 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
308 if (HasCleanups) {
309 // Make sure the line table doesn't jump back into the body for
310 // the ret after it's been at EndLoc.
311 if (CGDebugInfo *DI = getDebugInfo())
312 if (OnlySimpleReturnStmts)
313 DI->EmitLocation(Builder, EndLoc);
314
315 PopCleanupBlocks(PrologueCleanupDepth);
316 }
317
318 // Emit function epilog (to return).
319 llvm::DebugLoc Loc = EmitReturnBlock();
320
321 if (ShouldInstrumentFunction())
322 EmitFunctionInstrumentation("__cyg_profile_func_exit");
323
324 // Emit debug descriptor for function end.
325 if (CGDebugInfo *DI = getDebugInfo())
326 DI->EmitFunctionEnd(Builder);
327
328 // Reset the debug location to that of the simple 'return' expression, if any
329 // rather than that of the end of the function's scope '}'.
330 ApplyDebugLocation AL(*this, Loc);
331 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
332 EmitEndEHSpec(CurCodeDecl);
333
334 assert(EHStack.empty() &&
335 "did not remove all scopes from cleanup stack!");
336
337 // If someone did an indirect goto, emit the indirect goto block at the end of
338 // the function.
339 if (IndirectBranch) {
340 EmitBlock(IndirectBranch->getParent());
341 Builder.ClearInsertionPoint();
342 }
343
344 // If some of our locals escaped, insert a call to llvm.localescape in the
345 // entry block.
346 if (!EscapedLocals.empty()) {
347 // Invert the map from local to index into a simple vector. There should be
348 // no holes.
349 SmallVector<llvm::Value *, 4> EscapeArgs;
350 EscapeArgs.resize(EscapedLocals.size());
351 for (auto &Pair : EscapedLocals)
352 EscapeArgs[Pair.second] = Pair.first;
353 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
354 &CGM.getModule(), llvm::Intrinsic::localescape);
355 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
356 }
357
358 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
359 llvm::Instruction *Ptr = AllocaInsertPt;
360 AllocaInsertPt = nullptr;
361 Ptr->eraseFromParent();
362
363 // If someone took the address of a label but never did an indirect goto, we
364 // made a zero entry PHI node, which is illegal, zap it now.
365 if (IndirectBranch) {
366 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
367 if (PN->getNumIncomingValues() == 0) {
368 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
369 PN->eraseFromParent();
370 }
371 }
372
373 EmitIfUsed(*this, EHResumeBlock);
374 EmitIfUsed(*this, TerminateLandingPad);
375 EmitIfUsed(*this, TerminateHandler);
376 EmitIfUsed(*this, UnreachableBlock);
377
378 if (CGM.getCodeGenOpts().EmitDeclMetadata)
379 EmitDeclMetadata();
380
381 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
382 I = DeferredReplacements.begin(),
383 E = DeferredReplacements.end();
384 I != E; ++I) {
385 I->first->replaceAllUsesWith(I->second);
386 I->first->eraseFromParent();
387 }
388 }
389
390 /// ShouldInstrumentFunction - Return true if the current function should be
391 /// instrumented with __cyg_profile_func_* calls
ShouldInstrumentFunction()392 bool CodeGenFunction::ShouldInstrumentFunction() {
393 if (!CGM.getCodeGenOpts().InstrumentFunctions)
394 return false;
395 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
396 return false;
397 return true;
398 }
399
400 /// ShouldXRayInstrument - Return true if the current function should be
401 /// instrumented with XRay nop sleds.
ShouldXRayInstrumentFunction() const402 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
403 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
404 }
405
406 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
407 /// instrumentation function with the current function and the call site, if
408 /// function instrumentation is enabled.
EmitFunctionInstrumentation(const char * Fn)409 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
410 auto NL = ApplyDebugLocation::CreateArtificial(*this);
411 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
412 llvm::PointerType *PointerTy = Int8PtrTy;
413 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
414 llvm::FunctionType *FunctionTy =
415 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
416
417 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
418 llvm::CallInst *CallSite = Builder.CreateCall(
419 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
420 llvm::ConstantInt::get(Int32Ty, 0),
421 "callsite");
422
423 llvm::Value *args[] = {
424 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
425 CallSite
426 };
427
428 EmitNounwindRuntimeCall(F, args);
429 }
430
EmitMCountInstrumentation()431 void CodeGenFunction::EmitMCountInstrumentation() {
432 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
433
434 llvm::Constant *MCountFn =
435 CGM.CreateRuntimeFunction(FTy, getTarget().getMCountName());
436 EmitNounwindRuntimeCall(MCountFn);
437 }
438
439 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
440 // information in the program executable. The argument information stored
441 // includes the argument name, its type, the address and access qualifiers used.
GenOpenCLArgMetadata(const FunctionDecl * FD,llvm::Function * Fn,CodeGenModule & CGM,llvm::LLVMContext & Context,CGBuilderTy & Builder,ASTContext & ASTCtx)442 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
443 CodeGenModule &CGM, llvm::LLVMContext &Context,
444 CGBuilderTy &Builder, ASTContext &ASTCtx) {
445 // Create MDNodes that represent the kernel arg metadata.
446 // Each MDNode is a list in the form of "key", N number of values which is
447 // the same number of values as their are kernel arguments.
448
449 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
450
451 // MDNode for the kernel argument address space qualifiers.
452 SmallVector<llvm::Metadata *, 8> addressQuals;
453
454 // MDNode for the kernel argument access qualifiers (images only).
455 SmallVector<llvm::Metadata *, 8> accessQuals;
456
457 // MDNode for the kernel argument type names.
458 SmallVector<llvm::Metadata *, 8> argTypeNames;
459
460 // MDNode for the kernel argument base type names.
461 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
462
463 // MDNode for the kernel argument type qualifiers.
464 SmallVector<llvm::Metadata *, 8> argTypeQuals;
465
466 // MDNode for the kernel argument names.
467 SmallVector<llvm::Metadata *, 8> argNames;
468
469 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
470 const ParmVarDecl *parm = FD->getParamDecl(i);
471 QualType ty = parm->getType();
472 std::string typeQuals;
473
474 if (ty->isPointerType()) {
475 QualType pointeeTy = ty->getPointeeType();
476
477 // Get address qualifier.
478 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
479 ASTCtx.getTargetAddressSpace(pointeeTy.getAddressSpace()))));
480
481 // Get argument type name.
482 std::string typeName =
483 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
484
485 // Turn "unsigned type" to "utype"
486 std::string::size_type pos = typeName.find("unsigned");
487 if (pointeeTy.isCanonical() && pos != std::string::npos)
488 typeName.erase(pos+1, 8);
489
490 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
491
492 std::string baseTypeName =
493 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
494 Policy) +
495 "*";
496
497 // Turn "unsigned type" to "utype"
498 pos = baseTypeName.find("unsigned");
499 if (pos != std::string::npos)
500 baseTypeName.erase(pos+1, 8);
501
502 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
503
504 // Get argument type qualifiers:
505 if (ty.isRestrictQualified())
506 typeQuals = "restrict";
507 if (pointeeTy.isConstQualified() ||
508 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
509 typeQuals += typeQuals.empty() ? "const" : " const";
510 if (pointeeTy.isVolatileQualified())
511 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
512 } else {
513 uint32_t AddrSpc = 0;
514 bool isPipe = ty->isPipeType();
515 if (ty->isImageType() || isPipe)
516 AddrSpc =
517 CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
518
519 addressQuals.push_back(
520 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
521
522 // Get argument type name.
523 std::string typeName;
524 if (isPipe)
525 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
526 .getAsString(Policy);
527 else
528 typeName = ty.getUnqualifiedType().getAsString(Policy);
529
530 // Turn "unsigned type" to "utype"
531 std::string::size_type pos = typeName.find("unsigned");
532 if (ty.isCanonical() && pos != std::string::npos)
533 typeName.erase(pos+1, 8);
534
535 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
536
537 std::string baseTypeName;
538 if (isPipe)
539 baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
540 ->getElementType().getCanonicalType()
541 .getAsString(Policy);
542 else
543 baseTypeName =
544 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
545
546 // Turn "unsigned type" to "utype"
547 pos = baseTypeName.find("unsigned");
548 if (pos != std::string::npos)
549 baseTypeName.erase(pos+1, 8);
550
551 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
552
553 // Get argument type qualifiers:
554 if (ty.isConstQualified())
555 typeQuals = "const";
556 if (ty.isVolatileQualified())
557 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
558 if (isPipe)
559 typeQuals = "pipe";
560 }
561
562 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
563
564 // Get image and pipe access qualifier:
565 if (ty->isImageType()|| ty->isPipeType()) {
566 const OpenCLAccessAttr *A = parm->getAttr<OpenCLAccessAttr>();
567 if (A && A->isWriteOnly())
568 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
569 else if (A && A->isReadWrite())
570 accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
571 else
572 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
573 } else
574 accessQuals.push_back(llvm::MDString::get(Context, "none"));
575
576 // Get argument name.
577 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
578 }
579
580 Fn->setMetadata("kernel_arg_addr_space",
581 llvm::MDNode::get(Context, addressQuals));
582 Fn->setMetadata("kernel_arg_access_qual",
583 llvm::MDNode::get(Context, accessQuals));
584 Fn->setMetadata("kernel_arg_type",
585 llvm::MDNode::get(Context, argTypeNames));
586 Fn->setMetadata("kernel_arg_base_type",
587 llvm::MDNode::get(Context, argBaseTypeNames));
588 Fn->setMetadata("kernel_arg_type_qual",
589 llvm::MDNode::get(Context, argTypeQuals));
590 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
591 Fn->setMetadata("kernel_arg_name",
592 llvm::MDNode::get(Context, argNames));
593 }
594
EmitOpenCLKernelMetadata(const FunctionDecl * FD,llvm::Function * Fn)595 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
596 llvm::Function *Fn)
597 {
598 if (!FD->hasAttr<OpenCLKernelAttr>())
599 return;
600
601 llvm::LLVMContext &Context = getLLVMContext();
602
603 GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
604
605 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
606 QualType hintQTy = A->getTypeHint();
607 const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
608 bool isSignedInteger =
609 hintQTy->isSignedIntegerType() ||
610 (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
611 llvm::Metadata *attrMDArgs[] = {
612 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
613 CGM.getTypes().ConvertType(A->getTypeHint()))),
614 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
615 llvm::IntegerType::get(Context, 32),
616 llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
617 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, attrMDArgs));
618 }
619
620 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
621 llvm::Metadata *attrMDArgs[] = {
622 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
623 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
624 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
625 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, attrMDArgs));
626 }
627
628 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
629 llvm::Metadata *attrMDArgs[] = {
630 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
631 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
632 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
633 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, attrMDArgs));
634 }
635 }
636
637 /// Determine whether the function F ends with a return stmt.
endsWithReturn(const Decl * F)638 static bool endsWithReturn(const Decl* F) {
639 const Stmt *Body = nullptr;
640 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
641 Body = FD->getBody();
642 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
643 Body = OMD->getBody();
644
645 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
646 auto LastStmt = CS->body_rbegin();
647 if (LastStmt != CS->body_rend())
648 return isa<ReturnStmt>(*LastStmt);
649 }
650 return false;
651 }
652
StartFunction(GlobalDecl GD,QualType RetTy,llvm::Function * Fn,const CGFunctionInfo & FnInfo,const FunctionArgList & Args,SourceLocation Loc,SourceLocation StartLoc)653 void CodeGenFunction::StartFunction(GlobalDecl GD,
654 QualType RetTy,
655 llvm::Function *Fn,
656 const CGFunctionInfo &FnInfo,
657 const FunctionArgList &Args,
658 SourceLocation Loc,
659 SourceLocation StartLoc) {
660 assert(!CurFn &&
661 "Do not use a CodeGenFunction object for more than one function");
662
663 const Decl *D = GD.getDecl();
664
665 DidCallStackSave = false;
666 CurCodeDecl = D;
667 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
668 if (FD->usesSEHTry())
669 CurSEHParent = FD;
670 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
671 FnRetTy = RetTy;
672 CurFn = Fn;
673 CurFnInfo = &FnInfo;
674 assert(CurFn->isDeclaration() && "Function already has body?");
675
676 if (CGM.isInSanitizerBlacklist(Fn, Loc))
677 SanOpts.clear();
678
679 if (D) {
680 // Apply the no_sanitize* attributes to SanOpts.
681 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
682 SanOpts.Mask &= ~Attr->getMask();
683 }
684
685 // Apply sanitizer attributes to the function.
686 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
687 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
688 if (SanOpts.has(SanitizerKind::Thread))
689 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
690 if (SanOpts.has(SanitizerKind::Memory))
691 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
692 if (SanOpts.has(SanitizerKind::SafeStack))
693 Fn->addFnAttr(llvm::Attribute::SafeStack);
694
695 // Apply xray attributes to the function (as a string, for now)
696 if (D && ShouldXRayInstrumentFunction()) {
697 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
698 if (XRayAttr->alwaysXRayInstrument())
699 Fn->addFnAttr("function-instrument", "xray-always");
700 if (XRayAttr->neverXRayInstrument())
701 Fn->addFnAttr("function-instrument", "xray-never");
702 } else {
703 Fn->addFnAttr(
704 "xray-instruction-threshold",
705 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
706 }
707 }
708
709 // Pass inline keyword to optimizer if it appears explicitly on any
710 // declaration. Also, in the case of -fno-inline attach NoInline
711 // attribute to all functions that are not marked AlwaysInline, or
712 // to all functions that are not marked inline or implicitly inline
713 // in the case of -finline-hint-functions.
714 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
715 const CodeGenOptions& CodeGenOpts = CGM.getCodeGenOpts();
716 if (!CodeGenOpts.NoInline) {
717 for (auto RI : FD->redecls())
718 if (RI->isInlineSpecified()) {
719 Fn->addFnAttr(llvm::Attribute::InlineHint);
720 break;
721 }
722 if (CodeGenOpts.getInlining() == CodeGenOptions::OnlyHintInlining &&
723 !FD->isInlined() && !Fn->hasFnAttribute(llvm::Attribute::InlineHint))
724 Fn->addFnAttr(llvm::Attribute::NoInline);
725 } else if (!FD->hasAttr<AlwaysInlineAttr>())
726 Fn->addFnAttr(llvm::Attribute::NoInline);
727 if (CGM.getLangOpts().OpenMP && FD->hasAttr<OMPDeclareSimdDeclAttr>())
728 CGM.getOpenMPRuntime().emitDeclareSimdFunction(FD, Fn);
729 }
730
731 // Add no-jump-tables value.
732 Fn->addFnAttr("no-jump-tables",
733 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
734
735 if (getLangOpts().OpenCL) {
736 // Add metadata for a kernel function.
737 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
738 EmitOpenCLKernelMetadata(FD, Fn);
739 }
740
741 // If we are checking function types, emit a function type signature as
742 // prologue data.
743 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
744 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
745 if (llvm::Constant *PrologueSig =
746 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
747 llvm::Constant *FTRTTIConst =
748 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
749 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
750 llvm::Constant *PrologueStructConst =
751 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
752 Fn->setPrologueData(PrologueStructConst);
753 }
754 }
755 }
756
757 // If we're in C++ mode and the function name is "main", it is guaranteed
758 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
759 // used within a program").
760 if (getLangOpts().CPlusPlus)
761 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
762 if (FD->isMain())
763 Fn->addFnAttr(llvm::Attribute::NoRecurse);
764
765 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
766
767 // Create a marker to make it easy to insert allocas into the entryblock
768 // later. Don't create this with the builder, because we don't want it
769 // folded.
770 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
771 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
772
773 ReturnBlock = getJumpDestInCurrentScope("return");
774
775 Builder.SetInsertPoint(EntryBB);
776
777 // Emit subprogram debug descriptor.
778 if (CGDebugInfo *DI = getDebugInfo()) {
779 // Reconstruct the type from the argument list so that implicit parameters,
780 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
781 // convention.
782 CallingConv CC = CallingConv::CC_C;
783 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
784 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
785 CC = SrcFnTy->getCallConv();
786 SmallVector<QualType, 16> ArgTypes;
787 for (const VarDecl *VD : Args)
788 ArgTypes.push_back(VD->getType());
789 QualType FnType = getContext().getFunctionType(
790 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
791 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
792 }
793
794 if (ShouldInstrumentFunction())
795 EmitFunctionInstrumentation("__cyg_profile_func_enter");
796
797 if (CGM.getCodeGenOpts().InstrumentForProfiling)
798 EmitMCountInstrumentation();
799
800 if (RetTy->isVoidType()) {
801 // Void type; nothing to return.
802 ReturnValue = Address::invalid();
803
804 // Count the implicit return.
805 if (!endsWithReturn(D))
806 ++NumReturnExprs;
807 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
808 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
809 // Indirect aggregate return; emit returned value directly into sret slot.
810 // This reduces code size, and affects correctness in C++.
811 auto AI = CurFn->arg_begin();
812 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
813 ++AI;
814 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
815 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
816 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
817 // Load the sret pointer from the argument struct and return into that.
818 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
819 llvm::Function::arg_iterator EI = CurFn->arg_end();
820 --EI;
821 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
822 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
823 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
824 } else {
825 ReturnValue = CreateIRTemp(RetTy, "retval");
826
827 // Tell the epilog emitter to autorelease the result. We do this
828 // now so that various specialized functions can suppress it
829 // during their IR-generation.
830 if (getLangOpts().ObjCAutoRefCount &&
831 !CurFnInfo->isReturnsRetained() &&
832 RetTy->isObjCRetainableType())
833 AutoreleaseResult = true;
834 }
835
836 EmitStartEHSpec(CurCodeDecl);
837
838 PrologueCleanupDepth = EHStack.stable_begin();
839 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
840
841 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
842 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
843 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
844 if (MD->getParent()->isLambda() &&
845 MD->getOverloadedOperator() == OO_Call) {
846 // We're in a lambda; figure out the captures.
847 MD->getParent()->getCaptureFields(LambdaCaptureFields,
848 LambdaThisCaptureField);
849 if (LambdaThisCaptureField) {
850 // If the lambda captures the object referred to by '*this' - either by
851 // value or by reference, make sure CXXThisValue points to the correct
852 // object.
853
854 // Get the lvalue for the field (which is a copy of the enclosing object
855 // or contains the address of the enclosing object).
856 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
857 if (!LambdaThisCaptureField->getType()->isPointerType()) {
858 // If the enclosing object was captured by value, just use its address.
859 CXXThisValue = ThisFieldLValue.getAddress().getPointer();
860 } else {
861 // Load the lvalue pointed to by the field, since '*this' was captured
862 // by reference.
863 CXXThisValue =
864 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
865 }
866 }
867 for (auto *FD : MD->getParent()->fields()) {
868 if (FD->hasCapturedVLAType()) {
869 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
870 SourceLocation()).getScalarVal();
871 auto VAT = FD->getCapturedVLAType();
872 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
873 }
874 }
875 } else {
876 // Not in a lambda; just use 'this' from the method.
877 // FIXME: Should we generate a new load for each use of 'this'? The
878 // fast register allocator would be happier...
879 CXXThisValue = CXXABIThisValue;
880 }
881 }
882
883 // If any of the arguments have a variably modified type, make sure to
884 // emit the type size.
885 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
886 i != e; ++i) {
887 const VarDecl *VD = *i;
888
889 // Dig out the type as written from ParmVarDecls; it's unclear whether
890 // the standard (C99 6.9.1p10) requires this, but we're following the
891 // precedent set by gcc.
892 QualType Ty;
893 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
894 Ty = PVD->getOriginalType();
895 else
896 Ty = VD->getType();
897
898 if (Ty->isVariablyModifiedType())
899 EmitVariablyModifiedType(Ty);
900 }
901 // Emit a location at the end of the prologue.
902 if (CGDebugInfo *DI = getDebugInfo())
903 DI->EmitLocation(Builder, StartLoc);
904 }
905
EmitFunctionBody(FunctionArgList & Args,const Stmt * Body)906 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
907 const Stmt *Body) {
908 incrementProfileCounter(Body);
909 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
910 EmitCompoundStmtWithoutScope(*S);
911 else
912 EmitStmt(Body);
913 }
914
915 /// When instrumenting to collect profile data, the counts for some blocks
916 /// such as switch cases need to not include the fall-through counts, so
917 /// emit a branch around the instrumentation code. When not instrumenting,
918 /// this just calls EmitBlock().
EmitBlockWithFallThrough(llvm::BasicBlock * BB,const Stmt * S)919 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
920 const Stmt *S) {
921 llvm::BasicBlock *SkipCountBB = nullptr;
922 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
923 // When instrumenting for profiling, the fallthrough to certain
924 // statements needs to skip over the instrumentation code so that we
925 // get an accurate count.
926 SkipCountBB = createBasicBlock("skipcount");
927 EmitBranch(SkipCountBB);
928 }
929 EmitBlock(BB);
930 uint64_t CurrentCount = getCurrentProfileCount();
931 incrementProfileCounter(S);
932 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
933 if (SkipCountBB)
934 EmitBlock(SkipCountBB);
935 }
936
937 /// Tries to mark the given function nounwind based on the
938 /// non-existence of any throwing calls within it. We believe this is
939 /// lightweight enough to do at -O0.
TryMarkNoThrow(llvm::Function * F)940 static void TryMarkNoThrow(llvm::Function *F) {
941 // LLVM treats 'nounwind' on a function as part of the type, so we
942 // can't do this on functions that can be overwritten.
943 if (F->isInterposable()) return;
944
945 for (llvm::BasicBlock &BB : *F)
946 for (llvm::Instruction &I : BB)
947 if (I.mayThrow())
948 return;
949
950 F->setDoesNotThrow();
951 }
952
BuildFunctionArgList(GlobalDecl GD,FunctionArgList & Args)953 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
954 FunctionArgList &Args) {
955 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
956 QualType ResTy = FD->getReturnType();
957
958 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
959 if (MD && MD->isInstance()) {
960 if (CGM.getCXXABI().HasThisReturn(GD))
961 ResTy = MD->getThisType(getContext());
962 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
963 ResTy = CGM.getContext().VoidPtrTy;
964 CGM.getCXXABI().buildThisParam(*this, Args);
965 }
966
967 // The base version of an inheriting constructor whose constructed base is a
968 // virtual base is not passed any arguments (because it doesn't actually call
969 // the inherited constructor).
970 bool PassedParams = true;
971 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
972 if (auto Inherited = CD->getInheritedConstructor())
973 PassedParams =
974 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
975
976 if (PassedParams) {
977 for (auto *Param : FD->parameters()) {
978 Args.push_back(Param);
979 if (!Param->hasAttr<PassObjectSizeAttr>())
980 continue;
981
982 IdentifierInfo *NoID = nullptr;
983 auto *Implicit = ImplicitParamDecl::Create(
984 getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
985 getContext().getSizeType());
986 SizeArguments[Param] = Implicit;
987 Args.push_back(Implicit);
988 }
989 }
990
991 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
992 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
993
994 return ResTy;
995 }
996
GenerateCode(GlobalDecl GD,llvm::Function * Fn,const CGFunctionInfo & FnInfo)997 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
998 const CGFunctionInfo &FnInfo) {
999 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1000 CurGD = GD;
1001
1002 FunctionArgList Args;
1003 QualType ResTy = BuildFunctionArgList(GD, Args);
1004
1005 // Check if we should generate debug info for this function.
1006 if (FD->hasAttr<NoDebugAttr>())
1007 DebugInfo = nullptr; // disable debug info indefinitely for this function
1008
1009 SourceRange BodyRange;
1010 if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
1011 CurEHLocation = BodyRange.getEnd();
1012
1013 // Use the location of the start of the function to determine where
1014 // the function definition is located. By default use the location
1015 // of the declaration as the location for the subprogram. A function
1016 // may lack a declaration in the source code if it is created by code
1017 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1018 SourceLocation Loc = FD->getLocation();
1019
1020 // If this is a function specialization then use the pattern body
1021 // as the location for the function.
1022 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1023 if (SpecDecl->hasBody(SpecDecl))
1024 Loc = SpecDecl->getLocation();
1025
1026 // Emit the standard function prologue.
1027 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1028
1029 // Generate the body of the function.
1030 PGO.assignRegionCounters(GD, CurFn);
1031 if (isa<CXXDestructorDecl>(FD))
1032 EmitDestructorBody(Args);
1033 else if (isa<CXXConstructorDecl>(FD))
1034 EmitConstructorBody(Args);
1035 else if (getLangOpts().CUDA &&
1036 !getLangOpts().CUDAIsDevice &&
1037 FD->hasAttr<CUDAGlobalAttr>())
1038 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1039 else if (isa<CXXConversionDecl>(FD) &&
1040 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
1041 // The lambda conversion to block pointer is special; the semantics can't be
1042 // expressed in the AST, so IRGen needs to special-case it.
1043 EmitLambdaToBlockPointerBody(Args);
1044 } else if (isa<CXXMethodDecl>(FD) &&
1045 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1046 // The lambda static invoker function is special, because it forwards or
1047 // clones the body of the function call operator (but is actually static).
1048 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
1049 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1050 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1051 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1052 // Implicit copy-assignment gets the same special treatment as implicit
1053 // copy-constructors.
1054 emitImplicitAssignmentOperatorBody(Args);
1055 } else if (Stmt *Body = FD->getBody()) {
1056 EmitFunctionBody(Args, Body);
1057 } else
1058 llvm_unreachable("no definition for emitted function");
1059
1060 // C++11 [stmt.return]p2:
1061 // Flowing off the end of a function [...] results in undefined behavior in
1062 // a value-returning function.
1063 // C11 6.9.1p12:
1064 // If the '}' that terminates a function is reached, and the value of the
1065 // function call is used by the caller, the behavior is undefined.
1066 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1067 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1068 if (SanOpts.has(SanitizerKind::Return)) {
1069 SanitizerScope SanScope(this);
1070 llvm::Value *IsFalse = Builder.getFalse();
1071 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1072 "missing_return", EmitCheckSourceLocation(FD->getLocation()),
1073 None);
1074 } else if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
1075 EmitTrapCall(llvm::Intrinsic::trap);
1076 }
1077 Builder.CreateUnreachable();
1078 Builder.ClearInsertionPoint();
1079 }
1080
1081 // Emit the standard function epilogue.
1082 FinishFunction(BodyRange.getEnd());
1083
1084 // If we haven't marked the function nothrow through other means, do
1085 // a quick pass now to see if we can.
1086 if (!CurFn->doesNotThrow())
1087 TryMarkNoThrow(CurFn);
1088 }
1089
1090 /// ContainsLabel - Return true if the statement contains a label in it. If
1091 /// this statement is not executed normally, it not containing a label means
1092 /// that we can just remove the code.
ContainsLabel(const Stmt * S,bool IgnoreCaseStmts)1093 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1094 // Null statement, not a label!
1095 if (!S) return false;
1096
1097 // If this is a label, we have to emit the code, consider something like:
1098 // if (0) { ... foo: bar(); } goto foo;
1099 //
1100 // TODO: If anyone cared, we could track __label__'s, since we know that you
1101 // can't jump to one from outside their declared region.
1102 if (isa<LabelStmt>(S))
1103 return true;
1104
1105 // If this is a case/default statement, and we haven't seen a switch, we have
1106 // to emit the code.
1107 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1108 return true;
1109
1110 // If this is a switch statement, we want to ignore cases below it.
1111 if (isa<SwitchStmt>(S))
1112 IgnoreCaseStmts = true;
1113
1114 // Scan subexpressions for verboten labels.
1115 for (const Stmt *SubStmt : S->children())
1116 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1117 return true;
1118
1119 return false;
1120 }
1121
1122 /// containsBreak - Return true if the statement contains a break out of it.
1123 /// If the statement (recursively) contains a switch or loop with a break
1124 /// inside of it, this is fine.
containsBreak(const Stmt * S)1125 bool CodeGenFunction::containsBreak(const Stmt *S) {
1126 // Null statement, not a label!
1127 if (!S) return false;
1128
1129 // If this is a switch or loop that defines its own break scope, then we can
1130 // include it and anything inside of it.
1131 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1132 isa<ForStmt>(S))
1133 return false;
1134
1135 if (isa<BreakStmt>(S))
1136 return true;
1137
1138 // Scan subexpressions for verboten breaks.
1139 for (const Stmt *SubStmt : S->children())
1140 if (containsBreak(SubStmt))
1141 return true;
1142
1143 return false;
1144 }
1145
1146
1147 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1148 /// to a constant, or if it does but contains a label, return false. If it
1149 /// constant folds return true and set the boolean result in Result.
ConstantFoldsToSimpleInteger(const Expr * Cond,bool & ResultBool,bool AllowLabels)1150 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1151 bool &ResultBool,
1152 bool AllowLabels) {
1153 llvm::APSInt ResultInt;
1154 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1155 return false;
1156
1157 ResultBool = ResultInt.getBoolValue();
1158 return true;
1159 }
1160
1161 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1162 /// to a constant, or if it does but contains a label, return false. If it
1163 /// constant folds return true and set the folded value.
ConstantFoldsToSimpleInteger(const Expr * Cond,llvm::APSInt & ResultInt,bool AllowLabels)1164 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1165 llvm::APSInt &ResultInt,
1166 bool AllowLabels) {
1167 // FIXME: Rename and handle conversion of other evaluatable things
1168 // to bool.
1169 llvm::APSInt Int;
1170 if (!Cond->EvaluateAsInt(Int, getContext()))
1171 return false; // Not foldable, not integer or not fully evaluatable.
1172
1173 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1174 return false; // Contains a label.
1175
1176 ResultInt = Int;
1177 return true;
1178 }
1179
1180
1181
1182 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1183 /// statement) to the specified blocks. Based on the condition, this might try
1184 /// to simplify the codegen of the conditional based on the branch.
1185 ///
EmitBranchOnBoolExpr(const Expr * Cond,llvm::BasicBlock * TrueBlock,llvm::BasicBlock * FalseBlock,uint64_t TrueCount)1186 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1187 llvm::BasicBlock *TrueBlock,
1188 llvm::BasicBlock *FalseBlock,
1189 uint64_t TrueCount) {
1190 Cond = Cond->IgnoreParens();
1191
1192 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1193
1194 // Handle X && Y in a condition.
1195 if (CondBOp->getOpcode() == BO_LAnd) {
1196 // If we have "1 && X", simplify the code. "0 && X" would have constant
1197 // folded if the case was simple enough.
1198 bool ConstantBool = false;
1199 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1200 ConstantBool) {
1201 // br(1 && X) -> br(X).
1202 incrementProfileCounter(CondBOp);
1203 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1204 TrueCount);
1205 }
1206
1207 // If we have "X && 1", simplify the code to use an uncond branch.
1208 // "X && 0" would have been constant folded to 0.
1209 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1210 ConstantBool) {
1211 // br(X && 1) -> br(X).
1212 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1213 TrueCount);
1214 }
1215
1216 // Emit the LHS as a conditional. If the LHS conditional is false, we
1217 // want to jump to the FalseBlock.
1218 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1219 // The counter tells us how often we evaluate RHS, and all of TrueCount
1220 // can be propagated to that branch.
1221 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1222
1223 ConditionalEvaluation eval(*this);
1224 {
1225 ApplyDebugLocation DL(*this, Cond);
1226 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1227 EmitBlock(LHSTrue);
1228 }
1229
1230 incrementProfileCounter(CondBOp);
1231 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1232
1233 // Any temporaries created here are conditional.
1234 eval.begin(*this);
1235 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1236 eval.end(*this);
1237
1238 return;
1239 }
1240
1241 if (CondBOp->getOpcode() == BO_LOr) {
1242 // If we have "0 || X", simplify the code. "1 || X" would have constant
1243 // folded if the case was simple enough.
1244 bool ConstantBool = false;
1245 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1246 !ConstantBool) {
1247 // br(0 || X) -> br(X).
1248 incrementProfileCounter(CondBOp);
1249 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1250 TrueCount);
1251 }
1252
1253 // If we have "X || 0", simplify the code to use an uncond branch.
1254 // "X || 1" would have been constant folded to 1.
1255 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1256 !ConstantBool) {
1257 // br(X || 0) -> br(X).
1258 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1259 TrueCount);
1260 }
1261
1262 // Emit the LHS as a conditional. If the LHS conditional is true, we
1263 // want to jump to the TrueBlock.
1264 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1265 // We have the count for entry to the RHS and for the whole expression
1266 // being true, so we can divy up True count between the short circuit and
1267 // the RHS.
1268 uint64_t LHSCount =
1269 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1270 uint64_t RHSCount = TrueCount - LHSCount;
1271
1272 ConditionalEvaluation eval(*this);
1273 {
1274 ApplyDebugLocation DL(*this, Cond);
1275 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1276 EmitBlock(LHSFalse);
1277 }
1278
1279 incrementProfileCounter(CondBOp);
1280 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1281
1282 // Any temporaries created here are conditional.
1283 eval.begin(*this);
1284 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1285
1286 eval.end(*this);
1287
1288 return;
1289 }
1290 }
1291
1292 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1293 // br(!x, t, f) -> br(x, f, t)
1294 if (CondUOp->getOpcode() == UO_LNot) {
1295 // Negate the count.
1296 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1297 // Negate the condition and swap the destination blocks.
1298 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1299 FalseCount);
1300 }
1301 }
1302
1303 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1304 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1305 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1306 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1307
1308 ConditionalEvaluation cond(*this);
1309 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1310 getProfileCount(CondOp));
1311
1312 // When computing PGO branch weights, we only know the overall count for
1313 // the true block. This code is essentially doing tail duplication of the
1314 // naive code-gen, introducing new edges for which counts are not
1315 // available. Divide the counts proportionally between the LHS and RHS of
1316 // the conditional operator.
1317 uint64_t LHSScaledTrueCount = 0;
1318 if (TrueCount) {
1319 double LHSRatio =
1320 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1321 LHSScaledTrueCount = TrueCount * LHSRatio;
1322 }
1323
1324 cond.begin(*this);
1325 EmitBlock(LHSBlock);
1326 incrementProfileCounter(CondOp);
1327 {
1328 ApplyDebugLocation DL(*this, Cond);
1329 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1330 LHSScaledTrueCount);
1331 }
1332 cond.end(*this);
1333
1334 cond.begin(*this);
1335 EmitBlock(RHSBlock);
1336 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1337 TrueCount - LHSScaledTrueCount);
1338 cond.end(*this);
1339
1340 return;
1341 }
1342
1343 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1344 // Conditional operator handling can give us a throw expression as a
1345 // condition for a case like:
1346 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1347 // Fold this to:
1348 // br(c, throw x, br(y, t, f))
1349 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1350 return;
1351 }
1352
1353 // If the branch has a condition wrapped by __builtin_unpredictable,
1354 // create metadata that specifies that the branch is unpredictable.
1355 // Don't bother if not optimizing because that metadata would not be used.
1356 llvm::MDNode *Unpredictable = nullptr;
1357 auto *Call = dyn_cast<CallExpr>(Cond);
1358 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1359 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1360 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1361 llvm::MDBuilder MDHelper(getLLVMContext());
1362 Unpredictable = MDHelper.createUnpredictable();
1363 }
1364 }
1365
1366 // Create branch weights based on the number of times we get here and the
1367 // number of times the condition should be true.
1368 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1369 llvm::MDNode *Weights =
1370 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1371
1372 // Emit the code with the fully general case.
1373 llvm::Value *CondV;
1374 {
1375 ApplyDebugLocation DL(*this, Cond);
1376 CondV = EvaluateExprAsBool(Cond);
1377 }
1378 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1379 }
1380
1381 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1382 /// specified stmt yet.
ErrorUnsupported(const Stmt * S,const char * Type)1383 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1384 CGM.ErrorUnsupported(S, Type);
1385 }
1386
1387 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1388 /// variable-length array whose elements have a non-zero bit-pattern.
1389 ///
1390 /// \param baseType the inner-most element type of the array
1391 /// \param src - a char* pointing to the bit-pattern for a single
1392 /// base element of the array
1393 /// \param sizeInChars - the total size of the VLA, in chars
emitNonZeroVLAInit(CodeGenFunction & CGF,QualType baseType,Address dest,Address src,llvm::Value * sizeInChars)1394 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1395 Address dest, Address src,
1396 llvm::Value *sizeInChars) {
1397 CGBuilderTy &Builder = CGF.Builder;
1398
1399 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1400 llvm::Value *baseSizeInChars
1401 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1402
1403 Address begin =
1404 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1405 llvm::Value *end =
1406 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1407
1408 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1409 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1410 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1411
1412 // Make a loop over the VLA. C99 guarantees that the VLA element
1413 // count must be nonzero.
1414 CGF.EmitBlock(loopBB);
1415
1416 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1417 cur->addIncoming(begin.getPointer(), originBB);
1418
1419 CharUnits curAlign =
1420 dest.getAlignment().alignmentOfArrayElement(baseSize);
1421
1422 // memcpy the individual element bit-pattern.
1423 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1424 /*volatile*/ false);
1425
1426 // Go to the next element.
1427 llvm::Value *next =
1428 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1429
1430 // Leave if that's the end of the VLA.
1431 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1432 Builder.CreateCondBr(done, contBB, loopBB);
1433 cur->addIncoming(next, loopBB);
1434
1435 CGF.EmitBlock(contBB);
1436 }
1437
1438 void
EmitNullInitialization(Address DestPtr,QualType Ty)1439 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1440 // Ignore empty classes in C++.
1441 if (getLangOpts().CPlusPlus) {
1442 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1443 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1444 return;
1445 }
1446 }
1447
1448 // Cast the dest ptr to the appropriate i8 pointer type.
1449 if (DestPtr.getElementType() != Int8Ty)
1450 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1451
1452 // Get size and alignment info for this aggregate.
1453 CharUnits size = getContext().getTypeSizeInChars(Ty);
1454
1455 llvm::Value *SizeVal;
1456 const VariableArrayType *vla;
1457
1458 // Don't bother emitting a zero-byte memset.
1459 if (size.isZero()) {
1460 // But note that getTypeInfo returns 0 for a VLA.
1461 if (const VariableArrayType *vlaType =
1462 dyn_cast_or_null<VariableArrayType>(
1463 getContext().getAsArrayType(Ty))) {
1464 QualType eltType;
1465 llvm::Value *numElts;
1466 std::tie(numElts, eltType) = getVLASize(vlaType);
1467
1468 SizeVal = numElts;
1469 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1470 if (!eltSize.isOne())
1471 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1472 vla = vlaType;
1473 } else {
1474 return;
1475 }
1476 } else {
1477 SizeVal = CGM.getSize(size);
1478 vla = nullptr;
1479 }
1480
1481 // If the type contains a pointer to data member we can't memset it to zero.
1482 // Instead, create a null constant and copy it to the destination.
1483 // TODO: there are other patterns besides zero that we can usefully memset,
1484 // like -1, which happens to be the pattern used by member-pointers.
1485 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1486 // For a VLA, emit a single element, then splat that over the VLA.
1487 if (vla) Ty = getContext().getBaseElementType(vla);
1488
1489 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1490
1491 llvm::GlobalVariable *NullVariable =
1492 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1493 /*isConstant=*/true,
1494 llvm::GlobalVariable::PrivateLinkage,
1495 NullConstant, Twine());
1496 CharUnits NullAlign = DestPtr.getAlignment();
1497 NullVariable->setAlignment(NullAlign.getQuantity());
1498 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1499 NullAlign);
1500
1501 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1502
1503 // Get and call the appropriate llvm.memcpy overload.
1504 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1505 return;
1506 }
1507
1508 // Otherwise, just memset the whole thing to zero. This is legal
1509 // because in LLVM, all default initializers (other than the ones we just
1510 // handled above) are guaranteed to have a bit pattern of all zeros.
1511 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1512 }
1513
GetAddrOfLabel(const LabelDecl * L)1514 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1515 // Make sure that there is a block for the indirect goto.
1516 if (!IndirectBranch)
1517 GetIndirectGotoBlock();
1518
1519 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1520
1521 // Make sure the indirect branch includes all of the address-taken blocks.
1522 IndirectBranch->addDestination(BB);
1523 return llvm::BlockAddress::get(CurFn, BB);
1524 }
1525
GetIndirectGotoBlock()1526 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1527 // If we already made the indirect branch for indirect goto, return its block.
1528 if (IndirectBranch) return IndirectBranch->getParent();
1529
1530 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1531
1532 // Create the PHI node that indirect gotos will add entries to.
1533 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1534 "indirect.goto.dest");
1535
1536 // Create the indirect branch instruction.
1537 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1538 return IndirectBranch->getParent();
1539 }
1540
1541 /// Computes the length of an array in elements, as well as the base
1542 /// element type and a properly-typed first element pointer.
emitArrayLength(const ArrayType * origArrayType,QualType & baseType,Address & addr)1543 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1544 QualType &baseType,
1545 Address &addr) {
1546 const ArrayType *arrayType = origArrayType;
1547
1548 // If it's a VLA, we have to load the stored size. Note that
1549 // this is the size of the VLA in bytes, not its size in elements.
1550 llvm::Value *numVLAElements = nullptr;
1551 if (isa<VariableArrayType>(arrayType)) {
1552 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1553
1554 // Walk into all VLAs. This doesn't require changes to addr,
1555 // which has type T* where T is the first non-VLA element type.
1556 do {
1557 QualType elementType = arrayType->getElementType();
1558 arrayType = getContext().getAsArrayType(elementType);
1559
1560 // If we only have VLA components, 'addr' requires no adjustment.
1561 if (!arrayType) {
1562 baseType = elementType;
1563 return numVLAElements;
1564 }
1565 } while (isa<VariableArrayType>(arrayType));
1566
1567 // We get out here only if we find a constant array type
1568 // inside the VLA.
1569 }
1570
1571 // We have some number of constant-length arrays, so addr should
1572 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1573 // down to the first element of addr.
1574 SmallVector<llvm::Value*, 8> gepIndices;
1575
1576 // GEP down to the array type.
1577 llvm::ConstantInt *zero = Builder.getInt32(0);
1578 gepIndices.push_back(zero);
1579
1580 uint64_t countFromCLAs = 1;
1581 QualType eltType;
1582
1583 llvm::ArrayType *llvmArrayType =
1584 dyn_cast<llvm::ArrayType>(addr.getElementType());
1585 while (llvmArrayType) {
1586 assert(isa<ConstantArrayType>(arrayType));
1587 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1588 == llvmArrayType->getNumElements());
1589
1590 gepIndices.push_back(zero);
1591 countFromCLAs *= llvmArrayType->getNumElements();
1592 eltType = arrayType->getElementType();
1593
1594 llvmArrayType =
1595 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1596 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1597 assert((!llvmArrayType || arrayType) &&
1598 "LLVM and Clang types are out-of-synch");
1599 }
1600
1601 if (arrayType) {
1602 // From this point onwards, the Clang array type has been emitted
1603 // as some other type (probably a packed struct). Compute the array
1604 // size, and just emit the 'begin' expression as a bitcast.
1605 while (arrayType) {
1606 countFromCLAs *=
1607 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1608 eltType = arrayType->getElementType();
1609 arrayType = getContext().getAsArrayType(eltType);
1610 }
1611
1612 llvm::Type *baseType = ConvertType(eltType);
1613 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1614 } else {
1615 // Create the actual GEP.
1616 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1617 gepIndices, "array.begin"),
1618 addr.getAlignment());
1619 }
1620
1621 baseType = eltType;
1622
1623 llvm::Value *numElements
1624 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1625
1626 // If we had any VLA dimensions, factor them in.
1627 if (numVLAElements)
1628 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1629
1630 return numElements;
1631 }
1632
1633 std::pair<llvm::Value*, QualType>
getVLASize(QualType type)1634 CodeGenFunction::getVLASize(QualType type) {
1635 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1636 assert(vla && "type was not a variable array type!");
1637 return getVLASize(vla);
1638 }
1639
1640 std::pair<llvm::Value*, QualType>
getVLASize(const VariableArrayType * type)1641 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1642 // The number of elements so far; always size_t.
1643 llvm::Value *numElements = nullptr;
1644
1645 QualType elementType;
1646 do {
1647 elementType = type->getElementType();
1648 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1649 assert(vlaSize && "no size for VLA!");
1650 assert(vlaSize->getType() == SizeTy);
1651
1652 if (!numElements) {
1653 numElements = vlaSize;
1654 } else {
1655 // It's undefined behavior if this wraps around, so mark it that way.
1656 // FIXME: Teach -fsanitize=undefined to trap this.
1657 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1658 }
1659 } while ((type = getContext().getAsVariableArrayType(elementType)));
1660
1661 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1662 }
1663
EmitVariablyModifiedType(QualType type)1664 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1665 assert(type->isVariablyModifiedType() &&
1666 "Must pass variably modified type to EmitVLASizes!");
1667
1668 EnsureInsertPoint();
1669
1670 // We're going to walk down into the type and look for VLA
1671 // expressions.
1672 do {
1673 assert(type->isVariablyModifiedType());
1674
1675 const Type *ty = type.getTypePtr();
1676 switch (ty->getTypeClass()) {
1677
1678 #define TYPE(Class, Base)
1679 #define ABSTRACT_TYPE(Class, Base)
1680 #define NON_CANONICAL_TYPE(Class, Base)
1681 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1682 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1683 #include "clang/AST/TypeNodes.def"
1684 llvm_unreachable("unexpected dependent type!");
1685
1686 // These types are never variably-modified.
1687 case Type::Builtin:
1688 case Type::Complex:
1689 case Type::Vector:
1690 case Type::ExtVector:
1691 case Type::Record:
1692 case Type::Enum:
1693 case Type::Elaborated:
1694 case Type::TemplateSpecialization:
1695 case Type::ObjCObject:
1696 case Type::ObjCInterface:
1697 case Type::ObjCObjectPointer:
1698 llvm_unreachable("type class is never variably-modified!");
1699
1700 case Type::Adjusted:
1701 type = cast<AdjustedType>(ty)->getAdjustedType();
1702 break;
1703
1704 case Type::Decayed:
1705 type = cast<DecayedType>(ty)->getPointeeType();
1706 break;
1707
1708 case Type::Pointer:
1709 type = cast<PointerType>(ty)->getPointeeType();
1710 break;
1711
1712 case Type::BlockPointer:
1713 type = cast<BlockPointerType>(ty)->getPointeeType();
1714 break;
1715
1716 case Type::LValueReference:
1717 case Type::RValueReference:
1718 type = cast<ReferenceType>(ty)->getPointeeType();
1719 break;
1720
1721 case Type::MemberPointer:
1722 type = cast<MemberPointerType>(ty)->getPointeeType();
1723 break;
1724
1725 case Type::ConstantArray:
1726 case Type::IncompleteArray:
1727 // Losing element qualification here is fine.
1728 type = cast<ArrayType>(ty)->getElementType();
1729 break;
1730
1731 case Type::VariableArray: {
1732 // Losing element qualification here is fine.
1733 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1734
1735 // Unknown size indication requires no size computation.
1736 // Otherwise, evaluate and record it.
1737 if (const Expr *size = vat->getSizeExpr()) {
1738 // It's possible that we might have emitted this already,
1739 // e.g. with a typedef and a pointer to it.
1740 llvm::Value *&entry = VLASizeMap[size];
1741 if (!entry) {
1742 llvm::Value *Size = EmitScalarExpr(size);
1743
1744 // C11 6.7.6.2p5:
1745 // If the size is an expression that is not an integer constant
1746 // expression [...] each time it is evaluated it shall have a value
1747 // greater than zero.
1748 if (SanOpts.has(SanitizerKind::VLABound) &&
1749 size->getType()->isSignedIntegerType()) {
1750 SanitizerScope SanScope(this);
1751 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1752 llvm::Constant *StaticArgs[] = {
1753 EmitCheckSourceLocation(size->getLocStart()),
1754 EmitCheckTypeDescriptor(size->getType())
1755 };
1756 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1757 SanitizerKind::VLABound),
1758 "vla_bound_not_positive", StaticArgs, Size);
1759 }
1760
1761 // Always zexting here would be wrong if it weren't
1762 // undefined behavior to have a negative bound.
1763 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1764 }
1765 }
1766 type = vat->getElementType();
1767 break;
1768 }
1769
1770 case Type::FunctionProto:
1771 case Type::FunctionNoProto:
1772 type = cast<FunctionType>(ty)->getReturnType();
1773 break;
1774
1775 case Type::Paren:
1776 case Type::TypeOf:
1777 case Type::UnaryTransform:
1778 case Type::Attributed:
1779 case Type::SubstTemplateTypeParm:
1780 case Type::PackExpansion:
1781 // Keep walking after single level desugaring.
1782 type = type.getSingleStepDesugaredType(getContext());
1783 break;
1784
1785 case Type::Typedef:
1786 case Type::Decltype:
1787 case Type::Auto:
1788 // Stop walking: nothing to do.
1789 return;
1790
1791 case Type::TypeOfExpr:
1792 // Stop walking: emit typeof expression.
1793 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1794 return;
1795
1796 case Type::Atomic:
1797 type = cast<AtomicType>(ty)->getValueType();
1798 break;
1799
1800 case Type::Pipe:
1801 type = cast<PipeType>(ty)->getElementType();
1802 break;
1803 }
1804 } while (type->isVariablyModifiedType());
1805 }
1806
EmitVAListRef(const Expr * E)1807 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1808 if (getContext().getBuiltinVaListType()->isArrayType())
1809 return EmitPointerWithAlignment(E);
1810 return EmitLValue(E).getAddress();
1811 }
1812
EmitMSVAListRef(const Expr * E)1813 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1814 return EmitLValue(E).getAddress();
1815 }
1816
EmitDeclRefExprDbgValue(const DeclRefExpr * E,llvm::Constant * Init)1817 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1818 llvm::Constant *Init) {
1819 assert (Init && "Invalid DeclRefExpr initializer!");
1820 if (CGDebugInfo *Dbg = getDebugInfo())
1821 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
1822 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1823 }
1824
1825 CodeGenFunction::PeepholeProtection
protectFromPeepholes(RValue rvalue)1826 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1827 // At the moment, the only aggressive peephole we do in IR gen
1828 // is trunc(zext) folding, but if we add more, we can easily
1829 // extend this protection.
1830
1831 if (!rvalue.isScalar()) return PeepholeProtection();
1832 llvm::Value *value = rvalue.getScalarVal();
1833 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1834
1835 // Just make an extra bitcast.
1836 assert(HaveInsertPoint());
1837 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1838 Builder.GetInsertBlock());
1839
1840 PeepholeProtection protection;
1841 protection.Inst = inst;
1842 return protection;
1843 }
1844
unprotectFromPeepholes(PeepholeProtection protection)1845 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1846 if (!protection.Inst) return;
1847
1848 // In theory, we could try to duplicate the peepholes now, but whatever.
1849 protection.Inst->eraseFromParent();
1850 }
1851
EmitAnnotationCall(llvm::Value * AnnotationFn,llvm::Value * AnnotatedVal,StringRef AnnotationStr,SourceLocation Location)1852 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1853 llvm::Value *AnnotatedVal,
1854 StringRef AnnotationStr,
1855 SourceLocation Location) {
1856 llvm::Value *Args[4] = {
1857 AnnotatedVal,
1858 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1859 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1860 CGM.EmitAnnotationLineNo(Location)
1861 };
1862 return Builder.CreateCall(AnnotationFn, Args);
1863 }
1864
EmitVarAnnotations(const VarDecl * D,llvm::Value * V)1865 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1866 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1867 // FIXME We create a new bitcast for every annotation because that's what
1868 // llvm-gcc was doing.
1869 for (const auto *I : D->specific_attrs<AnnotateAttr>())
1870 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1871 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1872 I->getAnnotation(), D->getLocation());
1873 }
1874
EmitFieldAnnotations(const FieldDecl * D,Address Addr)1875 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1876 Address Addr) {
1877 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1878 llvm::Value *V = Addr.getPointer();
1879 llvm::Type *VTy = V->getType();
1880 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1881 CGM.Int8PtrTy);
1882
1883 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1884 // FIXME Always emit the cast inst so we can differentiate between
1885 // annotation on the first field of a struct and annotation on the struct
1886 // itself.
1887 if (VTy != CGM.Int8PtrTy)
1888 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1889 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1890 V = Builder.CreateBitCast(V, VTy);
1891 }
1892
1893 return Address(V, Addr.getAlignment());
1894 }
1895
~CGCapturedStmtInfo()1896 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
1897
SanitizerScope(CodeGenFunction * CGF)1898 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
1899 : CGF(CGF) {
1900 assert(!CGF->IsSanitizerScope);
1901 CGF->IsSanitizerScope = true;
1902 }
1903
~SanitizerScope()1904 CodeGenFunction::SanitizerScope::~SanitizerScope() {
1905 CGF->IsSanitizerScope = false;
1906 }
1907
InsertHelper(llvm::Instruction * I,const llvm::Twine & Name,llvm::BasicBlock * BB,llvm::BasicBlock::iterator InsertPt) const1908 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
1909 const llvm::Twine &Name,
1910 llvm::BasicBlock *BB,
1911 llvm::BasicBlock::iterator InsertPt) const {
1912 LoopStack.InsertHelper(I);
1913 if (IsSanitizerScope)
1914 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
1915 }
1916
InsertHelper(llvm::Instruction * I,const llvm::Twine & Name,llvm::BasicBlock * BB,llvm::BasicBlock::iterator InsertPt) const1917 void CGBuilderInserter::InsertHelper(
1918 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1919 llvm::BasicBlock::iterator InsertPt) const {
1920 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
1921 if (CGF)
1922 CGF->InsertHelper(I, Name, BB, InsertPt);
1923 }
1924
hasRequiredFeatures(const SmallVectorImpl<StringRef> & ReqFeatures,CodeGenModule & CGM,const FunctionDecl * FD,std::string & FirstMissing)1925 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
1926 CodeGenModule &CGM, const FunctionDecl *FD,
1927 std::string &FirstMissing) {
1928 // If there aren't any required features listed then go ahead and return.
1929 if (ReqFeatures.empty())
1930 return false;
1931
1932 // Now build up the set of caller features and verify that all the required
1933 // features are there.
1934 llvm::StringMap<bool> CallerFeatureMap;
1935 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
1936
1937 // If we have at least one of the features in the feature list return
1938 // true, otherwise return false.
1939 return std::all_of(
1940 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
1941 SmallVector<StringRef, 1> OrFeatures;
1942 Feature.split(OrFeatures, "|");
1943 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
1944 [&](StringRef Feature) {
1945 if (!CallerFeatureMap.lookup(Feature)) {
1946 FirstMissing = Feature.str();
1947 return false;
1948 }
1949 return true;
1950 });
1951 });
1952 }
1953
1954 // Emits an error if we don't have a valid set of target features for the
1955 // called function.
checkTargetFeatures(const CallExpr * E,const FunctionDecl * TargetDecl)1956 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
1957 const FunctionDecl *TargetDecl) {
1958 // Early exit if this is an indirect call.
1959 if (!TargetDecl)
1960 return;
1961
1962 // Get the current enclosing function if it exists. If it doesn't
1963 // we can't check the target features anyhow.
1964 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
1965 if (!FD)
1966 return;
1967
1968 // Grab the required features for the call. For a builtin this is listed in
1969 // the td file with the default cpu, for an always_inline function this is any
1970 // listed cpu and any listed features.
1971 unsigned BuiltinID = TargetDecl->getBuiltinID();
1972 std::string MissingFeature;
1973 if (BuiltinID) {
1974 SmallVector<StringRef, 1> ReqFeatures;
1975 const char *FeatureList =
1976 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1977 // Return if the builtin doesn't have any required features.
1978 if (!FeatureList || StringRef(FeatureList) == "")
1979 return;
1980 StringRef(FeatureList).split(ReqFeatures, ",");
1981 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1982 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
1983 << TargetDecl->getDeclName()
1984 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1985
1986 } else if (TargetDecl->hasAttr<TargetAttr>()) {
1987 // Get the required features for the callee.
1988 SmallVector<StringRef, 1> ReqFeatures;
1989 llvm::StringMap<bool> CalleeFeatureMap;
1990 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
1991 for (const auto &F : CalleeFeatureMap) {
1992 // Only positive features are "required".
1993 if (F.getValue())
1994 ReqFeatures.push_back(F.getKey());
1995 }
1996 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1997 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
1998 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
1999 }
2000 }
2001
EmitSanitizerStatReport(llvm::SanitizerStatKind SSK)2002 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2003 if (!CGM.getCodeGenOpts().SanitizeStats)
2004 return;
2005
2006 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2007 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2008 CGM.getSanStats().create(IRB, SSK);
2009 }
2010