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 return TEK_Scalar;
199
200 // Complexes.
201 case Type::Complex:
202 return TEK_Complex;
203
204 // Arrays, records, and Objective-C objects.
205 case Type::ConstantArray:
206 case Type::IncompleteArray:
207 case Type::VariableArray:
208 case Type::Record:
209 case Type::ObjCObject:
210 case Type::ObjCInterface:
211 return TEK_Aggregate;
212
213 // We operate on atomic values according to their underlying type.
214 case Type::Atomic:
215 type = cast<AtomicType>(type)->getValueType();
216 continue;
217 }
218 llvm_unreachable("unknown type kind!");
219 }
220 }
221
EmitReturnBlock()222 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
223 // For cleanliness, we try to avoid emitting the return block for
224 // simple cases.
225 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
226
227 if (CurBB) {
228 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
229
230 // We have a valid insert point, reuse it if it is empty or there are no
231 // explicit jumps to the return block.
232 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
233 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
234 delete ReturnBlock.getBlock();
235 } else
236 EmitBlock(ReturnBlock.getBlock());
237 return llvm::DebugLoc();
238 }
239
240 // Otherwise, if the return block is the target of a single direct
241 // branch then we can just put the code in that block instead. This
242 // cleans up functions which started with a unified return block.
243 if (ReturnBlock.getBlock()->hasOneUse()) {
244 llvm::BranchInst *BI =
245 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
246 if (BI && BI->isUnconditional() &&
247 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
248 // Record/return the DebugLoc of the simple 'return' expression to be used
249 // later by the actual 'ret' instruction.
250 llvm::DebugLoc Loc = BI->getDebugLoc();
251 Builder.SetInsertPoint(BI->getParent());
252 BI->eraseFromParent();
253 delete ReturnBlock.getBlock();
254 return Loc;
255 }
256 }
257
258 // FIXME: We are at an unreachable point, there is no reason to emit the block
259 // unless it has uses. However, we still need a place to put the debug
260 // region.end for now.
261
262 EmitBlock(ReturnBlock.getBlock());
263 return llvm::DebugLoc();
264 }
265
EmitIfUsed(CodeGenFunction & CGF,llvm::BasicBlock * BB)266 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
267 if (!BB) return;
268 if (!BB->use_empty())
269 return CGF.CurFn->getBasicBlockList().push_back(BB);
270 delete BB;
271 }
272
FinishFunction(SourceLocation EndLoc)273 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
274 assert(BreakContinueStack.empty() &&
275 "mismatched push/pop in break/continue stack!");
276
277 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
278 && NumSimpleReturnExprs == NumReturnExprs
279 && ReturnBlock.getBlock()->use_empty();
280 // Usually the return expression is evaluated before the cleanup
281 // code. If the function contains only a simple return statement,
282 // such as a constant, the location before the cleanup code becomes
283 // the last useful breakpoint in the function, because the simple
284 // return expression will be evaluated after the cleanup code. To be
285 // safe, set the debug location for cleanup code to the location of
286 // the return statement. Otherwise the cleanup code should be at the
287 // end of the function's lexical scope.
288 //
289 // If there are multiple branches to the return block, the branch
290 // instructions will get the location of the return statements and
291 // all will be fine.
292 if (CGDebugInfo *DI = getDebugInfo()) {
293 if (OnlySimpleReturnStmts)
294 DI->EmitLocation(Builder, LastStopPoint);
295 else
296 DI->EmitLocation(Builder, EndLoc);
297 }
298
299 // Pop any cleanups that might have been associated with the
300 // parameters. Do this in whatever block we're currently in; it's
301 // important to do this before we enter the return block or return
302 // edges will be *really* confused.
303 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
304 bool HasOnlyLifetimeMarkers =
305 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
306 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
307 if (HasCleanups) {
308 // Make sure the line table doesn't jump back into the body for
309 // the ret after it's been at EndLoc.
310 if (CGDebugInfo *DI = getDebugInfo())
311 if (OnlySimpleReturnStmts)
312 DI->EmitLocation(Builder, EndLoc);
313
314 PopCleanupBlocks(PrologueCleanupDepth);
315 }
316
317 // Emit function epilog (to return).
318 llvm::DebugLoc Loc = EmitReturnBlock();
319
320 if (ShouldInstrumentFunction())
321 EmitFunctionInstrumentation("__cyg_profile_func_exit");
322
323 // Emit debug descriptor for function end.
324 if (CGDebugInfo *DI = getDebugInfo())
325 DI->EmitFunctionEnd(Builder);
326
327 // Reset the debug location to that of the simple 'return' expression, if any
328 // rather than that of the end of the function's scope '}'.
329 ApplyDebugLocation AL(*this, Loc);
330 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
331 EmitEndEHSpec(CurCodeDecl);
332
333 assert(EHStack.empty() &&
334 "did not remove all scopes from cleanup stack!");
335
336 // If someone did an indirect goto, emit the indirect goto block at the end of
337 // the function.
338 if (IndirectBranch) {
339 EmitBlock(IndirectBranch->getParent());
340 Builder.ClearInsertionPoint();
341 }
342
343 // If some of our locals escaped, insert a call to llvm.localescape in the
344 // entry block.
345 if (!EscapedLocals.empty()) {
346 // Invert the map from local to index into a simple vector. There should be
347 // no holes.
348 SmallVector<llvm::Value *, 4> EscapeArgs;
349 EscapeArgs.resize(EscapedLocals.size());
350 for (auto &Pair : EscapedLocals)
351 EscapeArgs[Pair.second] = Pair.first;
352 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
353 &CGM.getModule(), llvm::Intrinsic::localescape);
354 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
355 }
356
357 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
358 llvm::Instruction *Ptr = AllocaInsertPt;
359 AllocaInsertPt = nullptr;
360 Ptr->eraseFromParent();
361
362 // If someone took the address of a label but never did an indirect goto, we
363 // made a zero entry PHI node, which is illegal, zap it now.
364 if (IndirectBranch) {
365 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
366 if (PN->getNumIncomingValues() == 0) {
367 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
368 PN->eraseFromParent();
369 }
370 }
371
372 EmitIfUsed(*this, EHResumeBlock);
373 EmitIfUsed(*this, TerminateLandingPad);
374 EmitIfUsed(*this, TerminateHandler);
375 EmitIfUsed(*this, UnreachableBlock);
376
377 if (CGM.getCodeGenOpts().EmitDeclMetadata)
378 EmitDeclMetadata();
379
380 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
381 I = DeferredReplacements.begin(),
382 E = DeferredReplacements.end();
383 I != E; ++I) {
384 I->first->replaceAllUsesWith(I->second);
385 I->first->eraseFromParent();
386 }
387 }
388
389 /// ShouldInstrumentFunction - Return true if the current function should be
390 /// instrumented with __cyg_profile_func_* calls
ShouldInstrumentFunction()391 bool CodeGenFunction::ShouldInstrumentFunction() {
392 if (!CGM.getCodeGenOpts().InstrumentFunctions)
393 return false;
394 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
395 return false;
396 return true;
397 }
398
399 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
400 /// instrumentation function with the current function and the call site, if
401 /// function instrumentation is enabled.
EmitFunctionInstrumentation(const char * Fn)402 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
403 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
404 llvm::PointerType *PointerTy = Int8PtrTy;
405 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
406 llvm::FunctionType *FunctionTy =
407 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
408
409 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
410 llvm::CallInst *CallSite = Builder.CreateCall(
411 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
412 llvm::ConstantInt::get(Int32Ty, 0),
413 "callsite");
414
415 llvm::Value *args[] = {
416 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
417 CallSite
418 };
419
420 EmitNounwindRuntimeCall(F, args);
421 }
422
EmitMCountInstrumentation()423 void CodeGenFunction::EmitMCountInstrumentation() {
424 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
425
426 llvm::Constant *MCountFn =
427 CGM.CreateRuntimeFunction(FTy, getTarget().getMCountName());
428 EmitNounwindRuntimeCall(MCountFn);
429 }
430
431 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
432 // information in the program executable. The argument information stored
433 // includes the argument name, its type, the address and access qualifiers used.
GenOpenCLArgMetadata(const FunctionDecl * FD,llvm::Function * Fn,CodeGenModule & CGM,llvm::LLVMContext & Context,SmallVector<llvm::Metadata *,5> & kernelMDArgs,CGBuilderTy & Builder,ASTContext & ASTCtx)434 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
435 CodeGenModule &CGM, llvm::LLVMContext &Context,
436 SmallVector<llvm::Metadata *, 5> &kernelMDArgs,
437 CGBuilderTy &Builder, ASTContext &ASTCtx) {
438 // Create MDNodes that represent the kernel arg metadata.
439 // Each MDNode is a list in the form of "key", N number of values which is
440 // the same number of values as their are kernel arguments.
441
442 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
443
444 // MDNode for the kernel argument address space qualifiers.
445 SmallVector<llvm::Metadata *, 8> addressQuals;
446 addressQuals.push_back(llvm::MDString::get(Context, "kernel_arg_addr_space"));
447
448 // MDNode for the kernel argument access qualifiers (images only).
449 SmallVector<llvm::Metadata *, 8> accessQuals;
450 accessQuals.push_back(llvm::MDString::get(Context, "kernel_arg_access_qual"));
451
452 // MDNode for the kernel argument type names.
453 SmallVector<llvm::Metadata *, 8> argTypeNames;
454 argTypeNames.push_back(llvm::MDString::get(Context, "kernel_arg_type"));
455
456 // MDNode for the kernel argument base type names.
457 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
458 argBaseTypeNames.push_back(
459 llvm::MDString::get(Context, "kernel_arg_base_type"));
460
461 // MDNode for the kernel argument type qualifiers.
462 SmallVector<llvm::Metadata *, 8> argTypeQuals;
463 argTypeQuals.push_back(llvm::MDString::get(Context, "kernel_arg_type_qual"));
464
465 // MDNode for the kernel argument names.
466 SmallVector<llvm::Metadata *, 8> argNames;
467 argNames.push_back(llvm::MDString::get(Context, "kernel_arg_name"));
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 if (ty->isImageType())
515 AddrSpc =
516 CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
517
518 addressQuals.push_back(
519 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
520
521 // Get argument type name.
522 std::string typeName = ty.getUnqualifiedType().getAsString(Policy);
523
524 // Turn "unsigned type" to "utype"
525 std::string::size_type pos = typeName.find("unsigned");
526 if (ty.isCanonical() && pos != std::string::npos)
527 typeName.erase(pos+1, 8);
528
529 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
530
531 std::string baseTypeName =
532 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
533
534 // Turn "unsigned type" to "utype"
535 pos = baseTypeName.find("unsigned");
536 if (pos != std::string::npos)
537 baseTypeName.erase(pos+1, 8);
538
539 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
540
541 // Get argument type qualifiers:
542 if (ty.isConstQualified())
543 typeQuals = "const";
544 if (ty.isVolatileQualified())
545 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
546 }
547
548 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
549
550 // Get image access qualifier:
551 if (ty->isImageType()) {
552 const OpenCLImageAccessAttr *A = parm->getAttr<OpenCLImageAccessAttr>();
553 if (A && A->isWriteOnly())
554 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
555 else
556 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
557 // FIXME: what about read_write?
558 } else
559 accessQuals.push_back(llvm::MDString::get(Context, "none"));
560
561 // Get argument name.
562 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
563 }
564
565 kernelMDArgs.push_back(llvm::MDNode::get(Context, addressQuals));
566 kernelMDArgs.push_back(llvm::MDNode::get(Context, accessQuals));
567 kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeNames));
568 kernelMDArgs.push_back(llvm::MDNode::get(Context, argBaseTypeNames));
569 kernelMDArgs.push_back(llvm::MDNode::get(Context, argTypeQuals));
570 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
571 kernelMDArgs.push_back(llvm::MDNode::get(Context, argNames));
572 }
573
EmitOpenCLKernelMetadata(const FunctionDecl * FD,llvm::Function * Fn)574 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
575 llvm::Function *Fn)
576 {
577 if (!FD->hasAttr<OpenCLKernelAttr>())
578 return;
579
580 llvm::LLVMContext &Context = getLLVMContext();
581
582 SmallVector<llvm::Metadata *, 5> kernelMDArgs;
583 kernelMDArgs.push_back(llvm::ConstantAsMetadata::get(Fn));
584
585 GenOpenCLArgMetadata(FD, Fn, CGM, Context, kernelMDArgs, Builder,
586 getContext());
587
588 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
589 QualType hintQTy = A->getTypeHint();
590 const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
591 bool isSignedInteger =
592 hintQTy->isSignedIntegerType() ||
593 (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
594 llvm::Metadata *attrMDArgs[] = {
595 llvm::MDString::get(Context, "vec_type_hint"),
596 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
597 CGM.getTypes().ConvertType(A->getTypeHint()))),
598 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
599 llvm::IntegerType::get(Context, 32),
600 llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
601 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
602 }
603
604 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
605 llvm::Metadata *attrMDArgs[] = {
606 llvm::MDString::get(Context, "work_group_size_hint"),
607 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
608 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
609 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
610 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
611 }
612
613 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
614 llvm::Metadata *attrMDArgs[] = {
615 llvm::MDString::get(Context, "reqd_work_group_size"),
616 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
617 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
618 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
619 kernelMDArgs.push_back(llvm::MDNode::get(Context, attrMDArgs));
620 }
621
622 llvm::MDNode *kernelMDNode = llvm::MDNode::get(Context, kernelMDArgs);
623 llvm::NamedMDNode *OpenCLKernelMetadata =
624 CGM.getModule().getOrInsertNamedMetadata("opencl.kernels");
625 OpenCLKernelMetadata->addOperand(kernelMDNode);
626 }
627
628 /// Determine whether the function F ends with a return stmt.
endsWithReturn(const Decl * F)629 static bool endsWithReturn(const Decl* F) {
630 const Stmt *Body = nullptr;
631 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
632 Body = FD->getBody();
633 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
634 Body = OMD->getBody();
635
636 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
637 auto LastStmt = CS->body_rbegin();
638 if (LastStmt != CS->body_rend())
639 return isa<ReturnStmt>(*LastStmt);
640 }
641 return false;
642 }
643
StartFunction(GlobalDecl GD,QualType RetTy,llvm::Function * Fn,const CGFunctionInfo & FnInfo,const FunctionArgList & Args,SourceLocation Loc,SourceLocation StartLoc)644 void CodeGenFunction::StartFunction(GlobalDecl GD,
645 QualType RetTy,
646 llvm::Function *Fn,
647 const CGFunctionInfo &FnInfo,
648 const FunctionArgList &Args,
649 SourceLocation Loc,
650 SourceLocation StartLoc) {
651 assert(!CurFn &&
652 "Do not use a CodeGenFunction object for more than one function");
653
654 const Decl *D = GD.getDecl();
655
656 DidCallStackSave = false;
657 CurCodeDecl = D;
658 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
659 FnRetTy = RetTy;
660 CurFn = Fn;
661 CurFnInfo = &FnInfo;
662 assert(CurFn->isDeclaration() && "Function already has body?");
663
664 if (CGM.isInSanitizerBlacklist(Fn, Loc))
665 SanOpts.clear();
666
667 if (D) {
668 // Apply the no_sanitize* attributes to SanOpts.
669 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
670 SanOpts.Mask &= ~Attr->getMask();
671 }
672
673 // Apply sanitizer attributes to the function.
674 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
675 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
676 if (SanOpts.has(SanitizerKind::Thread))
677 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
678 if (SanOpts.has(SanitizerKind::Memory))
679 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
680 if (SanOpts.has(SanitizerKind::SafeStack))
681 Fn->addFnAttr(llvm::Attribute::SafeStack);
682
683 // Pass inline keyword to optimizer if it appears explicitly on any
684 // declaration. Also, in the case of -fno-inline attach NoInline
685 // attribute to all function that are not marked AlwaysInline.
686 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
687 if (!CGM.getCodeGenOpts().NoInline) {
688 for (auto RI : FD->redecls())
689 if (RI->isInlineSpecified()) {
690 Fn->addFnAttr(llvm::Attribute::InlineHint);
691 break;
692 }
693 } else if (!FD->hasAttr<AlwaysInlineAttr>())
694 Fn->addFnAttr(llvm::Attribute::NoInline);
695 }
696
697 if (getLangOpts().OpenCL) {
698 // Add metadata for a kernel function.
699 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
700 EmitOpenCLKernelMetadata(FD, Fn);
701 }
702
703 // If we are checking function types, emit a function type signature as
704 // prologue data.
705 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
706 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
707 if (llvm::Constant *PrologueSig =
708 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
709 llvm::Constant *FTRTTIConst =
710 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
711 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
712 llvm::Constant *PrologueStructConst =
713 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
714 Fn->setPrologueData(PrologueStructConst);
715 }
716 }
717 }
718
719 // If we're in C++ mode and the function name is "main", it is guaranteed
720 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
721 // used within a program").
722 if (getLangOpts().CPlusPlus)
723 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
724 if (FD->isMain())
725 Fn->addFnAttr(llvm::Attribute::NoRecurse);
726
727 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
728
729 // Create a marker to make it easy to insert allocas into the entryblock
730 // later. Don't create this with the builder, because we don't want it
731 // folded.
732 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
733 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "", EntryBB);
734 if (Builder.isNamePreserving())
735 AllocaInsertPt->setName("allocapt");
736
737 ReturnBlock = getJumpDestInCurrentScope("return");
738
739 Builder.SetInsertPoint(EntryBB);
740
741 // Emit subprogram debug descriptor.
742 if (CGDebugInfo *DI = getDebugInfo()) {
743 SmallVector<QualType, 16> ArgTypes;
744 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
745 i != e; ++i) {
746 ArgTypes.push_back((*i)->getType());
747 }
748
749 QualType FnType =
750 getContext().getFunctionType(RetTy, ArgTypes,
751 FunctionProtoType::ExtProtoInfo());
752 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
753 }
754
755 if (ShouldInstrumentFunction())
756 EmitFunctionInstrumentation("__cyg_profile_func_enter");
757
758 if (CGM.getCodeGenOpts().InstrumentForProfiling)
759 EmitMCountInstrumentation();
760
761 if (RetTy->isVoidType()) {
762 // Void type; nothing to return.
763 ReturnValue = Address::invalid();
764
765 // Count the implicit return.
766 if (!endsWithReturn(D))
767 ++NumReturnExprs;
768 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
769 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
770 // Indirect aggregate return; emit returned value directly into sret slot.
771 // This reduces code size, and affects correctness in C++.
772 auto AI = CurFn->arg_begin();
773 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
774 ++AI;
775 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
776 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
777 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
778 // Load the sret pointer from the argument struct and return into that.
779 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
780 llvm::Function::arg_iterator EI = CurFn->arg_end();
781 --EI;
782 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
783 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
784 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
785 } else {
786 ReturnValue = CreateIRTemp(RetTy, "retval");
787
788 // Tell the epilog emitter to autorelease the result. We do this
789 // now so that various specialized functions can suppress it
790 // during their IR-generation.
791 if (getLangOpts().ObjCAutoRefCount &&
792 !CurFnInfo->isReturnsRetained() &&
793 RetTy->isObjCRetainableType())
794 AutoreleaseResult = true;
795 }
796
797 EmitStartEHSpec(CurCodeDecl);
798
799 PrologueCleanupDepth = EHStack.stable_begin();
800 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
801
802 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
803 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
804 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
805 if (MD->getParent()->isLambda() &&
806 MD->getOverloadedOperator() == OO_Call) {
807 // We're in a lambda; figure out the captures.
808 MD->getParent()->getCaptureFields(LambdaCaptureFields,
809 LambdaThisCaptureField);
810 if (LambdaThisCaptureField) {
811 // If this lambda captures this, load it.
812 LValue ThisLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
813 CXXThisValue = EmitLoadOfLValue(ThisLValue,
814 SourceLocation()).getScalarVal();
815 }
816 for (auto *FD : MD->getParent()->fields()) {
817 if (FD->hasCapturedVLAType()) {
818 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
819 SourceLocation()).getScalarVal();
820 auto VAT = FD->getCapturedVLAType();
821 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
822 }
823 }
824 } else {
825 // Not in a lambda; just use 'this' from the method.
826 // FIXME: Should we generate a new load for each use of 'this'? The
827 // fast register allocator would be happier...
828 CXXThisValue = CXXABIThisValue;
829 }
830 }
831
832 // If any of the arguments have a variably modified type, make sure to
833 // emit the type size.
834 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
835 i != e; ++i) {
836 const VarDecl *VD = *i;
837
838 // Dig out the type as written from ParmVarDecls; it's unclear whether
839 // the standard (C99 6.9.1p10) requires this, but we're following the
840 // precedent set by gcc.
841 QualType Ty;
842 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
843 Ty = PVD->getOriginalType();
844 else
845 Ty = VD->getType();
846
847 if (Ty->isVariablyModifiedType())
848 EmitVariablyModifiedType(Ty);
849 }
850 // Emit a location at the end of the prologue.
851 if (CGDebugInfo *DI = getDebugInfo())
852 DI->EmitLocation(Builder, StartLoc);
853 }
854
EmitFunctionBody(FunctionArgList & Args,const Stmt * Body)855 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
856 const Stmt *Body) {
857 incrementProfileCounter(Body);
858 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
859 EmitCompoundStmtWithoutScope(*S);
860 else
861 EmitStmt(Body);
862 }
863
864 /// When instrumenting to collect profile data, the counts for some blocks
865 /// such as switch cases need to not include the fall-through counts, so
866 /// emit a branch around the instrumentation code. When not instrumenting,
867 /// this just calls EmitBlock().
EmitBlockWithFallThrough(llvm::BasicBlock * BB,const Stmt * S)868 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
869 const Stmt *S) {
870 llvm::BasicBlock *SkipCountBB = nullptr;
871 if (HaveInsertPoint() && CGM.getCodeGenOpts().ProfileInstrGenerate) {
872 // When instrumenting for profiling, the fallthrough to certain
873 // statements needs to skip over the instrumentation code so that we
874 // get an accurate count.
875 SkipCountBB = createBasicBlock("skipcount");
876 EmitBranch(SkipCountBB);
877 }
878 EmitBlock(BB);
879 uint64_t CurrentCount = getCurrentProfileCount();
880 incrementProfileCounter(S);
881 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
882 if (SkipCountBB)
883 EmitBlock(SkipCountBB);
884 }
885
886 /// Tries to mark the given function nounwind based on the
887 /// non-existence of any throwing calls within it. We believe this is
888 /// lightweight enough to do at -O0.
TryMarkNoThrow(llvm::Function * F)889 static void TryMarkNoThrow(llvm::Function *F) {
890 // LLVM treats 'nounwind' on a function as part of the type, so we
891 // can't do this on functions that can be overwritten.
892 if (F->mayBeOverridden()) return;
893
894 for (llvm::BasicBlock &BB : *F)
895 for (llvm::Instruction &I : BB)
896 if (I.mayThrow())
897 return;
898
899 F->setDoesNotThrow();
900 }
901
GenerateCode(GlobalDecl GD,llvm::Function * Fn,const CGFunctionInfo & FnInfo)902 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
903 const CGFunctionInfo &FnInfo) {
904 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
905
906 // Check if we should generate debug info for this function.
907 if (FD->hasAttr<NoDebugAttr>())
908 DebugInfo = nullptr; // disable debug info indefinitely for this function
909
910 FunctionArgList Args;
911 QualType ResTy = FD->getReturnType();
912
913 CurGD = GD;
914 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
915 if (MD && MD->isInstance()) {
916 if (CGM.getCXXABI().HasThisReturn(GD))
917 ResTy = MD->getThisType(getContext());
918 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
919 ResTy = CGM.getContext().VoidPtrTy;
920 CGM.getCXXABI().buildThisParam(*this, Args);
921 }
922
923 for (auto *Param : FD->params()) {
924 Args.push_back(Param);
925 if (!Param->hasAttr<PassObjectSizeAttr>())
926 continue;
927
928 IdentifierInfo *NoID = nullptr;
929 auto *Implicit = ImplicitParamDecl::Create(
930 getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
931 getContext().getSizeType());
932 SizeArguments[Param] = Implicit;
933 Args.push_back(Implicit);
934 }
935
936 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
937 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
938
939 SourceRange BodyRange;
940 if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
941 CurEHLocation = BodyRange.getEnd();
942
943 // Use the location of the start of the function to determine where
944 // the function definition is located. By default use the location
945 // of the declaration as the location for the subprogram. A function
946 // may lack a declaration in the source code if it is created by code
947 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
948 SourceLocation Loc = FD->getLocation();
949
950 // If this is a function specialization then use the pattern body
951 // as the location for the function.
952 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
953 if (SpecDecl->hasBody(SpecDecl))
954 Loc = SpecDecl->getLocation();
955
956 // Emit the standard function prologue.
957 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
958
959 // Generate the body of the function.
960 PGO.assignRegionCounters(GD, CurFn);
961 if (isa<CXXDestructorDecl>(FD))
962 EmitDestructorBody(Args);
963 else if (isa<CXXConstructorDecl>(FD))
964 EmitConstructorBody(Args);
965 else if (getLangOpts().CUDA &&
966 !getLangOpts().CUDAIsDevice &&
967 FD->hasAttr<CUDAGlobalAttr>())
968 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
969 else if (isa<CXXConversionDecl>(FD) &&
970 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
971 // The lambda conversion to block pointer is special; the semantics can't be
972 // expressed in the AST, so IRGen needs to special-case it.
973 EmitLambdaToBlockPointerBody(Args);
974 } else if (isa<CXXMethodDecl>(FD) &&
975 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
976 // The lambda static invoker function is special, because it forwards or
977 // clones the body of the function call operator (but is actually static).
978 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
979 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
980 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
981 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
982 // Implicit copy-assignment gets the same special treatment as implicit
983 // copy-constructors.
984 emitImplicitAssignmentOperatorBody(Args);
985 } else if (Stmt *Body = FD->getBody()) {
986 EmitFunctionBody(Args, Body);
987 } else
988 llvm_unreachable("no definition for emitted function");
989
990 // C++11 [stmt.return]p2:
991 // Flowing off the end of a function [...] results in undefined behavior in
992 // a value-returning function.
993 // C11 6.9.1p12:
994 // If the '}' that terminates a function is reached, and the value of the
995 // function call is used by the caller, the behavior is undefined.
996 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
997 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
998 if (SanOpts.has(SanitizerKind::Return)) {
999 SanitizerScope SanScope(this);
1000 llvm::Value *IsFalse = Builder.getFalse();
1001 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1002 "missing_return", EmitCheckSourceLocation(FD->getLocation()),
1003 None);
1004 } else if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
1005 EmitTrapCall(llvm::Intrinsic::trap);
1006 }
1007 Builder.CreateUnreachable();
1008 Builder.ClearInsertionPoint();
1009 }
1010
1011 // Emit the standard function epilogue.
1012 FinishFunction(BodyRange.getEnd());
1013
1014 // If we haven't marked the function nothrow through other means, do
1015 // a quick pass now to see if we can.
1016 if (!CurFn->doesNotThrow())
1017 TryMarkNoThrow(CurFn);
1018 }
1019
1020 /// ContainsLabel - Return true if the statement contains a label in it. If
1021 /// this statement is not executed normally, it not containing a label means
1022 /// that we can just remove the code.
ContainsLabel(const Stmt * S,bool IgnoreCaseStmts)1023 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1024 // Null statement, not a label!
1025 if (!S) return false;
1026
1027 // If this is a label, we have to emit the code, consider something like:
1028 // if (0) { ... foo: bar(); } goto foo;
1029 //
1030 // TODO: If anyone cared, we could track __label__'s, since we know that you
1031 // can't jump to one from outside their declared region.
1032 if (isa<LabelStmt>(S))
1033 return true;
1034
1035 // If this is a case/default statement, and we haven't seen a switch, we have
1036 // to emit the code.
1037 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1038 return true;
1039
1040 // If this is a switch statement, we want to ignore cases below it.
1041 if (isa<SwitchStmt>(S))
1042 IgnoreCaseStmts = true;
1043
1044 // Scan subexpressions for verboten labels.
1045 for (const Stmt *SubStmt : S->children())
1046 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1047 return true;
1048
1049 return false;
1050 }
1051
1052 /// containsBreak - Return true if the statement contains a break out of it.
1053 /// If the statement (recursively) contains a switch or loop with a break
1054 /// inside of it, this is fine.
containsBreak(const Stmt * S)1055 bool CodeGenFunction::containsBreak(const Stmt *S) {
1056 // Null statement, not a label!
1057 if (!S) return false;
1058
1059 // If this is a switch or loop that defines its own break scope, then we can
1060 // include it and anything inside of it.
1061 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1062 isa<ForStmt>(S))
1063 return false;
1064
1065 if (isa<BreakStmt>(S))
1066 return true;
1067
1068 // Scan subexpressions for verboten breaks.
1069 for (const Stmt *SubStmt : S->children())
1070 if (containsBreak(SubStmt))
1071 return true;
1072
1073 return false;
1074 }
1075
1076
1077 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1078 /// to a constant, or if it does but contains a label, return false. If it
1079 /// constant folds return true and set the boolean result in Result.
ConstantFoldsToSimpleInteger(const Expr * Cond,bool & ResultBool)1080 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1081 bool &ResultBool) {
1082 llvm::APSInt ResultInt;
1083 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt))
1084 return false;
1085
1086 ResultBool = ResultInt.getBoolValue();
1087 return true;
1088 }
1089
1090 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1091 /// to a constant, or if it does but contains a label, return false. If it
1092 /// constant folds return true and set the folded value.
1093 bool CodeGenFunction::
ConstantFoldsToSimpleInteger(const Expr * Cond,llvm::APSInt & ResultInt)1094 ConstantFoldsToSimpleInteger(const Expr *Cond, llvm::APSInt &ResultInt) {
1095 // FIXME: Rename and handle conversion of other evaluatable things
1096 // to bool.
1097 llvm::APSInt Int;
1098 if (!Cond->EvaluateAsInt(Int, getContext()))
1099 return false; // Not foldable, not integer or not fully evaluatable.
1100
1101 if (CodeGenFunction::ContainsLabel(Cond))
1102 return false; // Contains a label.
1103
1104 ResultInt = Int;
1105 return true;
1106 }
1107
1108
1109
1110 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1111 /// statement) to the specified blocks. Based on the condition, this might try
1112 /// to simplify the codegen of the conditional based on the branch.
1113 ///
EmitBranchOnBoolExpr(const Expr * Cond,llvm::BasicBlock * TrueBlock,llvm::BasicBlock * FalseBlock,uint64_t TrueCount)1114 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1115 llvm::BasicBlock *TrueBlock,
1116 llvm::BasicBlock *FalseBlock,
1117 uint64_t TrueCount) {
1118 Cond = Cond->IgnoreParens();
1119
1120 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1121
1122 // Handle X && Y in a condition.
1123 if (CondBOp->getOpcode() == BO_LAnd) {
1124 // If we have "1 && X", simplify the code. "0 && X" would have constant
1125 // folded if the case was simple enough.
1126 bool ConstantBool = false;
1127 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1128 ConstantBool) {
1129 // br(1 && X) -> br(X).
1130 incrementProfileCounter(CondBOp);
1131 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1132 TrueCount);
1133 }
1134
1135 // If we have "X && 1", simplify the code to use an uncond branch.
1136 // "X && 0" would have been constant folded to 0.
1137 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1138 ConstantBool) {
1139 // br(X && 1) -> br(X).
1140 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1141 TrueCount);
1142 }
1143
1144 // Emit the LHS as a conditional. If the LHS conditional is false, we
1145 // want to jump to the FalseBlock.
1146 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1147 // The counter tells us how often we evaluate RHS, and all of TrueCount
1148 // can be propagated to that branch.
1149 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1150
1151 ConditionalEvaluation eval(*this);
1152 {
1153 ApplyDebugLocation DL(*this, Cond);
1154 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1155 EmitBlock(LHSTrue);
1156 }
1157
1158 incrementProfileCounter(CondBOp);
1159 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1160
1161 // Any temporaries created here are conditional.
1162 eval.begin(*this);
1163 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1164 eval.end(*this);
1165
1166 return;
1167 }
1168
1169 if (CondBOp->getOpcode() == BO_LOr) {
1170 // If we have "0 || X", simplify the code. "1 || X" would have constant
1171 // folded if the case was simple enough.
1172 bool ConstantBool = false;
1173 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1174 !ConstantBool) {
1175 // br(0 || X) -> br(X).
1176 incrementProfileCounter(CondBOp);
1177 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1178 TrueCount);
1179 }
1180
1181 // If we have "X || 0", simplify the code to use an uncond branch.
1182 // "X || 1" would have been constant folded to 1.
1183 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1184 !ConstantBool) {
1185 // br(X || 0) -> br(X).
1186 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1187 TrueCount);
1188 }
1189
1190 // Emit the LHS as a conditional. If the LHS conditional is true, we
1191 // want to jump to the TrueBlock.
1192 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1193 // We have the count for entry to the RHS and for the whole expression
1194 // being true, so we can divy up True count between the short circuit and
1195 // the RHS.
1196 uint64_t LHSCount =
1197 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1198 uint64_t RHSCount = TrueCount - LHSCount;
1199
1200 ConditionalEvaluation eval(*this);
1201 {
1202 ApplyDebugLocation DL(*this, Cond);
1203 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1204 EmitBlock(LHSFalse);
1205 }
1206
1207 incrementProfileCounter(CondBOp);
1208 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1209
1210 // Any temporaries created here are conditional.
1211 eval.begin(*this);
1212 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1213
1214 eval.end(*this);
1215
1216 return;
1217 }
1218 }
1219
1220 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1221 // br(!x, t, f) -> br(x, f, t)
1222 if (CondUOp->getOpcode() == UO_LNot) {
1223 // Negate the count.
1224 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1225 // Negate the condition and swap the destination blocks.
1226 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1227 FalseCount);
1228 }
1229 }
1230
1231 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1232 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1233 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1234 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1235
1236 ConditionalEvaluation cond(*this);
1237 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1238 getProfileCount(CondOp));
1239
1240 // When computing PGO branch weights, we only know the overall count for
1241 // the true block. This code is essentially doing tail duplication of the
1242 // naive code-gen, introducing new edges for which counts are not
1243 // available. Divide the counts proportionally between the LHS and RHS of
1244 // the conditional operator.
1245 uint64_t LHSScaledTrueCount = 0;
1246 if (TrueCount) {
1247 double LHSRatio =
1248 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1249 LHSScaledTrueCount = TrueCount * LHSRatio;
1250 }
1251
1252 cond.begin(*this);
1253 EmitBlock(LHSBlock);
1254 incrementProfileCounter(CondOp);
1255 {
1256 ApplyDebugLocation DL(*this, Cond);
1257 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1258 LHSScaledTrueCount);
1259 }
1260 cond.end(*this);
1261
1262 cond.begin(*this);
1263 EmitBlock(RHSBlock);
1264 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1265 TrueCount - LHSScaledTrueCount);
1266 cond.end(*this);
1267
1268 return;
1269 }
1270
1271 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1272 // Conditional operator handling can give us a throw expression as a
1273 // condition for a case like:
1274 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1275 // Fold this to:
1276 // br(c, throw x, br(y, t, f))
1277 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1278 return;
1279 }
1280
1281 // If the branch has a condition wrapped by __builtin_unpredictable,
1282 // create metadata that specifies that the branch is unpredictable.
1283 // Don't bother if not optimizing because that metadata would not be used.
1284 llvm::MDNode *Unpredictable = nullptr;
1285 if (CGM.getCodeGenOpts().OptimizationLevel != 0) {
1286 if (const CallExpr *Call = dyn_cast<CallExpr>(Cond)) {
1287 const Decl *TargetDecl = Call->getCalleeDecl();
1288 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl)) {
1289 if (FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1290 llvm::MDBuilder MDHelper(getLLVMContext());
1291 Unpredictable = MDHelper.createUnpredictable();
1292 }
1293 }
1294 }
1295 }
1296
1297 // Create branch weights based on the number of times we get here and the
1298 // number of times the condition should be true.
1299 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1300 llvm::MDNode *Weights =
1301 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1302
1303 // Emit the code with the fully general case.
1304 llvm::Value *CondV;
1305 {
1306 ApplyDebugLocation DL(*this, Cond);
1307 CondV = EvaluateExprAsBool(Cond);
1308 }
1309 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1310 }
1311
1312 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1313 /// specified stmt yet.
ErrorUnsupported(const Stmt * S,const char * Type)1314 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1315 CGM.ErrorUnsupported(S, Type);
1316 }
1317
1318 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1319 /// variable-length array whose elements have a non-zero bit-pattern.
1320 ///
1321 /// \param baseType the inner-most element type of the array
1322 /// \param src - a char* pointing to the bit-pattern for a single
1323 /// base element of the array
1324 /// \param sizeInChars - the total size of the VLA, in chars
emitNonZeroVLAInit(CodeGenFunction & CGF,QualType baseType,Address dest,Address src,llvm::Value * sizeInChars)1325 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1326 Address dest, Address src,
1327 llvm::Value *sizeInChars) {
1328 CGBuilderTy &Builder = CGF.Builder;
1329
1330 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1331 llvm::Value *baseSizeInChars
1332 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1333
1334 Address begin =
1335 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1336 llvm::Value *end =
1337 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1338
1339 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1340 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1341 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1342
1343 // Make a loop over the VLA. C99 guarantees that the VLA element
1344 // count must be nonzero.
1345 CGF.EmitBlock(loopBB);
1346
1347 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1348 cur->addIncoming(begin.getPointer(), originBB);
1349
1350 CharUnits curAlign =
1351 dest.getAlignment().alignmentOfArrayElement(baseSize);
1352
1353 // memcpy the individual element bit-pattern.
1354 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1355 /*volatile*/ false);
1356
1357 // Go to the next element.
1358 llvm::Value *next =
1359 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1360
1361 // Leave if that's the end of the VLA.
1362 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1363 Builder.CreateCondBr(done, contBB, loopBB);
1364 cur->addIncoming(next, loopBB);
1365
1366 CGF.EmitBlock(contBB);
1367 }
1368
1369 void
EmitNullInitialization(Address DestPtr,QualType Ty)1370 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1371 // Ignore empty classes in C++.
1372 if (getLangOpts().CPlusPlus) {
1373 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1374 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1375 return;
1376 }
1377 }
1378
1379 // Cast the dest ptr to the appropriate i8 pointer type.
1380 if (DestPtr.getElementType() != Int8Ty)
1381 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1382
1383 // Get size and alignment info for this aggregate.
1384 CharUnits size = getContext().getTypeSizeInChars(Ty);
1385
1386 llvm::Value *SizeVal;
1387 const VariableArrayType *vla;
1388
1389 // Don't bother emitting a zero-byte memset.
1390 if (size.isZero()) {
1391 // But note that getTypeInfo returns 0 for a VLA.
1392 if (const VariableArrayType *vlaType =
1393 dyn_cast_or_null<VariableArrayType>(
1394 getContext().getAsArrayType(Ty))) {
1395 QualType eltType;
1396 llvm::Value *numElts;
1397 std::tie(numElts, eltType) = getVLASize(vlaType);
1398
1399 SizeVal = numElts;
1400 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1401 if (!eltSize.isOne())
1402 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1403 vla = vlaType;
1404 } else {
1405 return;
1406 }
1407 } else {
1408 SizeVal = CGM.getSize(size);
1409 vla = nullptr;
1410 }
1411
1412 // If the type contains a pointer to data member we can't memset it to zero.
1413 // Instead, create a null constant and copy it to the destination.
1414 // TODO: there are other patterns besides zero that we can usefully memset,
1415 // like -1, which happens to be the pattern used by member-pointers.
1416 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1417 // For a VLA, emit a single element, then splat that over the VLA.
1418 if (vla) Ty = getContext().getBaseElementType(vla);
1419
1420 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1421
1422 llvm::GlobalVariable *NullVariable =
1423 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1424 /*isConstant=*/true,
1425 llvm::GlobalVariable::PrivateLinkage,
1426 NullConstant, Twine());
1427 CharUnits NullAlign = DestPtr.getAlignment();
1428 NullVariable->setAlignment(NullAlign.getQuantity());
1429 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1430 NullAlign);
1431
1432 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1433
1434 // Get and call the appropriate llvm.memcpy overload.
1435 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1436 return;
1437 }
1438
1439 // Otherwise, just memset the whole thing to zero. This is legal
1440 // because in LLVM, all default initializers (other than the ones we just
1441 // handled above) are guaranteed to have a bit pattern of all zeros.
1442 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1443 }
1444
GetAddrOfLabel(const LabelDecl * L)1445 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1446 // Make sure that there is a block for the indirect goto.
1447 if (!IndirectBranch)
1448 GetIndirectGotoBlock();
1449
1450 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1451
1452 // Make sure the indirect branch includes all of the address-taken blocks.
1453 IndirectBranch->addDestination(BB);
1454 return llvm::BlockAddress::get(CurFn, BB);
1455 }
1456
GetIndirectGotoBlock()1457 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1458 // If we already made the indirect branch for indirect goto, return its block.
1459 if (IndirectBranch) return IndirectBranch->getParent();
1460
1461 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1462
1463 // Create the PHI node that indirect gotos will add entries to.
1464 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1465 "indirect.goto.dest");
1466
1467 // Create the indirect branch instruction.
1468 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1469 return IndirectBranch->getParent();
1470 }
1471
1472 /// Computes the length of an array in elements, as well as the base
1473 /// element type and a properly-typed first element pointer.
emitArrayLength(const ArrayType * origArrayType,QualType & baseType,Address & addr)1474 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1475 QualType &baseType,
1476 Address &addr) {
1477 const ArrayType *arrayType = origArrayType;
1478
1479 // If it's a VLA, we have to load the stored size. Note that
1480 // this is the size of the VLA in bytes, not its size in elements.
1481 llvm::Value *numVLAElements = nullptr;
1482 if (isa<VariableArrayType>(arrayType)) {
1483 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1484
1485 // Walk into all VLAs. This doesn't require changes to addr,
1486 // which has type T* where T is the first non-VLA element type.
1487 do {
1488 QualType elementType = arrayType->getElementType();
1489 arrayType = getContext().getAsArrayType(elementType);
1490
1491 // If we only have VLA components, 'addr' requires no adjustment.
1492 if (!arrayType) {
1493 baseType = elementType;
1494 return numVLAElements;
1495 }
1496 } while (isa<VariableArrayType>(arrayType));
1497
1498 // We get out here only if we find a constant array type
1499 // inside the VLA.
1500 }
1501
1502 // We have some number of constant-length arrays, so addr should
1503 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1504 // down to the first element of addr.
1505 SmallVector<llvm::Value*, 8> gepIndices;
1506
1507 // GEP down to the array type.
1508 llvm::ConstantInt *zero = Builder.getInt32(0);
1509 gepIndices.push_back(zero);
1510
1511 uint64_t countFromCLAs = 1;
1512 QualType eltType;
1513
1514 llvm::ArrayType *llvmArrayType =
1515 dyn_cast<llvm::ArrayType>(addr.getElementType());
1516 while (llvmArrayType) {
1517 assert(isa<ConstantArrayType>(arrayType));
1518 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1519 == llvmArrayType->getNumElements());
1520
1521 gepIndices.push_back(zero);
1522 countFromCLAs *= llvmArrayType->getNumElements();
1523 eltType = arrayType->getElementType();
1524
1525 llvmArrayType =
1526 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1527 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1528 assert((!llvmArrayType || arrayType) &&
1529 "LLVM and Clang types are out-of-synch");
1530 }
1531
1532 if (arrayType) {
1533 // From this point onwards, the Clang array type has been emitted
1534 // as some other type (probably a packed struct). Compute the array
1535 // size, and just emit the 'begin' expression as a bitcast.
1536 while (arrayType) {
1537 countFromCLAs *=
1538 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1539 eltType = arrayType->getElementType();
1540 arrayType = getContext().getAsArrayType(eltType);
1541 }
1542
1543 llvm::Type *baseType = ConvertType(eltType);
1544 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1545 } else {
1546 // Create the actual GEP.
1547 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1548 gepIndices, "array.begin"),
1549 addr.getAlignment());
1550 }
1551
1552 baseType = eltType;
1553
1554 llvm::Value *numElements
1555 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1556
1557 // If we had any VLA dimensions, factor them in.
1558 if (numVLAElements)
1559 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1560
1561 return numElements;
1562 }
1563
1564 std::pair<llvm::Value*, QualType>
getVLASize(QualType type)1565 CodeGenFunction::getVLASize(QualType type) {
1566 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1567 assert(vla && "type was not a variable array type!");
1568 return getVLASize(vla);
1569 }
1570
1571 std::pair<llvm::Value*, QualType>
getVLASize(const VariableArrayType * type)1572 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1573 // The number of elements so far; always size_t.
1574 llvm::Value *numElements = nullptr;
1575
1576 QualType elementType;
1577 do {
1578 elementType = type->getElementType();
1579 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1580 assert(vlaSize && "no size for VLA!");
1581 assert(vlaSize->getType() == SizeTy);
1582
1583 if (!numElements) {
1584 numElements = vlaSize;
1585 } else {
1586 // It's undefined behavior if this wraps around, so mark it that way.
1587 // FIXME: Teach -fsanitize=undefined to trap this.
1588 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1589 }
1590 } while ((type = getContext().getAsVariableArrayType(elementType)));
1591
1592 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1593 }
1594
EmitVariablyModifiedType(QualType type)1595 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1596 assert(type->isVariablyModifiedType() &&
1597 "Must pass variably modified type to EmitVLASizes!");
1598
1599 EnsureInsertPoint();
1600
1601 // We're going to walk down into the type and look for VLA
1602 // expressions.
1603 do {
1604 assert(type->isVariablyModifiedType());
1605
1606 const Type *ty = type.getTypePtr();
1607 switch (ty->getTypeClass()) {
1608
1609 #define TYPE(Class, Base)
1610 #define ABSTRACT_TYPE(Class, Base)
1611 #define NON_CANONICAL_TYPE(Class, Base)
1612 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1613 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1614 #include "clang/AST/TypeNodes.def"
1615 llvm_unreachable("unexpected dependent type!");
1616
1617 // These types are never variably-modified.
1618 case Type::Builtin:
1619 case Type::Complex:
1620 case Type::Vector:
1621 case Type::ExtVector:
1622 case Type::Record:
1623 case Type::Enum:
1624 case Type::Elaborated:
1625 case Type::TemplateSpecialization:
1626 case Type::ObjCObject:
1627 case Type::ObjCInterface:
1628 case Type::ObjCObjectPointer:
1629 llvm_unreachable("type class is never variably-modified!");
1630
1631 case Type::Adjusted:
1632 type = cast<AdjustedType>(ty)->getAdjustedType();
1633 break;
1634
1635 case Type::Decayed:
1636 type = cast<DecayedType>(ty)->getPointeeType();
1637 break;
1638
1639 case Type::Pointer:
1640 type = cast<PointerType>(ty)->getPointeeType();
1641 break;
1642
1643 case Type::BlockPointer:
1644 type = cast<BlockPointerType>(ty)->getPointeeType();
1645 break;
1646
1647 case Type::LValueReference:
1648 case Type::RValueReference:
1649 type = cast<ReferenceType>(ty)->getPointeeType();
1650 break;
1651
1652 case Type::MemberPointer:
1653 type = cast<MemberPointerType>(ty)->getPointeeType();
1654 break;
1655
1656 case Type::ConstantArray:
1657 case Type::IncompleteArray:
1658 // Losing element qualification here is fine.
1659 type = cast<ArrayType>(ty)->getElementType();
1660 break;
1661
1662 case Type::VariableArray: {
1663 // Losing element qualification here is fine.
1664 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1665
1666 // Unknown size indication requires no size computation.
1667 // Otherwise, evaluate and record it.
1668 if (const Expr *size = vat->getSizeExpr()) {
1669 // It's possible that we might have emitted this already,
1670 // e.g. with a typedef and a pointer to it.
1671 llvm::Value *&entry = VLASizeMap[size];
1672 if (!entry) {
1673 llvm::Value *Size = EmitScalarExpr(size);
1674
1675 // C11 6.7.6.2p5:
1676 // If the size is an expression that is not an integer constant
1677 // expression [...] each time it is evaluated it shall have a value
1678 // greater than zero.
1679 if (SanOpts.has(SanitizerKind::VLABound) &&
1680 size->getType()->isSignedIntegerType()) {
1681 SanitizerScope SanScope(this);
1682 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1683 llvm::Constant *StaticArgs[] = {
1684 EmitCheckSourceLocation(size->getLocStart()),
1685 EmitCheckTypeDescriptor(size->getType())
1686 };
1687 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1688 SanitizerKind::VLABound),
1689 "vla_bound_not_positive", StaticArgs, Size);
1690 }
1691
1692 // Always zexting here would be wrong if it weren't
1693 // undefined behavior to have a negative bound.
1694 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1695 }
1696 }
1697 type = vat->getElementType();
1698 break;
1699 }
1700
1701 case Type::FunctionProto:
1702 case Type::FunctionNoProto:
1703 type = cast<FunctionType>(ty)->getReturnType();
1704 break;
1705
1706 case Type::Paren:
1707 case Type::TypeOf:
1708 case Type::UnaryTransform:
1709 case Type::Attributed:
1710 case Type::SubstTemplateTypeParm:
1711 case Type::PackExpansion:
1712 // Keep walking after single level desugaring.
1713 type = type.getSingleStepDesugaredType(getContext());
1714 break;
1715
1716 case Type::Typedef:
1717 case Type::Decltype:
1718 case Type::Auto:
1719 // Stop walking: nothing to do.
1720 return;
1721
1722 case Type::TypeOfExpr:
1723 // Stop walking: emit typeof expression.
1724 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1725 return;
1726
1727 case Type::Atomic:
1728 type = cast<AtomicType>(ty)->getValueType();
1729 break;
1730 }
1731 } while (type->isVariablyModifiedType());
1732 }
1733
EmitVAListRef(const Expr * E)1734 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1735 if (getContext().getBuiltinVaListType()->isArrayType())
1736 return EmitPointerWithAlignment(E);
1737 return EmitLValue(E).getAddress();
1738 }
1739
EmitMSVAListRef(const Expr * E)1740 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1741 return EmitLValue(E).getAddress();
1742 }
1743
EmitDeclRefExprDbgValue(const DeclRefExpr * E,llvm::Constant * Init)1744 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1745 llvm::Constant *Init) {
1746 assert (Init && "Invalid DeclRefExpr initializer!");
1747 if (CGDebugInfo *Dbg = getDebugInfo())
1748 if (CGM.getCodeGenOpts().getDebugInfo() >= CodeGenOptions::LimitedDebugInfo)
1749 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1750 }
1751
1752 CodeGenFunction::PeepholeProtection
protectFromPeepholes(RValue rvalue)1753 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1754 // At the moment, the only aggressive peephole we do in IR gen
1755 // is trunc(zext) folding, but if we add more, we can easily
1756 // extend this protection.
1757
1758 if (!rvalue.isScalar()) return PeepholeProtection();
1759 llvm::Value *value = rvalue.getScalarVal();
1760 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1761
1762 // Just make an extra bitcast.
1763 assert(HaveInsertPoint());
1764 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1765 Builder.GetInsertBlock());
1766
1767 PeepholeProtection protection;
1768 protection.Inst = inst;
1769 return protection;
1770 }
1771
unprotectFromPeepholes(PeepholeProtection protection)1772 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1773 if (!protection.Inst) return;
1774
1775 // In theory, we could try to duplicate the peepholes now, but whatever.
1776 protection.Inst->eraseFromParent();
1777 }
1778
EmitAnnotationCall(llvm::Value * AnnotationFn,llvm::Value * AnnotatedVal,StringRef AnnotationStr,SourceLocation Location)1779 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1780 llvm::Value *AnnotatedVal,
1781 StringRef AnnotationStr,
1782 SourceLocation Location) {
1783 llvm::Value *Args[4] = {
1784 AnnotatedVal,
1785 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1786 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1787 CGM.EmitAnnotationLineNo(Location)
1788 };
1789 return Builder.CreateCall(AnnotationFn, Args);
1790 }
1791
EmitVarAnnotations(const VarDecl * D,llvm::Value * V)1792 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1793 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1794 // FIXME We create a new bitcast for every annotation because that's what
1795 // llvm-gcc was doing.
1796 for (const auto *I : D->specific_attrs<AnnotateAttr>())
1797 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1798 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1799 I->getAnnotation(), D->getLocation());
1800 }
1801
EmitFieldAnnotations(const FieldDecl * D,Address Addr)1802 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1803 Address Addr) {
1804 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1805 llvm::Value *V = Addr.getPointer();
1806 llvm::Type *VTy = V->getType();
1807 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1808 CGM.Int8PtrTy);
1809
1810 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1811 // FIXME Always emit the cast inst so we can differentiate between
1812 // annotation on the first field of a struct and annotation on the struct
1813 // itself.
1814 if (VTy != CGM.Int8PtrTy)
1815 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1816 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1817 V = Builder.CreateBitCast(V, VTy);
1818 }
1819
1820 return Address(V, Addr.getAlignment());
1821 }
1822
~CGCapturedStmtInfo()1823 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
1824
SanitizerScope(CodeGenFunction * CGF)1825 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
1826 : CGF(CGF) {
1827 assert(!CGF->IsSanitizerScope);
1828 CGF->IsSanitizerScope = true;
1829 }
1830
~SanitizerScope()1831 CodeGenFunction::SanitizerScope::~SanitizerScope() {
1832 CGF->IsSanitizerScope = false;
1833 }
1834
InsertHelper(llvm::Instruction * I,const llvm::Twine & Name,llvm::BasicBlock * BB,llvm::BasicBlock::iterator InsertPt) const1835 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
1836 const llvm::Twine &Name,
1837 llvm::BasicBlock *BB,
1838 llvm::BasicBlock::iterator InsertPt) const {
1839 LoopStack.InsertHelper(I);
1840 if (IsSanitizerScope)
1841 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
1842 }
1843
1844 template <bool PreserveNames>
InsertHelper(llvm::Instruction * I,const llvm::Twine & Name,llvm::BasicBlock * BB,llvm::BasicBlock::iterator InsertPt) const1845 void CGBuilderInserter<PreserveNames>::InsertHelper(
1846 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1847 llvm::BasicBlock::iterator InsertPt) const {
1848 llvm::IRBuilderDefaultInserter<PreserveNames>::InsertHelper(I, Name, BB,
1849 InsertPt);
1850 if (CGF)
1851 CGF->InsertHelper(I, Name, BB, InsertPt);
1852 }
1853
1854 #ifdef NDEBUG
1855 #define PreserveNames false
1856 #else
1857 #define PreserveNames true
1858 #endif
1859 template void CGBuilderInserter<PreserveNames>::InsertHelper(
1860 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
1861 llvm::BasicBlock::iterator InsertPt) const;
1862 #undef PreserveNames
1863
hasRequiredFeatures(const SmallVectorImpl<StringRef> & ReqFeatures,CodeGenModule & CGM,const FunctionDecl * FD,std::string & FirstMissing)1864 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
1865 CodeGenModule &CGM, const FunctionDecl *FD,
1866 std::string &FirstMissing) {
1867 // If there aren't any required features listed then go ahead and return.
1868 if (ReqFeatures.empty())
1869 return false;
1870
1871 // Now build up the set of caller features and verify that all the required
1872 // features are there.
1873 llvm::StringMap<bool> CallerFeatureMap;
1874 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
1875
1876 // If we have at least one of the features in the feature list return
1877 // true, otherwise return false.
1878 return std::all_of(
1879 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
1880 SmallVector<StringRef, 1> OrFeatures;
1881 Feature.split(OrFeatures, "|");
1882 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
1883 [&](StringRef Feature) {
1884 if (!CallerFeatureMap.lookup(Feature)) {
1885 FirstMissing = Feature.str();
1886 return false;
1887 }
1888 return true;
1889 });
1890 });
1891 }
1892
1893 // Emits an error if we don't have a valid set of target features for the
1894 // called function.
checkTargetFeatures(const CallExpr * E,const FunctionDecl * TargetDecl)1895 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
1896 const FunctionDecl *TargetDecl) {
1897 // Early exit if this is an indirect call.
1898 if (!TargetDecl)
1899 return;
1900
1901 // Get the current enclosing function if it exists. If it doesn't
1902 // we can't check the target features anyhow.
1903 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
1904 if (!FD)
1905 return;
1906
1907 // Grab the required features for the call. For a builtin this is listed in
1908 // the td file with the default cpu, for an always_inline function this is any
1909 // listed cpu and any listed features.
1910 unsigned BuiltinID = TargetDecl->getBuiltinID();
1911 std::string MissingFeature;
1912 if (BuiltinID) {
1913 SmallVector<StringRef, 1> ReqFeatures;
1914 const char *FeatureList =
1915 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1916 // Return if the builtin doesn't have any required features.
1917 if (!FeatureList || StringRef(FeatureList) == "")
1918 return;
1919 StringRef(FeatureList).split(ReqFeatures, ",");
1920 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1921 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
1922 << TargetDecl->getDeclName()
1923 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
1924
1925 } else if (TargetDecl->hasAttr<TargetAttr>()) {
1926 // Get the required features for the callee.
1927 SmallVector<StringRef, 1> ReqFeatures;
1928 llvm::StringMap<bool> CalleeFeatureMap;
1929 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
1930 for (const auto &F : CalleeFeatureMap) {
1931 // Only positive features are "required".
1932 if (F.getValue())
1933 ReqFeatures.push_back(F.getKey());
1934 }
1935 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
1936 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
1937 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
1938 }
1939 }
1940