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/ASTLambda.h" 26 #include "clang/AST/Decl.h" 27 #include "clang/AST/DeclCXX.h" 28 #include "clang/AST/StmtCXX.h" 29 #include "clang/AST/StmtObjC.h" 30 #include "clang/Basic/Builtins.h" 31 #include "clang/Basic/TargetInfo.h" 32 #include "clang/CodeGen/CGFunctionInfo.h" 33 #include "clang/Frontend/CodeGenOptions.h" 34 #include "clang/Sema/SemaDiagnostic.h" 35 #include "llvm/IR/DataLayout.h" 36 #include "llvm/IR/Dominators.h" 37 #include "llvm/IR/Intrinsics.h" 38 #include "llvm/IR/MDBuilder.h" 39 #include "llvm/IR/Operator.h" 40 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 41 using namespace clang; 42 using namespace CodeGen; 43 44 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time 45 /// markers. 46 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts, 47 const LangOptions &LangOpts) { 48 if (CGOpts.DisableLifetimeMarkers) 49 return false; 50 51 // Disable lifetime markers in msan builds. 52 // FIXME: Remove this when msan works with lifetime markers. 53 if (LangOpts.Sanitize.has(SanitizerKind::Memory)) 54 return false; 55 56 // Asan uses markers for use-after-scope checks. 57 if (CGOpts.SanitizeAddressUseAfterScope) 58 return true; 59 60 // For now, only in optimized builds. 61 return CGOpts.OptimizationLevel != 0; 62 } 63 64 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext) 65 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()), 66 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(), 67 CGBuilderInserterTy(this)), 68 CurFn(nullptr), ReturnValue(Address::invalid()), 69 CapturedStmtInfo(nullptr), SanOpts(CGM.getLangOpts().Sanitize), 70 IsSanitizerScope(false), CurFuncIsThunk(false), AutoreleaseResult(false), 71 SawAsmBlock(false), IsOutlinedSEHHelper(false), BlockInfo(nullptr), 72 BlockPointer(nullptr), LambdaThisCaptureField(nullptr), 73 NormalCleanupDest(Address::invalid()), NextCleanupDestIndex(1), 74 FirstBlockInfo(nullptr), EHResumeBlock(nullptr), ExceptionSlot(nullptr), 75 EHSelectorSlot(nullptr), DebugInfo(CGM.getModuleDebugInfo()), 76 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr), 77 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr), 78 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0), 79 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr), 80 CXXABIThisValue(nullptr), CXXThisValue(nullptr), 81 CXXStructorImplicitParamDecl(nullptr), 82 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr), 83 CurLexicalScope(nullptr), TerminateLandingPad(nullptr), 84 TerminateHandler(nullptr), TrapBB(nullptr), 85 ShouldEmitLifetimeMarkers( 86 shouldEmitLifetimeMarkers(CGM.getCodeGenOpts(), CGM.getLangOpts())) { 87 if (!suppressNewContext) 88 CGM.getCXXABI().getMangleContext().startNewFunction(); 89 90 llvm::FastMathFlags FMF; 91 if (CGM.getLangOpts().FastMath) 92 FMF.setFast(); 93 if (CGM.getLangOpts().FiniteMathOnly) { 94 FMF.setNoNaNs(); 95 FMF.setNoInfs(); 96 } 97 if (CGM.getCodeGenOpts().NoNaNsFPMath) { 98 FMF.setNoNaNs(); 99 } 100 if (CGM.getCodeGenOpts().NoSignedZeros) { 101 FMF.setNoSignedZeros(); 102 } 103 if (CGM.getCodeGenOpts().ReciprocalMath) { 104 FMF.setAllowReciprocal(); 105 } 106 if (CGM.getCodeGenOpts().Reassociate) { 107 FMF.setAllowReassoc(); 108 } 109 Builder.setFastMathFlags(FMF); 110 } 111 112 CodeGenFunction::~CodeGenFunction() { 113 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup"); 114 115 // If there are any unclaimed block infos, go ahead and destroy them 116 // now. This can happen if IR-gen gets clever and skips evaluating 117 // something. 118 if (FirstBlockInfo) 119 destroyBlockInfos(FirstBlockInfo); 120 121 if (getLangOpts().OpenMP && CurFn) 122 CGM.getOpenMPRuntime().functionFinished(*this); 123 } 124 125 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T, 126 LValueBaseInfo *BaseInfo, 127 TBAAAccessInfo *TBAAInfo) { 128 return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo, TBAAInfo, 129 /* forPointeeType= */ true); 130 } 131 132 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T, 133 LValueBaseInfo *BaseInfo, 134 TBAAAccessInfo *TBAAInfo, 135 bool forPointeeType) { 136 if (TBAAInfo) 137 *TBAAInfo = CGM.getTBAAAccessInfo(T); 138 139 // Honor alignment typedef attributes even on incomplete types. 140 // We also honor them straight for C++ class types, even as pointees; 141 // there's an expressivity gap here. 142 if (auto TT = T->getAs<TypedefType>()) { 143 if (auto Align = TT->getDecl()->getMaxAlignment()) { 144 if (BaseInfo) 145 *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType); 146 return getContext().toCharUnitsFromBits(Align); 147 } 148 } 149 150 if (BaseInfo) 151 *BaseInfo = LValueBaseInfo(AlignmentSource::Type); 152 153 CharUnits Alignment; 154 if (T->isIncompleteType()) { 155 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best. 156 } else { 157 // For C++ class pointees, we don't know whether we're pointing at a 158 // base or a complete object, so we generally need to use the 159 // non-virtual alignment. 160 const CXXRecordDecl *RD; 161 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) { 162 Alignment = CGM.getClassPointerAlignment(RD); 163 } else { 164 Alignment = getContext().getTypeAlignInChars(T); 165 if (T.getQualifiers().hasUnaligned()) 166 Alignment = CharUnits::One(); 167 } 168 169 // Cap to the global maximum type alignment unless the alignment 170 // was somehow explicit on the type. 171 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) { 172 if (Alignment.getQuantity() > MaxAlign && 173 !getContext().isAlignmentRequired(T)) 174 Alignment = CharUnits::fromQuantity(MaxAlign); 175 } 176 } 177 return Alignment; 178 } 179 180 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) { 181 LValueBaseInfo BaseInfo; 182 TBAAAccessInfo TBAAInfo; 183 CharUnits Alignment = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo); 184 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo, 185 TBAAInfo); 186 } 187 188 /// Given a value of type T* that may not be to a complete object, 189 /// construct an l-value with the natural pointee alignment of T. 190 LValue 191 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) { 192 LValueBaseInfo BaseInfo; 193 TBAAAccessInfo TBAAInfo; 194 CharUnits Align = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo, 195 /* forPointeeType= */ true); 196 return MakeAddrLValue(Address(V, Align), T, BaseInfo, TBAAInfo); 197 } 198 199 200 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) { 201 return CGM.getTypes().ConvertTypeForMem(T); 202 } 203 204 llvm::Type *CodeGenFunction::ConvertType(QualType T) { 205 return CGM.getTypes().ConvertType(T); 206 } 207 208 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) { 209 type = type.getCanonicalType(); 210 while (true) { 211 switch (type->getTypeClass()) { 212 #define TYPE(name, parent) 213 #define ABSTRACT_TYPE(name, parent) 214 #define NON_CANONICAL_TYPE(name, parent) case Type::name: 215 #define DEPENDENT_TYPE(name, parent) case Type::name: 216 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name: 217 #include "clang/AST/TypeNodes.def" 218 llvm_unreachable("non-canonical or dependent type in IR-generation"); 219 220 case Type::Auto: 221 case Type::DeducedTemplateSpecialization: 222 llvm_unreachable("undeduced type in IR-generation"); 223 224 // Various scalar types. 225 case Type::Builtin: 226 case Type::Pointer: 227 case Type::BlockPointer: 228 case Type::LValueReference: 229 case Type::RValueReference: 230 case Type::MemberPointer: 231 case Type::Vector: 232 case Type::ExtVector: 233 case Type::FunctionProto: 234 case Type::FunctionNoProto: 235 case Type::Enum: 236 case Type::ObjCObjectPointer: 237 case Type::Pipe: 238 return TEK_Scalar; 239 240 // Complexes. 241 case Type::Complex: 242 return TEK_Complex; 243 244 // Arrays, records, and Objective-C objects. 245 case Type::ConstantArray: 246 case Type::IncompleteArray: 247 case Type::VariableArray: 248 case Type::Record: 249 case Type::ObjCObject: 250 case Type::ObjCInterface: 251 return TEK_Aggregate; 252 253 // We operate on atomic values according to their underlying type. 254 case Type::Atomic: 255 type = cast<AtomicType>(type)->getValueType(); 256 continue; 257 } 258 llvm_unreachable("unknown type kind!"); 259 } 260 } 261 262 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() { 263 // For cleanliness, we try to avoid emitting the return block for 264 // simple cases. 265 llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); 266 267 if (CurBB) { 268 assert(!CurBB->getTerminator() && "Unexpected terminated block."); 269 270 // We have a valid insert point, reuse it if it is empty or there are no 271 // explicit jumps to the return block. 272 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) { 273 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB); 274 delete ReturnBlock.getBlock(); 275 } else 276 EmitBlock(ReturnBlock.getBlock()); 277 return llvm::DebugLoc(); 278 } 279 280 // Otherwise, if the return block is the target of a single direct 281 // branch then we can just put the code in that block instead. This 282 // cleans up functions which started with a unified return block. 283 if (ReturnBlock.getBlock()->hasOneUse()) { 284 llvm::BranchInst *BI = 285 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin()); 286 if (BI && BI->isUnconditional() && 287 BI->getSuccessor(0) == ReturnBlock.getBlock()) { 288 // Record/return the DebugLoc of the simple 'return' expression to be used 289 // later by the actual 'ret' instruction. 290 llvm::DebugLoc Loc = BI->getDebugLoc(); 291 Builder.SetInsertPoint(BI->getParent()); 292 BI->eraseFromParent(); 293 delete ReturnBlock.getBlock(); 294 return Loc; 295 } 296 } 297 298 // FIXME: We are at an unreachable point, there is no reason to emit the block 299 // unless it has uses. However, we still need a place to put the debug 300 // region.end for now. 301 302 EmitBlock(ReturnBlock.getBlock()); 303 return llvm::DebugLoc(); 304 } 305 306 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) { 307 if (!BB) return; 308 if (!BB->use_empty()) 309 return CGF.CurFn->getBasicBlockList().push_back(BB); 310 delete BB; 311 } 312 313 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) { 314 assert(BreakContinueStack.empty() && 315 "mismatched push/pop in break/continue stack!"); 316 317 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0 318 && NumSimpleReturnExprs == NumReturnExprs 319 && ReturnBlock.getBlock()->use_empty(); 320 // Usually the return expression is evaluated before the cleanup 321 // code. If the function contains only a simple return statement, 322 // such as a constant, the location before the cleanup code becomes 323 // the last useful breakpoint in the function, because the simple 324 // return expression will be evaluated after the cleanup code. To be 325 // safe, set the debug location for cleanup code to the location of 326 // the return statement. Otherwise the cleanup code should be at the 327 // end of the function's lexical scope. 328 // 329 // If there are multiple branches to the return block, the branch 330 // instructions will get the location of the return statements and 331 // all will be fine. 332 if (CGDebugInfo *DI = getDebugInfo()) { 333 if (OnlySimpleReturnStmts) 334 DI->EmitLocation(Builder, LastStopPoint); 335 else 336 DI->EmitLocation(Builder, EndLoc); 337 } 338 339 // Pop any cleanups that might have been associated with the 340 // parameters. Do this in whatever block we're currently in; it's 341 // important to do this before we enter the return block or return 342 // edges will be *really* confused. 343 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth; 344 bool HasOnlyLifetimeMarkers = 345 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth); 346 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers; 347 if (HasCleanups) { 348 // Make sure the line table doesn't jump back into the body for 349 // the ret after it's been at EndLoc. 350 if (CGDebugInfo *DI = getDebugInfo()) 351 if (OnlySimpleReturnStmts) 352 DI->EmitLocation(Builder, EndLoc); 353 354 PopCleanupBlocks(PrologueCleanupDepth); 355 } 356 357 // Emit function epilog (to return). 358 llvm::DebugLoc Loc = EmitReturnBlock(); 359 360 if (ShouldInstrumentFunction()) { 361 if (CGM.getCodeGenOpts().InstrumentFunctions) 362 CurFn->addFnAttr("instrument-function-exit", "__cyg_profile_func_exit"); 363 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining) 364 CurFn->addFnAttr("instrument-function-exit-inlined", 365 "__cyg_profile_func_exit"); 366 } 367 368 // Emit debug descriptor for function end. 369 if (CGDebugInfo *DI = getDebugInfo()) 370 DI->EmitFunctionEnd(Builder, CurFn); 371 372 // Reset the debug location to that of the simple 'return' expression, if any 373 // rather than that of the end of the function's scope '}'. 374 ApplyDebugLocation AL(*this, Loc); 375 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc); 376 EmitEndEHSpec(CurCodeDecl); 377 378 assert(EHStack.empty() && 379 "did not remove all scopes from cleanup stack!"); 380 381 // If someone did an indirect goto, emit the indirect goto block at the end of 382 // the function. 383 if (IndirectBranch) { 384 EmitBlock(IndirectBranch->getParent()); 385 Builder.ClearInsertionPoint(); 386 } 387 388 // If some of our locals escaped, insert a call to llvm.localescape in the 389 // entry block. 390 if (!EscapedLocals.empty()) { 391 // Invert the map from local to index into a simple vector. There should be 392 // no holes. 393 SmallVector<llvm::Value *, 4> EscapeArgs; 394 EscapeArgs.resize(EscapedLocals.size()); 395 for (auto &Pair : EscapedLocals) 396 EscapeArgs[Pair.second] = Pair.first; 397 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration( 398 &CGM.getModule(), llvm::Intrinsic::localescape); 399 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs); 400 } 401 402 // Remove the AllocaInsertPt instruction, which is just a convenience for us. 403 llvm::Instruction *Ptr = AllocaInsertPt; 404 AllocaInsertPt = nullptr; 405 Ptr->eraseFromParent(); 406 407 // If someone took the address of a label but never did an indirect goto, we 408 // made a zero entry PHI node, which is illegal, zap it now. 409 if (IndirectBranch) { 410 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress()); 411 if (PN->getNumIncomingValues() == 0) { 412 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType())); 413 PN->eraseFromParent(); 414 } 415 } 416 417 EmitIfUsed(*this, EHResumeBlock); 418 EmitIfUsed(*this, TerminateLandingPad); 419 EmitIfUsed(*this, TerminateHandler); 420 EmitIfUsed(*this, UnreachableBlock); 421 422 for (const auto &FuncletAndParent : TerminateFunclets) 423 EmitIfUsed(*this, FuncletAndParent.second); 424 425 if (CGM.getCodeGenOpts().EmitDeclMetadata) 426 EmitDeclMetadata(); 427 428 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator 429 I = DeferredReplacements.begin(), 430 E = DeferredReplacements.end(); 431 I != E; ++I) { 432 I->first->replaceAllUsesWith(I->second); 433 I->first->eraseFromParent(); 434 } 435 436 // Eliminate CleanupDestSlot alloca by replacing it with SSA values and 437 // PHIs if the current function is a coroutine. We don't do it for all 438 // functions as it may result in slight increase in numbers of instructions 439 // if compiled with no optimizations. We do it for coroutine as the lifetime 440 // of CleanupDestSlot alloca make correct coroutine frame building very 441 // difficult. 442 if (NormalCleanupDest.isValid() && isCoroutine()) { 443 llvm::DominatorTree DT(*CurFn); 444 llvm::PromoteMemToReg( 445 cast<llvm::AllocaInst>(NormalCleanupDest.getPointer()), DT); 446 NormalCleanupDest = Address::invalid(); 447 } 448 } 449 450 /// ShouldInstrumentFunction - Return true if the current function should be 451 /// instrumented with __cyg_profile_func_* calls 452 bool CodeGenFunction::ShouldInstrumentFunction() { 453 if (!CGM.getCodeGenOpts().InstrumentFunctions && 454 !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining && 455 !CGM.getCodeGenOpts().InstrumentFunctionEntryBare) 456 return false; 457 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) 458 return false; 459 return true; 460 } 461 462 /// ShouldXRayInstrument - Return true if the current function should be 463 /// instrumented with XRay nop sleds. 464 bool CodeGenFunction::ShouldXRayInstrumentFunction() const { 465 return CGM.getCodeGenOpts().XRayInstrumentFunctions; 466 } 467 468 /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to 469 /// the __xray_customevent(...) builin calls, when doing XRay instrumentation. 470 bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const { 471 return CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents; 472 } 473 474 llvm::Constant * 475 CodeGenFunction::EncodeAddrForUseInPrologue(llvm::Function *F, 476 llvm::Constant *Addr) { 477 // Addresses stored in prologue data can't require run-time fixups and must 478 // be PC-relative. Run-time fixups are undesirable because they necessitate 479 // writable text segments, which are unsafe. And absolute addresses are 480 // undesirable because they break PIE mode. 481 482 // Add a layer of indirection through a private global. Taking its address 483 // won't result in a run-time fixup, even if Addr has linkonce_odr linkage. 484 auto *GV = new llvm::GlobalVariable(CGM.getModule(), Addr->getType(), 485 /*isConstant=*/true, 486 llvm::GlobalValue::PrivateLinkage, Addr); 487 488 // Create a PC-relative address. 489 auto *GOTAsInt = llvm::ConstantExpr::getPtrToInt(GV, IntPtrTy); 490 auto *FuncAsInt = llvm::ConstantExpr::getPtrToInt(F, IntPtrTy); 491 auto *PCRelAsInt = llvm::ConstantExpr::getSub(GOTAsInt, FuncAsInt); 492 return (IntPtrTy == Int32Ty) 493 ? PCRelAsInt 494 : llvm::ConstantExpr::getTrunc(PCRelAsInt, Int32Ty); 495 } 496 497 llvm::Value * 498 CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F, 499 llvm::Value *EncodedAddr) { 500 // Reconstruct the address of the global. 501 auto *PCRelAsInt = Builder.CreateSExt(EncodedAddr, IntPtrTy); 502 auto *FuncAsInt = Builder.CreatePtrToInt(F, IntPtrTy, "func_addr.int"); 503 auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int"); 504 auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr"); 505 506 // Load the original pointer through the global. 507 return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()), 508 "decoded_addr"); 509 } 510 511 static void removeImageAccessQualifier(std::string& TyName) { 512 std::string ReadOnlyQual("__read_only"); 513 std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual); 514 if (ReadOnlyPos != std::string::npos) 515 // "+ 1" for the space after access qualifier. 516 TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1); 517 else { 518 std::string WriteOnlyQual("__write_only"); 519 std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual); 520 if (WriteOnlyPos != std::string::npos) 521 TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1); 522 else { 523 std::string ReadWriteQual("__read_write"); 524 std::string::size_type ReadWritePos = TyName.find(ReadWriteQual); 525 if (ReadWritePos != std::string::npos) 526 TyName.erase(ReadWritePos, ReadWriteQual.size() + 1); 527 } 528 } 529 } 530 531 // Returns the address space id that should be produced to the 532 // kernel_arg_addr_space metadata. This is always fixed to the ids 533 // as specified in the SPIR 2.0 specification in order to differentiate 534 // for example in clGetKernelArgInfo() implementation between the address 535 // spaces with targets without unique mapping to the OpenCL address spaces 536 // (basically all single AS CPUs). 537 static unsigned ArgInfoAddressSpace(LangAS AS) { 538 switch (AS) { 539 case LangAS::opencl_global: return 1; 540 case LangAS::opencl_constant: return 2; 541 case LangAS::opencl_local: return 3; 542 case LangAS::opencl_generic: return 4; // Not in SPIR 2.0 specs. 543 default: 544 return 0; // Assume private. 545 } 546 } 547 548 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument 549 // information in the program executable. The argument information stored 550 // includes the argument name, its type, the address and access qualifiers used. 551 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn, 552 CodeGenModule &CGM, llvm::LLVMContext &Context, 553 CGBuilderTy &Builder, ASTContext &ASTCtx) { 554 // Create MDNodes that represent the kernel arg metadata. 555 // Each MDNode is a list in the form of "key", N number of values which is 556 // the same number of values as their are kernel arguments. 557 558 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy(); 559 560 // MDNode for the kernel argument address space qualifiers. 561 SmallVector<llvm::Metadata *, 8> addressQuals; 562 563 // MDNode for the kernel argument access qualifiers (images only). 564 SmallVector<llvm::Metadata *, 8> accessQuals; 565 566 // MDNode for the kernel argument type names. 567 SmallVector<llvm::Metadata *, 8> argTypeNames; 568 569 // MDNode for the kernel argument base type names. 570 SmallVector<llvm::Metadata *, 8> argBaseTypeNames; 571 572 // MDNode for the kernel argument type qualifiers. 573 SmallVector<llvm::Metadata *, 8> argTypeQuals; 574 575 // MDNode for the kernel argument names. 576 SmallVector<llvm::Metadata *, 8> argNames; 577 578 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) { 579 const ParmVarDecl *parm = FD->getParamDecl(i); 580 QualType ty = parm->getType(); 581 std::string typeQuals; 582 583 if (ty->isPointerType()) { 584 QualType pointeeTy = ty->getPointeeType(); 585 586 // Get address qualifier. 587 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32( 588 ArgInfoAddressSpace(pointeeTy.getAddressSpace())))); 589 590 // Get argument type name. 591 std::string typeName = 592 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*"; 593 594 // Turn "unsigned type" to "utype" 595 std::string::size_type pos = typeName.find("unsigned"); 596 if (pointeeTy.isCanonical() && pos != std::string::npos) 597 typeName.erase(pos+1, 8); 598 599 argTypeNames.push_back(llvm::MDString::get(Context, typeName)); 600 601 std::string baseTypeName = 602 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString( 603 Policy) + 604 "*"; 605 606 // Turn "unsigned type" to "utype" 607 pos = baseTypeName.find("unsigned"); 608 if (pos != std::string::npos) 609 baseTypeName.erase(pos+1, 8); 610 611 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName)); 612 613 // Get argument type qualifiers: 614 if (ty.isRestrictQualified()) 615 typeQuals = "restrict"; 616 if (pointeeTy.isConstQualified() || 617 (pointeeTy.getAddressSpace() == LangAS::opencl_constant)) 618 typeQuals += typeQuals.empty() ? "const" : " const"; 619 if (pointeeTy.isVolatileQualified()) 620 typeQuals += typeQuals.empty() ? "volatile" : " volatile"; 621 } else { 622 uint32_t AddrSpc = 0; 623 bool isPipe = ty->isPipeType(); 624 if (ty->isImageType() || isPipe) 625 AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global); 626 627 addressQuals.push_back( 628 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc))); 629 630 // Get argument type name. 631 std::string typeName; 632 if (isPipe) 633 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType() 634 .getAsString(Policy); 635 else 636 typeName = ty.getUnqualifiedType().getAsString(Policy); 637 638 // Turn "unsigned type" to "utype" 639 std::string::size_type pos = typeName.find("unsigned"); 640 if (ty.isCanonical() && pos != std::string::npos) 641 typeName.erase(pos+1, 8); 642 643 std::string baseTypeName; 644 if (isPipe) 645 baseTypeName = ty.getCanonicalType()->getAs<PipeType>() 646 ->getElementType().getCanonicalType() 647 .getAsString(Policy); 648 else 649 baseTypeName = 650 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy); 651 652 // Remove access qualifiers on images 653 // (as they are inseparable from type in clang implementation, 654 // but OpenCL spec provides a special query to get access qualifier 655 // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER): 656 if (ty->isImageType()) { 657 removeImageAccessQualifier(typeName); 658 removeImageAccessQualifier(baseTypeName); 659 } 660 661 argTypeNames.push_back(llvm::MDString::get(Context, typeName)); 662 663 // Turn "unsigned type" to "utype" 664 pos = baseTypeName.find("unsigned"); 665 if (pos != std::string::npos) 666 baseTypeName.erase(pos+1, 8); 667 668 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName)); 669 670 if (isPipe) 671 typeQuals = "pipe"; 672 } 673 674 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals)); 675 676 // Get image and pipe access qualifier: 677 if (ty->isImageType()|| ty->isPipeType()) { 678 const Decl *PDecl = parm; 679 if (auto *TD = dyn_cast<TypedefType>(ty)) 680 PDecl = TD->getDecl(); 681 const OpenCLAccessAttr *A = PDecl->getAttr<OpenCLAccessAttr>(); 682 if (A && A->isWriteOnly()) 683 accessQuals.push_back(llvm::MDString::get(Context, "write_only")); 684 else if (A && A->isReadWrite()) 685 accessQuals.push_back(llvm::MDString::get(Context, "read_write")); 686 else 687 accessQuals.push_back(llvm::MDString::get(Context, "read_only")); 688 } else 689 accessQuals.push_back(llvm::MDString::get(Context, "none")); 690 691 // Get argument name. 692 argNames.push_back(llvm::MDString::get(Context, parm->getName())); 693 } 694 695 Fn->setMetadata("kernel_arg_addr_space", 696 llvm::MDNode::get(Context, addressQuals)); 697 Fn->setMetadata("kernel_arg_access_qual", 698 llvm::MDNode::get(Context, accessQuals)); 699 Fn->setMetadata("kernel_arg_type", 700 llvm::MDNode::get(Context, argTypeNames)); 701 Fn->setMetadata("kernel_arg_base_type", 702 llvm::MDNode::get(Context, argBaseTypeNames)); 703 Fn->setMetadata("kernel_arg_type_qual", 704 llvm::MDNode::get(Context, argTypeQuals)); 705 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata) 706 Fn->setMetadata("kernel_arg_name", 707 llvm::MDNode::get(Context, argNames)); 708 } 709 710 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD, 711 llvm::Function *Fn) 712 { 713 if (!FD->hasAttr<OpenCLKernelAttr>()) 714 return; 715 716 llvm::LLVMContext &Context = getLLVMContext(); 717 718 GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext()); 719 720 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) { 721 QualType HintQTy = A->getTypeHint(); 722 const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>(); 723 bool IsSignedInteger = 724 HintQTy->isSignedIntegerType() || 725 (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType()); 726 llvm::Metadata *AttrMDArgs[] = { 727 llvm::ConstantAsMetadata::get(llvm::UndefValue::get( 728 CGM.getTypes().ConvertType(A->getTypeHint()))), 729 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get( 730 llvm::IntegerType::get(Context, 32), 731 llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))}; 732 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs)); 733 } 734 735 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) { 736 llvm::Metadata *AttrMDArgs[] = { 737 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())), 738 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())), 739 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))}; 740 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs)); 741 } 742 743 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) { 744 llvm::Metadata *AttrMDArgs[] = { 745 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())), 746 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())), 747 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))}; 748 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs)); 749 } 750 751 if (const OpenCLIntelReqdSubGroupSizeAttr *A = 752 FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) { 753 llvm::Metadata *AttrMDArgs[] = { 754 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))}; 755 Fn->setMetadata("intel_reqd_sub_group_size", 756 llvm::MDNode::get(Context, AttrMDArgs)); 757 } 758 } 759 760 /// Determine whether the function F ends with a return stmt. 761 static bool endsWithReturn(const Decl* F) { 762 const Stmt *Body = nullptr; 763 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F)) 764 Body = FD->getBody(); 765 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F)) 766 Body = OMD->getBody(); 767 768 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) { 769 auto LastStmt = CS->body_rbegin(); 770 if (LastStmt != CS->body_rend()) 771 return isa<ReturnStmt>(*LastStmt); 772 } 773 return false; 774 } 775 776 static void markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) { 777 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time"); 778 Fn->removeFnAttr(llvm::Attribute::SanitizeThread); 779 } 780 781 static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) { 782 auto *MD = dyn_cast_or_null<CXXMethodDecl>(D); 783 if (!MD || !MD->getDeclName().getAsIdentifierInfo() || 784 !MD->getDeclName().getAsIdentifierInfo()->isStr("allocate") || 785 (MD->getNumParams() != 1 && MD->getNumParams() != 2)) 786 return false; 787 788 if (MD->parameters()[0]->getType().getCanonicalType() != Ctx.getSizeType()) 789 return false; 790 791 if (MD->getNumParams() == 2) { 792 auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>(); 793 if (!PT || !PT->isVoidPointerType() || 794 !PT->getPointeeType().isConstQualified()) 795 return false; 796 } 797 798 return true; 799 } 800 801 /// Return the UBSan prologue signature for \p FD if one is available. 802 static llvm::Constant *getPrologueSignature(CodeGenModule &CGM, 803 const FunctionDecl *FD) { 804 if (const auto *MD = dyn_cast<CXXMethodDecl>(FD)) 805 if (!MD->isStatic()) 806 return nullptr; 807 return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM); 808 } 809 810 void CodeGenFunction::StartFunction(GlobalDecl GD, 811 QualType RetTy, 812 llvm::Function *Fn, 813 const CGFunctionInfo &FnInfo, 814 const FunctionArgList &Args, 815 SourceLocation Loc, 816 SourceLocation StartLoc) { 817 assert(!CurFn && 818 "Do not use a CodeGenFunction object for more than one function"); 819 820 const Decl *D = GD.getDecl(); 821 822 DidCallStackSave = false; 823 CurCodeDecl = D; 824 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) 825 if (FD->usesSEHTry()) 826 CurSEHParent = FD; 827 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr); 828 FnRetTy = RetTy; 829 CurFn = Fn; 830 CurFnInfo = &FnInfo; 831 assert(CurFn->isDeclaration() && "Function already has body?"); 832 833 // If this function has been blacklisted for any of the enabled sanitizers, 834 // disable the sanitizer for the function. 835 do { 836 #define SANITIZER(NAME, ID) \ 837 if (SanOpts.empty()) \ 838 break; \ 839 if (SanOpts.has(SanitizerKind::ID)) \ 840 if (CGM.isInSanitizerBlacklist(SanitizerKind::ID, Fn, Loc)) \ 841 SanOpts.set(SanitizerKind::ID, false); 842 843 #include "clang/Basic/Sanitizers.def" 844 #undef SANITIZER 845 } while (0); 846 847 if (D) { 848 // Apply the no_sanitize* attributes to SanOpts. 849 for (auto Attr : D->specific_attrs<NoSanitizeAttr>()) 850 SanOpts.Mask &= ~Attr->getMask(); 851 } 852 853 // Apply sanitizer attributes to the function. 854 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress)) 855 Fn->addFnAttr(llvm::Attribute::SanitizeAddress); 856 if (SanOpts.hasOneOf(SanitizerKind::HWAddress)) 857 Fn->addFnAttr(llvm::Attribute::SanitizeHWAddress); 858 if (SanOpts.has(SanitizerKind::Thread)) 859 Fn->addFnAttr(llvm::Attribute::SanitizeThread); 860 if (SanOpts.has(SanitizerKind::Memory)) 861 Fn->addFnAttr(llvm::Attribute::SanitizeMemory); 862 if (SanOpts.has(SanitizerKind::SafeStack)) 863 Fn->addFnAttr(llvm::Attribute::SafeStack); 864 865 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize, 866 // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time. 867 if (SanOpts.has(SanitizerKind::Thread)) { 868 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) { 869 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0); 870 if (OMD->getMethodFamily() == OMF_dealloc || 871 OMD->getMethodFamily() == OMF_initialize || 872 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) { 873 markAsIgnoreThreadCheckingAtRuntime(Fn); 874 } 875 } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) { 876 IdentifierInfo *II = FD->getIdentifier(); 877 if (II && II->isStr("__destroy_helper_block_")) 878 markAsIgnoreThreadCheckingAtRuntime(Fn); 879 } 880 } 881 882 // Ignore unrelated casts in STL allocate() since the allocator must cast 883 // from void* to T* before object initialization completes. Don't match on the 884 // namespace because not all allocators are in std:: 885 if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) { 886 if (matchesStlAllocatorFn(D, getContext())) 887 SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast; 888 } 889 890 // Apply xray attributes to the function (as a string, for now) 891 bool InstrumentXray = ShouldXRayInstrumentFunction(); 892 if (D && InstrumentXray) { 893 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) { 894 if (XRayAttr->alwaysXRayInstrument()) 895 Fn->addFnAttr("function-instrument", "xray-always"); 896 if (XRayAttr->neverXRayInstrument()) 897 Fn->addFnAttr("function-instrument", "xray-never"); 898 if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) { 899 Fn->addFnAttr("xray-log-args", 900 llvm::utostr(LogArgs->getArgumentCount())); 901 } 902 } else { 903 if (!CGM.imbueXRayAttrs(Fn, Loc)) 904 Fn->addFnAttr( 905 "xray-instruction-threshold", 906 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold)); 907 } 908 } 909 910 // Add no-jump-tables value. 911 Fn->addFnAttr("no-jump-tables", 912 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables)); 913 914 // Add profile-sample-accurate value. 915 if (CGM.getCodeGenOpts().ProfileSampleAccurate) 916 Fn->addFnAttr("profile-sample-accurate"); 917 918 if (getLangOpts().OpenCL) { 919 // Add metadata for a kernel function. 920 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) 921 EmitOpenCLKernelMetadata(FD, Fn); 922 } 923 924 // If we are checking function types, emit a function type signature as 925 // prologue data. 926 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) { 927 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) { 928 if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) { 929 // Remove any (C++17) exception specifications, to allow calling e.g. a 930 // noexcept function through a non-noexcept pointer. 931 auto ProtoTy = 932 getContext().getFunctionTypeWithExceptionSpec(FD->getType(), 933 EST_None); 934 llvm::Constant *FTRTTIConst = 935 CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true); 936 llvm::Constant *FTRTTIConstEncoded = 937 EncodeAddrForUseInPrologue(Fn, FTRTTIConst); 938 llvm::Constant *PrologueStructElems[] = {PrologueSig, 939 FTRTTIConstEncoded}; 940 llvm::Constant *PrologueStructConst = 941 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true); 942 Fn->setPrologueData(PrologueStructConst); 943 } 944 } 945 } 946 947 // If we're checking nullability, we need to know whether we can check the 948 // return value. Initialize the flag to 'true' and refine it in EmitParmDecl. 949 if (SanOpts.has(SanitizerKind::NullabilityReturn)) { 950 auto Nullability = FnRetTy->getNullability(getContext()); 951 if (Nullability && *Nullability == NullabilityKind::NonNull) { 952 if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) && 953 CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>())) 954 RetValNullabilityPrecondition = 955 llvm::ConstantInt::getTrue(getLLVMContext()); 956 } 957 } 958 959 // If we're in C++ mode and the function name is "main", it is guaranteed 960 // to be norecurse by the standard (3.6.1.3 "The function main shall not be 961 // used within a program"). 962 if (getLangOpts().CPlusPlus) 963 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) 964 if (FD->isMain()) 965 Fn->addFnAttr(llvm::Attribute::NoRecurse); 966 967 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn); 968 969 // Create a marker to make it easy to insert allocas into the entryblock 970 // later. Don't create this with the builder, because we don't want it 971 // folded. 972 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty); 973 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB); 974 975 ReturnBlock = getJumpDestInCurrentScope("return"); 976 977 Builder.SetInsertPoint(EntryBB); 978 979 // If we're checking the return value, allocate space for a pointer to a 980 // precise source location of the checked return statement. 981 if (requiresReturnValueCheck()) { 982 ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr"); 983 InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy)); 984 } 985 986 // Emit subprogram debug descriptor. 987 if (CGDebugInfo *DI = getDebugInfo()) { 988 // Reconstruct the type from the argument list so that implicit parameters, 989 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling 990 // convention. 991 CallingConv CC = CallingConv::CC_C; 992 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D)) 993 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>()) 994 CC = SrcFnTy->getCallConv(); 995 SmallVector<QualType, 16> ArgTypes; 996 for (const VarDecl *VD : Args) 997 ArgTypes.push_back(VD->getType()); 998 QualType FnType = getContext().getFunctionType( 999 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC)); 1000 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder); 1001 } 1002 1003 if (ShouldInstrumentFunction()) { 1004 if (CGM.getCodeGenOpts().InstrumentFunctions) 1005 CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter"); 1006 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining) 1007 CurFn->addFnAttr("instrument-function-entry-inlined", 1008 "__cyg_profile_func_enter"); 1009 if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare) 1010 CurFn->addFnAttr("instrument-function-entry-inlined", 1011 "__cyg_profile_func_enter_bare"); 1012 } 1013 1014 // Since emitting the mcount call here impacts optimizations such as function 1015 // inlining, we just add an attribute to insert a mcount call in backend. 1016 // The attribute "counting-function" is set to mcount function name which is 1017 // architecture dependent. 1018 if (CGM.getCodeGenOpts().InstrumentForProfiling) { 1019 if (CGM.getCodeGenOpts().CallFEntry) 1020 Fn->addFnAttr("fentry-call", "true"); 1021 else { 1022 if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) { 1023 Fn->addFnAttr("instrument-function-entry-inlined", 1024 getTarget().getMCountName()); 1025 } 1026 } 1027 } 1028 1029 if (RetTy->isVoidType()) { 1030 // Void type; nothing to return. 1031 ReturnValue = Address::invalid(); 1032 1033 // Count the implicit return. 1034 if (!endsWithReturn(D)) 1035 ++NumReturnExprs; 1036 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect && 1037 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) { 1038 // Indirect aggregate return; emit returned value directly into sret slot. 1039 // This reduces code size, and affects correctness in C++. 1040 auto AI = CurFn->arg_begin(); 1041 if (CurFnInfo->getReturnInfo().isSRetAfterThis()) 1042 ++AI; 1043 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign()); 1044 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca && 1045 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) { 1046 // Load the sret pointer from the argument struct and return into that. 1047 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex(); 1048 llvm::Function::arg_iterator EI = CurFn->arg_end(); 1049 --EI; 1050 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx); 1051 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result"); 1052 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy)); 1053 } else { 1054 ReturnValue = CreateIRTemp(RetTy, "retval"); 1055 1056 // Tell the epilog emitter to autorelease the result. We do this 1057 // now so that various specialized functions can suppress it 1058 // during their IR-generation. 1059 if (getLangOpts().ObjCAutoRefCount && 1060 !CurFnInfo->isReturnsRetained() && 1061 RetTy->isObjCRetainableType()) 1062 AutoreleaseResult = true; 1063 } 1064 1065 EmitStartEHSpec(CurCodeDecl); 1066 1067 PrologueCleanupDepth = EHStack.stable_begin(); 1068 EmitFunctionProlog(*CurFnInfo, CurFn, Args); 1069 1070 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) { 1071 CGM.getCXXABI().EmitInstanceFunctionProlog(*this); 1072 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D); 1073 if (MD->getParent()->isLambda() && 1074 MD->getOverloadedOperator() == OO_Call) { 1075 // We're in a lambda; figure out the captures. 1076 MD->getParent()->getCaptureFields(LambdaCaptureFields, 1077 LambdaThisCaptureField); 1078 if (LambdaThisCaptureField) { 1079 // If the lambda captures the object referred to by '*this' - either by 1080 // value or by reference, make sure CXXThisValue points to the correct 1081 // object. 1082 1083 // Get the lvalue for the field (which is a copy of the enclosing object 1084 // or contains the address of the enclosing object). 1085 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField); 1086 if (!LambdaThisCaptureField->getType()->isPointerType()) { 1087 // If the enclosing object was captured by value, just use its address. 1088 CXXThisValue = ThisFieldLValue.getAddress().getPointer(); 1089 } else { 1090 // Load the lvalue pointed to by the field, since '*this' was captured 1091 // by reference. 1092 CXXThisValue = 1093 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal(); 1094 } 1095 } 1096 for (auto *FD : MD->getParent()->fields()) { 1097 if (FD->hasCapturedVLAType()) { 1098 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD), 1099 SourceLocation()).getScalarVal(); 1100 auto VAT = FD->getCapturedVLAType(); 1101 VLASizeMap[VAT->getSizeExpr()] = ExprArg; 1102 } 1103 } 1104 } else { 1105 // Not in a lambda; just use 'this' from the method. 1106 // FIXME: Should we generate a new load for each use of 'this'? The 1107 // fast register allocator would be happier... 1108 CXXThisValue = CXXABIThisValue; 1109 } 1110 1111 // Check the 'this' pointer once per function, if it's available. 1112 if (CXXABIThisValue) { 1113 SanitizerSet SkippedChecks; 1114 SkippedChecks.set(SanitizerKind::ObjectSize, true); 1115 QualType ThisTy = MD->getThisType(getContext()); 1116 1117 // If this is the call operator of a lambda with no capture-default, it 1118 // may have a static invoker function, which may call this operator with 1119 // a null 'this' pointer. 1120 if (isLambdaCallOperator(MD) && 1121 cast<CXXRecordDecl>(MD->getParent())->getLambdaCaptureDefault() == 1122 LCD_None) 1123 SkippedChecks.set(SanitizerKind::Null, true); 1124 1125 EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall 1126 : TCK_MemberCall, 1127 Loc, CXXABIThisValue, ThisTy, 1128 getContext().getTypeAlignInChars(ThisTy->getPointeeType()), 1129 SkippedChecks); 1130 } 1131 } 1132 1133 // If any of the arguments have a variably modified type, make sure to 1134 // emit the type size. 1135 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 1136 i != e; ++i) { 1137 const VarDecl *VD = *i; 1138 1139 // Dig out the type as written from ParmVarDecls; it's unclear whether 1140 // the standard (C99 6.9.1p10) requires this, but we're following the 1141 // precedent set by gcc. 1142 QualType Ty; 1143 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD)) 1144 Ty = PVD->getOriginalType(); 1145 else 1146 Ty = VD->getType(); 1147 1148 if (Ty->isVariablyModifiedType()) 1149 EmitVariablyModifiedType(Ty); 1150 } 1151 // Emit a location at the end of the prologue. 1152 if (CGDebugInfo *DI = getDebugInfo()) 1153 DI->EmitLocation(Builder, StartLoc); 1154 } 1155 1156 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args, 1157 const Stmt *Body) { 1158 incrementProfileCounter(Body); 1159 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body)) 1160 EmitCompoundStmtWithoutScope(*S); 1161 else 1162 EmitStmt(Body); 1163 } 1164 1165 /// When instrumenting to collect profile data, the counts for some blocks 1166 /// such as switch cases need to not include the fall-through counts, so 1167 /// emit a branch around the instrumentation code. When not instrumenting, 1168 /// this just calls EmitBlock(). 1169 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB, 1170 const Stmt *S) { 1171 llvm::BasicBlock *SkipCountBB = nullptr; 1172 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) { 1173 // When instrumenting for profiling, the fallthrough to certain 1174 // statements needs to skip over the instrumentation code so that we 1175 // get an accurate count. 1176 SkipCountBB = createBasicBlock("skipcount"); 1177 EmitBranch(SkipCountBB); 1178 } 1179 EmitBlock(BB); 1180 uint64_t CurrentCount = getCurrentProfileCount(); 1181 incrementProfileCounter(S); 1182 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount); 1183 if (SkipCountBB) 1184 EmitBlock(SkipCountBB); 1185 } 1186 1187 /// Tries to mark the given function nounwind based on the 1188 /// non-existence of any throwing calls within it. We believe this is 1189 /// lightweight enough to do at -O0. 1190 static void TryMarkNoThrow(llvm::Function *F) { 1191 // LLVM treats 'nounwind' on a function as part of the type, so we 1192 // can't do this on functions that can be overwritten. 1193 if (F->isInterposable()) return; 1194 1195 for (llvm::BasicBlock &BB : *F) 1196 for (llvm::Instruction &I : BB) 1197 if (I.mayThrow()) 1198 return; 1199 1200 F->setDoesNotThrow(); 1201 } 1202 1203 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD, 1204 FunctionArgList &Args) { 1205 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl()); 1206 QualType ResTy = FD->getReturnType(); 1207 1208 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD); 1209 if (MD && MD->isInstance()) { 1210 if (CGM.getCXXABI().HasThisReturn(GD)) 1211 ResTy = MD->getThisType(getContext()); 1212 else if (CGM.getCXXABI().hasMostDerivedReturn(GD)) 1213 ResTy = CGM.getContext().VoidPtrTy; 1214 CGM.getCXXABI().buildThisParam(*this, Args); 1215 } 1216 1217 // The base version of an inheriting constructor whose constructed base is a 1218 // virtual base is not passed any arguments (because it doesn't actually call 1219 // the inherited constructor). 1220 bool PassedParams = true; 1221 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD)) 1222 if (auto Inherited = CD->getInheritedConstructor()) 1223 PassedParams = 1224 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType()); 1225 1226 if (PassedParams) { 1227 for (auto *Param : FD->parameters()) { 1228 Args.push_back(Param); 1229 if (!Param->hasAttr<PassObjectSizeAttr>()) 1230 continue; 1231 1232 auto *Implicit = ImplicitParamDecl::Create( 1233 getContext(), Param->getDeclContext(), Param->getLocation(), 1234 /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other); 1235 SizeArguments[Param] = Implicit; 1236 Args.push_back(Implicit); 1237 } 1238 } 1239 1240 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD))) 1241 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args); 1242 1243 return ResTy; 1244 } 1245 1246 static bool 1247 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD, 1248 const ASTContext &Context) { 1249 QualType T = FD->getReturnType(); 1250 // Avoid the optimization for functions that return a record type with a 1251 // trivial destructor or another trivially copyable type. 1252 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) { 1253 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl())) 1254 return !ClassDecl->hasTrivialDestructor(); 1255 } 1256 return !T.isTriviallyCopyableType(Context); 1257 } 1258 1259 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn, 1260 const CGFunctionInfo &FnInfo) { 1261 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl()); 1262 CurGD = GD; 1263 1264 FunctionArgList Args; 1265 QualType ResTy = BuildFunctionArgList(GD, Args); 1266 1267 // Check if we should generate debug info for this function. 1268 if (FD->hasAttr<NoDebugAttr>()) 1269 DebugInfo = nullptr; // disable debug info indefinitely for this function 1270 1271 // The function might not have a body if we're generating thunks for a 1272 // function declaration. 1273 SourceRange BodyRange; 1274 if (Stmt *Body = FD->getBody()) 1275 BodyRange = Body->getSourceRange(); 1276 else 1277 BodyRange = FD->getLocation(); 1278 CurEHLocation = BodyRange.getEnd(); 1279 1280 // Use the location of the start of the function to determine where 1281 // the function definition is located. By default use the location 1282 // of the declaration as the location for the subprogram. A function 1283 // may lack a declaration in the source code if it is created by code 1284 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk). 1285 SourceLocation Loc = FD->getLocation(); 1286 1287 // If this is a function specialization then use the pattern body 1288 // as the location for the function. 1289 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern()) 1290 if (SpecDecl->hasBody(SpecDecl)) 1291 Loc = SpecDecl->getLocation(); 1292 1293 Stmt *Body = FD->getBody(); 1294 1295 // Initialize helper which will detect jumps which can cause invalid lifetime 1296 // markers. 1297 if (Body && ShouldEmitLifetimeMarkers) 1298 Bypasses.Init(Body); 1299 1300 // Emit the standard function prologue. 1301 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin()); 1302 1303 // Generate the body of the function. 1304 PGO.assignRegionCounters(GD, CurFn); 1305 if (isa<CXXDestructorDecl>(FD)) 1306 EmitDestructorBody(Args); 1307 else if (isa<CXXConstructorDecl>(FD)) 1308 EmitConstructorBody(Args); 1309 else if (getLangOpts().CUDA && 1310 !getLangOpts().CUDAIsDevice && 1311 FD->hasAttr<CUDAGlobalAttr>()) 1312 CGM.getCUDARuntime().emitDeviceStub(*this, Args); 1313 else if (isa<CXXMethodDecl>(FD) && 1314 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) { 1315 // The lambda static invoker function is special, because it forwards or 1316 // clones the body of the function call operator (but is actually static). 1317 EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD)); 1318 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) && 1319 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() || 1320 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) { 1321 // Implicit copy-assignment gets the same special treatment as implicit 1322 // copy-constructors. 1323 emitImplicitAssignmentOperatorBody(Args); 1324 } else if (Body) { 1325 EmitFunctionBody(Args, Body); 1326 } else 1327 llvm_unreachable("no definition for emitted function"); 1328 1329 // C++11 [stmt.return]p2: 1330 // Flowing off the end of a function [...] results in undefined behavior in 1331 // a value-returning function. 1332 // C11 6.9.1p12: 1333 // If the '}' that terminates a function is reached, and the value of the 1334 // function call is used by the caller, the behavior is undefined. 1335 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock && 1336 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) { 1337 bool ShouldEmitUnreachable = 1338 CGM.getCodeGenOpts().StrictReturn || 1339 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext()); 1340 if (SanOpts.has(SanitizerKind::Return)) { 1341 SanitizerScope SanScope(this); 1342 llvm::Value *IsFalse = Builder.getFalse(); 1343 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return), 1344 SanitizerHandler::MissingReturn, 1345 EmitCheckSourceLocation(FD->getLocation()), None); 1346 } else if (ShouldEmitUnreachable) { 1347 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 1348 EmitTrapCall(llvm::Intrinsic::trap); 1349 } 1350 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) { 1351 Builder.CreateUnreachable(); 1352 Builder.ClearInsertionPoint(); 1353 } 1354 } 1355 1356 // Emit the standard function epilogue. 1357 FinishFunction(BodyRange.getEnd()); 1358 1359 // If we haven't marked the function nothrow through other means, do 1360 // a quick pass now to see if we can. 1361 if (!CurFn->doesNotThrow()) 1362 TryMarkNoThrow(CurFn); 1363 } 1364 1365 /// ContainsLabel - Return true if the statement contains a label in it. If 1366 /// this statement is not executed normally, it not containing a label means 1367 /// that we can just remove the code. 1368 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) { 1369 // Null statement, not a label! 1370 if (!S) return false; 1371 1372 // If this is a label, we have to emit the code, consider something like: 1373 // if (0) { ... foo: bar(); } goto foo; 1374 // 1375 // TODO: If anyone cared, we could track __label__'s, since we know that you 1376 // can't jump to one from outside their declared region. 1377 if (isa<LabelStmt>(S)) 1378 return true; 1379 1380 // If this is a case/default statement, and we haven't seen a switch, we have 1381 // to emit the code. 1382 if (isa<SwitchCase>(S) && !IgnoreCaseStmts) 1383 return true; 1384 1385 // If this is a switch statement, we want to ignore cases below it. 1386 if (isa<SwitchStmt>(S)) 1387 IgnoreCaseStmts = true; 1388 1389 // Scan subexpressions for verboten labels. 1390 for (const Stmt *SubStmt : S->children()) 1391 if (ContainsLabel(SubStmt, IgnoreCaseStmts)) 1392 return true; 1393 1394 return false; 1395 } 1396 1397 /// containsBreak - Return true if the statement contains a break out of it. 1398 /// If the statement (recursively) contains a switch or loop with a break 1399 /// inside of it, this is fine. 1400 bool CodeGenFunction::containsBreak(const Stmt *S) { 1401 // Null statement, not a label! 1402 if (!S) return false; 1403 1404 // If this is a switch or loop that defines its own break scope, then we can 1405 // include it and anything inside of it. 1406 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) || 1407 isa<ForStmt>(S)) 1408 return false; 1409 1410 if (isa<BreakStmt>(S)) 1411 return true; 1412 1413 // Scan subexpressions for verboten breaks. 1414 for (const Stmt *SubStmt : S->children()) 1415 if (containsBreak(SubStmt)) 1416 return true; 1417 1418 return false; 1419 } 1420 1421 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) { 1422 if (!S) return false; 1423 1424 // Some statement kinds add a scope and thus never add a decl to the current 1425 // scope. Note, this list is longer than the list of statements that might 1426 // have an unscoped decl nested within them, but this way is conservatively 1427 // correct even if more statement kinds are added. 1428 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) || 1429 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) || 1430 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) || 1431 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S)) 1432 return false; 1433 1434 if (isa<DeclStmt>(S)) 1435 return true; 1436 1437 for (const Stmt *SubStmt : S->children()) 1438 if (mightAddDeclToScope(SubStmt)) 1439 return true; 1440 1441 return false; 1442 } 1443 1444 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold 1445 /// to a constant, or if it does but contains a label, return false. If it 1446 /// constant folds return true and set the boolean result in Result. 1447 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond, 1448 bool &ResultBool, 1449 bool AllowLabels) { 1450 llvm::APSInt ResultInt; 1451 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels)) 1452 return false; 1453 1454 ResultBool = ResultInt.getBoolValue(); 1455 return true; 1456 } 1457 1458 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold 1459 /// to a constant, or if it does but contains a label, return false. If it 1460 /// constant folds return true and set the folded value. 1461 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond, 1462 llvm::APSInt &ResultInt, 1463 bool AllowLabels) { 1464 // FIXME: Rename and handle conversion of other evaluatable things 1465 // to bool. 1466 llvm::APSInt Int; 1467 if (!Cond->EvaluateAsInt(Int, getContext())) 1468 return false; // Not foldable, not integer or not fully evaluatable. 1469 1470 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond)) 1471 return false; // Contains a label. 1472 1473 ResultInt = Int; 1474 return true; 1475 } 1476 1477 1478 1479 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if 1480 /// statement) to the specified blocks. Based on the condition, this might try 1481 /// to simplify the codegen of the conditional based on the branch. 1482 /// 1483 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond, 1484 llvm::BasicBlock *TrueBlock, 1485 llvm::BasicBlock *FalseBlock, 1486 uint64_t TrueCount) { 1487 Cond = Cond->IgnoreParens(); 1488 1489 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) { 1490 1491 // Handle X && Y in a condition. 1492 if (CondBOp->getOpcode() == BO_LAnd) { 1493 // If we have "1 && X", simplify the code. "0 && X" would have constant 1494 // folded if the case was simple enough. 1495 bool ConstantBool = false; 1496 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) && 1497 ConstantBool) { 1498 // br(1 && X) -> br(X). 1499 incrementProfileCounter(CondBOp); 1500 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, 1501 TrueCount); 1502 } 1503 1504 // If we have "X && 1", simplify the code to use an uncond branch. 1505 // "X && 0" would have been constant folded to 0. 1506 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) && 1507 ConstantBool) { 1508 // br(X && 1) -> br(X). 1509 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock, 1510 TrueCount); 1511 } 1512 1513 // Emit the LHS as a conditional. If the LHS conditional is false, we 1514 // want to jump to the FalseBlock. 1515 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true"); 1516 // The counter tells us how often we evaluate RHS, and all of TrueCount 1517 // can be propagated to that branch. 1518 uint64_t RHSCount = getProfileCount(CondBOp->getRHS()); 1519 1520 ConditionalEvaluation eval(*this); 1521 { 1522 ApplyDebugLocation DL(*this, Cond); 1523 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount); 1524 EmitBlock(LHSTrue); 1525 } 1526 1527 incrementProfileCounter(CondBOp); 1528 setCurrentProfileCount(getProfileCount(CondBOp->getRHS())); 1529 1530 // Any temporaries created here are conditional. 1531 eval.begin(*this); 1532 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount); 1533 eval.end(*this); 1534 1535 return; 1536 } 1537 1538 if (CondBOp->getOpcode() == BO_LOr) { 1539 // If we have "0 || X", simplify the code. "1 || X" would have constant 1540 // folded if the case was simple enough. 1541 bool ConstantBool = false; 1542 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) && 1543 !ConstantBool) { 1544 // br(0 || X) -> br(X). 1545 incrementProfileCounter(CondBOp); 1546 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, 1547 TrueCount); 1548 } 1549 1550 // If we have "X || 0", simplify the code to use an uncond branch. 1551 // "X || 1" would have been constant folded to 1. 1552 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) && 1553 !ConstantBool) { 1554 // br(X || 0) -> br(X). 1555 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock, 1556 TrueCount); 1557 } 1558 1559 // Emit the LHS as a conditional. If the LHS conditional is true, we 1560 // want to jump to the TrueBlock. 1561 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false"); 1562 // We have the count for entry to the RHS and for the whole expression 1563 // being true, so we can divy up True count between the short circuit and 1564 // the RHS. 1565 uint64_t LHSCount = 1566 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS()); 1567 uint64_t RHSCount = TrueCount - LHSCount; 1568 1569 ConditionalEvaluation eval(*this); 1570 { 1571 ApplyDebugLocation DL(*this, Cond); 1572 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount); 1573 EmitBlock(LHSFalse); 1574 } 1575 1576 incrementProfileCounter(CondBOp); 1577 setCurrentProfileCount(getProfileCount(CondBOp->getRHS())); 1578 1579 // Any temporaries created here are conditional. 1580 eval.begin(*this); 1581 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount); 1582 1583 eval.end(*this); 1584 1585 return; 1586 } 1587 } 1588 1589 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) { 1590 // br(!x, t, f) -> br(x, f, t) 1591 if (CondUOp->getOpcode() == UO_LNot) { 1592 // Negate the count. 1593 uint64_t FalseCount = getCurrentProfileCount() - TrueCount; 1594 // Negate the condition and swap the destination blocks. 1595 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock, 1596 FalseCount); 1597 } 1598 } 1599 1600 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) { 1601 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f)) 1602 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true"); 1603 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false"); 1604 1605 ConditionalEvaluation cond(*this); 1606 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock, 1607 getProfileCount(CondOp)); 1608 1609 // When computing PGO branch weights, we only know the overall count for 1610 // the true block. This code is essentially doing tail duplication of the 1611 // naive code-gen, introducing new edges for which counts are not 1612 // available. Divide the counts proportionally between the LHS and RHS of 1613 // the conditional operator. 1614 uint64_t LHSScaledTrueCount = 0; 1615 if (TrueCount) { 1616 double LHSRatio = 1617 getProfileCount(CondOp) / (double)getCurrentProfileCount(); 1618 LHSScaledTrueCount = TrueCount * LHSRatio; 1619 } 1620 1621 cond.begin(*this); 1622 EmitBlock(LHSBlock); 1623 incrementProfileCounter(CondOp); 1624 { 1625 ApplyDebugLocation DL(*this, Cond); 1626 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock, 1627 LHSScaledTrueCount); 1628 } 1629 cond.end(*this); 1630 1631 cond.begin(*this); 1632 EmitBlock(RHSBlock); 1633 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock, 1634 TrueCount - LHSScaledTrueCount); 1635 cond.end(*this); 1636 1637 return; 1638 } 1639 1640 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) { 1641 // Conditional operator handling can give us a throw expression as a 1642 // condition for a case like: 1643 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f) 1644 // Fold this to: 1645 // br(c, throw x, br(y, t, f)) 1646 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false); 1647 return; 1648 } 1649 1650 // If the branch has a condition wrapped by __builtin_unpredictable, 1651 // create metadata that specifies that the branch is unpredictable. 1652 // Don't bother if not optimizing because that metadata would not be used. 1653 llvm::MDNode *Unpredictable = nullptr; 1654 auto *Call = dyn_cast<CallExpr>(Cond); 1655 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) { 1656 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl()); 1657 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) { 1658 llvm::MDBuilder MDHelper(getLLVMContext()); 1659 Unpredictable = MDHelper.createUnpredictable(); 1660 } 1661 } 1662 1663 // Create branch weights based on the number of times we get here and the 1664 // number of times the condition should be true. 1665 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount); 1666 llvm::MDNode *Weights = 1667 createProfileWeights(TrueCount, CurrentCount - TrueCount); 1668 1669 // Emit the code with the fully general case. 1670 llvm::Value *CondV; 1671 { 1672 ApplyDebugLocation DL(*this, Cond); 1673 CondV = EvaluateExprAsBool(Cond); 1674 } 1675 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable); 1676 } 1677 1678 /// ErrorUnsupported - Print out an error that codegen doesn't support the 1679 /// specified stmt yet. 1680 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) { 1681 CGM.ErrorUnsupported(S, Type); 1682 } 1683 1684 /// emitNonZeroVLAInit - Emit the "zero" initialization of a 1685 /// variable-length array whose elements have a non-zero bit-pattern. 1686 /// 1687 /// \param baseType the inner-most element type of the array 1688 /// \param src - a char* pointing to the bit-pattern for a single 1689 /// base element of the array 1690 /// \param sizeInChars - the total size of the VLA, in chars 1691 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType, 1692 Address dest, Address src, 1693 llvm::Value *sizeInChars) { 1694 CGBuilderTy &Builder = CGF.Builder; 1695 1696 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType); 1697 llvm::Value *baseSizeInChars 1698 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity()); 1699 1700 Address begin = 1701 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin"); 1702 llvm::Value *end = 1703 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end"); 1704 1705 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock(); 1706 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop"); 1707 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont"); 1708 1709 // Make a loop over the VLA. C99 guarantees that the VLA element 1710 // count must be nonzero. 1711 CGF.EmitBlock(loopBB); 1712 1713 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur"); 1714 cur->addIncoming(begin.getPointer(), originBB); 1715 1716 CharUnits curAlign = 1717 dest.getAlignment().alignmentOfArrayElement(baseSize); 1718 1719 // memcpy the individual element bit-pattern. 1720 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars, 1721 /*volatile*/ false); 1722 1723 // Go to the next element. 1724 llvm::Value *next = 1725 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next"); 1726 1727 // Leave if that's the end of the VLA. 1728 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone"); 1729 Builder.CreateCondBr(done, contBB, loopBB); 1730 cur->addIncoming(next, loopBB); 1731 1732 CGF.EmitBlock(contBB); 1733 } 1734 1735 void 1736 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) { 1737 // Ignore empty classes in C++. 1738 if (getLangOpts().CPlusPlus) { 1739 if (const RecordType *RT = Ty->getAs<RecordType>()) { 1740 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty()) 1741 return; 1742 } 1743 } 1744 1745 // Cast the dest ptr to the appropriate i8 pointer type. 1746 if (DestPtr.getElementType() != Int8Ty) 1747 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty); 1748 1749 // Get size and alignment info for this aggregate. 1750 CharUnits size = getContext().getTypeSizeInChars(Ty); 1751 1752 llvm::Value *SizeVal; 1753 const VariableArrayType *vla; 1754 1755 // Don't bother emitting a zero-byte memset. 1756 if (size.isZero()) { 1757 // But note that getTypeInfo returns 0 for a VLA. 1758 if (const VariableArrayType *vlaType = 1759 dyn_cast_or_null<VariableArrayType>( 1760 getContext().getAsArrayType(Ty))) { 1761 auto VlaSize = getVLASize(vlaType); 1762 SizeVal = VlaSize.NumElts; 1763 CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type); 1764 if (!eltSize.isOne()) 1765 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize)); 1766 vla = vlaType; 1767 } else { 1768 return; 1769 } 1770 } else { 1771 SizeVal = CGM.getSize(size); 1772 vla = nullptr; 1773 } 1774 1775 // If the type contains a pointer to data member we can't memset it to zero. 1776 // Instead, create a null constant and copy it to the destination. 1777 // TODO: there are other patterns besides zero that we can usefully memset, 1778 // like -1, which happens to be the pattern used by member-pointers. 1779 if (!CGM.getTypes().isZeroInitializable(Ty)) { 1780 // For a VLA, emit a single element, then splat that over the VLA. 1781 if (vla) Ty = getContext().getBaseElementType(vla); 1782 1783 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty); 1784 1785 llvm::GlobalVariable *NullVariable = 1786 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(), 1787 /*isConstant=*/true, 1788 llvm::GlobalVariable::PrivateLinkage, 1789 NullConstant, Twine()); 1790 CharUnits NullAlign = DestPtr.getAlignment(); 1791 NullVariable->setAlignment(NullAlign.getQuantity()); 1792 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()), 1793 NullAlign); 1794 1795 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal); 1796 1797 // Get and call the appropriate llvm.memcpy overload. 1798 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false); 1799 return; 1800 } 1801 1802 // Otherwise, just memset the whole thing to zero. This is legal 1803 // because in LLVM, all default initializers (other than the ones we just 1804 // handled above) are guaranteed to have a bit pattern of all zeros. 1805 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false); 1806 } 1807 1808 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) { 1809 // Make sure that there is a block for the indirect goto. 1810 if (!IndirectBranch) 1811 GetIndirectGotoBlock(); 1812 1813 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock(); 1814 1815 // Make sure the indirect branch includes all of the address-taken blocks. 1816 IndirectBranch->addDestination(BB); 1817 return llvm::BlockAddress::get(CurFn, BB); 1818 } 1819 1820 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() { 1821 // If we already made the indirect branch for indirect goto, return its block. 1822 if (IndirectBranch) return IndirectBranch->getParent(); 1823 1824 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto")); 1825 1826 // Create the PHI node that indirect gotos will add entries to. 1827 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0, 1828 "indirect.goto.dest"); 1829 1830 // Create the indirect branch instruction. 1831 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal); 1832 return IndirectBranch->getParent(); 1833 } 1834 1835 /// Computes the length of an array in elements, as well as the base 1836 /// element type and a properly-typed first element pointer. 1837 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType, 1838 QualType &baseType, 1839 Address &addr) { 1840 const ArrayType *arrayType = origArrayType; 1841 1842 // If it's a VLA, we have to load the stored size. Note that 1843 // this is the size of the VLA in bytes, not its size in elements. 1844 llvm::Value *numVLAElements = nullptr; 1845 if (isa<VariableArrayType>(arrayType)) { 1846 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts; 1847 1848 // Walk into all VLAs. This doesn't require changes to addr, 1849 // which has type T* where T is the first non-VLA element type. 1850 do { 1851 QualType elementType = arrayType->getElementType(); 1852 arrayType = getContext().getAsArrayType(elementType); 1853 1854 // If we only have VLA components, 'addr' requires no adjustment. 1855 if (!arrayType) { 1856 baseType = elementType; 1857 return numVLAElements; 1858 } 1859 } while (isa<VariableArrayType>(arrayType)); 1860 1861 // We get out here only if we find a constant array type 1862 // inside the VLA. 1863 } 1864 1865 // We have some number of constant-length arrays, so addr should 1866 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks 1867 // down to the first element of addr. 1868 SmallVector<llvm::Value*, 8> gepIndices; 1869 1870 // GEP down to the array type. 1871 llvm::ConstantInt *zero = Builder.getInt32(0); 1872 gepIndices.push_back(zero); 1873 1874 uint64_t countFromCLAs = 1; 1875 QualType eltType; 1876 1877 llvm::ArrayType *llvmArrayType = 1878 dyn_cast<llvm::ArrayType>(addr.getElementType()); 1879 while (llvmArrayType) { 1880 assert(isa<ConstantArrayType>(arrayType)); 1881 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue() 1882 == llvmArrayType->getNumElements()); 1883 1884 gepIndices.push_back(zero); 1885 countFromCLAs *= llvmArrayType->getNumElements(); 1886 eltType = arrayType->getElementType(); 1887 1888 llvmArrayType = 1889 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType()); 1890 arrayType = getContext().getAsArrayType(arrayType->getElementType()); 1891 assert((!llvmArrayType || arrayType) && 1892 "LLVM and Clang types are out-of-synch"); 1893 } 1894 1895 if (arrayType) { 1896 // From this point onwards, the Clang array type has been emitted 1897 // as some other type (probably a packed struct). Compute the array 1898 // size, and just emit the 'begin' expression as a bitcast. 1899 while (arrayType) { 1900 countFromCLAs *= 1901 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue(); 1902 eltType = arrayType->getElementType(); 1903 arrayType = getContext().getAsArrayType(eltType); 1904 } 1905 1906 llvm::Type *baseType = ConvertType(eltType); 1907 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin"); 1908 } else { 1909 // Create the actual GEP. 1910 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(), 1911 gepIndices, "array.begin"), 1912 addr.getAlignment()); 1913 } 1914 1915 baseType = eltType; 1916 1917 llvm::Value *numElements 1918 = llvm::ConstantInt::get(SizeTy, countFromCLAs); 1919 1920 // If we had any VLA dimensions, factor them in. 1921 if (numVLAElements) 1922 numElements = Builder.CreateNUWMul(numVLAElements, numElements); 1923 1924 return numElements; 1925 } 1926 1927 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) { 1928 const VariableArrayType *vla = getContext().getAsVariableArrayType(type); 1929 assert(vla && "type was not a variable array type!"); 1930 return getVLASize(vla); 1931 } 1932 1933 CodeGenFunction::VlaSizePair 1934 CodeGenFunction::getVLASize(const VariableArrayType *type) { 1935 // The number of elements so far; always size_t. 1936 llvm::Value *numElements = nullptr; 1937 1938 QualType elementType; 1939 do { 1940 elementType = type->getElementType(); 1941 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()]; 1942 assert(vlaSize && "no size for VLA!"); 1943 assert(vlaSize->getType() == SizeTy); 1944 1945 if (!numElements) { 1946 numElements = vlaSize; 1947 } else { 1948 // It's undefined behavior if this wraps around, so mark it that way. 1949 // FIXME: Teach -fsanitize=undefined to trap this. 1950 numElements = Builder.CreateNUWMul(numElements, vlaSize); 1951 } 1952 } while ((type = getContext().getAsVariableArrayType(elementType))); 1953 1954 return { numElements, elementType }; 1955 } 1956 1957 CodeGenFunction::VlaSizePair 1958 CodeGenFunction::getVLAElements1D(QualType type) { 1959 const VariableArrayType *vla = getContext().getAsVariableArrayType(type); 1960 assert(vla && "type was not a variable array type!"); 1961 return getVLAElements1D(vla); 1962 } 1963 1964 CodeGenFunction::VlaSizePair 1965 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) { 1966 llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()]; 1967 assert(VlaSize && "no size for VLA!"); 1968 assert(VlaSize->getType() == SizeTy); 1969 return { VlaSize, Vla->getElementType() }; 1970 } 1971 1972 void CodeGenFunction::EmitVariablyModifiedType(QualType type) { 1973 assert(type->isVariablyModifiedType() && 1974 "Must pass variably modified type to EmitVLASizes!"); 1975 1976 EnsureInsertPoint(); 1977 1978 // We're going to walk down into the type and look for VLA 1979 // expressions. 1980 do { 1981 assert(type->isVariablyModifiedType()); 1982 1983 const Type *ty = type.getTypePtr(); 1984 switch (ty->getTypeClass()) { 1985 1986 #define TYPE(Class, Base) 1987 #define ABSTRACT_TYPE(Class, Base) 1988 #define NON_CANONICAL_TYPE(Class, Base) 1989 #define DEPENDENT_TYPE(Class, Base) case Type::Class: 1990 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base) 1991 #include "clang/AST/TypeNodes.def" 1992 llvm_unreachable("unexpected dependent type!"); 1993 1994 // These types are never variably-modified. 1995 case Type::Builtin: 1996 case Type::Complex: 1997 case Type::Vector: 1998 case Type::ExtVector: 1999 case Type::Record: 2000 case Type::Enum: 2001 case Type::Elaborated: 2002 case Type::TemplateSpecialization: 2003 case Type::ObjCTypeParam: 2004 case Type::ObjCObject: 2005 case Type::ObjCInterface: 2006 case Type::ObjCObjectPointer: 2007 llvm_unreachable("type class is never variably-modified!"); 2008 2009 case Type::Adjusted: 2010 type = cast<AdjustedType>(ty)->getAdjustedType(); 2011 break; 2012 2013 case Type::Decayed: 2014 type = cast<DecayedType>(ty)->getPointeeType(); 2015 break; 2016 2017 case Type::Pointer: 2018 type = cast<PointerType>(ty)->getPointeeType(); 2019 break; 2020 2021 case Type::BlockPointer: 2022 type = cast<BlockPointerType>(ty)->getPointeeType(); 2023 break; 2024 2025 case Type::LValueReference: 2026 case Type::RValueReference: 2027 type = cast<ReferenceType>(ty)->getPointeeType(); 2028 break; 2029 2030 case Type::MemberPointer: 2031 type = cast<MemberPointerType>(ty)->getPointeeType(); 2032 break; 2033 2034 case Type::ConstantArray: 2035 case Type::IncompleteArray: 2036 // Losing element qualification here is fine. 2037 type = cast<ArrayType>(ty)->getElementType(); 2038 break; 2039 2040 case Type::VariableArray: { 2041 // Losing element qualification here is fine. 2042 const VariableArrayType *vat = cast<VariableArrayType>(ty); 2043 2044 // Unknown size indication requires no size computation. 2045 // Otherwise, evaluate and record it. 2046 if (const Expr *size = vat->getSizeExpr()) { 2047 // It's possible that we might have emitted this already, 2048 // e.g. with a typedef and a pointer to it. 2049 llvm::Value *&entry = VLASizeMap[size]; 2050 if (!entry) { 2051 llvm::Value *Size = EmitScalarExpr(size); 2052 2053 // C11 6.7.6.2p5: 2054 // If the size is an expression that is not an integer constant 2055 // expression [...] each time it is evaluated it shall have a value 2056 // greater than zero. 2057 if (SanOpts.has(SanitizerKind::VLABound) && 2058 size->getType()->isSignedIntegerType()) { 2059 SanitizerScope SanScope(this); 2060 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType()); 2061 llvm::Constant *StaticArgs[] = { 2062 EmitCheckSourceLocation(size->getLocStart()), 2063 EmitCheckTypeDescriptor(size->getType()) 2064 }; 2065 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero), 2066 SanitizerKind::VLABound), 2067 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size); 2068 } 2069 2070 // Always zexting here would be wrong if it weren't 2071 // undefined behavior to have a negative bound. 2072 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false); 2073 } 2074 } 2075 type = vat->getElementType(); 2076 break; 2077 } 2078 2079 case Type::FunctionProto: 2080 case Type::FunctionNoProto: 2081 type = cast<FunctionType>(ty)->getReturnType(); 2082 break; 2083 2084 case Type::Paren: 2085 case Type::TypeOf: 2086 case Type::UnaryTransform: 2087 case Type::Attributed: 2088 case Type::SubstTemplateTypeParm: 2089 case Type::PackExpansion: 2090 // Keep walking after single level desugaring. 2091 type = type.getSingleStepDesugaredType(getContext()); 2092 break; 2093 2094 case Type::Typedef: 2095 case Type::Decltype: 2096 case Type::Auto: 2097 case Type::DeducedTemplateSpecialization: 2098 // Stop walking: nothing to do. 2099 return; 2100 2101 case Type::TypeOfExpr: 2102 // Stop walking: emit typeof expression. 2103 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr()); 2104 return; 2105 2106 case Type::Atomic: 2107 type = cast<AtomicType>(ty)->getValueType(); 2108 break; 2109 2110 case Type::Pipe: 2111 type = cast<PipeType>(ty)->getElementType(); 2112 break; 2113 } 2114 } while (type->isVariablyModifiedType()); 2115 } 2116 2117 Address CodeGenFunction::EmitVAListRef(const Expr* E) { 2118 if (getContext().getBuiltinVaListType()->isArrayType()) 2119 return EmitPointerWithAlignment(E); 2120 return EmitLValue(E).getAddress(); 2121 } 2122 2123 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) { 2124 return EmitLValue(E).getAddress(); 2125 } 2126 2127 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E, 2128 const APValue &Init) { 2129 assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!"); 2130 if (CGDebugInfo *Dbg = getDebugInfo()) 2131 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo) 2132 Dbg->EmitGlobalVariable(E->getDecl(), Init); 2133 } 2134 2135 CodeGenFunction::PeepholeProtection 2136 CodeGenFunction::protectFromPeepholes(RValue rvalue) { 2137 // At the moment, the only aggressive peephole we do in IR gen 2138 // is trunc(zext) folding, but if we add more, we can easily 2139 // extend this protection. 2140 2141 if (!rvalue.isScalar()) return PeepholeProtection(); 2142 llvm::Value *value = rvalue.getScalarVal(); 2143 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection(); 2144 2145 // Just make an extra bitcast. 2146 assert(HaveInsertPoint()); 2147 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "", 2148 Builder.GetInsertBlock()); 2149 2150 PeepholeProtection protection; 2151 protection.Inst = inst; 2152 return protection; 2153 } 2154 2155 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) { 2156 if (!protection.Inst) return; 2157 2158 // In theory, we could try to duplicate the peepholes now, but whatever. 2159 protection.Inst->eraseFromParent(); 2160 } 2161 2162 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn, 2163 llvm::Value *AnnotatedVal, 2164 StringRef AnnotationStr, 2165 SourceLocation Location) { 2166 llvm::Value *Args[4] = { 2167 AnnotatedVal, 2168 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy), 2169 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy), 2170 CGM.EmitAnnotationLineNo(Location) 2171 }; 2172 return Builder.CreateCall(AnnotationFn, Args); 2173 } 2174 2175 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) { 2176 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute"); 2177 // FIXME We create a new bitcast for every annotation because that's what 2178 // llvm-gcc was doing. 2179 for (const auto *I : D->specific_attrs<AnnotateAttr>()) 2180 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation), 2181 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()), 2182 I->getAnnotation(), D->getLocation()); 2183 } 2184 2185 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D, 2186 Address Addr) { 2187 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute"); 2188 llvm::Value *V = Addr.getPointer(); 2189 llvm::Type *VTy = V->getType(); 2190 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation, 2191 CGM.Int8PtrTy); 2192 2193 for (const auto *I : D->specific_attrs<AnnotateAttr>()) { 2194 // FIXME Always emit the cast inst so we can differentiate between 2195 // annotation on the first field of a struct and annotation on the struct 2196 // itself. 2197 if (VTy != CGM.Int8PtrTy) 2198 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy)); 2199 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation()); 2200 V = Builder.CreateBitCast(V, VTy); 2201 } 2202 2203 return Address(V, Addr.getAlignment()); 2204 } 2205 2206 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { } 2207 2208 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF) 2209 : CGF(CGF) { 2210 assert(!CGF->IsSanitizerScope); 2211 CGF->IsSanitizerScope = true; 2212 } 2213 2214 CodeGenFunction::SanitizerScope::~SanitizerScope() { 2215 CGF->IsSanitizerScope = false; 2216 } 2217 2218 void CodeGenFunction::InsertHelper(llvm::Instruction *I, 2219 const llvm::Twine &Name, 2220 llvm::BasicBlock *BB, 2221 llvm::BasicBlock::iterator InsertPt) const { 2222 LoopStack.InsertHelper(I); 2223 if (IsSanitizerScope) 2224 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I); 2225 } 2226 2227 void CGBuilderInserter::InsertHelper( 2228 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB, 2229 llvm::BasicBlock::iterator InsertPt) const { 2230 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt); 2231 if (CGF) 2232 CGF->InsertHelper(I, Name, BB, InsertPt); 2233 } 2234 2235 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures, 2236 CodeGenModule &CGM, const FunctionDecl *FD, 2237 std::string &FirstMissing) { 2238 // If there aren't any required features listed then go ahead and return. 2239 if (ReqFeatures.empty()) 2240 return false; 2241 2242 // Now build up the set of caller features and verify that all the required 2243 // features are there. 2244 llvm::StringMap<bool> CallerFeatureMap; 2245 CGM.getFunctionFeatureMap(CallerFeatureMap, FD); 2246 2247 // If we have at least one of the features in the feature list return 2248 // true, otherwise return false. 2249 return std::all_of( 2250 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) { 2251 SmallVector<StringRef, 1> OrFeatures; 2252 Feature.split(OrFeatures, "|"); 2253 return std::any_of(OrFeatures.begin(), OrFeatures.end(), 2254 [&](StringRef Feature) { 2255 if (!CallerFeatureMap.lookup(Feature)) { 2256 FirstMissing = Feature.str(); 2257 return false; 2258 } 2259 return true; 2260 }); 2261 }); 2262 } 2263 2264 // Emits an error if we don't have a valid set of target features for the 2265 // called function. 2266 void CodeGenFunction::checkTargetFeatures(const CallExpr *E, 2267 const FunctionDecl *TargetDecl) { 2268 // Early exit if this is an indirect call. 2269 if (!TargetDecl) 2270 return; 2271 2272 // Get the current enclosing function if it exists. If it doesn't 2273 // we can't check the target features anyhow. 2274 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl); 2275 if (!FD) 2276 return; 2277 2278 // Grab the required features for the call. For a builtin this is listed in 2279 // the td file with the default cpu, for an always_inline function this is any 2280 // listed cpu and any listed features. 2281 unsigned BuiltinID = TargetDecl->getBuiltinID(); 2282 std::string MissingFeature; 2283 if (BuiltinID) { 2284 SmallVector<StringRef, 1> ReqFeatures; 2285 const char *FeatureList = 2286 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID); 2287 // Return if the builtin doesn't have any required features. 2288 if (!FeatureList || StringRef(FeatureList) == "") 2289 return; 2290 StringRef(FeatureList).split(ReqFeatures, ","); 2291 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature)) 2292 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature) 2293 << TargetDecl->getDeclName() 2294 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID); 2295 2296 } else if (TargetDecl->hasAttr<TargetAttr>()) { 2297 // Get the required features for the callee. 2298 SmallVector<StringRef, 1> ReqFeatures; 2299 llvm::StringMap<bool> CalleeFeatureMap; 2300 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl); 2301 for (const auto &F : CalleeFeatureMap) { 2302 // Only positive features are "required". 2303 if (F.getValue()) 2304 ReqFeatures.push_back(F.getKey()); 2305 } 2306 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature)) 2307 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature) 2308 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature; 2309 } 2310 } 2311 2312 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) { 2313 if (!CGM.getCodeGenOpts().SanitizeStats) 2314 return; 2315 2316 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint()); 2317 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation()); 2318 CGM.getSanStats().create(IRB, SSK); 2319 } 2320 2321 llvm::Value * 2322 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) { 2323 llvm::Value *TrueCondition = nullptr; 2324 if (!RO.ParsedAttribute.Architecture.empty()) 2325 TrueCondition = EmitX86CpuIs(RO.ParsedAttribute.Architecture); 2326 2327 if (!RO.ParsedAttribute.Features.empty()) { 2328 SmallVector<StringRef, 8> FeatureList; 2329 llvm::for_each(RO.ParsedAttribute.Features, 2330 [&FeatureList](const std::string &Feature) { 2331 FeatureList.push_back(StringRef{Feature}.substr(1)); 2332 }); 2333 llvm::Value *FeatureCmp = EmitX86CpuSupports(FeatureList); 2334 TrueCondition = TrueCondition ? Builder.CreateAnd(TrueCondition, FeatureCmp) 2335 : FeatureCmp; 2336 } 2337 return TrueCondition; 2338 } 2339 2340 void CodeGenFunction::EmitMultiVersionResolver( 2341 llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) { 2342 assert((getContext().getTargetInfo().getTriple().getArch() == 2343 llvm::Triple::x86 || 2344 getContext().getTargetInfo().getTriple().getArch() == 2345 llvm::Triple::x86_64) && 2346 "Only implemented for x86 targets"); 2347 2348 // Main function's basic block. 2349 llvm::BasicBlock *CurBlock = createBasicBlock("entry", Resolver); 2350 Builder.SetInsertPoint(CurBlock); 2351 EmitX86CpuInit(); 2352 2353 llvm::Function *DefaultFunc = nullptr; 2354 for (const MultiVersionResolverOption &RO : Options) { 2355 Builder.SetInsertPoint(CurBlock); 2356 llvm::Value *TrueCondition = FormResolverCondition(RO); 2357 2358 if (!TrueCondition) { 2359 DefaultFunc = RO.Function; 2360 } else { 2361 llvm::BasicBlock *RetBlock = createBasicBlock("ro_ret", Resolver); 2362 llvm::IRBuilder<> RetBuilder(RetBlock); 2363 RetBuilder.CreateRet(RO.Function); 2364 CurBlock = createBasicBlock("ro_else", Resolver); 2365 Builder.CreateCondBr(TrueCondition, RetBlock, CurBlock); 2366 } 2367 } 2368 2369 assert(DefaultFunc && "No default version?"); 2370 // Emit return from the 'else-ist' block. 2371 Builder.SetInsertPoint(CurBlock); 2372 Builder.CreateRet(DefaultFunc); 2373 } 2374 2375 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) { 2376 if (CGDebugInfo *DI = getDebugInfo()) 2377 return DI->SourceLocToDebugLoc(Location); 2378 2379 return llvm::DebugLoc(); 2380 } 2381