| 1 | //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This coordinates the per-function state used while generating code. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "CodeGenFunction.h" |
| 14 | #include "CGBlocks.h" |
| 15 | #include "CGCUDARuntime.h" |
| 16 | #include "CGCXXABI.h" |
| 17 | #include "CGCleanup.h" |
| 18 | #include "CGDebugInfo.h" |
| 19 | #include "CGHLSLRuntime.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/Attr.h" |
| 27 | #include "clang/AST/Decl.h" |
| 28 | #include "clang/AST/DeclCXX.h" |
| 29 | #include "clang/AST/Expr.h" |
| 30 | #include "clang/AST/IgnoreExpr.h" |
| 31 | #include "clang/AST/StmtCXX.h" |
| 32 | #include "clang/AST/StmtObjC.h" |
| 33 | #include "clang/Basic/Builtins.h" |
| 34 | #include "clang/Basic/CodeGenOptions.h" |
| 35 | #include "clang/Basic/DiagnosticFrontend.h" |
| 36 | #include "clang/Basic/TargetBuiltins.h" |
| 37 | #include "clang/Basic/TargetInfo.h" |
| 38 | #include "clang/CodeGen/CGFunctionInfo.h" |
| 39 | #include "llvm/ADT/ArrayRef.h" |
| 40 | #include "llvm/ADT/ScopeExit.h" |
| 41 | #include "llvm/Frontend/OpenMP/OMPIRBuilder.h" |
| 42 | #include "llvm/IR/DataLayout.h" |
| 43 | #include "llvm/IR/Dominators.h" |
| 44 | #include "llvm/IR/FPEnv.h" |
| 45 | #include "llvm/IR/Instruction.h" |
| 46 | #include "llvm/IR/IntrinsicInst.h" |
| 47 | #include "llvm/IR/Intrinsics.h" |
| 48 | #include "llvm/IR/MDBuilder.h" |
| 49 | #include "llvm/Support/CRC.h" |
| 50 | #include "llvm/Support/SipHash.h" |
| 51 | #include "llvm/Support/xxhash.h" |
| 52 | #include "llvm/Transforms/Scalar/LowerExpectIntrinsic.h" |
| 53 | #include "llvm/Transforms/Utils/PromoteMemToReg.h" |
| 54 | #include <optional> |
| 55 | |
| 56 | using namespace clang; |
| 57 | using namespace CodeGen; |
| 58 | |
| 59 | /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time |
| 60 | /// markers. |
| 61 | static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts, |
| 62 | const LangOptions &LangOpts) { |
| 63 | if (CGOpts.DisableLifetimeMarkers) |
| 64 | return false; |
| 65 | |
| 66 | // Sanitizers may use markers. |
| 67 | if (CGOpts.SanitizeAddressUseAfterScope || |
| 68 | LangOpts.Sanitize.has(K: SanitizerKind::HWAddress) || |
| 69 | LangOpts.Sanitize.has(K: SanitizerKind::Memory) || |
| 70 | LangOpts.Sanitize.has(K: SanitizerKind::MemtagStack)) |
| 71 | return true; |
| 72 | |
| 73 | // For now, only in optimized builds. |
| 74 | return CGOpts.OptimizationLevel != 0; |
| 75 | } |
| 76 | |
| 77 | CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext) |
| 78 | : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()), |
| 79 | Builder(cgm, cgm.getModule().getContext(), CGBuilderInserterTy(this)), |
| 80 | SanOpts(CGM.getLangOpts().Sanitize), CurFPFeatures(CGM.getLangOpts()), |
| 81 | DebugInfo(CGM.getModuleDebugInfo()), |
| 82 | PGO(std::make_unique<CodeGenPGO>(args&: cgm)), |
| 83 | ShouldEmitLifetimeMarkers( |
| 84 | shouldEmitLifetimeMarkers(CGOpts: CGM.getCodeGenOpts(), LangOpts: CGM.getLangOpts())) { |
| 85 | if (!suppressNewContext) |
| 86 | CGM.getCXXABI().getMangleContext().startNewFunction(); |
| 87 | EHStack.setCGF(this); |
| 88 | |
| 89 | SetFastMathFlags(CurFPFeatures); |
| 90 | } |
| 91 | |
| 92 | CodeGenFunction::~CodeGenFunction() { |
| 93 | assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup" ); |
| 94 | assert(DeferredDeactivationCleanupStack.empty() && |
| 95 | "missed to deactivate a cleanup" ); |
| 96 | |
| 97 | if (getLangOpts().OpenMP && CurFn) |
| 98 | CGM.getOpenMPRuntime().functionFinished(CGF&: *this); |
| 99 | |
| 100 | // If we have an OpenMPIRBuilder we want to finalize functions (incl. |
| 101 | // outlining etc) at some point. Doing it once the function codegen is done |
| 102 | // seems to be a reasonable spot. We do it here, as opposed to the deletion |
| 103 | // time of the CodeGenModule, because we have to ensure the IR has not yet |
| 104 | // been "emitted" to the outside, thus, modifications are still sensible. |
| 105 | if (CGM.getLangOpts().OpenMPIRBuilder && CurFn) |
| 106 | CGM.getOpenMPRuntime().getOMPBuilder().finalize(Fn: CurFn); |
| 107 | } |
| 108 | |
| 109 | // Map the LangOption for exception behavior into |
| 110 | // the corresponding enum in the IR. |
| 111 | llvm::fp::ExceptionBehavior |
| 112 | clang::ToConstrainedExceptMD(LangOptions::FPExceptionModeKind Kind) { |
| 113 | |
| 114 | switch (Kind) { |
| 115 | case LangOptions::FPE_Ignore: return llvm::fp::ebIgnore; |
| 116 | case LangOptions::FPE_MayTrap: return llvm::fp::ebMayTrap; |
| 117 | case LangOptions::FPE_Strict: return llvm::fp::ebStrict; |
| 118 | default: |
| 119 | llvm_unreachable("Unsupported FP Exception Behavior" ); |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | void CodeGenFunction::SetFastMathFlags(FPOptions FPFeatures) { |
| 124 | llvm::FastMathFlags FMF; |
| 125 | FMF.setAllowReassoc(FPFeatures.getAllowFPReassociate()); |
| 126 | FMF.setNoNaNs(FPFeatures.getNoHonorNaNs()); |
| 127 | FMF.setNoInfs(FPFeatures.getNoHonorInfs()); |
| 128 | FMF.setNoSignedZeros(FPFeatures.getNoSignedZero()); |
| 129 | FMF.setAllowReciprocal(FPFeatures.getAllowReciprocal()); |
| 130 | FMF.setApproxFunc(FPFeatures.getAllowApproxFunc()); |
| 131 | FMF.setAllowContract(FPFeatures.allowFPContractAcrossStatement()); |
| 132 | Builder.setFastMathFlags(FMF); |
| 133 | } |
| 134 | |
| 135 | CodeGenFunction::CGFPOptionsRAII::CGFPOptionsRAII(CodeGenFunction &CGF, |
| 136 | const Expr *E) |
| 137 | : CGF(CGF) { |
| 138 | ConstructorHelper(FPFeatures: E->getFPFeaturesInEffect(LO: CGF.getLangOpts())); |
| 139 | } |
| 140 | |
| 141 | CodeGenFunction::CGFPOptionsRAII::CGFPOptionsRAII(CodeGenFunction &CGF, |
| 142 | FPOptions FPFeatures) |
| 143 | : CGF(CGF) { |
| 144 | ConstructorHelper(FPFeatures); |
| 145 | } |
| 146 | |
| 147 | void CodeGenFunction::CGFPOptionsRAII::ConstructorHelper(FPOptions FPFeatures) { |
| 148 | OldFPFeatures = CGF.CurFPFeatures; |
| 149 | CGF.CurFPFeatures = FPFeatures; |
| 150 | |
| 151 | OldExcept = CGF.Builder.getDefaultConstrainedExcept(); |
| 152 | OldRounding = CGF.Builder.getDefaultConstrainedRounding(); |
| 153 | |
| 154 | if (OldFPFeatures == FPFeatures) |
| 155 | return; |
| 156 | |
| 157 | FMFGuard.emplace(args&: CGF.Builder); |
| 158 | |
| 159 | llvm::RoundingMode NewRoundingBehavior = FPFeatures.getRoundingMode(); |
| 160 | CGF.Builder.setDefaultConstrainedRounding(NewRoundingBehavior); |
| 161 | auto NewExceptionBehavior = |
| 162 | ToConstrainedExceptMD(Kind: FPFeatures.getExceptionMode()); |
| 163 | CGF.Builder.setDefaultConstrainedExcept(NewExceptionBehavior); |
| 164 | |
| 165 | CGF.SetFastMathFlags(FPFeatures); |
| 166 | |
| 167 | assert((CGF.CurFuncDecl == nullptr || CGF.Builder.getIsFPConstrained() || |
| 168 | isa<CXXConstructorDecl>(CGF.CurFuncDecl) || |
| 169 | isa<CXXDestructorDecl>(CGF.CurFuncDecl) || |
| 170 | (NewExceptionBehavior == llvm::fp::ebIgnore && |
| 171 | NewRoundingBehavior == llvm::RoundingMode::NearestTiesToEven)) && |
| 172 | "FPConstrained should be enabled on entire function" ); |
| 173 | |
| 174 | auto mergeFnAttrValue = [&](StringRef Name, bool Value) { |
| 175 | auto OldValue = |
| 176 | CGF.CurFn->getFnAttribute(Kind: Name).getValueAsBool(); |
| 177 | auto NewValue = OldValue & Value; |
| 178 | if (OldValue != NewValue) |
| 179 | CGF.CurFn->addFnAttr(Kind: Name, Val: llvm::toStringRef(B: NewValue)); |
| 180 | }; |
| 181 | mergeFnAttrValue("no-signed-zeros-fp-math" , FPFeatures.getNoSignedZero()); |
| 182 | } |
| 183 | |
| 184 | CodeGenFunction::CGFPOptionsRAII::~CGFPOptionsRAII() { |
| 185 | CGF.CurFPFeatures = OldFPFeatures; |
| 186 | CGF.Builder.setDefaultConstrainedExcept(OldExcept); |
| 187 | CGF.Builder.setDefaultConstrainedRounding(OldRounding); |
| 188 | } |
| 189 | |
| 190 | static LValue |
| 191 | makeNaturalAlignAddrLValue(llvm::Value *V, QualType T, bool ForPointeeType, |
| 192 | bool MightBeSigned, CodeGenFunction &CGF, |
| 193 | KnownNonNull_t IsKnownNonNull = NotKnownNonNull) { |
| 194 | LValueBaseInfo BaseInfo; |
| 195 | TBAAAccessInfo TBAAInfo; |
| 196 | CharUnits Alignment = |
| 197 | CGF.CGM.getNaturalTypeAlignment(T, BaseInfo: &BaseInfo, TBAAInfo: &TBAAInfo, forPointeeType: ForPointeeType); |
| 198 | Address Addr = |
| 199 | MightBeSigned |
| 200 | ? CGF.makeNaturalAddressForPointer(Ptr: V, T, Alignment, ForPointeeType: false, BaseInfo: nullptr, |
| 201 | TBAAInfo: nullptr, IsKnownNonNull) |
| 202 | : Address(V, CGF.ConvertTypeForMem(T), Alignment, IsKnownNonNull); |
| 203 | return CGF.MakeAddrLValue(Addr, T, BaseInfo, TBAAInfo); |
| 204 | } |
| 205 | |
| 206 | LValue |
| 207 | CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T, |
| 208 | KnownNonNull_t IsKnownNonNull) { |
| 209 | return ::makeNaturalAlignAddrLValue(V, T, /*ForPointeeType*/ false, |
| 210 | /*MightBeSigned*/ true, CGF&: *this, |
| 211 | IsKnownNonNull); |
| 212 | } |
| 213 | |
| 214 | LValue |
| 215 | CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) { |
| 216 | return ::makeNaturalAlignAddrLValue(V, T, /*ForPointeeType*/ true, |
| 217 | /*MightBeSigned*/ true, CGF&: *this); |
| 218 | } |
| 219 | |
| 220 | LValue CodeGenFunction::MakeNaturalAlignRawAddrLValue(llvm::Value *V, |
| 221 | QualType T) { |
| 222 | return ::makeNaturalAlignAddrLValue(V, T, /*ForPointeeType*/ false, |
| 223 | /*MightBeSigned*/ false, CGF&: *this); |
| 224 | } |
| 225 | |
| 226 | LValue CodeGenFunction::MakeNaturalAlignPointeeRawAddrLValue(llvm::Value *V, |
| 227 | QualType T) { |
| 228 | return ::makeNaturalAlignAddrLValue(V, T, /*ForPointeeType*/ true, |
| 229 | /*MightBeSigned*/ false, CGF&: *this); |
| 230 | } |
| 231 | |
| 232 | llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) { |
| 233 | return CGM.getTypes().ConvertTypeForMem(T); |
| 234 | } |
| 235 | |
| 236 | llvm::Type *CodeGenFunction::ConvertType(QualType T) { |
| 237 | return CGM.getTypes().ConvertType(T); |
| 238 | } |
| 239 | |
| 240 | llvm::Type *CodeGenFunction::convertTypeForLoadStore(QualType ASTTy, |
| 241 | llvm::Type *LLVMTy) { |
| 242 | return CGM.getTypes().convertTypeForLoadStore(T: ASTTy, LLVMTy); |
| 243 | } |
| 244 | |
| 245 | TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) { |
| 246 | type = type.getCanonicalType(); |
| 247 | while (true) { |
| 248 | switch (type->getTypeClass()) { |
| 249 | #define TYPE(name, parent) |
| 250 | #define ABSTRACT_TYPE(name, parent) |
| 251 | #define NON_CANONICAL_TYPE(name, parent) case Type::name: |
| 252 | #define DEPENDENT_TYPE(name, parent) case Type::name: |
| 253 | #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name: |
| 254 | #include "clang/AST/TypeNodes.inc" |
| 255 | llvm_unreachable("non-canonical or dependent type in IR-generation" ); |
| 256 | |
| 257 | case Type::Auto: |
| 258 | case Type::DeducedTemplateSpecialization: |
| 259 | llvm_unreachable("undeduced type in IR-generation" ); |
| 260 | |
| 261 | // Various scalar types. |
| 262 | case Type::Builtin: |
| 263 | case Type::Pointer: |
| 264 | case Type::BlockPointer: |
| 265 | case Type::LValueReference: |
| 266 | case Type::RValueReference: |
| 267 | case Type::MemberPointer: |
| 268 | case Type::Vector: |
| 269 | case Type::ExtVector: |
| 270 | case Type::ConstantMatrix: |
| 271 | case Type::FunctionProto: |
| 272 | case Type::FunctionNoProto: |
| 273 | case Type::Enum: |
| 274 | case Type::ObjCObjectPointer: |
| 275 | case Type::Pipe: |
| 276 | case Type::BitInt: |
| 277 | case Type::HLSLAttributedResource: |
| 278 | case Type::HLSLInlineSpirv: |
| 279 | case Type::OverflowBehavior: |
| 280 | return TEK_Scalar; |
| 281 | |
| 282 | // Complexes. |
| 283 | case Type::Complex: |
| 284 | return TEK_Complex; |
| 285 | |
| 286 | // Arrays, records, and Objective-C objects. |
| 287 | case Type::ConstantArray: |
| 288 | case Type::IncompleteArray: |
| 289 | case Type::VariableArray: |
| 290 | case Type::Record: |
| 291 | case Type::ObjCObject: |
| 292 | case Type::ObjCInterface: |
| 293 | case Type::ArrayParameter: |
| 294 | return TEK_Aggregate; |
| 295 | |
| 296 | // We operate on atomic values according to their underlying type. |
| 297 | case Type::Atomic: |
| 298 | type = cast<AtomicType>(Val&: type)->getValueType(); |
| 299 | continue; |
| 300 | } |
| 301 | llvm_unreachable("unknown type kind!" ); |
| 302 | } |
| 303 | } |
| 304 | |
| 305 | llvm::DebugLoc CodeGenFunction::EmitReturnBlock() { |
| 306 | // For cleanliness, we try to avoid emitting the return block for |
| 307 | // simple cases. |
| 308 | llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); |
| 309 | |
| 310 | if (CurBB) { |
| 311 | assert(!CurBB->getTerminator() && "Unexpected terminated block." ); |
| 312 | |
| 313 | // We have a valid insert point, reuse it if it is empty or there are no |
| 314 | // explicit jumps to the return block. |
| 315 | if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) { |
| 316 | ReturnBlock.getBlock()->replaceAllUsesWith(V: CurBB); |
| 317 | delete ReturnBlock.getBlock(); |
| 318 | ReturnBlock = JumpDest(); |
| 319 | } else |
| 320 | EmitBlock(BB: ReturnBlock.getBlock()); |
| 321 | return llvm::DebugLoc(); |
| 322 | } |
| 323 | |
| 324 | // Otherwise, if the return block is the target of a single direct |
| 325 | // branch then we can just put the code in that block instead. This |
| 326 | // cleans up functions which started with a unified return block. |
| 327 | if (ReturnBlock.getBlock()->hasOneUse()) { |
| 328 | llvm::BranchInst *BI = |
| 329 | dyn_cast<llvm::BranchInst>(Val: *ReturnBlock.getBlock()->user_begin()); |
| 330 | if (BI && BI->isUnconditional() && |
| 331 | BI->getSuccessor(Idx: 0) == ReturnBlock.getBlock()) { |
| 332 | // Record/return the DebugLoc of the simple 'return' expression to be used |
| 333 | // later by the actual 'ret' instruction. |
| 334 | llvm::DebugLoc Loc = BI->getDebugLoc(); |
| 335 | Builder.SetInsertPoint(BI->getParent()); |
| 336 | BI->eraseFromParent(); |
| 337 | delete ReturnBlock.getBlock(); |
| 338 | ReturnBlock = JumpDest(); |
| 339 | return Loc; |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | // FIXME: We are at an unreachable point, there is no reason to emit the block |
| 344 | // unless it has uses. However, we still need a place to put the debug |
| 345 | // region.end for now. |
| 346 | |
| 347 | EmitBlock(BB: ReturnBlock.getBlock()); |
| 348 | return llvm::DebugLoc(); |
| 349 | } |
| 350 | |
| 351 | static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) { |
| 352 | if (!BB) return; |
| 353 | if (!BB->use_empty()) { |
| 354 | CGF.CurFn->insert(Position: CGF.CurFn->end(), BB); |
| 355 | return; |
| 356 | } |
| 357 | delete BB; |
| 358 | } |
| 359 | |
| 360 | void CodeGenFunction::FinishFunction(SourceLocation EndLoc) { |
| 361 | assert(BreakContinueStack.empty() && |
| 362 | "mismatched push/pop in break/continue stack!" ); |
| 363 | assert(LifetimeExtendedCleanupStack.empty() && |
| 364 | "mismatched push/pop of cleanups in EHStack!" ); |
| 365 | assert(DeferredDeactivationCleanupStack.empty() && |
| 366 | "mismatched activate/deactivate of cleanups!" ); |
| 367 | |
| 368 | if (CGM.shouldEmitConvergenceTokens()) { |
| 369 | ConvergenceTokenStack.pop_back(); |
| 370 | assert(ConvergenceTokenStack.empty() && |
| 371 | "mismatched push/pop in convergence stack!" ); |
| 372 | } |
| 373 | |
| 374 | bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0 |
| 375 | && NumSimpleReturnExprs == NumReturnExprs |
| 376 | && ReturnBlock.getBlock()->use_empty(); |
| 377 | // Usually the return expression is evaluated before the cleanup |
| 378 | // code. If the function contains only a simple return statement, |
| 379 | // such as a constant, the location before the cleanup code becomes |
| 380 | // the last useful breakpoint in the function, because the simple |
| 381 | // return expression will be evaluated after the cleanup code. To be |
| 382 | // safe, set the debug location for cleanup code to the location of |
| 383 | // the return statement. Otherwise the cleanup code should be at the |
| 384 | // end of the function's lexical scope. |
| 385 | // |
| 386 | // If there are multiple branches to the return block, the branch |
| 387 | // instructions will get the location of the return statements and |
| 388 | // all will be fine. |
| 389 | if (CGDebugInfo *DI = getDebugInfo()) { |
| 390 | if (OnlySimpleReturnStmts) |
| 391 | DI->EmitLocation(Builder, Loc: LastStopPoint); |
| 392 | else |
| 393 | DI->EmitLocation(Builder, Loc: EndLoc); |
| 394 | } |
| 395 | |
| 396 | // Pop any cleanups that might have been associated with the |
| 397 | // parameters. Do this in whatever block we're currently in; it's |
| 398 | // important to do this before we enter the return block or return |
| 399 | // edges will be *really* confused. |
| 400 | bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth; |
| 401 | bool HasOnlyNoopCleanups = |
| 402 | HasCleanups && EHStack.containsOnlyNoopCleanups(Old: PrologueCleanupDepth); |
| 403 | bool EmitRetDbgLoc = !HasCleanups || HasOnlyNoopCleanups; |
| 404 | |
| 405 | std::optional<ApplyDebugLocation> OAL; |
| 406 | if (HasCleanups) { |
| 407 | // Make sure the line table doesn't jump back into the body for |
| 408 | // the ret after it's been at EndLoc. |
| 409 | if (CGDebugInfo *DI = getDebugInfo()) { |
| 410 | if (OnlySimpleReturnStmts) |
| 411 | DI->EmitLocation(Builder, Loc: EndLoc); |
| 412 | else |
| 413 | // We may not have a valid end location. Try to apply it anyway, and |
| 414 | // fall back to an artificial location if needed. |
| 415 | OAL = ApplyDebugLocation::CreateDefaultArtificial(CGF&: *this, TemporaryLocation: EndLoc); |
| 416 | } |
| 417 | |
| 418 | PopCleanupBlocks(OldCleanupStackSize: PrologueCleanupDepth); |
| 419 | } |
| 420 | |
| 421 | // Emit function epilog (to return). |
| 422 | llvm::DebugLoc Loc = EmitReturnBlock(); |
| 423 | |
| 424 | if (ShouldInstrumentFunction()) { |
| 425 | if (CGM.getCodeGenOpts().InstrumentFunctions) |
| 426 | CurFn->addFnAttr(Kind: "instrument-function-exit" , Val: "__cyg_profile_func_exit" ); |
| 427 | if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining) |
| 428 | CurFn->addFnAttr(Kind: "instrument-function-exit-inlined" , |
| 429 | Val: "__cyg_profile_func_exit" ); |
| 430 | } |
| 431 | |
| 432 | // Emit debug descriptor for function end. |
| 433 | if (CGDebugInfo *DI = getDebugInfo()) |
| 434 | DI->EmitFunctionEnd(Builder, Fn: CurFn); |
| 435 | |
| 436 | // Reset the debug location to that of the simple 'return' expression, if any |
| 437 | // rather than that of the end of the function's scope '}'. |
| 438 | uint64_t RetKeyInstructionsAtomGroup = Loc ? Loc->getAtomGroup() : 0; |
| 439 | ApplyDebugLocation AL(*this, Loc); |
| 440 | EmitFunctionEpilog(FI: *CurFnInfo, EmitRetDbgLoc, EndLoc, |
| 441 | RetKeyInstructionsSourceAtom: RetKeyInstructionsAtomGroup); |
| 442 | EmitEndEHSpec(D: CurCodeDecl); |
| 443 | |
| 444 | assert(EHStack.empty() && |
| 445 | "did not remove all scopes from cleanup stack!" ); |
| 446 | |
| 447 | // If someone did an indirect goto, emit the indirect goto block at the end of |
| 448 | // the function. |
| 449 | if (IndirectBranch) { |
| 450 | EmitBlock(BB: IndirectBranch->getParent()); |
| 451 | Builder.ClearInsertionPoint(); |
| 452 | } |
| 453 | |
| 454 | // If some of our locals escaped, insert a call to llvm.localescape in the |
| 455 | // entry block. |
| 456 | if (!EscapedLocals.empty()) { |
| 457 | // Invert the map from local to index into a simple vector. There should be |
| 458 | // no holes. |
| 459 | SmallVector<llvm::Value *, 4> EscapeArgs; |
| 460 | EscapeArgs.resize(N: EscapedLocals.size()); |
| 461 | for (auto &Pair : EscapedLocals) |
| 462 | EscapeArgs[Pair.second] = Pair.first; |
| 463 | llvm::Function *FrameEscapeFn = llvm::Intrinsic::getOrInsertDeclaration( |
| 464 | M: &CGM.getModule(), id: llvm::Intrinsic::localescape); |
| 465 | CGBuilderTy(CGM, AllocaInsertPt).CreateCall(Callee: FrameEscapeFn, Args: EscapeArgs); |
| 466 | } |
| 467 | |
| 468 | // Remove the AllocaInsertPt instruction, which is just a convenience for us. |
| 469 | llvm::Instruction *Ptr = AllocaInsertPt; |
| 470 | AllocaInsertPt = nullptr; |
| 471 | Ptr->eraseFromParent(); |
| 472 | |
| 473 | // PostAllocaInsertPt, if created, was lazily created when it was required, |
| 474 | // remove it now since it was just created for our own convenience. |
| 475 | if (PostAllocaInsertPt) { |
| 476 | llvm::Instruction *PostPtr = PostAllocaInsertPt; |
| 477 | PostAllocaInsertPt = nullptr; |
| 478 | PostPtr->eraseFromParent(); |
| 479 | } |
| 480 | |
| 481 | // If someone took the address of a label but never did an indirect goto, we |
| 482 | // made a zero entry PHI node, which is illegal, zap it now. |
| 483 | if (IndirectBranch) { |
| 484 | llvm::PHINode *PN = cast<llvm::PHINode>(Val: IndirectBranch->getAddress()); |
| 485 | if (PN->getNumIncomingValues() == 0) { |
| 486 | PN->replaceAllUsesWith(V: llvm::PoisonValue::get(T: PN->getType())); |
| 487 | PN->eraseFromParent(); |
| 488 | } |
| 489 | } |
| 490 | |
| 491 | EmitIfUsed(CGF&: *this, BB: EHResumeBlock); |
| 492 | EmitIfUsed(CGF&: *this, BB: TerminateLandingPad); |
| 493 | EmitIfUsed(CGF&: *this, BB: TerminateHandler); |
| 494 | EmitIfUsed(CGF&: *this, BB: UnreachableBlock); |
| 495 | |
| 496 | for (const auto &FuncletAndParent : TerminateFunclets) |
| 497 | EmitIfUsed(CGF&: *this, BB: FuncletAndParent.second); |
| 498 | |
| 499 | if (CGM.getCodeGenOpts().EmitDeclMetadata) |
| 500 | EmitDeclMetadata(); |
| 501 | |
| 502 | for (const auto &R : DeferredReplacements) { |
| 503 | if (llvm::Value *Old = R.first) { |
| 504 | Old->replaceAllUsesWith(V: R.second); |
| 505 | cast<llvm::Instruction>(Val: Old)->eraseFromParent(); |
| 506 | } |
| 507 | } |
| 508 | DeferredReplacements.clear(); |
| 509 | |
| 510 | // Eliminate CleanupDestSlot alloca by replacing it with SSA values and |
| 511 | // PHIs if the current function is a coroutine. We don't do it for all |
| 512 | // functions as it may result in slight increase in numbers of instructions |
| 513 | // if compiled with no optimizations. We do it for coroutine as the lifetime |
| 514 | // of CleanupDestSlot alloca make correct coroutine frame building very |
| 515 | // difficult. |
| 516 | if (NormalCleanupDest.isValid() && isCoroutine()) { |
| 517 | llvm::DominatorTree DT(*CurFn); |
| 518 | llvm::PromoteMemToReg( |
| 519 | Allocas: cast<llvm::AllocaInst>(Val: NormalCleanupDest.getPointer()), DT); |
| 520 | NormalCleanupDest = Address::invalid(); |
| 521 | } |
| 522 | |
| 523 | // Scan function arguments for vector width. |
| 524 | for (llvm::Argument &A : CurFn->args()) |
| 525 | if (auto *VT = dyn_cast<llvm::VectorType>(Val: A.getType())) |
| 526 | LargestVectorWidth = |
| 527 | std::max(a: (uint64_t)LargestVectorWidth, |
| 528 | b: VT->getPrimitiveSizeInBits().getKnownMinValue()); |
| 529 | |
| 530 | // Update vector width based on return type. |
| 531 | if (auto *VT = dyn_cast<llvm::VectorType>(Val: CurFn->getReturnType())) |
| 532 | LargestVectorWidth = |
| 533 | std::max(a: (uint64_t)LargestVectorWidth, |
| 534 | b: VT->getPrimitiveSizeInBits().getKnownMinValue()); |
| 535 | |
| 536 | if (CurFnInfo->getMaxVectorWidth() > LargestVectorWidth) |
| 537 | LargestVectorWidth = CurFnInfo->getMaxVectorWidth(); |
| 538 | |
| 539 | // Add the min-legal-vector-width attribute. This contains the max width from: |
| 540 | // 1. min-vector-width attribute used in the source program. |
| 541 | // 2. Any builtins used that have a vector width specified. |
| 542 | // 3. Values passed in and out of inline assembly. |
| 543 | // 4. Width of vector arguments and return types for this function. |
| 544 | // 5. Width of vector arguments and return types for functions called by this |
| 545 | // function. |
| 546 | if (getContext().getTargetInfo().getTriple().isX86()) |
| 547 | CurFn->addFnAttr(Kind: "min-legal-vector-width" , |
| 548 | Val: llvm::utostr(X: LargestVectorWidth)); |
| 549 | |
| 550 | // If we generated an unreachable return block, delete it now. |
| 551 | if (ReturnBlock.isValid() && ReturnBlock.getBlock()->use_empty()) { |
| 552 | Builder.ClearInsertionPoint(); |
| 553 | ReturnBlock.getBlock()->eraseFromParent(); |
| 554 | } |
| 555 | if (ReturnValue.isValid()) { |
| 556 | auto *RetAlloca = |
| 557 | dyn_cast<llvm::AllocaInst>(Val: ReturnValue.emitRawPointer(CGF&: *this)); |
| 558 | if (RetAlloca && RetAlloca->use_empty()) { |
| 559 | RetAlloca->eraseFromParent(); |
| 560 | ReturnValue = Address::invalid(); |
| 561 | } |
| 562 | } |
| 563 | } |
| 564 | |
| 565 | /// ShouldInstrumentFunction - Return true if the current function should be |
| 566 | /// instrumented with __cyg_profile_func_* calls |
| 567 | bool CodeGenFunction::ShouldInstrumentFunction() { |
| 568 | if (!CGM.getCodeGenOpts().InstrumentFunctions && |
| 569 | !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining && |
| 570 | !CGM.getCodeGenOpts().InstrumentFunctionEntryBare) |
| 571 | return false; |
| 572 | if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) |
| 573 | return false; |
| 574 | return true; |
| 575 | } |
| 576 | |
| 577 | bool CodeGenFunction::ShouldSkipSanitizerInstrumentation() { |
| 578 | if (!CurFuncDecl) |
| 579 | return false; |
| 580 | return CurFuncDecl->hasAttr<DisableSanitizerInstrumentationAttr>(); |
| 581 | } |
| 582 | |
| 583 | /// ShouldXRayInstrument - Return true if the current function should be |
| 584 | /// instrumented with XRay nop sleds. |
| 585 | bool CodeGenFunction::ShouldXRayInstrumentFunction() const { |
| 586 | return CGM.getCodeGenOpts().XRayInstrumentFunctions; |
| 587 | } |
| 588 | |
| 589 | /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to |
| 590 | /// the __xray_customevent(...) builtin calls, when doing XRay instrumentation. |
| 591 | bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const { |
| 592 | return CGM.getCodeGenOpts().XRayInstrumentFunctions && |
| 593 | (CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents || |
| 594 | CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask == |
| 595 | XRayInstrKind::Custom); |
| 596 | } |
| 597 | |
| 598 | bool CodeGenFunction::AlwaysEmitXRayTypedEvents() const { |
| 599 | return CGM.getCodeGenOpts().XRayInstrumentFunctions && |
| 600 | (CGM.getCodeGenOpts().XRayAlwaysEmitTypedEvents || |
| 601 | CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask == |
| 602 | XRayInstrKind::Typed); |
| 603 | } |
| 604 | |
| 605 | llvm::ConstantInt * |
| 606 | CodeGenFunction::getUBSanFunctionTypeHash(QualType Ty) const { |
| 607 | // Remove any (C++17) exception specifications, to allow calling e.g. a |
| 608 | // noexcept function through a non-noexcept pointer. |
| 609 | if (!Ty->isFunctionNoProtoType()) |
| 610 | Ty = getContext().getFunctionTypeWithExceptionSpec(Orig: Ty, ESI: EST_None); |
| 611 | std::string Mangled; |
| 612 | llvm::raw_string_ostream Out(Mangled); |
| 613 | CGM.getCXXABI().getMangleContext().mangleCanonicalTypeName(T: Ty, Out, NormalizeIntegers: false); |
| 614 | return llvm::ConstantInt::get( |
| 615 | Ty: CGM.Int32Ty, V: static_cast<uint32_t>(llvm::xxh3_64bits(data: Mangled))); |
| 616 | } |
| 617 | |
| 618 | void CodeGenFunction::EmitKernelMetadata(const FunctionDecl *FD, |
| 619 | llvm::Function *Fn) { |
| 620 | if (!FD->hasAttr<DeviceKernelAttr>() && !FD->hasAttr<CUDAGlobalAttr>()) |
| 621 | return; |
| 622 | |
| 623 | llvm::LLVMContext &Context = getLLVMContext(); |
| 624 | |
| 625 | CGM.GenKernelArgMetadata(FN: Fn, FD, CGF: this); |
| 626 | |
| 627 | if (!(getLangOpts().OpenCL || |
| 628 | (getLangOpts().CUDA && |
| 629 | getContext().getTargetInfo().getTriple().isSPIRV()))) |
| 630 | return; |
| 631 | |
| 632 | if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) { |
| 633 | QualType HintQTy = A->getTypeHint(); |
| 634 | const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>(); |
| 635 | bool IsSignedInteger = |
| 636 | HintQTy->isSignedIntegerType() || |
| 637 | (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType()); |
| 638 | llvm::Metadata *AttrMDArgs[] = { |
| 639 | llvm::ConstantAsMetadata::get(C: llvm::PoisonValue::get( |
| 640 | T: CGM.getTypes().ConvertType(T: A->getTypeHint()))), |
| 641 | llvm::ConstantAsMetadata::get(C: llvm::ConstantInt::get( |
| 642 | Ty: llvm::IntegerType::get(C&: Context, NumBits: 32), |
| 643 | V: llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))}; |
| 644 | Fn->setMetadata(Kind: "vec_type_hint" , Node: llvm::MDNode::get(Context, MDs: AttrMDArgs)); |
| 645 | } |
| 646 | |
| 647 | if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) { |
| 648 | auto Eval = [&](Expr *E) { |
| 649 | return E->EvaluateKnownConstInt(Ctx: FD->getASTContext()).getExtValue(); |
| 650 | }; |
| 651 | llvm::Metadata *AttrMDArgs[] = { |
| 652 | llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: Eval(A->getXDim()))), |
| 653 | llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: Eval(A->getYDim()))), |
| 654 | llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: Eval(A->getZDim())))}; |
| 655 | Fn->setMetadata(Kind: "work_group_size_hint" , Node: llvm::MDNode::get(Context, MDs: AttrMDArgs)); |
| 656 | } |
| 657 | |
| 658 | if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) { |
| 659 | auto Eval = [&](Expr *E) { |
| 660 | return E->EvaluateKnownConstInt(Ctx: FD->getASTContext()).getExtValue(); |
| 661 | }; |
| 662 | llvm::Metadata *AttrMDArgs[] = { |
| 663 | llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: Eval(A->getXDim()))), |
| 664 | llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: Eval(A->getYDim()))), |
| 665 | llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: Eval(A->getZDim())))}; |
| 666 | Fn->setMetadata(Kind: "reqd_work_group_size" , Node: llvm::MDNode::get(Context, MDs: AttrMDArgs)); |
| 667 | } |
| 668 | |
| 669 | if (const OpenCLIntelReqdSubGroupSizeAttr *A = |
| 670 | FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) { |
| 671 | llvm::Metadata *AttrMDArgs[] = { |
| 672 | llvm::ConstantAsMetadata::get(C: Builder.getInt32(C: A->getSubGroupSize()))}; |
| 673 | Fn->setMetadata(Kind: "intel_reqd_sub_group_size" , |
| 674 | Node: llvm::MDNode::get(Context, MDs: AttrMDArgs)); |
| 675 | } |
| 676 | } |
| 677 | |
| 678 | /// Determine whether the function F ends with a return stmt. |
| 679 | static bool endsWithReturn(const Decl* F) { |
| 680 | const Stmt *Body = nullptr; |
| 681 | if (auto *FD = dyn_cast_or_null<FunctionDecl>(Val: F)) |
| 682 | Body = FD->getBody(); |
| 683 | else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(Val: F)) |
| 684 | Body = OMD->getBody(); |
| 685 | |
| 686 | if (auto *CS = dyn_cast_or_null<CompoundStmt>(Val: Body)) { |
| 687 | auto LastStmt = CS->body_rbegin(); |
| 688 | if (LastStmt != CS->body_rend()) |
| 689 | return isa<ReturnStmt>(Val: *LastStmt); |
| 690 | } |
| 691 | return false; |
| 692 | } |
| 693 | |
| 694 | void CodeGenFunction::markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) { |
| 695 | if (SanOpts.has(K: SanitizerKind::Thread)) { |
| 696 | Fn->addFnAttr(Kind: "sanitize_thread_no_checking_at_run_time" ); |
| 697 | Fn->removeFnAttr(Kind: llvm::Attribute::SanitizeThread); |
| 698 | } |
| 699 | } |
| 700 | |
| 701 | /// Check if the return value of this function requires sanitization. |
| 702 | bool CodeGenFunction::requiresReturnValueCheck() const { |
| 703 | return requiresReturnValueNullabilityCheck() || |
| 704 | (SanOpts.has(K: SanitizerKind::ReturnsNonnullAttribute) && CurCodeDecl && |
| 705 | CurCodeDecl->getAttr<ReturnsNonNullAttr>()); |
| 706 | } |
| 707 | |
| 708 | static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) { |
| 709 | auto *MD = dyn_cast_or_null<CXXMethodDecl>(Val: D); |
| 710 | if (!MD || !MD->getDeclName().getAsIdentifierInfo() || |
| 711 | !MD->getDeclName().getAsIdentifierInfo()->isStr(Str: "allocate" ) || |
| 712 | (MD->getNumParams() != 1 && MD->getNumParams() != 2)) |
| 713 | return false; |
| 714 | |
| 715 | if (!Ctx.hasSameType(T1: MD->parameters()[0]->getType(), T2: Ctx.getSizeType())) |
| 716 | return false; |
| 717 | |
| 718 | if (MD->getNumParams() == 2) { |
| 719 | auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>(); |
| 720 | if (!PT || !PT->isVoidPointerType() || |
| 721 | !PT->getPointeeType().isConstQualified()) |
| 722 | return false; |
| 723 | } |
| 724 | |
| 725 | return true; |
| 726 | } |
| 727 | |
| 728 | bool CodeGenFunction::isInAllocaArgument(CGCXXABI &ABI, QualType Ty) { |
| 729 | const CXXRecordDecl *RD = Ty->getAsCXXRecordDecl(); |
| 730 | return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory; |
| 731 | } |
| 732 | |
| 733 | bool CodeGenFunction::hasInAllocaArg(const CXXMethodDecl *MD) { |
| 734 | return getTarget().getTriple().getArch() == llvm::Triple::x86 && |
| 735 | getTarget().getCXXABI().isMicrosoft() && |
| 736 | llvm::any_of(Range: MD->parameters(), P: [&](ParmVarDecl *P) { |
| 737 | return isInAllocaArgument(ABI&: CGM.getCXXABI(), Ty: P->getType()); |
| 738 | }); |
| 739 | } |
| 740 | |
| 741 | /// Return the UBSan prologue signature for \p FD if one is available. |
| 742 | static llvm::Constant *getPrologueSignature(CodeGenModule &CGM, |
| 743 | const FunctionDecl *FD) { |
| 744 | if (const auto *MD = dyn_cast<CXXMethodDecl>(Val: FD)) |
| 745 | if (!MD->isStatic()) |
| 746 | return nullptr; |
| 747 | return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM); |
| 748 | } |
| 749 | |
| 750 | void CodeGenFunction::StartFunction(GlobalDecl GD, QualType RetTy, |
| 751 | llvm::Function *Fn, |
| 752 | const CGFunctionInfo &FnInfo, |
| 753 | const FunctionArgList &Args, |
| 754 | SourceLocation Loc, |
| 755 | SourceLocation StartLoc) { |
| 756 | assert(!CurFn && |
| 757 | "Do not use a CodeGenFunction object for more than one function" ); |
| 758 | |
| 759 | const Decl *D = GD.getDecl(); |
| 760 | |
| 761 | DidCallStackSave = false; |
| 762 | CurCodeDecl = D; |
| 763 | const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D); |
| 764 | if (FD && FD->usesSEHTry()) |
| 765 | CurSEHParent = GD; |
| 766 | CurFuncDecl = (D ? D->getNonClosureContext() : nullptr); |
| 767 | FnRetTy = RetTy; |
| 768 | CurFn = Fn; |
| 769 | CurFnInfo = &FnInfo; |
| 770 | assert(CurFn->isDeclaration() && "Function already has body?" ); |
| 771 | |
| 772 | // If this function is ignored for any of the enabled sanitizers, |
| 773 | // disable the sanitizer for the function. |
| 774 | do { |
| 775 | #define SANITIZER(NAME, ID) \ |
| 776 | if (SanOpts.empty()) \ |
| 777 | break; \ |
| 778 | if (SanOpts.has(SanitizerKind::ID)) \ |
| 779 | if (CGM.isInNoSanitizeList(SanitizerKind::ID, Fn, Loc)) \ |
| 780 | SanOpts.set(SanitizerKind::ID, false); |
| 781 | |
| 782 | #include "clang/Basic/Sanitizers.def" |
| 783 | #undef SANITIZER |
| 784 | } while (false); |
| 785 | |
| 786 | if (D) { |
| 787 | const bool SanitizeBounds = SanOpts.hasOneOf(K: SanitizerKind::Bounds); |
| 788 | SanitizerMask no_sanitize_mask; |
| 789 | bool NoSanitizeCoverage = false; |
| 790 | |
| 791 | for (auto *Attr : D->specific_attrs<NoSanitizeAttr>()) { |
| 792 | no_sanitize_mask |= Attr->getMask(); |
| 793 | // SanitizeCoverage is not handled by SanOpts. |
| 794 | if (Attr->hasCoverage()) |
| 795 | NoSanitizeCoverage = true; |
| 796 | } |
| 797 | |
| 798 | // Apply the no_sanitize* attributes to SanOpts. |
| 799 | SanOpts.Mask &= ~no_sanitize_mask; |
| 800 | if (no_sanitize_mask & SanitizerKind::Address) |
| 801 | SanOpts.set(K: SanitizerKind::KernelAddress, Value: false); |
| 802 | if (no_sanitize_mask & SanitizerKind::KernelAddress) |
| 803 | SanOpts.set(K: SanitizerKind::Address, Value: false); |
| 804 | if (no_sanitize_mask & SanitizerKind::HWAddress) |
| 805 | SanOpts.set(K: SanitizerKind::KernelHWAddress, Value: false); |
| 806 | if (no_sanitize_mask & SanitizerKind::KernelHWAddress) |
| 807 | SanOpts.set(K: SanitizerKind::HWAddress, Value: false); |
| 808 | |
| 809 | if (SanitizeBounds && !SanOpts.hasOneOf(K: SanitizerKind::Bounds)) |
| 810 | Fn->addFnAttr(Kind: llvm::Attribute::NoSanitizeBounds); |
| 811 | |
| 812 | if (NoSanitizeCoverage && CGM.getCodeGenOpts().hasSanitizeCoverage()) |
| 813 | Fn->addFnAttr(Kind: llvm::Attribute::NoSanitizeCoverage); |
| 814 | |
| 815 | // Some passes need the non-negated no_sanitize attribute. Pass them on. |
| 816 | if (CGM.getCodeGenOpts().hasSanitizeBinaryMetadata()) { |
| 817 | if (no_sanitize_mask & SanitizerKind::Thread) |
| 818 | Fn->addFnAttr(Kind: "no_sanitize_thread" ); |
| 819 | } |
| 820 | } |
| 821 | |
| 822 | if (ShouldSkipSanitizerInstrumentation()) { |
| 823 | CurFn->addFnAttr(Kind: llvm::Attribute::DisableSanitizerInstrumentation); |
| 824 | } else { |
| 825 | // Apply sanitizer attributes to the function. |
| 826 | if (SanOpts.hasOneOf(K: SanitizerKind::Address | SanitizerKind::KernelAddress)) |
| 827 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeAddress); |
| 828 | if (SanOpts.hasOneOf(K: SanitizerKind::HWAddress | |
| 829 | SanitizerKind::KernelHWAddress)) |
| 830 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeHWAddress); |
| 831 | if (SanOpts.has(K: SanitizerKind::MemtagStack)) |
| 832 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeMemTag); |
| 833 | if (SanOpts.has(K: SanitizerKind::Thread)) |
| 834 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeThread); |
| 835 | if (SanOpts.has(K: SanitizerKind::Type)) |
| 836 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeType); |
| 837 | if (SanOpts.has(K: SanitizerKind::NumericalStability)) |
| 838 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeNumericalStability); |
| 839 | if (SanOpts.hasOneOf(K: SanitizerKind::Memory | SanitizerKind::KernelMemory)) |
| 840 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeMemory); |
| 841 | if (SanOpts.has(K: SanitizerKind::AllocToken)) |
| 842 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeAllocToken); |
| 843 | } |
| 844 | if (SanOpts.has(K: SanitizerKind::SafeStack)) |
| 845 | Fn->addFnAttr(Kind: llvm::Attribute::SafeStack); |
| 846 | if (SanOpts.has(K: SanitizerKind::ShadowCallStack)) |
| 847 | Fn->addFnAttr(Kind: llvm::Attribute::ShadowCallStack); |
| 848 | |
| 849 | if (SanOpts.has(K: SanitizerKind::Realtime)) |
| 850 | if (FD && FD->getASTContext().hasAnyFunctionEffects()) |
| 851 | for (const FunctionEffectWithCondition &Fe : FD->getFunctionEffects()) { |
| 852 | if (Fe.Effect.kind() == FunctionEffect::Kind::NonBlocking) |
| 853 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeRealtime); |
| 854 | else if (Fe.Effect.kind() == FunctionEffect::Kind::Blocking) |
| 855 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeRealtimeBlocking); |
| 856 | } |
| 857 | |
| 858 | // Apply fuzzing attribute to the function. |
| 859 | if (SanOpts.hasOneOf(K: SanitizerKind::Fuzzer | SanitizerKind::FuzzerNoLink)) |
| 860 | Fn->addFnAttr(Kind: llvm::Attribute::OptForFuzzing); |
| 861 | |
| 862 | // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize, |
| 863 | // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time. |
| 864 | if (SanOpts.has(K: SanitizerKind::Thread)) { |
| 865 | if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(Val: D)) { |
| 866 | const IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(argIndex: 0); |
| 867 | if (OMD->getMethodFamily() == OMF_dealloc || |
| 868 | OMD->getMethodFamily() == OMF_initialize || |
| 869 | (OMD->getSelector().isUnarySelector() && II->isStr(Str: ".cxx_destruct" ))) { |
| 870 | markAsIgnoreThreadCheckingAtRuntime(Fn); |
| 871 | } |
| 872 | } |
| 873 | } |
| 874 | |
| 875 | // Ignore unrelated casts in STL allocate() since the allocator must cast |
| 876 | // from void* to T* before object initialization completes. Don't match on the |
| 877 | // namespace because not all allocators are in std:: |
| 878 | if (D && SanOpts.has(K: SanitizerKind::CFIUnrelatedCast)) { |
| 879 | if (matchesStlAllocatorFn(D, Ctx: getContext())) |
| 880 | SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast; |
| 881 | } |
| 882 | |
| 883 | // Ignore null checks in coroutine functions since the coroutines passes |
| 884 | // are not aware of how to move the extra UBSan instructions across the split |
| 885 | // coroutine boundaries. |
| 886 | if (D && SanOpts.has(K: SanitizerKind::Null)) |
| 887 | if (FD && FD->getBody() && |
| 888 | FD->getBody()->getStmtClass() == Stmt::CoroutineBodyStmtClass) |
| 889 | SanOpts.Mask &= ~SanitizerKind::Null; |
| 890 | |
| 891 | // Apply xray attributes to the function (as a string, for now) |
| 892 | bool AlwaysXRayAttr = false; |
| 893 | if (const auto *XRayAttr = D ? D->getAttr<XRayInstrumentAttr>() : nullptr) { |
| 894 | if (CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
| 895 | K: XRayInstrKind::FunctionEntry) || |
| 896 | CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
| 897 | K: XRayInstrKind::FunctionExit)) { |
| 898 | if (XRayAttr->alwaysXRayInstrument() && ShouldXRayInstrumentFunction()) { |
| 899 | Fn->addFnAttr(Kind: "function-instrument" , Val: "xray-always" ); |
| 900 | AlwaysXRayAttr = true; |
| 901 | } |
| 902 | if (XRayAttr->neverXRayInstrument()) |
| 903 | Fn->addFnAttr(Kind: "function-instrument" , Val: "xray-never" ); |
| 904 | if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) |
| 905 | if (ShouldXRayInstrumentFunction()) |
| 906 | Fn->addFnAttr(Kind: "xray-log-args" , |
| 907 | Val: llvm::utostr(X: LogArgs->getArgumentCount())); |
| 908 | } |
| 909 | } else { |
| 910 | if (ShouldXRayInstrumentFunction() && !CGM.imbueXRayAttrs(Fn, Loc)) |
| 911 | Fn->addFnAttr( |
| 912 | Kind: "xray-instruction-threshold" , |
| 913 | Val: llvm::itostr(X: CGM.getCodeGenOpts().XRayInstructionThreshold)); |
| 914 | } |
| 915 | |
| 916 | if (ShouldXRayInstrumentFunction()) { |
| 917 | if (CGM.getCodeGenOpts().XRayIgnoreLoops) |
| 918 | Fn->addFnAttr(Kind: "xray-ignore-loops" ); |
| 919 | |
| 920 | if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
| 921 | K: XRayInstrKind::FunctionExit)) |
| 922 | Fn->addFnAttr(Kind: "xray-skip-exit" ); |
| 923 | |
| 924 | if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
| 925 | K: XRayInstrKind::FunctionEntry)) |
| 926 | Fn->addFnAttr(Kind: "xray-skip-entry" ); |
| 927 | |
| 928 | auto FuncGroups = CGM.getCodeGenOpts().XRayTotalFunctionGroups; |
| 929 | if (FuncGroups > 1) { |
| 930 | auto FuncName = llvm::ArrayRef<uint8_t>(CurFn->getName().bytes_begin(), |
| 931 | CurFn->getName().bytes_end()); |
| 932 | auto Group = crc32(Data: FuncName) % FuncGroups; |
| 933 | if (Group != CGM.getCodeGenOpts().XRaySelectedFunctionGroup && |
| 934 | !AlwaysXRayAttr) |
| 935 | Fn->addFnAttr(Kind: "function-instrument" , Val: "xray-never" ); |
| 936 | } |
| 937 | } |
| 938 | |
| 939 | if (CGM.getCodeGenOpts().getProfileInstr() != |
| 940 | llvm::driver::ProfileInstrKind::ProfileNone) { |
| 941 | switch (CGM.isFunctionBlockedFromProfileInstr(Fn, Loc)) { |
| 942 | case ProfileList::Skip: |
| 943 | Fn->addFnAttr(Kind: llvm::Attribute::SkipProfile); |
| 944 | break; |
| 945 | case ProfileList::Forbid: |
| 946 | Fn->addFnAttr(Kind: llvm::Attribute::NoProfile); |
| 947 | break; |
| 948 | case ProfileList::Allow: |
| 949 | break; |
| 950 | } |
| 951 | } |
| 952 | |
| 953 | unsigned Count, Offset; |
| 954 | StringRef Section; |
| 955 | if (const auto *Attr = |
| 956 | D ? D->getAttr<PatchableFunctionEntryAttr>() : nullptr) { |
| 957 | Count = Attr->getCount(); |
| 958 | Offset = Attr->getOffset(); |
| 959 | Section = Attr->getSection(); |
| 960 | } else { |
| 961 | Count = CGM.getCodeGenOpts().PatchableFunctionEntryCount; |
| 962 | Offset = CGM.getCodeGenOpts().PatchableFunctionEntryOffset; |
| 963 | } |
| 964 | if (Section.empty()) |
| 965 | Section = CGM.getCodeGenOpts().PatchableFunctionEntrySection; |
| 966 | if (Count && Offset <= Count) { |
| 967 | Fn->addFnAttr(Kind: "patchable-function-entry" , Val: std::to_string(val: Count - Offset)); |
| 968 | if (Offset) |
| 969 | Fn->addFnAttr(Kind: "patchable-function-prefix" , Val: std::to_string(val: Offset)); |
| 970 | if (!Section.empty()) |
| 971 | Fn->addFnAttr(Kind: "patchable-function-entry-section" , Val: Section); |
| 972 | } |
| 973 | // Instruct that functions for COFF/CodeView targets should start with a |
| 974 | // patchable instruction, but only on x86/x64. Don't forward this to ARM/ARM64 |
| 975 | // backends as they don't need it -- instructions on these architectures are |
| 976 | // always atomically patchable at runtime. |
| 977 | if (CGM.getCodeGenOpts().HotPatch && |
| 978 | getContext().getTargetInfo().getTriple().isX86() && |
| 979 | getContext().getTargetInfo().getTriple().getEnvironment() != |
| 980 | llvm::Triple::CODE16) |
| 981 | Fn->addFnAttr(Kind: "patchable-function" , Val: "prologue-short-redirect" ); |
| 982 | |
| 983 | // Add no-jump-tables value. |
| 984 | if (CGM.getCodeGenOpts().NoUseJumpTables) |
| 985 | Fn->addFnAttr(Kind: "no-jump-tables" , Val: "true" ); |
| 986 | |
| 987 | // Add no-inline-line-tables value. |
| 988 | if (CGM.getCodeGenOpts().NoInlineLineTables) |
| 989 | Fn->addFnAttr(Kind: "no-inline-line-tables" ); |
| 990 | |
| 991 | // Add profile-sample-accurate value. |
| 992 | if (CGM.getCodeGenOpts().ProfileSampleAccurate) |
| 993 | Fn->addFnAttr(Kind: "profile-sample-accurate" ); |
| 994 | |
| 995 | if (!CGM.getCodeGenOpts().SampleProfileFile.empty()) |
| 996 | Fn->addFnAttr(Kind: "use-sample-profile" ); |
| 997 | |
| 998 | if (D && D->hasAttr<CFICanonicalJumpTableAttr>()) |
| 999 | Fn->addFnAttr(Kind: "cfi-canonical-jump-table" ); |
| 1000 | |
| 1001 | if (D && D->hasAttr<NoProfileFunctionAttr>()) |
| 1002 | Fn->addFnAttr(Kind: llvm::Attribute::NoProfile); |
| 1003 | |
| 1004 | if (D && D->hasAttr<HybridPatchableAttr>()) |
| 1005 | Fn->addFnAttr(Kind: llvm::Attribute::HybridPatchable); |
| 1006 | |
| 1007 | if (D) { |
| 1008 | // Function attributes take precedence over command line flags. |
| 1009 | if (auto *A = D->getAttr<FunctionReturnThunksAttr>()) { |
| 1010 | switch (A->getThunkType()) { |
| 1011 | case FunctionReturnThunksAttr::Kind::Keep: |
| 1012 | break; |
| 1013 | case FunctionReturnThunksAttr::Kind::Extern: |
| 1014 | Fn->addFnAttr(Kind: llvm::Attribute::FnRetThunkExtern); |
| 1015 | break; |
| 1016 | } |
| 1017 | } else if (CGM.getCodeGenOpts().FunctionReturnThunks) |
| 1018 | Fn->addFnAttr(Kind: llvm::Attribute::FnRetThunkExtern); |
| 1019 | } |
| 1020 | |
| 1021 | if (FD && (getLangOpts().OpenCL || |
| 1022 | (getLangOpts().CUDA && |
| 1023 | getContext().getTargetInfo().getTriple().isSPIRV()) || |
| 1024 | ((getLangOpts().HIP || getLangOpts().OffloadViaLLVM) && |
| 1025 | getLangOpts().CUDAIsDevice))) { |
| 1026 | // Add metadata for a kernel function. |
| 1027 | EmitKernelMetadata(FD, Fn); |
| 1028 | } |
| 1029 | |
| 1030 | if (FD && FD->hasAttr<ClspvLibclcBuiltinAttr>()) { |
| 1031 | Fn->setMetadata(Kind: "clspv_libclc_builtin" , |
| 1032 | Node: llvm::MDNode::get(Context&: getLLVMContext(), MDs: {})); |
| 1033 | } |
| 1034 | |
| 1035 | // If we are checking function types, emit a function type signature as |
| 1036 | // prologue data. |
| 1037 | if (FD && SanOpts.has(K: SanitizerKind::Function) && |
| 1038 | !FD->getType()->isCFIUncheckedCalleeFunctionType()) { |
| 1039 | if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) { |
| 1040 | llvm::LLVMContext &Ctx = Fn->getContext(); |
| 1041 | llvm::MDBuilder MDB(Ctx); |
| 1042 | Fn->setMetadata( |
| 1043 | KindID: llvm::LLVMContext::MD_func_sanitize, |
| 1044 | Node: MDB.createRTTIPointerPrologue( |
| 1045 | PrologueSig, RTTI: getUBSanFunctionTypeHash(Ty: FD->getType()))); |
| 1046 | } |
| 1047 | } |
| 1048 | |
| 1049 | // If we're checking nullability, we need to know whether we can check the |
| 1050 | // return value. Initialize the flag to 'true' and refine it in EmitParmDecl. |
| 1051 | if (SanOpts.has(K: SanitizerKind::NullabilityReturn)) { |
| 1052 | auto Nullability = FnRetTy->getNullability(); |
| 1053 | if (Nullability && *Nullability == NullabilityKind::NonNull && |
| 1054 | !FnRetTy->isRecordType()) { |
| 1055 | if (!(SanOpts.has(K: SanitizerKind::ReturnsNonnullAttribute) && |
| 1056 | CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>())) |
| 1057 | RetValNullabilityPrecondition = |
| 1058 | llvm::ConstantInt::getTrue(Context&: getLLVMContext()); |
| 1059 | } |
| 1060 | } |
| 1061 | |
| 1062 | // If we're in C++ mode and the function name is "main", it is guaranteed |
| 1063 | // to be norecurse by the standard (3.6.1.3 "The function main shall not be |
| 1064 | // used within a program"). |
| 1065 | // |
| 1066 | // OpenCL C 2.0 v2.2-11 s6.9.i: |
| 1067 | // Recursion is not supported. |
| 1068 | // |
| 1069 | // HLSL |
| 1070 | // Recursion is not supported. |
| 1071 | // |
| 1072 | // SYCL v1.2.1 s3.10: |
| 1073 | // kernels cannot include RTTI information, exception classes, |
| 1074 | // recursive code, virtual functions or make use of C++ libraries that |
| 1075 | // are not compiled for the device. |
| 1076 | if (FD && |
| 1077 | ((getLangOpts().CPlusPlus && FD->isMain()) || getLangOpts().OpenCL || |
| 1078 | getLangOpts().HLSL || getLangOpts().SYCLIsDevice || |
| 1079 | (getLangOpts().CUDA && FD->hasAttr<CUDAGlobalAttr>()))) |
| 1080 | Fn->addFnAttr(Kind: llvm::Attribute::NoRecurse); |
| 1081 | |
| 1082 | llvm::RoundingMode RM = getLangOpts().getDefaultRoundingMode(); |
| 1083 | llvm::fp::ExceptionBehavior FPExceptionBehavior = |
| 1084 | ToConstrainedExceptMD(Kind: getLangOpts().getDefaultExceptionMode()); |
| 1085 | Builder.setDefaultConstrainedRounding(RM); |
| 1086 | Builder.setDefaultConstrainedExcept(FPExceptionBehavior); |
| 1087 | if ((FD && (FD->UsesFPIntrin() || FD->hasAttr<StrictFPAttr>())) || |
| 1088 | (!FD && (FPExceptionBehavior != llvm::fp::ebIgnore || |
| 1089 | RM != llvm::RoundingMode::NearestTiesToEven))) { |
| 1090 | Builder.setIsFPConstrained(true); |
| 1091 | Fn->addFnAttr(Kind: llvm::Attribute::StrictFP); |
| 1092 | } |
| 1093 | |
| 1094 | // If a custom alignment is used, force realigning to this alignment on |
| 1095 | // any main function which certainly will need it. |
| 1096 | if (FD && ((FD->isMain() || FD->isMSVCRTEntryPoint()) && |
| 1097 | CGM.getCodeGenOpts().StackAlignment)) |
| 1098 | Fn->addFnAttr(Kind: "stackrealign" ); |
| 1099 | |
| 1100 | // "main" doesn't need to zero out call-used registers. |
| 1101 | if (FD && FD->isMain()) |
| 1102 | Fn->removeFnAttr(Kind: "zero-call-used-regs" ); |
| 1103 | |
| 1104 | // Add vscale_range attribute if appropriate. |
| 1105 | llvm::StringMap<bool> FeatureMap; |
| 1106 | auto IsArmStreaming = TargetInfo::ArmStreamingKind::NotStreaming; |
| 1107 | if (FD) { |
| 1108 | getContext().getFunctionFeatureMap(FeatureMap, FD); |
| 1109 | if (const auto *T = FD->getType()->getAs<FunctionProtoType>()) |
| 1110 | if (T->getAArch64SMEAttributes() & |
| 1111 | FunctionType::SME_PStateSMCompatibleMask) |
| 1112 | IsArmStreaming = TargetInfo::ArmStreamingKind::StreamingCompatible; |
| 1113 | |
| 1114 | if (IsArmStreamingFunction(FD, IncludeLocallyStreaming: true)) |
| 1115 | IsArmStreaming = TargetInfo::ArmStreamingKind::Streaming; |
| 1116 | } |
| 1117 | std::optional<std::pair<unsigned, unsigned>> VScaleRange = |
| 1118 | getContext().getTargetInfo().getVScaleRange(LangOpts: getLangOpts(), Mode: IsArmStreaming, |
| 1119 | FeatureMap: &FeatureMap); |
| 1120 | if (VScaleRange) { |
| 1121 | CurFn->addFnAttr(Attr: llvm::Attribute::getWithVScaleRangeArgs( |
| 1122 | Context&: getLLVMContext(), MinValue: VScaleRange->first, MaxValue: VScaleRange->second)); |
| 1123 | } |
| 1124 | |
| 1125 | llvm::BasicBlock *EntryBB = createBasicBlock(name: "entry" , parent: CurFn); |
| 1126 | |
| 1127 | // Create a marker to make it easy to insert allocas into the entryblock |
| 1128 | // later. Don't create this with the builder, because we don't want it |
| 1129 | // folded. |
| 1130 | llvm::Value *Poison = llvm::PoisonValue::get(T: Int32Ty); |
| 1131 | AllocaInsertPt = new llvm::BitCastInst(Poison, Int32Ty, "allocapt" , EntryBB); |
| 1132 | |
| 1133 | ReturnBlock = getJumpDestInCurrentScope(Name: "return" ); |
| 1134 | |
| 1135 | Builder.SetInsertPoint(EntryBB); |
| 1136 | |
| 1137 | // If we're checking the return value, allocate space for a pointer to a |
| 1138 | // precise source location of the checked return statement. |
| 1139 | if (requiresReturnValueCheck()) { |
| 1140 | ReturnLocation = CreateDefaultAlignTempAlloca(Ty: Int8PtrTy, Name: "return.sloc.ptr" ); |
| 1141 | Builder.CreateStore(Val: llvm::ConstantPointerNull::get(T: Int8PtrTy), |
| 1142 | Addr: ReturnLocation); |
| 1143 | } |
| 1144 | |
| 1145 | // Emit subprogram debug descriptor. |
| 1146 | if (CGDebugInfo *DI = getDebugInfo()) { |
| 1147 | // Reconstruct the type from the argument list so that implicit parameters, |
| 1148 | // such as 'this' and 'vtt', show up in the debug info. Preserve the calling |
| 1149 | // convention. |
| 1150 | DI->emitFunctionStart(GD, Loc, ScopeLoc: StartLoc, |
| 1151 | FnType: DI->getFunctionType(FD, RetTy, Args), Fn: CurFn, |
| 1152 | CurFnIsThunk: CurFuncIsThunk); |
| 1153 | } |
| 1154 | |
| 1155 | if (ShouldInstrumentFunction()) { |
| 1156 | if (CGM.getCodeGenOpts().InstrumentFunctions) |
| 1157 | CurFn->addFnAttr(Kind: "instrument-function-entry" , Val: "__cyg_profile_func_enter" ); |
| 1158 | if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining) |
| 1159 | CurFn->addFnAttr(Kind: "instrument-function-entry-inlined" , |
| 1160 | Val: "__cyg_profile_func_enter" ); |
| 1161 | if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare) |
| 1162 | CurFn->addFnAttr(Kind: "instrument-function-entry-inlined" , |
| 1163 | Val: "__cyg_profile_func_enter_bare" ); |
| 1164 | } |
| 1165 | |
| 1166 | // Since emitting the mcount call here impacts optimizations such as function |
| 1167 | // inlining, we just add an attribute to insert a mcount call in backend. |
| 1168 | // The attribute "counting-function" is set to mcount function name which is |
| 1169 | // architecture dependent. |
| 1170 | if (CGM.getCodeGenOpts().InstrumentForProfiling) { |
| 1171 | // Calls to fentry/mcount should not be generated if function has |
| 1172 | // the no_instrument_function attribute. |
| 1173 | if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) { |
| 1174 | if (CGM.getCodeGenOpts().CallFEntry) |
| 1175 | Fn->addFnAttr(Kind: "fentry-call" , Val: "true" ); |
| 1176 | else { |
| 1177 | Fn->addFnAttr(Kind: "instrument-function-entry-inlined" , |
| 1178 | Val: getTarget().getMCountName()); |
| 1179 | } |
| 1180 | if (CGM.getCodeGenOpts().MNopMCount) { |
| 1181 | if (!CGM.getCodeGenOpts().CallFEntry) |
| 1182 | CGM.getDiags().Report(DiagID: diag::err_opt_not_valid_without_opt) |
| 1183 | << "-mnop-mcount" << "-mfentry" ; |
| 1184 | Fn->addFnAttr(Kind: "mnop-mcount" ); |
| 1185 | } |
| 1186 | |
| 1187 | if (CGM.getCodeGenOpts().RecordMCount) { |
| 1188 | if (!CGM.getCodeGenOpts().CallFEntry) |
| 1189 | CGM.getDiags().Report(DiagID: diag::err_opt_not_valid_without_opt) |
| 1190 | << "-mrecord-mcount" << "-mfentry" ; |
| 1191 | Fn->addFnAttr(Kind: "mrecord-mcount" ); |
| 1192 | } |
| 1193 | } |
| 1194 | } |
| 1195 | |
| 1196 | if (CGM.getCodeGenOpts().PackedStack) { |
| 1197 | if (getContext().getTargetInfo().getTriple().getArch() != |
| 1198 | llvm::Triple::systemz) |
| 1199 | CGM.getDiags().Report(DiagID: diag::err_opt_not_valid_on_target) |
| 1200 | << "-mpacked-stack" ; |
| 1201 | Fn->addFnAttr(Kind: "packed-stack" ); |
| 1202 | } |
| 1203 | |
| 1204 | if (CGM.getCodeGenOpts().WarnStackSize != UINT_MAX && |
| 1205 | !CGM.getDiags().isIgnored(DiagID: diag::warn_fe_backend_frame_larger_than, Loc)) |
| 1206 | Fn->addFnAttr(Kind: "warn-stack-size" , |
| 1207 | Val: std::to_string(val: CGM.getCodeGenOpts().WarnStackSize)); |
| 1208 | |
| 1209 | if (RetTy->isVoidType()) { |
| 1210 | // Void type; nothing to return. |
| 1211 | ReturnValue = Address::invalid(); |
| 1212 | |
| 1213 | // Count the implicit return. |
| 1214 | if (!endsWithReturn(F: D)) |
| 1215 | ++NumReturnExprs; |
| 1216 | } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect) { |
| 1217 | // Indirect return; emit returned value directly into sret slot. |
| 1218 | // This reduces code size, and affects correctness in C++. |
| 1219 | auto AI = CurFn->arg_begin(); |
| 1220 | if (CurFnInfo->getReturnInfo().isSRetAfterThis()) |
| 1221 | ++AI; |
| 1222 | ReturnValue = makeNaturalAddressForPointer( |
| 1223 | Ptr: &*AI, T: RetTy, Alignment: CurFnInfo->getReturnInfo().getIndirectAlign(), ForPointeeType: false, |
| 1224 | BaseInfo: nullptr, TBAAInfo: nullptr, IsKnownNonNull: KnownNonNull); |
| 1225 | if (!CurFnInfo->getReturnInfo().getIndirectByVal()) { |
| 1226 | ReturnValuePointer = |
| 1227 | CreateDefaultAlignTempAlloca(Ty: ReturnValue.getType(), Name: "result.ptr" ); |
| 1228 | Builder.CreateStore(Val: ReturnValue.emitRawPointer(CGF&: *this), |
| 1229 | Addr: ReturnValuePointer); |
| 1230 | } |
| 1231 | } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca && |
| 1232 | !hasScalarEvaluationKind(T: CurFnInfo->getReturnType())) { |
| 1233 | // Load the sret pointer from the argument struct and return into that. |
| 1234 | unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex(); |
| 1235 | llvm::Function::arg_iterator EI = CurFn->arg_end(); |
| 1236 | --EI; |
| 1237 | llvm::Value *Addr = Builder.CreateStructGEP( |
| 1238 | Ty: CurFnInfo->getArgStruct(), Ptr: &*EI, Idx); |
| 1239 | llvm::Type *Ty = |
| 1240 | cast<llvm::GetElementPtrInst>(Val: Addr)->getResultElementType(); |
| 1241 | ReturnValuePointer = Address(Addr, Ty, getPointerAlign()); |
| 1242 | Addr = Builder.CreateAlignedLoad(Ty, Addr, Align: getPointerAlign(), Name: "agg.result" ); |
| 1243 | ReturnValue = Address(Addr, ConvertType(T: RetTy), |
| 1244 | CGM.getNaturalTypeAlignment(T: RetTy), KnownNonNull); |
| 1245 | } else { |
| 1246 | ReturnValue = CreateIRTempWithoutCast(T: RetTy, Name: "retval" ); |
| 1247 | |
| 1248 | // Tell the epilog emitter to autorelease the result. We do this |
| 1249 | // now so that various specialized functions can suppress it |
| 1250 | // during their IR-generation. |
| 1251 | if (getLangOpts().ObjCAutoRefCount && |
| 1252 | !CurFnInfo->isReturnsRetained() && |
| 1253 | RetTy->isObjCRetainableType()) |
| 1254 | AutoreleaseResult = true; |
| 1255 | } |
| 1256 | |
| 1257 | EmitStartEHSpec(D: CurCodeDecl); |
| 1258 | |
| 1259 | PrologueCleanupDepth = EHStack.stable_begin(); |
| 1260 | |
| 1261 | // Emit OpenMP specific initialization of the device functions. |
| 1262 | if (getLangOpts().OpenMP && CurCodeDecl) |
| 1263 | CGM.getOpenMPRuntime().emitFunctionProlog(CGF&: *this, D: CurCodeDecl); |
| 1264 | |
| 1265 | if (FD && getLangOpts().HLSL) { |
| 1266 | // Handle emitting HLSL entry functions. |
| 1267 | if (FD->hasAttr<HLSLShaderAttr>()) { |
| 1268 | CGM.getHLSLRuntime().emitEntryFunction(FD, Fn); |
| 1269 | } |
| 1270 | } |
| 1271 | |
| 1272 | EmitFunctionProlog(FI: *CurFnInfo, Fn: CurFn, Args); |
| 1273 | |
| 1274 | if (const CXXMethodDecl *MD = dyn_cast_if_present<CXXMethodDecl>(Val: D); |
| 1275 | MD && !MD->isStatic()) { |
| 1276 | bool IsInLambda = |
| 1277 | MD->getParent()->isLambda() && MD->getOverloadedOperator() == OO_Call; |
| 1278 | if (MD->isImplicitObjectMemberFunction()) |
| 1279 | CGM.getCXXABI().EmitInstanceFunctionProlog(CGF&: *this); |
| 1280 | if (IsInLambda) { |
| 1281 | // We're in a lambda; figure out the captures. |
| 1282 | MD->getParent()->getCaptureFields(Captures&: LambdaCaptureFields, |
| 1283 | ThisCapture&: LambdaThisCaptureField); |
| 1284 | if (LambdaThisCaptureField) { |
| 1285 | // If the lambda captures the object referred to by '*this' - either by |
| 1286 | // value or by reference, make sure CXXThisValue points to the correct |
| 1287 | // object. |
| 1288 | |
| 1289 | // Get the lvalue for the field (which is a copy of the enclosing object |
| 1290 | // or contains the address of the enclosing object). |
| 1291 | LValue ThisFieldLValue = EmitLValueForLambdaField(Field: LambdaThisCaptureField); |
| 1292 | if (!LambdaThisCaptureField->getType()->isPointerType()) { |
| 1293 | // If the enclosing object was captured by value, just use its |
| 1294 | // address. Sign this pointer. |
| 1295 | CXXThisValue = ThisFieldLValue.getPointer(CGF&: *this); |
| 1296 | } else { |
| 1297 | // Load the lvalue pointed to by the field, since '*this' was captured |
| 1298 | // by reference. |
| 1299 | CXXThisValue = |
| 1300 | EmitLoadOfLValue(V: ThisFieldLValue, Loc: SourceLocation()).getScalarVal(); |
| 1301 | } |
| 1302 | } |
| 1303 | for (auto *FD : MD->getParent()->fields()) { |
| 1304 | if (FD->hasCapturedVLAType()) { |
| 1305 | auto *ExprArg = EmitLoadOfLValue(V: EmitLValueForLambdaField(Field: FD), |
| 1306 | Loc: SourceLocation()).getScalarVal(); |
| 1307 | auto VAT = FD->getCapturedVLAType(); |
| 1308 | VLASizeMap[VAT->getSizeExpr()] = ExprArg; |
| 1309 | } |
| 1310 | } |
| 1311 | } else if (MD->isImplicitObjectMemberFunction()) { |
| 1312 | // Not in a lambda; just use 'this' from the method. |
| 1313 | // FIXME: Should we generate a new load for each use of 'this'? The |
| 1314 | // fast register allocator would be happier... |
| 1315 | CXXThisValue = CXXABIThisValue; |
| 1316 | } |
| 1317 | |
| 1318 | // Check the 'this' pointer once per function, if it's available. |
| 1319 | if (CXXABIThisValue) { |
| 1320 | SanitizerSet SkippedChecks; |
| 1321 | SkippedChecks.set(K: SanitizerKind::ObjectSize, Value: true); |
| 1322 | QualType ThisTy = MD->getThisType(); |
| 1323 | |
| 1324 | // If this is the call operator of a lambda with no captures, it |
| 1325 | // may have a static invoker function, which may call this operator with |
| 1326 | // a null 'this' pointer. |
| 1327 | if (isLambdaCallOperator(MD) && MD->getParent()->isCapturelessLambda()) |
| 1328 | SkippedChecks.set(K: SanitizerKind::Null, Value: true); |
| 1329 | |
| 1330 | EmitTypeCheck( |
| 1331 | TCK: isa<CXXConstructorDecl>(Val: MD) ? TCK_ConstructorCall : TCK_MemberCall, |
| 1332 | Loc, V: CXXABIThisValue, Type: ThisTy, Alignment: CXXABIThisAlignment, SkippedChecks); |
| 1333 | } |
| 1334 | } |
| 1335 | |
| 1336 | // If any of the arguments have a variably modified type, make sure to |
| 1337 | // emit the type size, but only if the function is not naked. Naked functions |
| 1338 | // have no prolog to run this evaluation. |
| 1339 | if (!FD || !FD->hasAttr<NakedAttr>()) { |
| 1340 | for (const VarDecl *VD : Args) { |
| 1341 | // Dig out the type as written from ParmVarDecls; it's unclear whether |
| 1342 | // the standard (C99 6.9.1p10) requires this, but we're following the |
| 1343 | // precedent set by gcc. |
| 1344 | QualType Ty; |
| 1345 | if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Val: VD)) |
| 1346 | Ty = PVD->getOriginalType(); |
| 1347 | else |
| 1348 | Ty = VD->getType(); |
| 1349 | |
| 1350 | if (Ty->isVariablyModifiedType()) |
| 1351 | EmitVariablyModifiedType(Ty); |
| 1352 | } |
| 1353 | } |
| 1354 | // Emit a location at the end of the prologue. |
| 1355 | if (CGDebugInfo *DI = getDebugInfo()) |
| 1356 | DI->EmitLocation(Builder, Loc: StartLoc); |
| 1357 | // TODO: Do we need to handle this in two places like we do with |
| 1358 | // target-features/target-cpu? |
| 1359 | if (CurFuncDecl) |
| 1360 | if (const auto *VecWidth = CurFuncDecl->getAttr<MinVectorWidthAttr>()) |
| 1361 | LargestVectorWidth = VecWidth->getVectorWidth(); |
| 1362 | |
| 1363 | if (CGM.shouldEmitConvergenceTokens()) |
| 1364 | ConvergenceTokenStack.push_back(Elt: getOrEmitConvergenceEntryToken(F: CurFn)); |
| 1365 | } |
| 1366 | |
| 1367 | void CodeGenFunction::EmitFunctionBody(const Stmt *Body) { |
| 1368 | incrementProfileCounter(S: Body); |
| 1369 | maybeCreateMCDCCondBitmap(); |
| 1370 | if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Val: Body)) |
| 1371 | EmitCompoundStmtWithoutScope(S: *S); |
| 1372 | else |
| 1373 | EmitStmt(S: Body); |
| 1374 | } |
| 1375 | |
| 1376 | /// When instrumenting to collect profile data, the counts for some blocks |
| 1377 | /// such as switch cases need to not include the fall-through counts, so |
| 1378 | /// emit a branch around the instrumentation code. When not instrumenting, |
| 1379 | /// this just calls EmitBlock(). |
| 1380 | void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB, |
| 1381 | const Stmt *S) { |
| 1382 | llvm::BasicBlock *SkipCountBB = nullptr; |
| 1383 | if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) { |
| 1384 | // When instrumenting for profiling, the fallthrough to certain |
| 1385 | // statements needs to skip over the instrumentation code so that we |
| 1386 | // get an accurate count. |
| 1387 | SkipCountBB = createBasicBlock(name: "skipcount" ); |
| 1388 | EmitBranch(Block: SkipCountBB); |
| 1389 | } |
| 1390 | EmitBlock(BB); |
| 1391 | uint64_t CurrentCount = getCurrentProfileCount(); |
| 1392 | incrementProfileCounter(ExecSkip: UseExecPath, S); |
| 1393 | setCurrentProfileCount(getCurrentProfileCount() + CurrentCount); |
| 1394 | if (SkipCountBB) |
| 1395 | EmitBlock(BB: SkipCountBB); |
| 1396 | } |
| 1397 | |
| 1398 | /// Tries to mark the given function nounwind based on the |
| 1399 | /// non-existence of any throwing calls within it. We believe this is |
| 1400 | /// lightweight enough to do at -O0. |
| 1401 | static void TryMarkNoThrow(llvm::Function *F) { |
| 1402 | // LLVM treats 'nounwind' on a function as part of the type, so we |
| 1403 | // can't do this on functions that can be overwritten. |
| 1404 | if (F->isInterposable()) return; |
| 1405 | |
| 1406 | for (llvm::BasicBlock &BB : *F) |
| 1407 | for (llvm::Instruction &I : BB) |
| 1408 | if (I.mayThrow()) |
| 1409 | return; |
| 1410 | |
| 1411 | F->setDoesNotThrow(); |
| 1412 | } |
| 1413 | |
| 1414 | QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD, |
| 1415 | FunctionArgList &Args) { |
| 1416 | const FunctionDecl *FD = cast<FunctionDecl>(Val: GD.getDecl()); |
| 1417 | QualType ResTy = FD->getReturnType(); |
| 1418 | |
| 1419 | const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Val: FD); |
| 1420 | if (MD && MD->isImplicitObjectMemberFunction()) { |
| 1421 | if (CGM.getCXXABI().HasThisReturn(GD)) |
| 1422 | ResTy = MD->getThisType(); |
| 1423 | else if (CGM.getCXXABI().hasMostDerivedReturn(GD)) |
| 1424 | ResTy = CGM.getContext().VoidPtrTy; |
| 1425 | CGM.getCXXABI().buildThisParam(CGF&: *this, Params&: Args); |
| 1426 | } |
| 1427 | |
| 1428 | // The base version of an inheriting constructor whose constructed base is a |
| 1429 | // virtual base is not passed any arguments (because it doesn't actually call |
| 1430 | // the inherited constructor). |
| 1431 | bool PassedParams = true; |
| 1432 | if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(Val: FD)) |
| 1433 | if (auto Inherited = CD->getInheritedConstructor()) |
| 1434 | PassedParams = |
| 1435 | getTypes().inheritingCtorHasParams(Inherited, Type: GD.getCtorType()); |
| 1436 | |
| 1437 | if (PassedParams) { |
| 1438 | for (auto *Param : FD->parameters()) { |
| 1439 | Args.push_back(Elt: Param); |
| 1440 | if (!Param->hasAttr<PassObjectSizeAttr>()) |
| 1441 | continue; |
| 1442 | |
| 1443 | auto *Implicit = ImplicitParamDecl::Create( |
| 1444 | C&: getContext(), DC: Param->getDeclContext(), IdLoc: Param->getLocation(), |
| 1445 | /*Id=*/nullptr, T: getContext().getSizeType(), ParamKind: ImplicitParamKind::Other); |
| 1446 | SizeArguments[Param] = Implicit; |
| 1447 | Args.push_back(Elt: Implicit); |
| 1448 | } |
| 1449 | } |
| 1450 | |
| 1451 | if (MD && (isa<CXXConstructorDecl>(Val: MD) || isa<CXXDestructorDecl>(Val: MD))) |
| 1452 | CGM.getCXXABI().addImplicitStructorParams(CGF&: *this, ResTy, Params&: Args); |
| 1453 | |
| 1454 | return ResTy; |
| 1455 | } |
| 1456 | |
| 1457 | void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn, |
| 1458 | const CGFunctionInfo &FnInfo) { |
| 1459 | assert(Fn && "generating code for null Function" ); |
| 1460 | const FunctionDecl *FD = cast<FunctionDecl>(Val: GD.getDecl()); |
| 1461 | CurGD = GD; |
| 1462 | |
| 1463 | FunctionArgList Args; |
| 1464 | QualType ResTy = BuildFunctionArgList(GD, Args); |
| 1465 | |
| 1466 | CGM.getTargetCodeGenInfo().checkFunctionABI(CGM, Decl: FD); |
| 1467 | |
| 1468 | if (FD->isInlineBuiltinDeclaration()) { |
| 1469 | // When generating code for a builtin with an inline declaration, use a |
| 1470 | // mangled name to hold the actual body, while keeping an external |
| 1471 | // definition in case the function pointer is referenced somewhere. |
| 1472 | std::string FDInlineName = (Fn->getName() + ".inline" ).str(); |
| 1473 | llvm::Module *M = Fn->getParent(); |
| 1474 | llvm::Function *Clone = M->getFunction(Name: FDInlineName); |
| 1475 | if (!Clone) { |
| 1476 | Clone = llvm::Function::Create(Ty: Fn->getFunctionType(), |
| 1477 | Linkage: llvm::GlobalValue::InternalLinkage, |
| 1478 | AddrSpace: Fn->getAddressSpace(), N: FDInlineName, M); |
| 1479 | Clone->addFnAttr(Kind: llvm::Attribute::AlwaysInline); |
| 1480 | } |
| 1481 | Fn->setLinkage(llvm::GlobalValue::ExternalLinkage); |
| 1482 | Fn = Clone; |
| 1483 | } else { |
| 1484 | // Detect the unusual situation where an inline version is shadowed by a |
| 1485 | // non-inline version. In that case we should pick the external one |
| 1486 | // everywhere. That's GCC behavior too. Unfortunately, I cannot find a way |
| 1487 | // to detect that situation before we reach codegen, so do some late |
| 1488 | // replacement. |
| 1489 | for (const FunctionDecl *PD = FD->getPreviousDecl(); PD; |
| 1490 | PD = PD->getPreviousDecl()) { |
| 1491 | if (LLVM_UNLIKELY(PD->isInlineBuiltinDeclaration())) { |
| 1492 | std::string FDInlineName = (Fn->getName() + ".inline" ).str(); |
| 1493 | llvm::Module *M = Fn->getParent(); |
| 1494 | if (llvm::Function *Clone = M->getFunction(Name: FDInlineName)) { |
| 1495 | Clone->replaceAllUsesWith(V: Fn); |
| 1496 | Clone->eraseFromParent(); |
| 1497 | } |
| 1498 | break; |
| 1499 | } |
| 1500 | } |
| 1501 | } |
| 1502 | |
| 1503 | // Check if we should generate debug info for this function. |
| 1504 | if (FD->hasAttr<NoDebugAttr>()) { |
| 1505 | // Clear non-distinct debug info that was possibly attached to the function |
| 1506 | // due to an earlier declaration without the nodebug attribute |
| 1507 | Fn->setSubprogram(nullptr); |
| 1508 | // Disable debug info indefinitely for this function |
| 1509 | DebugInfo = nullptr; |
| 1510 | } |
| 1511 | // Finalize function debug info on exit. |
| 1512 | llvm::scope_exit Cleanup([this] { |
| 1513 | if (CGDebugInfo *DI = getDebugInfo()) |
| 1514 | DI->completeFunction(); |
| 1515 | }); |
| 1516 | |
| 1517 | // The function might not have a body if we're generating thunks for a |
| 1518 | // function declaration. |
| 1519 | SourceRange BodyRange; |
| 1520 | if (Stmt *Body = FD->getBody()) |
| 1521 | BodyRange = Body->getSourceRange(); |
| 1522 | else |
| 1523 | BodyRange = FD->getLocation(); |
| 1524 | CurEHLocation = BodyRange.getEnd(); |
| 1525 | |
| 1526 | // Use the location of the start of the function to determine where |
| 1527 | // the function definition is located. By default use the location |
| 1528 | // of the declaration as the location for the subprogram. A function |
| 1529 | // may lack a declaration in the source code if it is created by code |
| 1530 | // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk). |
| 1531 | SourceLocation Loc = FD->getLocation(); |
| 1532 | |
| 1533 | // If this is a function specialization then use the pattern body |
| 1534 | // as the location for the function. |
| 1535 | if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern()) |
| 1536 | if (SpecDecl->hasBody(Definition&: SpecDecl)) |
| 1537 | Loc = SpecDecl->getLocation(); |
| 1538 | |
| 1539 | Stmt *Body = FD->getBody(); |
| 1540 | |
| 1541 | if (Body) { |
| 1542 | // Coroutines always emit lifetime markers. |
| 1543 | if (isa<CoroutineBodyStmt>(Val: Body)) |
| 1544 | ShouldEmitLifetimeMarkers = true; |
| 1545 | |
| 1546 | // Initialize helper which will detect jumps which can cause invalid |
| 1547 | // lifetime markers. |
| 1548 | if (ShouldEmitLifetimeMarkers) |
| 1549 | Bypasses.Init(CGM, Body); |
| 1550 | } |
| 1551 | |
| 1552 | // Emit the standard function prologue. |
| 1553 | StartFunction(GD, RetTy: ResTy, Fn, FnInfo, Args, Loc, StartLoc: BodyRange.getBegin()); |
| 1554 | |
| 1555 | // Save parameters for coroutine function. |
| 1556 | if (Body && isa_and_nonnull<CoroutineBodyStmt>(Val: Body)) |
| 1557 | llvm::append_range(C&: FnArgs, R: FD->parameters()); |
| 1558 | |
| 1559 | // Ensure that the function adheres to the forward progress guarantee, which |
| 1560 | // is required by certain optimizations. |
| 1561 | // In C++11 and up, the attribute will be removed if the body contains a |
| 1562 | // trivial empty loop. |
| 1563 | if (checkIfFunctionMustProgress()) |
| 1564 | CurFn->addFnAttr(Kind: llvm::Attribute::MustProgress); |
| 1565 | |
| 1566 | // Generate the body of the function. |
| 1567 | PGO->assignRegionCounters(GD, Fn: CurFn); |
| 1568 | if (isa<CXXDestructorDecl>(Val: FD)) |
| 1569 | EmitDestructorBody(Args); |
| 1570 | else if (isa<CXXConstructorDecl>(Val: FD)) |
| 1571 | EmitConstructorBody(Args); |
| 1572 | else if (getLangOpts().CUDA && |
| 1573 | !getLangOpts().CUDAIsDevice && |
| 1574 | FD->hasAttr<CUDAGlobalAttr>()) |
| 1575 | CGM.getCUDARuntime().emitDeviceStub(CGF&: *this, Args); |
| 1576 | else if (isa<CXXMethodDecl>(Val: FD) && |
| 1577 | cast<CXXMethodDecl>(Val: FD)->isLambdaStaticInvoker()) { |
| 1578 | // The lambda static invoker function is special, because it forwards or |
| 1579 | // clones the body of the function call operator (but is actually static). |
| 1580 | EmitLambdaStaticInvokeBody(MD: cast<CXXMethodDecl>(Val: FD)); |
| 1581 | } else if (isa<CXXMethodDecl>(Val: FD) && |
| 1582 | isLambdaCallOperator(MD: cast<CXXMethodDecl>(Val: FD)) && |
| 1583 | !FnInfo.isDelegateCall() && |
| 1584 | cast<CXXMethodDecl>(Val: FD)->getParent()->getLambdaStaticInvoker() && |
| 1585 | hasInAllocaArg(MD: cast<CXXMethodDecl>(Val: FD))) { |
| 1586 | // If emitting a lambda with static invoker on X86 Windows, change |
| 1587 | // the call operator body. |
| 1588 | // Make sure that this is a call operator with an inalloca arg and check |
| 1589 | // for delegate call to make sure this is the original call op and not the |
| 1590 | // new forwarding function for the static invoker. |
| 1591 | EmitLambdaInAllocaCallOpBody(MD: cast<CXXMethodDecl>(Val: FD)); |
| 1592 | } else if (FD->isDefaulted() && isa<CXXMethodDecl>(Val: FD) && |
| 1593 | (cast<CXXMethodDecl>(Val: FD)->isCopyAssignmentOperator() || |
| 1594 | cast<CXXMethodDecl>(Val: FD)->isMoveAssignmentOperator())) { |
| 1595 | // Implicit copy-assignment gets the same special treatment as implicit |
| 1596 | // copy-constructors. |
| 1597 | emitImplicitAssignmentOperatorBody(Args); |
| 1598 | } else if (DeviceKernelAttr::isOpenCLSpelling( |
| 1599 | A: FD->getAttr<DeviceKernelAttr>()) && |
| 1600 | GD.getKernelReferenceKind() == KernelReferenceKind::Kernel) { |
| 1601 | CallArgList CallArgs; |
| 1602 | for (unsigned i = 0; i < Args.size(); ++i) { |
| 1603 | Address ArgAddr = GetAddrOfLocalVar(VD: Args[i]); |
| 1604 | QualType ArgQualType = Args[i]->getType(); |
| 1605 | RValue ArgRValue = convertTempToRValue(addr: ArgAddr, type: ArgQualType, Loc); |
| 1606 | CallArgs.add(rvalue: ArgRValue, type: ArgQualType); |
| 1607 | } |
| 1608 | GlobalDecl GDStub = GlobalDecl(FD, KernelReferenceKind::Stub); |
| 1609 | const FunctionType *FT = cast<FunctionType>(Val: FD->getType()); |
| 1610 | CGM.getTargetCodeGenInfo().setOCLKernelStubCallingConvention(FT); |
| 1611 | const CGFunctionInfo &FnInfo = CGM.getTypes().arrangeFreeFunctionCall( |
| 1612 | Args: CallArgs, Ty: FT, /*ChainCall=*/false); |
| 1613 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(Info: FnInfo); |
| 1614 | llvm::Constant *GDStubFunctionPointer = |
| 1615 | CGM.getRawFunctionPointer(GD: GDStub, Ty: FTy); |
| 1616 | CGCallee GDStubCallee = CGCallee::forDirect(functionPtr: GDStubFunctionPointer, abstractInfo: GDStub); |
| 1617 | EmitCall(CallInfo: FnInfo, Callee: GDStubCallee, ReturnValue: ReturnValueSlot(), Args: CallArgs, CallOrInvoke: nullptr, IsMustTail: false, |
| 1618 | Loc); |
| 1619 | } else if (Body) { |
| 1620 | EmitFunctionBody(Body); |
| 1621 | } else |
| 1622 | llvm_unreachable("no definition for emitted function" ); |
| 1623 | |
| 1624 | // C++11 [stmt.return]p2: |
| 1625 | // Flowing off the end of a function [...] results in undefined behavior in |
| 1626 | // a value-returning function. |
| 1627 | // C11 6.9.1p12: |
| 1628 | // If the '}' that terminates a function is reached, and the value of the |
| 1629 | // function call is used by the caller, the behavior is undefined. |
| 1630 | if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock && |
| 1631 | !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) { |
| 1632 | bool ShouldEmitUnreachable = |
| 1633 | CGM.getCodeGenOpts().StrictReturn || |
| 1634 | !CGM.MayDropFunctionReturn(Context: FD->getASTContext(), ReturnType: FD->getReturnType()); |
| 1635 | if (SanOpts.has(K: SanitizerKind::Return)) { |
| 1636 | auto CheckOrdinal = SanitizerKind::SO_Return; |
| 1637 | auto CheckHandler = SanitizerHandler::MissingReturn; |
| 1638 | SanitizerDebugLocation SanScope(this, {CheckOrdinal}, CheckHandler); |
| 1639 | llvm::Value *IsFalse = Builder.getFalse(); |
| 1640 | EmitCheck(Checked: std::make_pair(x&: IsFalse, y&: CheckOrdinal), Check: CheckHandler, |
| 1641 | StaticArgs: EmitCheckSourceLocation(Loc: FD->getLocation()), DynamicArgs: {}); |
| 1642 | } else if (ShouldEmitUnreachable) { |
| 1643 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
| 1644 | EmitTrapCall(IntrID: llvm::Intrinsic::trap); |
| 1645 | } |
| 1646 | if (SanOpts.has(K: SanitizerKind::Return) || ShouldEmitUnreachable) { |
| 1647 | Builder.CreateUnreachable(); |
| 1648 | Builder.ClearInsertionPoint(); |
| 1649 | } |
| 1650 | } |
| 1651 | |
| 1652 | // Emit the standard function epilogue. |
| 1653 | FinishFunction(EndLoc: BodyRange.getEnd()); |
| 1654 | |
| 1655 | PGO->verifyCounterMap(); |
| 1656 | |
| 1657 | if (CurCodeDecl->hasAttr<PersonalityAttr>()) { |
| 1658 | StringRef Identifier = |
| 1659 | CurCodeDecl->getAttr<PersonalityAttr>()->getRoutine()->getName(); |
| 1660 | llvm::FunctionCallee PersonalityRoutine = |
| 1661 | CGM.CreateRuntimeFunction(Ty: llvm::FunctionType::get(Result: CGM.Int32Ty, isVarArg: true), |
| 1662 | Name: Identifier, ExtraAttrs: {}, /*local=*/Local: true); |
| 1663 | Fn->setPersonalityFn(cast<llvm::Constant>(Val: PersonalityRoutine.getCallee())); |
| 1664 | } |
| 1665 | |
| 1666 | // If we haven't marked the function nothrow through other means, do |
| 1667 | // a quick pass now to see if we can. |
| 1668 | if (!CurFn->doesNotThrow()) |
| 1669 | TryMarkNoThrow(F: CurFn); |
| 1670 | } |
| 1671 | |
| 1672 | /// ContainsLabel - Return true if the statement contains a label in it. If |
| 1673 | /// this statement is not executed normally, it not containing a label means |
| 1674 | /// that we can just remove the code. |
| 1675 | bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) { |
| 1676 | // Null statement, not a label! |
| 1677 | if (!S) return false; |
| 1678 | |
| 1679 | // If this is a label, we have to emit the code, consider something like: |
| 1680 | // if (0) { ... foo: bar(); } goto foo; |
| 1681 | // |
| 1682 | // TODO: If anyone cared, we could track __label__'s, since we know that you |
| 1683 | // can't jump to one from outside their declared region. |
| 1684 | if (isa<LabelStmt>(Val: S)) |
| 1685 | return true; |
| 1686 | |
| 1687 | // If this is a case/default statement, and we haven't seen a switch, we have |
| 1688 | // to emit the code. |
| 1689 | if (isa<SwitchCase>(Val: S) && !IgnoreCaseStmts) |
| 1690 | return true; |
| 1691 | |
| 1692 | // If this is a switch statement, we want to ignore cases below it. |
| 1693 | if (isa<SwitchStmt>(Val: S)) |
| 1694 | IgnoreCaseStmts = true; |
| 1695 | |
| 1696 | // Scan subexpressions for verboten labels. |
| 1697 | for (const Stmt *SubStmt : S->children()) |
| 1698 | if (ContainsLabel(S: SubStmt, IgnoreCaseStmts)) |
| 1699 | return true; |
| 1700 | |
| 1701 | return false; |
| 1702 | } |
| 1703 | |
| 1704 | /// containsBreak - Return true if the statement contains a break out of it. |
| 1705 | /// If the statement (recursively) contains a switch or loop with a break |
| 1706 | /// inside of it, this is fine. |
| 1707 | bool CodeGenFunction::containsBreak(const Stmt *S) { |
| 1708 | // Null statement, not a label! |
| 1709 | if (!S) return false; |
| 1710 | |
| 1711 | // If this is a switch or loop that defines its own break scope, then we can |
| 1712 | // include it and anything inside of it. |
| 1713 | if (isa<SwitchStmt>(Val: S) || isa<WhileStmt>(Val: S) || isa<DoStmt>(Val: S) || |
| 1714 | isa<ForStmt>(Val: S)) |
| 1715 | return false; |
| 1716 | |
| 1717 | if (isa<BreakStmt>(Val: S)) |
| 1718 | return true; |
| 1719 | |
| 1720 | // Scan subexpressions for verboten breaks. |
| 1721 | for (const Stmt *SubStmt : S->children()) |
| 1722 | if (containsBreak(S: SubStmt)) |
| 1723 | return true; |
| 1724 | |
| 1725 | return false; |
| 1726 | } |
| 1727 | |
| 1728 | bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) { |
| 1729 | if (!S) return false; |
| 1730 | |
| 1731 | // Some statement kinds add a scope and thus never add a decl to the current |
| 1732 | // scope. Note, this list is longer than the list of statements that might |
| 1733 | // have an unscoped decl nested within them, but this way is conservatively |
| 1734 | // correct even if more statement kinds are added. |
| 1735 | if (isa<IfStmt>(Val: S) || isa<SwitchStmt>(Val: S) || isa<WhileStmt>(Val: S) || |
| 1736 | isa<DoStmt>(Val: S) || isa<ForStmt>(Val: S) || isa<CompoundStmt>(Val: S) || |
| 1737 | isa<CXXForRangeStmt>(Val: S) || isa<CXXTryStmt>(Val: S) || |
| 1738 | isa<ObjCForCollectionStmt>(Val: S) || isa<ObjCAtTryStmt>(Val: S)) |
| 1739 | return false; |
| 1740 | |
| 1741 | if (isa<DeclStmt>(Val: S)) |
| 1742 | return true; |
| 1743 | |
| 1744 | for (const Stmt *SubStmt : S->children()) |
| 1745 | if (mightAddDeclToScope(S: SubStmt)) |
| 1746 | return true; |
| 1747 | |
| 1748 | return false; |
| 1749 | } |
| 1750 | |
| 1751 | /// ConstantFoldsToSimpleInteger - If the specified expression does not fold |
| 1752 | /// to a constant, or if it does but contains a label, return false. If it |
| 1753 | /// constant folds return true and set the boolean result in Result. |
| 1754 | bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond, |
| 1755 | bool &ResultBool, |
| 1756 | bool AllowLabels) { |
| 1757 | // If MC/DC is enabled, disable folding so that we can instrument all |
| 1758 | // conditions to yield complete test vectors. We still keep track of |
| 1759 | // folded conditions during region mapping and visualization. |
| 1760 | if (!AllowLabels && CGM.getCodeGenOpts().hasProfileClangInstr() && |
| 1761 | CGM.getCodeGenOpts().MCDCCoverage) |
| 1762 | return false; |
| 1763 | |
| 1764 | llvm::APSInt ResultInt; |
| 1765 | if (!ConstantFoldsToSimpleInteger(Cond, Result&: ResultInt, AllowLabels)) |
| 1766 | return false; |
| 1767 | |
| 1768 | ResultBool = ResultInt.getBoolValue(); |
| 1769 | return true; |
| 1770 | } |
| 1771 | |
| 1772 | /// ConstantFoldsToSimpleInteger - If the specified expression does not fold |
| 1773 | /// to a constant, or if it does but contains a label, return false. If it |
| 1774 | /// constant folds return true and set the folded value. |
| 1775 | bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond, |
| 1776 | llvm::APSInt &ResultInt, |
| 1777 | bool AllowLabels) { |
| 1778 | // FIXME: Rename and handle conversion of other evaluatable things |
| 1779 | // to bool. |
| 1780 | Expr::EvalResult Result; |
| 1781 | if (!Cond->EvaluateAsInt(Result, Ctx: getContext())) |
| 1782 | return false; // Not foldable, not integer or not fully evaluatable. |
| 1783 | |
| 1784 | llvm::APSInt Int = Result.Val.getInt(); |
| 1785 | if (!AllowLabels && CodeGenFunction::ContainsLabel(S: Cond)) |
| 1786 | return false; // Contains a label. |
| 1787 | |
| 1788 | PGO->markStmtMaybeUsed(S: Cond); |
| 1789 | ResultInt = std::move(Int); |
| 1790 | return true; |
| 1791 | } |
| 1792 | |
| 1793 | /// Strip parentheses and simplistic logical-NOT operators. |
| 1794 | const Expr *CodeGenFunction::stripCond(const Expr *C) { |
| 1795 | while (true) { |
| 1796 | const Expr *SC = IgnoreExprNodes( |
| 1797 | E: C, Fns&: IgnoreParensSingleStep, Fns&: IgnoreUOpLNotSingleStep, |
| 1798 | Fns&: IgnoreBuiltinExpectSingleStep, Fns&: IgnoreImplicitCastsSingleStep); |
| 1799 | if (C == SC) |
| 1800 | return SC; |
| 1801 | C = SC; |
| 1802 | } |
| 1803 | } |
| 1804 | |
| 1805 | /// Determine whether the given condition is an instrumentable condition |
| 1806 | /// (i.e. no "&&" or "||"). |
| 1807 | bool CodeGenFunction::isInstrumentedCondition(const Expr *C) { |
| 1808 | const BinaryOperator *BOp = dyn_cast<BinaryOperator>(Val: stripCond(C)); |
| 1809 | return (!BOp || !BOp->isLogicalOp()); |
| 1810 | } |
| 1811 | |
| 1812 | /// EmitBranchToCounterBlock - Emit a conditional branch to a new block that |
| 1813 | /// increments a profile counter based on the semantics of the given logical |
| 1814 | /// operator opcode. This is used to instrument branch condition coverage for |
| 1815 | /// logical operators. |
| 1816 | void CodeGenFunction::EmitBranchToCounterBlock( |
| 1817 | const Expr *Cond, BinaryOperator::Opcode LOp, llvm::BasicBlock *TrueBlock, |
| 1818 | llvm::BasicBlock *FalseBlock, uint64_t TrueCount /* = 0 */, |
| 1819 | Stmt::Likelihood LH /* =None */, const Expr *CntrIdx /* = nullptr */) { |
| 1820 | // If not instrumenting, just emit a branch. |
| 1821 | bool InstrumentRegions = CGM.getCodeGenOpts().hasProfileClangInstr(); |
| 1822 | if (!InstrumentRegions || !isInstrumentedCondition(C: Cond)) |
| 1823 | return EmitBranchOnBoolExpr(Cond, TrueBlock, FalseBlock, TrueCount, LH); |
| 1824 | |
| 1825 | const Stmt *CntrStmt = (CntrIdx ? CntrIdx : Cond); |
| 1826 | |
| 1827 | llvm::BasicBlock *ThenBlock = nullptr; |
| 1828 | llvm::BasicBlock *ElseBlock = nullptr; |
| 1829 | llvm::BasicBlock *NextBlock = nullptr; |
| 1830 | |
| 1831 | // Create the block we'll use to increment the appropriate counter. |
| 1832 | llvm::BasicBlock *CounterIncrBlock = createBasicBlock(name: "lop.rhscnt" ); |
| 1833 | |
| 1834 | llvm::BasicBlock *SkipIncrBlock = |
| 1835 | (hasSkipCounter(S: CntrStmt) ? createBasicBlock(name: "lop.rhsskip" ) : nullptr); |
| 1836 | llvm::BasicBlock *SkipNextBlock = nullptr; |
| 1837 | |
| 1838 | // Set block pointers according to Logical-AND (BO_LAnd) semantics. This |
| 1839 | // means we need to evaluate the condition and increment the counter on TRUE: |
| 1840 | // |
| 1841 | // if (Cond) |
| 1842 | // goto CounterIncrBlock; |
| 1843 | // else |
| 1844 | // goto FalseBlock; |
| 1845 | // |
| 1846 | // CounterIncrBlock: |
| 1847 | // Counter++; |
| 1848 | // goto TrueBlock; |
| 1849 | |
| 1850 | if (LOp == BO_LAnd) { |
| 1851 | SkipNextBlock = FalseBlock; |
| 1852 | ThenBlock = CounterIncrBlock; |
| 1853 | ElseBlock = (SkipIncrBlock ? SkipIncrBlock : SkipNextBlock); |
| 1854 | NextBlock = TrueBlock; |
| 1855 | } |
| 1856 | |
| 1857 | // Set block pointers according to Logical-OR (BO_LOr) semantics. This means |
| 1858 | // we need to evaluate the condition and increment the counter on FALSE: |
| 1859 | // |
| 1860 | // if (Cond) |
| 1861 | // goto TrueBlock; |
| 1862 | // else |
| 1863 | // goto CounterIncrBlock; |
| 1864 | // |
| 1865 | // CounterIncrBlock: |
| 1866 | // Counter++; |
| 1867 | // goto FalseBlock; |
| 1868 | |
| 1869 | else if (LOp == BO_LOr) { |
| 1870 | SkipNextBlock = TrueBlock; |
| 1871 | ThenBlock = (SkipIncrBlock ? SkipIncrBlock : SkipNextBlock); |
| 1872 | ElseBlock = CounterIncrBlock; |
| 1873 | NextBlock = FalseBlock; |
| 1874 | } else { |
| 1875 | llvm_unreachable("Expected Opcode must be that of a Logical Operator" ); |
| 1876 | } |
| 1877 | |
| 1878 | // Emit Branch based on condition. |
| 1879 | EmitBranchOnBoolExpr(Cond, TrueBlock: ThenBlock, FalseBlock: ElseBlock, TrueCount, LH); |
| 1880 | |
| 1881 | if (SkipIncrBlock) { |
| 1882 | EmitBlock(BB: SkipIncrBlock); |
| 1883 | incrementProfileCounter(ExecSkip: UseSkipPath, S: CntrStmt); |
| 1884 | EmitBranch(Block: SkipNextBlock); |
| 1885 | } |
| 1886 | |
| 1887 | // Emit the block containing the counter increment(s). |
| 1888 | EmitBlock(BB: CounterIncrBlock); |
| 1889 | |
| 1890 | // Increment corresponding counter; if index not provided, use Cond as index. |
| 1891 | incrementProfileCounter(ExecSkip: UseExecPath, S: CntrStmt); |
| 1892 | |
| 1893 | // Go to the next block. |
| 1894 | EmitBranch(Block: NextBlock); |
| 1895 | } |
| 1896 | |
| 1897 | /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if |
| 1898 | /// statement) to the specified blocks. Based on the condition, this might try |
| 1899 | /// to simplify the codegen of the conditional based on the branch. |
| 1900 | /// \param LH The value of the likelihood attribute on the True branch. |
| 1901 | /// \param ConditionalOp Used by MC/DC code coverage to track the result of the |
| 1902 | /// ConditionalOperator (ternary) through a recursive call for the operator's |
| 1903 | /// LHS and RHS nodes. |
| 1904 | void CodeGenFunction::EmitBranchOnBoolExpr( |
| 1905 | const Expr *Cond, llvm::BasicBlock *TrueBlock, llvm::BasicBlock *FalseBlock, |
| 1906 | uint64_t TrueCount, Stmt::Likelihood LH, const Expr *ConditionalOp, |
| 1907 | const VarDecl *ConditionalDecl) { |
| 1908 | Cond = Cond->IgnoreParens(); |
| 1909 | |
| 1910 | if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Val: Cond)) { |
| 1911 | bool HasSkip = hasSkipCounter(S: CondBOp); |
| 1912 | |
| 1913 | // Handle X && Y in a condition. |
| 1914 | if (CondBOp->getOpcode() == BO_LAnd) { |
| 1915 | // If we have "1 && X", simplify the code. "0 && X" would have constant |
| 1916 | // folded if the case was simple enough. |
| 1917 | bool ConstantBool = false; |
| 1918 | if (ConstantFoldsToSimpleInteger(Cond: CondBOp->getLHS(), ResultBool&: ConstantBool) && |
| 1919 | ConstantBool) { |
| 1920 | // br(1 && X) -> br(X). |
| 1921 | incrementProfileCounter(S: CondBOp); |
| 1922 | EmitBranchToCounterBlock(Cond: CondBOp->getRHS(), LOp: BO_LAnd, TrueBlock, |
| 1923 | FalseBlock, TrueCount, LH); |
| 1924 | return; |
| 1925 | } |
| 1926 | |
| 1927 | // If we have "X && 1", simplify the code to use an uncond branch. |
| 1928 | // "X && 0" would have been constant folded to 0. |
| 1929 | if (ConstantFoldsToSimpleInteger(Cond: CondBOp->getRHS(), ResultBool&: ConstantBool) && |
| 1930 | ConstantBool) { |
| 1931 | // br(X && 1) -> br(X). |
| 1932 | EmitBranchToCounterBlock(Cond: CondBOp->getLHS(), LOp: BO_LAnd, TrueBlock, |
| 1933 | FalseBlock, TrueCount, LH, CntrIdx: CondBOp); |
| 1934 | return; |
| 1935 | } |
| 1936 | |
| 1937 | // Emit the LHS as a conditional. If the LHS conditional is false, we |
| 1938 | // want to jump to the FalseBlock. |
| 1939 | llvm::BasicBlock *LHSTrue = createBasicBlock(name: "land.lhs.true" ); |
| 1940 | llvm::BasicBlock *LHSFalse = |
| 1941 | (HasSkip ? createBasicBlock(name: "land.lhsskip" ) : FalseBlock); |
| 1942 | // The counter tells us how often we evaluate RHS, and all of TrueCount |
| 1943 | // can be propagated to that branch. |
| 1944 | uint64_t RHSCount = getProfileCount(S: CondBOp->getRHS()); |
| 1945 | |
| 1946 | ConditionalEvaluation eval(*this); |
| 1947 | { |
| 1948 | ApplyDebugLocation DL(*this, Cond); |
| 1949 | // Propagate the likelihood attribute like __builtin_expect |
| 1950 | // __builtin_expect(X && Y, 1) -> X and Y are likely |
| 1951 | // __builtin_expect(X && Y, 0) -> only Y is unlikely |
| 1952 | EmitBranchOnBoolExpr(Cond: CondBOp->getLHS(), TrueBlock: LHSTrue, FalseBlock: LHSFalse, TrueCount: RHSCount, |
| 1953 | LH: LH == Stmt::LH_Unlikely ? Stmt::LH_None : LH); |
| 1954 | if (HasSkip) { |
| 1955 | EmitBlock(BB: LHSFalse); |
| 1956 | incrementProfileCounter(ExecSkip: UseSkipPath, S: CondBOp); |
| 1957 | EmitBranch(Block: FalseBlock); |
| 1958 | } |
| 1959 | EmitBlock(BB: LHSTrue); |
| 1960 | } |
| 1961 | |
| 1962 | incrementProfileCounter(ExecSkip: UseExecPath, S: CondBOp); |
| 1963 | setCurrentProfileCount(getProfileCount(S: CondBOp->getRHS())); |
| 1964 | |
| 1965 | // Any temporaries created here are conditional. |
| 1966 | eval.begin(CGF&: *this); |
| 1967 | EmitBranchToCounterBlock(Cond: CondBOp->getRHS(), LOp: BO_LAnd, TrueBlock, |
| 1968 | FalseBlock, TrueCount, LH); |
| 1969 | eval.end(CGF&: *this); |
| 1970 | return; |
| 1971 | } |
| 1972 | |
| 1973 | if (CondBOp->getOpcode() == BO_LOr) { |
| 1974 | // If we have "0 || X", simplify the code. "1 || X" would have constant |
| 1975 | // folded if the case was simple enough. |
| 1976 | bool ConstantBool = false; |
| 1977 | if (ConstantFoldsToSimpleInteger(Cond: CondBOp->getLHS(), ResultBool&: ConstantBool) && |
| 1978 | !ConstantBool) { |
| 1979 | // br(0 || X) -> br(X). |
| 1980 | incrementProfileCounter(S: CondBOp); |
| 1981 | EmitBranchToCounterBlock(Cond: CondBOp->getRHS(), LOp: BO_LOr, TrueBlock, |
| 1982 | FalseBlock, TrueCount, LH); |
| 1983 | return; |
| 1984 | } |
| 1985 | |
| 1986 | // If we have "X || 0", simplify the code to use an uncond branch. |
| 1987 | // "X || 1" would have been constant folded to 1. |
| 1988 | if (ConstantFoldsToSimpleInteger(Cond: CondBOp->getRHS(), ResultBool&: ConstantBool) && |
| 1989 | !ConstantBool) { |
| 1990 | // br(X || 0) -> br(X). |
| 1991 | EmitBranchToCounterBlock(Cond: CondBOp->getLHS(), LOp: BO_LOr, TrueBlock, |
| 1992 | FalseBlock, TrueCount, LH, CntrIdx: CondBOp); |
| 1993 | return; |
| 1994 | } |
| 1995 | // Emit the LHS as a conditional. If the LHS conditional is true, we |
| 1996 | // want to jump to the TrueBlock. |
| 1997 | llvm::BasicBlock *LHSTrue = |
| 1998 | (HasSkip ? createBasicBlock(name: "lor.lhsskip" ) : TrueBlock); |
| 1999 | llvm::BasicBlock *LHSFalse = createBasicBlock(name: "lor.lhs.false" ); |
| 2000 | // We have the count for entry to the RHS and for the whole expression |
| 2001 | // being true, so we can divy up True count between the short circuit and |
| 2002 | // the RHS. |
| 2003 | uint64_t LHSCount = |
| 2004 | getCurrentProfileCount() - getProfileCount(S: CondBOp->getRHS()); |
| 2005 | uint64_t RHSCount = TrueCount - LHSCount; |
| 2006 | |
| 2007 | ConditionalEvaluation eval(*this); |
| 2008 | { |
| 2009 | // Propagate the likelihood attribute like __builtin_expect |
| 2010 | // __builtin_expect(X || Y, 1) -> only Y is likely |
| 2011 | // __builtin_expect(X || Y, 0) -> both X and Y are unlikely |
| 2012 | ApplyDebugLocation DL(*this, Cond); |
| 2013 | EmitBranchOnBoolExpr(Cond: CondBOp->getLHS(), TrueBlock: LHSTrue, FalseBlock: LHSFalse, TrueCount: LHSCount, |
| 2014 | LH: LH == Stmt::LH_Likely ? Stmt::LH_None : LH); |
| 2015 | if (HasSkip) { |
| 2016 | EmitBlock(BB: LHSTrue); |
| 2017 | incrementProfileCounter(ExecSkip: UseSkipPath, S: CondBOp); |
| 2018 | EmitBranch(Block: TrueBlock); |
| 2019 | } |
| 2020 | EmitBlock(BB: LHSFalse); |
| 2021 | } |
| 2022 | |
| 2023 | incrementProfileCounter(ExecSkip: UseExecPath, S: CondBOp); |
| 2024 | setCurrentProfileCount(getProfileCount(S: CondBOp->getRHS())); |
| 2025 | |
| 2026 | // Any temporaries created here are conditional. |
| 2027 | eval.begin(CGF&: *this); |
| 2028 | EmitBranchToCounterBlock(Cond: CondBOp->getRHS(), LOp: BO_LOr, TrueBlock, FalseBlock, |
| 2029 | TrueCount: RHSCount, LH); |
| 2030 | |
| 2031 | eval.end(CGF&: *this); |
| 2032 | return; |
| 2033 | } |
| 2034 | } |
| 2035 | |
| 2036 | if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Val: Cond)) { |
| 2037 | // br(!x, t, f) -> br(x, f, t) |
| 2038 | // Avoid doing this optimization when instrumenting a condition for MC/DC. |
| 2039 | // LNot is taken as part of the condition for simplicity, and changing its |
| 2040 | // sense negatively impacts test vector tracking. |
| 2041 | bool MCDCCondition = CGM.getCodeGenOpts().hasProfileClangInstr() && |
| 2042 | CGM.getCodeGenOpts().MCDCCoverage && |
| 2043 | isInstrumentedCondition(C: Cond); |
| 2044 | if (CondUOp->getOpcode() == UO_LNot && !MCDCCondition) { |
| 2045 | // Negate the count. |
| 2046 | uint64_t FalseCount = getCurrentProfileCount() - TrueCount; |
| 2047 | // The values of the enum are chosen to make this negation possible. |
| 2048 | LH = static_cast<Stmt::Likelihood>(-LH); |
| 2049 | // Negate the condition and swap the destination blocks. |
| 2050 | return EmitBranchOnBoolExpr(Cond: CondUOp->getSubExpr(), TrueBlock: FalseBlock, FalseBlock: TrueBlock, |
| 2051 | TrueCount: FalseCount, LH); |
| 2052 | } |
| 2053 | } |
| 2054 | |
| 2055 | if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Val: Cond)) { |
| 2056 | // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f)) |
| 2057 | llvm::BasicBlock *LHSBlock = createBasicBlock(name: "cond.true" ); |
| 2058 | llvm::BasicBlock *RHSBlock = createBasicBlock(name: "cond.false" ); |
| 2059 | |
| 2060 | // The ConditionalOperator itself has no likelihood information for its |
| 2061 | // true and false branches. This matches the behavior of __builtin_expect. |
| 2062 | ConditionalEvaluation cond(*this); |
| 2063 | EmitBranchOnBoolExpr(Cond: CondOp->getCond(), TrueBlock: LHSBlock, FalseBlock: RHSBlock, |
| 2064 | TrueCount: getProfileCount(S: CondOp), LH: Stmt::LH_None); |
| 2065 | |
| 2066 | // When computing PGO branch weights, we only know the overall count for |
| 2067 | // the true block. This code is essentially doing tail duplication of the |
| 2068 | // naive code-gen, introducing new edges for which counts are not |
| 2069 | // available. Divide the counts proportionally between the LHS and RHS of |
| 2070 | // the conditional operator. |
| 2071 | uint64_t LHSScaledTrueCount = 0; |
| 2072 | if (TrueCount) { |
| 2073 | double LHSRatio = |
| 2074 | getProfileCount(S: CondOp) / (double)getCurrentProfileCount(); |
| 2075 | LHSScaledTrueCount = TrueCount * LHSRatio; |
| 2076 | } |
| 2077 | |
| 2078 | cond.begin(CGF&: *this); |
| 2079 | EmitBlock(BB: LHSBlock); |
| 2080 | incrementProfileCounter(ExecSkip: UseExecPath, S: CondOp); |
| 2081 | { |
| 2082 | ApplyDebugLocation DL(*this, Cond); |
| 2083 | EmitBranchOnBoolExpr(Cond: CondOp->getLHS(), TrueBlock, FalseBlock, |
| 2084 | TrueCount: LHSScaledTrueCount, LH, ConditionalOp: CondOp); |
| 2085 | } |
| 2086 | cond.end(CGF&: *this); |
| 2087 | |
| 2088 | cond.begin(CGF&: *this); |
| 2089 | EmitBlock(BB: RHSBlock); |
| 2090 | incrementProfileCounter(ExecSkip: UseSkipPath, S: CondOp); |
| 2091 | EmitBranchOnBoolExpr(Cond: CondOp->getRHS(), TrueBlock, FalseBlock, |
| 2092 | TrueCount: TrueCount - LHSScaledTrueCount, LH, ConditionalOp: CondOp); |
| 2093 | cond.end(CGF&: *this); |
| 2094 | |
| 2095 | return; |
| 2096 | } |
| 2097 | |
| 2098 | if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Val: Cond)) { |
| 2099 | // Conditional operator handling can give us a throw expression as a |
| 2100 | // condition for a case like: |
| 2101 | // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f) |
| 2102 | // Fold this to: |
| 2103 | // br(c, throw x, br(y, t, f)) |
| 2104 | EmitCXXThrowExpr(E: Throw, /*KeepInsertionPoint*/false); |
| 2105 | return; |
| 2106 | } |
| 2107 | |
| 2108 | // Emit the code with the fully general case. |
| 2109 | llvm::Value *CondV; |
| 2110 | { |
| 2111 | ApplyDebugLocation DL(*this, Cond); |
| 2112 | CondV = EvaluateExprAsBool(E: Cond); |
| 2113 | } |
| 2114 | |
| 2115 | MaybeEmitDeferredVarDeclInit(var: ConditionalDecl); |
| 2116 | |
| 2117 | // If not at the top of the logical operator nest, update MCDC temp with the |
| 2118 | // boolean result of the evaluated condition. |
| 2119 | { |
| 2120 | const Expr *MCDCBaseExpr = Cond; |
| 2121 | // When a nested ConditionalOperator (ternary) is encountered in a boolean |
| 2122 | // expression, MC/DC tracks the result of the ternary, and this is tied to |
| 2123 | // the ConditionalOperator expression and not the ternary's LHS or RHS. If |
| 2124 | // this is the case, the ConditionalOperator expression is passed through |
| 2125 | // the ConditionalOp parameter and then used as the MCDC base expression. |
| 2126 | if (ConditionalOp) |
| 2127 | MCDCBaseExpr = ConditionalOp; |
| 2128 | |
| 2129 | if (isMCDCBranchExpr(E: stripCond(C: MCDCBaseExpr)) && |
| 2130 | !isMCDCDecisionExpr(E: stripCond(C: Cond))) |
| 2131 | maybeUpdateMCDCCondBitmap(E: MCDCBaseExpr, Val: CondV); |
| 2132 | } |
| 2133 | |
| 2134 | llvm::MDNode *Weights = nullptr; |
| 2135 | llvm::MDNode *Unpredictable = nullptr; |
| 2136 | |
| 2137 | // If the branch has a condition wrapped by __builtin_unpredictable, |
| 2138 | // create metadata that specifies that the branch is unpredictable. |
| 2139 | // Don't bother if not optimizing because that metadata would not be used. |
| 2140 | auto *Call = dyn_cast<CallExpr>(Val: Cond->IgnoreImpCasts()); |
| 2141 | if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) { |
| 2142 | auto *FD = dyn_cast_or_null<FunctionDecl>(Val: Call->getCalleeDecl()); |
| 2143 | if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) { |
| 2144 | llvm::MDBuilder MDHelper(getLLVMContext()); |
| 2145 | Unpredictable = MDHelper.createUnpredictable(); |
| 2146 | } |
| 2147 | } |
| 2148 | |
| 2149 | // If there is a Likelihood knowledge for the cond, lower it. |
| 2150 | // Note that if not optimizing this won't emit anything. |
| 2151 | llvm::Value *NewCondV = emitCondLikelihoodViaExpectIntrinsic(Cond: CondV, LH); |
| 2152 | if (CondV != NewCondV) |
| 2153 | CondV = NewCondV; |
| 2154 | else { |
| 2155 | // Otherwise, lower profile counts. Note that we do this even at -O0. |
| 2156 | uint64_t CurrentCount = std::max(a: getCurrentProfileCount(), b: TrueCount); |
| 2157 | Weights = createProfileWeights(TrueCount, FalseCount: CurrentCount - TrueCount); |
| 2158 | } |
| 2159 | |
| 2160 | llvm::Instruction *BrInst = Builder.CreateCondBr(Cond: CondV, True: TrueBlock, False: FalseBlock, |
| 2161 | BranchWeights: Weights, Unpredictable); |
| 2162 | addInstToNewSourceAtom(KeyInstruction: BrInst, Backup: CondV); |
| 2163 | |
| 2164 | switch (HLSLControlFlowAttr) { |
| 2165 | case HLSLControlFlowHintAttr::Microsoft_branch: |
| 2166 | case HLSLControlFlowHintAttr::Microsoft_flatten: { |
| 2167 | llvm::MDBuilder MDHelper(CGM.getLLVMContext()); |
| 2168 | |
| 2169 | llvm::ConstantInt *BranchHintConstant = |
| 2170 | HLSLControlFlowAttr == |
| 2171 | HLSLControlFlowHintAttr::Spelling::Microsoft_branch |
| 2172 | ? llvm::ConstantInt::get(Ty: CGM.Int32Ty, V: 1) |
| 2173 | : llvm::ConstantInt::get(Ty: CGM.Int32Ty, V: 2); |
| 2174 | |
| 2175 | SmallVector<llvm::Metadata *, 2> Vals( |
| 2176 | {MDHelper.createString(Str: "hlsl.controlflow.hint" ), |
| 2177 | MDHelper.createConstant(C: BranchHintConstant)}); |
| 2178 | BrInst->setMetadata(Kind: "hlsl.controlflow.hint" , |
| 2179 | Node: llvm::MDNode::get(Context&: CGM.getLLVMContext(), MDs: Vals)); |
| 2180 | break; |
| 2181 | } |
| 2182 | // This is required to avoid warnings during compilation |
| 2183 | case HLSLControlFlowHintAttr::SpellingNotCalculated: |
| 2184 | break; |
| 2185 | } |
| 2186 | } |
| 2187 | |
| 2188 | llvm::Value *CodeGenFunction::EmitScalarOrConstFoldImmArg(unsigned ICEArguments, |
| 2189 | unsigned Idx, |
| 2190 | const CallExpr *E) { |
| 2191 | llvm::Value *Arg = nullptr; |
| 2192 | if ((ICEArguments & (1 << Idx)) == 0) { |
| 2193 | Arg = EmitScalarExpr(E: E->getArg(Arg: Idx)); |
| 2194 | } else { |
| 2195 | // If this is required to be a constant, constant fold it so that we |
| 2196 | // know that the generated intrinsic gets a ConstantInt. |
| 2197 | std::optional<llvm::APSInt> Result = |
| 2198 | E->getArg(Arg: Idx)->getIntegerConstantExpr(Ctx: getContext()); |
| 2199 | assert(Result && "Expected argument to be a constant" ); |
| 2200 | Arg = llvm::ConstantInt::get(Context&: getLLVMContext(), V: *Result); |
| 2201 | } |
| 2202 | return Arg; |
| 2203 | } |
| 2204 | |
| 2205 | /// ErrorUnsupported - Print out an error that codegen doesn't support the |
| 2206 | /// specified stmt yet. |
| 2207 | void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) { |
| 2208 | CGM.ErrorUnsupported(S, Type); |
| 2209 | } |
| 2210 | |
| 2211 | /// emitNonZeroVLAInit - Emit the "zero" initialization of a |
| 2212 | /// variable-length array whose elements have a non-zero bit-pattern. |
| 2213 | /// |
| 2214 | /// \param baseType the inner-most element type of the array |
| 2215 | /// \param src - a char* pointing to the bit-pattern for a single |
| 2216 | /// base element of the array |
| 2217 | /// \param sizeInChars - the total size of the VLA, in chars |
| 2218 | static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType, |
| 2219 | Address dest, Address src, |
| 2220 | llvm::Value *sizeInChars) { |
| 2221 | CGBuilderTy &Builder = CGF.Builder; |
| 2222 | |
| 2223 | CharUnits baseSize = CGF.getContext().getTypeSizeInChars(T: baseType); |
| 2224 | llvm::Value *baseSizeInChars |
| 2225 | = llvm::ConstantInt::get(Ty: CGF.IntPtrTy, V: baseSize.getQuantity()); |
| 2226 | |
| 2227 | Address begin = dest.withElementType(ElemTy: CGF.Int8Ty); |
| 2228 | llvm::Value *end = Builder.CreateInBoundsGEP(Ty: begin.getElementType(), |
| 2229 | Ptr: begin.emitRawPointer(CGF), |
| 2230 | IdxList: sizeInChars, Name: "vla.end" ); |
| 2231 | |
| 2232 | llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock(); |
| 2233 | llvm::BasicBlock *loopBB = CGF.createBasicBlock(name: "vla-init.loop" ); |
| 2234 | llvm::BasicBlock *contBB = CGF.createBasicBlock(name: "vla-init.cont" ); |
| 2235 | |
| 2236 | // Make a loop over the VLA. C99 guarantees that the VLA element |
| 2237 | // count must be nonzero. |
| 2238 | CGF.EmitBlock(BB: loopBB); |
| 2239 | |
| 2240 | llvm::PHINode *cur = Builder.CreatePHI(Ty: begin.getType(), NumReservedValues: 2, Name: "vla.cur" ); |
| 2241 | cur->addIncoming(V: begin.emitRawPointer(CGF), BB: originBB); |
| 2242 | |
| 2243 | CharUnits curAlign = |
| 2244 | dest.getAlignment().alignmentOfArrayElement(elementSize: baseSize); |
| 2245 | |
| 2246 | // memcpy the individual element bit-pattern. |
| 2247 | Builder.CreateMemCpy(Dest: Address(cur, CGF.Int8Ty, curAlign), Src: src, Size: baseSizeInChars, |
| 2248 | /*volatile*/ IsVolatile: false); |
| 2249 | |
| 2250 | // Go to the next element. |
| 2251 | llvm::Value *next = |
| 2252 | Builder.CreateInBoundsGEP(Ty: CGF.Int8Ty, Ptr: cur, IdxList: baseSizeInChars, Name: "vla.next" ); |
| 2253 | |
| 2254 | // Leave if that's the end of the VLA. |
| 2255 | llvm::Value *done = Builder.CreateICmpEQ(LHS: next, RHS: end, Name: "vla-init.isdone" ); |
| 2256 | Builder.CreateCondBr(Cond: done, True: contBB, False: loopBB); |
| 2257 | cur->addIncoming(V: next, BB: loopBB); |
| 2258 | |
| 2259 | CGF.EmitBlock(BB: contBB); |
| 2260 | } |
| 2261 | |
| 2262 | Address CodeGenFunction::EmitAddressOfPFPField(Address RecordPtr, |
| 2263 | const PFPField &Field) { |
| 2264 | return EmitAddressOfPFPField( |
| 2265 | RecordPtr, |
| 2266 | FieldPtr: Builder.CreateConstInBoundsByteGEP(Addr: RecordPtr.withElementType(ElemTy: Int8Ty), |
| 2267 | Offset: Field.Offset), |
| 2268 | Field: Field.Field); |
| 2269 | } |
| 2270 | |
| 2271 | Address CodeGenFunction::EmitAddressOfPFPField(Address RecordPtr, |
| 2272 | Address PtrPtr, |
| 2273 | const FieldDecl *Field) { |
| 2274 | llvm::Value *Disc; |
| 2275 | if (CGM.getContext().arePFPFieldsTriviallyCopyable(RD: Field->getParent())) { |
| 2276 | uint64_t FieldSignature = |
| 2277 | llvm::getPointerAuthStableSipHash(S: CGM.getPFPFieldName(FD: Field)); |
| 2278 | Disc = llvm::ConstantInt::get(Ty: CGM.Int64Ty, V: FieldSignature); |
| 2279 | } else |
| 2280 | Disc = Builder.CreatePtrToInt(V: RecordPtr.getBasePointer(), DestTy: CGM.Int64Ty); |
| 2281 | |
| 2282 | llvm::GlobalValue *DS = CGM.getPFPDeactivationSymbol(FD: Field); |
| 2283 | llvm::OperandBundleDef DSBundle("deactivation-symbol" , DS); |
| 2284 | llvm::Value *Args[] = {PtrPtr.getBasePointer(), Disc, Builder.getTrue()}; |
| 2285 | return Address( |
| 2286 | Builder.CreateCall(Callee: CGM.getIntrinsic(IID: llvm::Intrinsic::protected_field_ptr, |
| 2287 | Tys: PtrPtr.getType()), |
| 2288 | Args, OpBundles: DSBundle), |
| 2289 | VoidPtrTy, PtrPtr.getAlignment()); |
| 2290 | } |
| 2291 | |
| 2292 | void |
| 2293 | CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) { |
| 2294 | // Ignore empty classes in C++. |
| 2295 | if (getLangOpts().CPlusPlus) |
| 2296 | if (const auto *RD = Ty->getAsCXXRecordDecl(); RD && RD->isEmpty()) |
| 2297 | return; |
| 2298 | |
| 2299 | if (DestPtr.getElementType() != Int8Ty) |
| 2300 | DestPtr = DestPtr.withElementType(ElemTy: Int8Ty); |
| 2301 | |
| 2302 | // Get size and alignment info for this aggregate. |
| 2303 | CharUnits size = getContext().getTypeSizeInChars(T: Ty); |
| 2304 | |
| 2305 | llvm::Value *SizeVal; |
| 2306 | const VariableArrayType *vla; |
| 2307 | |
| 2308 | // Don't bother emitting a zero-byte memset. |
| 2309 | if (size.isZero()) { |
| 2310 | // But note that getTypeInfo returns 0 for a VLA. |
| 2311 | if (const VariableArrayType *vlaType = |
| 2312 | dyn_cast_or_null<VariableArrayType>( |
| 2313 | Val: getContext().getAsArrayType(T: Ty))) { |
| 2314 | auto VlaSize = getVLASize(vla: vlaType); |
| 2315 | SizeVal = VlaSize.NumElts; |
| 2316 | CharUnits eltSize = getContext().getTypeSizeInChars(T: VlaSize.Type); |
| 2317 | if (!eltSize.isOne()) |
| 2318 | SizeVal = Builder.CreateNUWMul(LHS: SizeVal, RHS: CGM.getSize(numChars: eltSize)); |
| 2319 | vla = vlaType; |
| 2320 | } else { |
| 2321 | return; |
| 2322 | } |
| 2323 | } else { |
| 2324 | SizeVal = CGM.getSize(numChars: size); |
| 2325 | vla = nullptr; |
| 2326 | } |
| 2327 | |
| 2328 | // If the type contains a pointer to data member we can't memset it to zero. |
| 2329 | // Instead, create a null constant and copy it to the destination. |
| 2330 | // TODO: there are other patterns besides zero that we can usefully memset, |
| 2331 | // like -1, which happens to be the pattern used by member-pointers. |
| 2332 | if (!CGM.getTypes().isZeroInitializable(T: Ty)) { |
| 2333 | // For a VLA, emit a single element, then splat that over the VLA. |
| 2334 | if (vla) Ty = getContext().getBaseElementType(VAT: vla); |
| 2335 | |
| 2336 | llvm::Constant *NullConstant = CGM.EmitNullConstant(T: Ty); |
| 2337 | |
| 2338 | llvm::GlobalVariable *NullVariable = |
| 2339 | new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(), |
| 2340 | /*isConstant=*/true, |
| 2341 | llvm::GlobalVariable::PrivateLinkage, |
| 2342 | NullConstant, Twine()); |
| 2343 | CharUnits NullAlign = DestPtr.getAlignment(); |
| 2344 | NullVariable->setAlignment(NullAlign.getAsAlign()); |
| 2345 | Address SrcPtr(NullVariable, Builder.getInt8Ty(), NullAlign); |
| 2346 | |
| 2347 | if (vla) return emitNonZeroVLAInit(CGF&: *this, baseType: Ty, dest: DestPtr, src: SrcPtr, sizeInChars: SizeVal); |
| 2348 | |
| 2349 | // Get and call the appropriate llvm.memcpy overload. |
| 2350 | Builder.CreateMemCpy(Dest: DestPtr, Src: SrcPtr, Size: SizeVal, IsVolatile: false); |
| 2351 | } else { |
| 2352 | // Otherwise, just memset the whole thing to zero. This is legal |
| 2353 | // because in LLVM, all default initializers (other than the ones we just |
| 2354 | // handled above, and the case handled below) are guaranteed to have a bit |
| 2355 | // pattern of all zeros. |
| 2356 | Builder.CreateMemSet(Dest: DestPtr, Value: Builder.getInt8(C: 0), Size: SizeVal, IsVolatile: false); |
| 2357 | } |
| 2358 | |
| 2359 | // With the pointer field protection feature, null pointers do not have a bit |
| 2360 | // pattern of zero in memory, so we must initialize them separately. |
| 2361 | for (auto &Field : getContext().findPFPFields(Ty)) { |
| 2362 | auto addr = EmitAddressOfPFPField(RecordPtr: DestPtr, Field); |
| 2363 | Builder.CreateStore(Val: llvm::ConstantPointerNull::get(T: VoidPtrTy), Addr: addr); |
| 2364 | } |
| 2365 | } |
| 2366 | |
| 2367 | llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) { |
| 2368 | // Make sure that there is a block for the indirect goto. |
| 2369 | if (!IndirectBranch) |
| 2370 | GetIndirectGotoBlock(); |
| 2371 | |
| 2372 | llvm::BasicBlock *BB = getJumpDestForLabel(S: L).getBlock(); |
| 2373 | |
| 2374 | // Make sure the indirect branch includes all of the address-taken blocks. |
| 2375 | IndirectBranch->addDestination(Dest: BB); |
| 2376 | return llvm::BlockAddress::get(Ty: CurFn->getType(), BB); |
| 2377 | } |
| 2378 | |
| 2379 | llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() { |
| 2380 | // If we already made the indirect branch for indirect goto, return its block. |
| 2381 | if (IndirectBranch) return IndirectBranch->getParent(); |
| 2382 | |
| 2383 | CGBuilderTy TmpBuilder(CGM, createBasicBlock(name: "indirectgoto" )); |
| 2384 | |
| 2385 | // Create the PHI node that indirect gotos will add entries to. |
| 2386 | llvm::Value *DestVal = TmpBuilder.CreatePHI(Ty: Int8PtrTy, NumReservedValues: 0, |
| 2387 | Name: "indirect.goto.dest" ); |
| 2388 | |
| 2389 | // Create the indirect branch instruction. |
| 2390 | IndirectBranch = TmpBuilder.CreateIndirectBr(Addr: DestVal); |
| 2391 | return IndirectBranch->getParent(); |
| 2392 | } |
| 2393 | |
| 2394 | /// Computes the length of an array in elements, as well as the base |
| 2395 | /// element type and a properly-typed first element pointer. |
| 2396 | llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType, |
| 2397 | QualType &baseType, |
| 2398 | Address &addr) { |
| 2399 | const ArrayType *arrayType = origArrayType; |
| 2400 | |
| 2401 | // If it's a VLA, we have to load the stored size. Note that |
| 2402 | // this is the size of the VLA in bytes, not its size in elements. |
| 2403 | llvm::Value *numVLAElements = nullptr; |
| 2404 | if (isa<VariableArrayType>(Val: arrayType)) { |
| 2405 | numVLAElements = getVLASize(vla: cast<VariableArrayType>(Val: arrayType)).NumElts; |
| 2406 | |
| 2407 | // Walk into all VLAs. This doesn't require changes to addr, |
| 2408 | // which has type T* where T is the first non-VLA element type. |
| 2409 | do { |
| 2410 | QualType elementType = arrayType->getElementType(); |
| 2411 | arrayType = getContext().getAsArrayType(T: elementType); |
| 2412 | |
| 2413 | // If we only have VLA components, 'addr' requires no adjustment. |
| 2414 | if (!arrayType) { |
| 2415 | baseType = elementType; |
| 2416 | return numVLAElements; |
| 2417 | } |
| 2418 | } while (isa<VariableArrayType>(Val: arrayType)); |
| 2419 | |
| 2420 | // We get out here only if we find a constant array type |
| 2421 | // inside the VLA. |
| 2422 | } |
| 2423 | |
| 2424 | // We have some number of constant-length arrays, so addr should |
| 2425 | // have LLVM type [M x [N x [...]]]*. Build a GEP that walks |
| 2426 | // down to the first element of addr. |
| 2427 | SmallVector<llvm::Value*, 8> gepIndices; |
| 2428 | |
| 2429 | // GEP down to the array type. |
| 2430 | llvm::ConstantInt *zero = Builder.getInt32(C: 0); |
| 2431 | gepIndices.push_back(Elt: zero); |
| 2432 | |
| 2433 | uint64_t countFromCLAs = 1; |
| 2434 | QualType eltType; |
| 2435 | |
| 2436 | llvm::ArrayType *llvmArrayType = |
| 2437 | dyn_cast<llvm::ArrayType>(Val: addr.getElementType()); |
| 2438 | while (llvmArrayType) { |
| 2439 | assert(isa<ConstantArrayType>(arrayType)); |
| 2440 | assert(cast<ConstantArrayType>(arrayType)->getZExtSize() == |
| 2441 | llvmArrayType->getNumElements()); |
| 2442 | |
| 2443 | gepIndices.push_back(Elt: zero); |
| 2444 | countFromCLAs *= llvmArrayType->getNumElements(); |
| 2445 | eltType = arrayType->getElementType(); |
| 2446 | |
| 2447 | llvmArrayType = |
| 2448 | dyn_cast<llvm::ArrayType>(Val: llvmArrayType->getElementType()); |
| 2449 | arrayType = getContext().getAsArrayType(T: arrayType->getElementType()); |
| 2450 | assert((!llvmArrayType || arrayType) && |
| 2451 | "LLVM and Clang types are out-of-synch" ); |
| 2452 | } |
| 2453 | |
| 2454 | if (arrayType) { |
| 2455 | // From this point onwards, the Clang array type has been emitted |
| 2456 | // as some other type (probably a packed struct). Compute the array |
| 2457 | // size, and just emit the 'begin' expression as a bitcast. |
| 2458 | while (arrayType) { |
| 2459 | countFromCLAs *= cast<ConstantArrayType>(Val: arrayType)->getZExtSize(); |
| 2460 | eltType = arrayType->getElementType(); |
| 2461 | arrayType = getContext().getAsArrayType(T: eltType); |
| 2462 | } |
| 2463 | |
| 2464 | llvm::Type *baseType = ConvertType(T: eltType); |
| 2465 | addr = addr.withElementType(ElemTy: baseType); |
| 2466 | } else { |
| 2467 | // Create the actual GEP. |
| 2468 | addr = Address(Builder.CreateInBoundsGEP(Ty: addr.getElementType(), |
| 2469 | Ptr: addr.emitRawPointer(CGF&: *this), |
| 2470 | IdxList: gepIndices, Name: "array.begin" ), |
| 2471 | ConvertTypeForMem(T: eltType), addr.getAlignment()); |
| 2472 | } |
| 2473 | |
| 2474 | baseType = eltType; |
| 2475 | |
| 2476 | llvm::Value *numElements |
| 2477 | = llvm::ConstantInt::get(Ty: SizeTy, V: countFromCLAs); |
| 2478 | |
| 2479 | // If we had any VLA dimensions, factor them in. |
| 2480 | if (numVLAElements) |
| 2481 | numElements = Builder.CreateNUWMul(LHS: numVLAElements, RHS: numElements); |
| 2482 | |
| 2483 | return numElements; |
| 2484 | } |
| 2485 | |
| 2486 | CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) { |
| 2487 | const VariableArrayType *vla = getContext().getAsVariableArrayType(T: type); |
| 2488 | assert(vla && "type was not a variable array type!" ); |
| 2489 | return getVLASize(vla); |
| 2490 | } |
| 2491 | |
| 2492 | CodeGenFunction::VlaSizePair |
| 2493 | CodeGenFunction::getVLASize(const VariableArrayType *type) { |
| 2494 | // The number of elements so far; always size_t. |
| 2495 | llvm::Value *numElements = nullptr; |
| 2496 | |
| 2497 | QualType elementType; |
| 2498 | do { |
| 2499 | elementType = type->getElementType(); |
| 2500 | llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()]; |
| 2501 | assert(vlaSize && "no size for VLA!" ); |
| 2502 | assert(vlaSize->getType() == SizeTy); |
| 2503 | |
| 2504 | if (!numElements) { |
| 2505 | numElements = vlaSize; |
| 2506 | } else { |
| 2507 | // It's undefined behavior if this wraps around, so mark it that way. |
| 2508 | // FIXME: Teach -fsanitize=undefined to trap this. |
| 2509 | numElements = Builder.CreateNUWMul(LHS: numElements, RHS: vlaSize); |
| 2510 | } |
| 2511 | } while ((type = getContext().getAsVariableArrayType(T: elementType))); |
| 2512 | |
| 2513 | return { numElements, elementType }; |
| 2514 | } |
| 2515 | |
| 2516 | CodeGenFunction::VlaSizePair |
| 2517 | CodeGenFunction::getVLAElements1D(QualType type) { |
| 2518 | const VariableArrayType *vla = getContext().getAsVariableArrayType(T: type); |
| 2519 | assert(vla && "type was not a variable array type!" ); |
| 2520 | return getVLAElements1D(vla); |
| 2521 | } |
| 2522 | |
| 2523 | CodeGenFunction::VlaSizePair |
| 2524 | CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) { |
| 2525 | llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()]; |
| 2526 | assert(VlaSize && "no size for VLA!" ); |
| 2527 | assert(VlaSize->getType() == SizeTy); |
| 2528 | return { VlaSize, Vla->getElementType() }; |
| 2529 | } |
| 2530 | |
| 2531 | void CodeGenFunction::EmitVariablyModifiedType(QualType type) { |
| 2532 | assert(type->isVariablyModifiedType() && |
| 2533 | "Must pass variably modified type to EmitVLASizes!" ); |
| 2534 | |
| 2535 | EnsureInsertPoint(); |
| 2536 | |
| 2537 | // We're going to walk down into the type and look for VLA |
| 2538 | // expressions. |
| 2539 | do { |
| 2540 | assert(type->isVariablyModifiedType()); |
| 2541 | |
| 2542 | const Type *ty = type.getTypePtr(); |
| 2543 | switch (ty->getTypeClass()) { |
| 2544 | |
| 2545 | #define TYPE(Class, Base) |
| 2546 | #define ABSTRACT_TYPE(Class, Base) |
| 2547 | #define NON_CANONICAL_TYPE(Class, Base) |
| 2548 | #define DEPENDENT_TYPE(Class, Base) case Type::Class: |
| 2549 | #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base) |
| 2550 | #include "clang/AST/TypeNodes.inc" |
| 2551 | llvm_unreachable("unexpected dependent type!" ); |
| 2552 | |
| 2553 | // These types are never variably-modified. |
| 2554 | case Type::Builtin: |
| 2555 | case Type::Complex: |
| 2556 | case Type::Vector: |
| 2557 | case Type::ExtVector: |
| 2558 | case Type::ConstantMatrix: |
| 2559 | case Type::Record: |
| 2560 | case Type::Enum: |
| 2561 | case Type::Using: |
| 2562 | case Type::TemplateSpecialization: |
| 2563 | case Type::ObjCTypeParam: |
| 2564 | case Type::ObjCObject: |
| 2565 | case Type::ObjCInterface: |
| 2566 | case Type::ObjCObjectPointer: |
| 2567 | case Type::BitInt: |
| 2568 | case Type::HLSLInlineSpirv: |
| 2569 | case Type::PredefinedSugar: |
| 2570 | llvm_unreachable("type class is never variably-modified!" ); |
| 2571 | |
| 2572 | case Type::Adjusted: |
| 2573 | type = cast<AdjustedType>(Val: ty)->getAdjustedType(); |
| 2574 | break; |
| 2575 | |
| 2576 | case Type::Decayed: |
| 2577 | type = cast<DecayedType>(Val: ty)->getPointeeType(); |
| 2578 | break; |
| 2579 | |
| 2580 | case Type::Pointer: |
| 2581 | type = cast<PointerType>(Val: ty)->getPointeeType(); |
| 2582 | break; |
| 2583 | |
| 2584 | case Type::BlockPointer: |
| 2585 | type = cast<BlockPointerType>(Val: ty)->getPointeeType(); |
| 2586 | break; |
| 2587 | |
| 2588 | case Type::LValueReference: |
| 2589 | case Type::RValueReference: |
| 2590 | type = cast<ReferenceType>(Val: ty)->getPointeeType(); |
| 2591 | break; |
| 2592 | |
| 2593 | case Type::MemberPointer: |
| 2594 | type = cast<MemberPointerType>(Val: ty)->getPointeeType(); |
| 2595 | break; |
| 2596 | |
| 2597 | case Type::ArrayParameter: |
| 2598 | case Type::ConstantArray: |
| 2599 | case Type::IncompleteArray: |
| 2600 | // Losing element qualification here is fine. |
| 2601 | type = cast<ArrayType>(Val: ty)->getElementType(); |
| 2602 | break; |
| 2603 | |
| 2604 | case Type::VariableArray: { |
| 2605 | // Losing element qualification here is fine. |
| 2606 | const VariableArrayType *vat = cast<VariableArrayType>(Val: ty); |
| 2607 | |
| 2608 | // Unknown size indication requires no size computation. |
| 2609 | // Otherwise, evaluate and record it. |
| 2610 | if (const Expr *sizeExpr = vat->getSizeExpr()) { |
| 2611 | // It's possible that we might have emitted this already, |
| 2612 | // e.g. with a typedef and a pointer to it. |
| 2613 | llvm::Value *&entry = VLASizeMap[sizeExpr]; |
| 2614 | if (!entry) { |
| 2615 | llvm::Value *size = EmitScalarExpr(E: sizeExpr); |
| 2616 | |
| 2617 | // C11 6.7.6.2p5: |
| 2618 | // If the size is an expression that is not an integer constant |
| 2619 | // expression [...] each time it is evaluated it shall have a value |
| 2620 | // greater than zero. |
| 2621 | if (SanOpts.has(K: SanitizerKind::VLABound)) { |
| 2622 | auto CheckOrdinal = SanitizerKind::SO_VLABound; |
| 2623 | auto CheckHandler = SanitizerHandler::VLABoundNotPositive; |
| 2624 | SanitizerDebugLocation SanScope(this, {CheckOrdinal}, CheckHandler); |
| 2625 | llvm::Value *Zero = llvm::Constant::getNullValue(Ty: size->getType()); |
| 2626 | clang::QualType SEType = sizeExpr->getType(); |
| 2627 | llvm::Value *CheckCondition = |
| 2628 | SEType->isSignedIntegerType() |
| 2629 | ? Builder.CreateICmpSGT(LHS: size, RHS: Zero) |
| 2630 | : Builder.CreateICmpUGT(LHS: size, RHS: Zero); |
| 2631 | llvm::Constant *StaticArgs[] = { |
| 2632 | EmitCheckSourceLocation(Loc: sizeExpr->getBeginLoc()), |
| 2633 | EmitCheckTypeDescriptor(T: SEType)}; |
| 2634 | EmitCheck(Checked: std::make_pair(x&: CheckCondition, y&: CheckOrdinal), |
| 2635 | Check: CheckHandler, StaticArgs, DynamicArgs: size); |
| 2636 | } |
| 2637 | |
| 2638 | // Always zexting here would be wrong if it weren't |
| 2639 | // undefined behavior to have a negative bound. |
| 2640 | // FIXME: What about when size's type is larger than size_t? |
| 2641 | entry = Builder.CreateIntCast(V: size, DestTy: SizeTy, /*signed*/ isSigned: false); |
| 2642 | } |
| 2643 | } |
| 2644 | type = vat->getElementType(); |
| 2645 | break; |
| 2646 | } |
| 2647 | |
| 2648 | case Type::FunctionProto: |
| 2649 | case Type::FunctionNoProto: |
| 2650 | type = cast<FunctionType>(Val: ty)->getReturnType(); |
| 2651 | break; |
| 2652 | |
| 2653 | case Type::Paren: |
| 2654 | case Type::TypeOf: |
| 2655 | case Type::UnaryTransform: |
| 2656 | case Type::Attributed: |
| 2657 | case Type::BTFTagAttributed: |
| 2658 | case Type::OverflowBehavior: |
| 2659 | case Type::HLSLAttributedResource: |
| 2660 | case Type::SubstTemplateTypeParm: |
| 2661 | case Type::MacroQualified: |
| 2662 | case Type::CountAttributed: |
| 2663 | // Keep walking after single level desugaring. |
| 2664 | type = type.getSingleStepDesugaredType(Context: getContext()); |
| 2665 | break; |
| 2666 | |
| 2667 | case Type::Typedef: |
| 2668 | case Type::Decltype: |
| 2669 | case Type::Auto: |
| 2670 | case Type::DeducedTemplateSpecialization: |
| 2671 | case Type::PackIndexing: |
| 2672 | // Stop walking: nothing to do. |
| 2673 | return; |
| 2674 | |
| 2675 | case Type::TypeOfExpr: |
| 2676 | // Stop walking: emit typeof expression. |
| 2677 | EmitIgnoredExpr(E: cast<TypeOfExprType>(Val: ty)->getUnderlyingExpr()); |
| 2678 | return; |
| 2679 | |
| 2680 | case Type::Atomic: |
| 2681 | type = cast<AtomicType>(Val: ty)->getValueType(); |
| 2682 | break; |
| 2683 | |
| 2684 | case Type::Pipe: |
| 2685 | type = cast<PipeType>(Val: ty)->getElementType(); |
| 2686 | break; |
| 2687 | } |
| 2688 | } while (type->isVariablyModifiedType()); |
| 2689 | } |
| 2690 | |
| 2691 | Address CodeGenFunction::EmitVAListRef(const Expr* E) { |
| 2692 | if (getContext().getBuiltinVaListType()->isArrayType()) |
| 2693 | return EmitPointerWithAlignment(Addr: E); |
| 2694 | return EmitLValue(E).getAddress(); |
| 2695 | } |
| 2696 | |
| 2697 | Address CodeGenFunction::EmitMSVAListRef(const Expr *E) { |
| 2698 | return EmitLValue(E).getAddress(); |
| 2699 | } |
| 2700 | |
| 2701 | void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E, |
| 2702 | const APValue &Init) { |
| 2703 | assert(Init.hasValue() && "Invalid DeclRefExpr initializer!" ); |
| 2704 | if (CGDebugInfo *Dbg = getDebugInfo()) |
| 2705 | if (CGM.getCodeGenOpts().hasReducedDebugInfo()) |
| 2706 | Dbg->EmitGlobalVariable(VD: E->getDecl(), Init); |
| 2707 | } |
| 2708 | |
| 2709 | CodeGenFunction::PeepholeProtection |
| 2710 | CodeGenFunction::protectFromPeepholes(RValue rvalue) { |
| 2711 | // At the moment, the only aggressive peephole we do in IR gen |
| 2712 | // is trunc(zext) folding, but if we add more, we can easily |
| 2713 | // extend this protection. |
| 2714 | |
| 2715 | if (!rvalue.isScalar()) return PeepholeProtection(); |
| 2716 | llvm::Value *value = rvalue.getScalarVal(); |
| 2717 | if (!isa<llvm::ZExtInst>(Val: value)) return PeepholeProtection(); |
| 2718 | |
| 2719 | // Just make an extra bitcast. |
| 2720 | assert(HaveInsertPoint()); |
| 2721 | llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "" , |
| 2722 | Builder.GetInsertBlock()); |
| 2723 | |
| 2724 | PeepholeProtection protection; |
| 2725 | protection.Inst = inst; |
| 2726 | return protection; |
| 2727 | } |
| 2728 | |
| 2729 | void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) { |
| 2730 | if (!protection.Inst) return; |
| 2731 | |
| 2732 | // In theory, we could try to duplicate the peepholes now, but whatever. |
| 2733 | protection.Inst->eraseFromParent(); |
| 2734 | } |
| 2735 | |
| 2736 | void CodeGenFunction::emitAlignmentAssumption(llvm::Value *PtrValue, |
| 2737 | QualType Ty, SourceLocation Loc, |
| 2738 | SourceLocation AssumptionLoc, |
| 2739 | llvm::Value *Alignment, |
| 2740 | llvm::Value *OffsetValue) { |
| 2741 | if (Alignment->getType() != IntPtrTy) |
| 2742 | Alignment = |
| 2743 | Builder.CreateIntCast(V: Alignment, DestTy: IntPtrTy, isSigned: false, Name: "casted.align" ); |
| 2744 | if (OffsetValue && OffsetValue->getType() != IntPtrTy) |
| 2745 | OffsetValue = |
| 2746 | Builder.CreateIntCast(V: OffsetValue, DestTy: IntPtrTy, isSigned: true, Name: "casted.offset" ); |
| 2747 | llvm::Value *TheCheck = nullptr; |
| 2748 | if (SanOpts.has(K: SanitizerKind::Alignment)) { |
| 2749 | llvm::Value *PtrIntValue = |
| 2750 | Builder.CreatePtrToInt(V: PtrValue, DestTy: IntPtrTy, Name: "ptrint" ); |
| 2751 | |
| 2752 | if (OffsetValue) { |
| 2753 | bool IsOffsetZero = false; |
| 2754 | if (const auto *CI = dyn_cast<llvm::ConstantInt>(Val: OffsetValue)) |
| 2755 | IsOffsetZero = CI->isZero(); |
| 2756 | |
| 2757 | if (!IsOffsetZero) |
| 2758 | PtrIntValue = Builder.CreateSub(LHS: PtrIntValue, RHS: OffsetValue, Name: "offsetptr" ); |
| 2759 | } |
| 2760 | |
| 2761 | llvm::Value *Zero = llvm::ConstantInt::get(Ty: IntPtrTy, V: 0); |
| 2762 | llvm::Value *Mask = |
| 2763 | Builder.CreateSub(LHS: Alignment, RHS: llvm::ConstantInt::get(Ty: IntPtrTy, V: 1)); |
| 2764 | llvm::Value *MaskedPtr = Builder.CreateAnd(LHS: PtrIntValue, RHS: Mask, Name: "maskedptr" ); |
| 2765 | TheCheck = Builder.CreateICmpEQ(LHS: MaskedPtr, RHS: Zero, Name: "maskcond" ); |
| 2766 | } |
| 2767 | llvm::Instruction *Assumption = Builder.CreateAlignmentAssumption( |
| 2768 | DL: CGM.getDataLayout(), PtrValue, Alignment, OffsetValue); |
| 2769 | |
| 2770 | if (!SanOpts.has(K: SanitizerKind::Alignment)) |
| 2771 | return; |
| 2772 | emitAlignmentAssumptionCheck(Ptr: PtrValue, Ty, Loc, AssumptionLoc, Alignment, |
| 2773 | OffsetValue, TheCheck, Assumption); |
| 2774 | } |
| 2775 | |
| 2776 | void CodeGenFunction::emitAlignmentAssumption(llvm::Value *PtrValue, |
| 2777 | const Expr *E, |
| 2778 | SourceLocation AssumptionLoc, |
| 2779 | llvm::Value *Alignment, |
| 2780 | llvm::Value *OffsetValue) { |
| 2781 | QualType Ty = E->getType(); |
| 2782 | SourceLocation Loc = E->getExprLoc(); |
| 2783 | |
| 2784 | emitAlignmentAssumption(PtrValue, Ty, Loc, AssumptionLoc, Alignment, |
| 2785 | OffsetValue); |
| 2786 | } |
| 2787 | |
| 2788 | llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Function *AnnotationFn, |
| 2789 | llvm::Value *AnnotatedVal, |
| 2790 | StringRef AnnotationStr, |
| 2791 | SourceLocation Location, |
| 2792 | const AnnotateAttr *Attr) { |
| 2793 | SmallVector<llvm::Value *, 5> Args = { |
| 2794 | AnnotatedVal, |
| 2795 | CGM.EmitAnnotationString(Str: AnnotationStr), |
| 2796 | CGM.EmitAnnotationUnit(Loc: Location), |
| 2797 | CGM.EmitAnnotationLineNo(L: Location), |
| 2798 | }; |
| 2799 | if (Attr) |
| 2800 | Args.push_back(Elt: CGM.EmitAnnotationArgs(Attr)); |
| 2801 | return Builder.CreateCall(Callee: AnnotationFn, Args); |
| 2802 | } |
| 2803 | |
| 2804 | void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) { |
| 2805 | assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute" ); |
| 2806 | for (const auto *I : D->specific_attrs<AnnotateAttr>()) |
| 2807 | EmitAnnotationCall(AnnotationFn: CGM.getIntrinsic(IID: llvm::Intrinsic::var_annotation, |
| 2808 | Tys: {V->getType(), CGM.ConstGlobalsPtrTy}), |
| 2809 | AnnotatedVal: V, AnnotationStr: I->getAnnotation(), Location: D->getLocation(), Attr: I); |
| 2810 | } |
| 2811 | |
| 2812 | Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D, |
| 2813 | Address Addr) { |
| 2814 | assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute" ); |
| 2815 | llvm::Value *V = Addr.emitRawPointer(CGF&: *this); |
| 2816 | llvm::Type *VTy = V->getType(); |
| 2817 | auto *PTy = dyn_cast<llvm::PointerType>(Val: VTy); |
| 2818 | unsigned AS = PTy ? PTy->getAddressSpace() : 0; |
| 2819 | llvm::PointerType *IntrinTy = |
| 2820 | llvm::PointerType::get(C&: CGM.getLLVMContext(), AddressSpace: AS); |
| 2821 | llvm::Function *F = CGM.getIntrinsic(IID: llvm::Intrinsic::ptr_annotation, |
| 2822 | Tys: {IntrinTy, CGM.ConstGlobalsPtrTy}); |
| 2823 | |
| 2824 | for (const auto *I : D->specific_attrs<AnnotateAttr>()) { |
| 2825 | // FIXME Always emit the cast inst so we can differentiate between |
| 2826 | // annotation on the first field of a struct and annotation on the struct |
| 2827 | // itself. |
| 2828 | if (VTy != IntrinTy) |
| 2829 | V = Builder.CreateBitCast(V, DestTy: IntrinTy); |
| 2830 | V = EmitAnnotationCall(AnnotationFn: F, AnnotatedVal: V, AnnotationStr: I->getAnnotation(), Location: D->getLocation(), Attr: I); |
| 2831 | V = Builder.CreateBitCast(V, DestTy: VTy); |
| 2832 | } |
| 2833 | |
| 2834 | return Address(V, Addr.getElementType(), Addr.getAlignment()); |
| 2835 | } |
| 2836 | |
| 2837 | CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { } |
| 2838 | |
| 2839 | CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF) |
| 2840 | : CGF(CGF) { |
| 2841 | assert(!CGF->IsSanitizerScope); |
| 2842 | CGF->IsSanitizerScope = true; |
| 2843 | } |
| 2844 | |
| 2845 | CodeGenFunction::SanitizerScope::~SanitizerScope() { |
| 2846 | CGF->IsSanitizerScope = false; |
| 2847 | } |
| 2848 | |
| 2849 | void CodeGenFunction::InsertHelper(llvm::Instruction *I, |
| 2850 | const llvm::Twine &Name, |
| 2851 | llvm::BasicBlock::iterator InsertPt) const { |
| 2852 | LoopStack.InsertHelper(I); |
| 2853 | if (IsSanitizerScope) |
| 2854 | I->setNoSanitizeMetadata(); |
| 2855 | } |
| 2856 | |
| 2857 | void CGBuilderInserter::InsertHelper( |
| 2858 | llvm::Instruction *I, const llvm::Twine &Name, |
| 2859 | llvm::BasicBlock::iterator InsertPt) const { |
| 2860 | llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, InsertPt); |
| 2861 | if (CGF) |
| 2862 | CGF->InsertHelper(I, Name, InsertPt); |
| 2863 | } |
| 2864 | |
| 2865 | // Emits an error if we don't have a valid set of target features for the |
| 2866 | // called function. |
| 2867 | void CodeGenFunction::checkTargetFeatures(const CallExpr *E, |
| 2868 | const FunctionDecl *TargetDecl) { |
| 2869 | // SemaChecking cannot handle below x86 builtins because they have different |
| 2870 | // parameter ranges with different TargetAttribute of caller. |
| 2871 | if (CGM.getContext().getTargetInfo().getTriple().isX86()) { |
| 2872 | unsigned BuiltinID = TargetDecl->getBuiltinID(); |
| 2873 | if (BuiltinID == X86::BI__builtin_ia32_cmpps || |
| 2874 | BuiltinID == X86::BI__builtin_ia32_cmpss || |
| 2875 | BuiltinID == X86::BI__builtin_ia32_cmppd || |
| 2876 | BuiltinID == X86::BI__builtin_ia32_cmpsd) { |
| 2877 | const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: CurCodeDecl); |
| 2878 | llvm::StringMap<bool> TargetFetureMap; |
| 2879 | CGM.getContext().getFunctionFeatureMap(FeatureMap&: TargetFetureMap, FD); |
| 2880 | llvm::APSInt Result = |
| 2881 | *(E->getArg(Arg: 2)->getIntegerConstantExpr(Ctx: CGM.getContext())); |
| 2882 | if (Result.getSExtValue() > 7 && !TargetFetureMap.lookup(Key: "avx" )) |
| 2883 | CGM.getDiags().Report(Loc: E->getBeginLoc(), DiagID: diag::err_builtin_needs_feature) |
| 2884 | << TargetDecl->getDeclName() << "avx" ; |
| 2885 | } |
| 2886 | } |
| 2887 | return checkTargetFeatures(Loc: E->getBeginLoc(), TargetDecl); |
| 2888 | } |
| 2889 | |
| 2890 | // Emits an error if we don't have a valid set of target features for the |
| 2891 | // called function. |
| 2892 | void CodeGenFunction::checkTargetFeatures(SourceLocation Loc, |
| 2893 | const FunctionDecl *TargetDecl) { |
| 2894 | // Early exit if this is an indirect call. |
| 2895 | if (!TargetDecl) |
| 2896 | return; |
| 2897 | |
| 2898 | // Get the current enclosing function if it exists. If it doesn't |
| 2899 | // we can't check the target features anyhow. |
| 2900 | const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: CurCodeDecl); |
| 2901 | if (!FD) |
| 2902 | return; |
| 2903 | |
| 2904 | bool IsAlwaysInline = TargetDecl->hasAttr<AlwaysInlineAttr>(); |
| 2905 | bool IsFlatten = FD && FD->hasAttr<FlattenAttr>(); |
| 2906 | |
| 2907 | // Grab the required features for the call. For a builtin this is listed in |
| 2908 | // the td file with the default cpu, for an always_inline function this is any |
| 2909 | // listed cpu and any listed features. |
| 2910 | unsigned BuiltinID = TargetDecl->getBuiltinID(); |
| 2911 | std::string MissingFeature; |
| 2912 | llvm::StringMap<bool> CallerFeatureMap; |
| 2913 | CGM.getContext().getFunctionFeatureMap(FeatureMap&: CallerFeatureMap, FD); |
| 2914 | // When compiling in HipStdPar mode we have to be conservative in rejecting |
| 2915 | // target specific features in the FE, and defer the possible error to the |
| 2916 | // AcceleratorCodeSelection pass, wherein iff an unsupported target builtin is |
| 2917 | // referenced by an accelerator executable function, we emit an error. |
| 2918 | bool IsHipStdPar = getLangOpts().HIPStdPar && getLangOpts().CUDAIsDevice; |
| 2919 | if (BuiltinID) { |
| 2920 | StringRef FeatureList(CGM.getContext().BuiltinInfo.getRequiredFeatures(ID: BuiltinID)); |
| 2921 | if (!Builtin::evaluateRequiredTargetFeatures( |
| 2922 | RequiredFatures: FeatureList, TargetFetureMap: CallerFeatureMap) && !IsHipStdPar) { |
| 2923 | CGM.getDiags().Report(Loc, DiagID: diag::err_builtin_needs_feature) |
| 2924 | << TargetDecl->getDeclName() |
| 2925 | << FeatureList; |
| 2926 | } |
| 2927 | } else if (!TargetDecl->isMultiVersion() && |
| 2928 | TargetDecl->hasAttr<TargetAttr>()) { |
| 2929 | // Get the required features for the callee. |
| 2930 | |
| 2931 | const TargetAttr *TD = TargetDecl->getAttr<TargetAttr>(); |
| 2932 | ParsedTargetAttr ParsedAttr = |
| 2933 | CGM.getContext().filterFunctionTargetAttrs(TD); |
| 2934 | |
| 2935 | SmallVector<StringRef, 1> ReqFeatures; |
| 2936 | llvm::StringMap<bool> CalleeFeatureMap; |
| 2937 | CGM.getContext().getFunctionFeatureMap(FeatureMap&: CalleeFeatureMap, TargetDecl); |
| 2938 | |
| 2939 | for (const auto &F : ParsedAttr.Features) { |
| 2940 | if (F[0] == '+' && CalleeFeatureMap.lookup(Key: F.substr(pos: 1))) |
| 2941 | ReqFeatures.push_back(Elt: StringRef(F).substr(Start: 1)); |
| 2942 | } |
| 2943 | |
| 2944 | for (const auto &F : CalleeFeatureMap) { |
| 2945 | // Only positive features are "required". |
| 2946 | if (F.getValue()) |
| 2947 | ReqFeatures.push_back(Elt: F.getKey()); |
| 2948 | } |
| 2949 | if (!llvm::all_of(Range&: ReqFeatures, |
| 2950 | P: [&](StringRef Feature) { |
| 2951 | if (!CallerFeatureMap.lookup(Key: Feature)) { |
| 2952 | MissingFeature = Feature.str(); |
| 2953 | return false; |
| 2954 | } |
| 2955 | return true; |
| 2956 | }) && |
| 2957 | !IsHipStdPar) { |
| 2958 | if (IsAlwaysInline) |
| 2959 | CGM.getDiags().Report(Loc, DiagID: diag::err_function_needs_feature) |
| 2960 | << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature; |
| 2961 | else if (IsFlatten) |
| 2962 | CGM.getDiags().Report(Loc, DiagID: diag::err_flatten_function_needs_feature) |
| 2963 | << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature; |
| 2964 | } |
| 2965 | |
| 2966 | } else if (!FD->isMultiVersion() && FD->hasAttr<TargetAttr>()) { |
| 2967 | llvm::StringMap<bool> CalleeFeatureMap; |
| 2968 | CGM.getContext().getFunctionFeatureMap(FeatureMap&: CalleeFeatureMap, TargetDecl); |
| 2969 | |
| 2970 | for (const auto &F : CalleeFeatureMap) { |
| 2971 | if (F.getValue() && |
| 2972 | (!CallerFeatureMap.lookup(Key: F.getKey()) || |
| 2973 | !CallerFeatureMap.find(Key: F.getKey())->getValue()) && |
| 2974 | !IsHipStdPar) { |
| 2975 | if (IsAlwaysInline) |
| 2976 | CGM.getDiags().Report(Loc, DiagID: diag::err_function_needs_feature) |
| 2977 | << FD->getDeclName() << TargetDecl->getDeclName() << F.getKey(); |
| 2978 | else if (IsFlatten) |
| 2979 | CGM.getDiags().Report(Loc, DiagID: diag::err_flatten_function_needs_feature) |
| 2980 | << FD->getDeclName() << TargetDecl->getDeclName() << F.getKey(); |
| 2981 | } |
| 2982 | } |
| 2983 | } |
| 2984 | } |
| 2985 | |
| 2986 | void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) { |
| 2987 | if (!CGM.getCodeGenOpts().SanitizeStats) |
| 2988 | return; |
| 2989 | |
| 2990 | llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint()); |
| 2991 | IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation()); |
| 2992 | CGM.getSanStats().create(B&: IRB, SK: SSK); |
| 2993 | } |
| 2994 | |
| 2995 | void CodeGenFunction::EmitKCFIOperandBundle( |
| 2996 | const CGCallee &Callee, SmallVectorImpl<llvm::OperandBundleDef> &Bundles) { |
| 2997 | const CGCalleeInfo &CI = Callee.getAbstractInfo(); |
| 2998 | const FunctionProtoType *FP = CI.getCalleeFunctionProtoType(); |
| 2999 | if (!FP) |
| 3000 | return; |
| 3001 | |
| 3002 | StringRef Salt; |
| 3003 | if (const auto &Info = FP->getExtraAttributeInfo()) |
| 3004 | Salt = Info.CFISalt; |
| 3005 | |
| 3006 | Bundles.emplace_back(Args: "kcfi" , Args: CGM.CreateKCFITypeId(T: FP->desugar(), Salt)); |
| 3007 | } |
| 3008 | |
| 3009 | llvm::Value * |
| 3010 | CodeGenFunction::FormAArch64ResolverCondition(const FMVResolverOption &RO) { |
| 3011 | return RO.Features.empty() ? nullptr : EmitAArch64CpuSupports(FeatureStrs: RO.Features); |
| 3012 | } |
| 3013 | |
| 3014 | llvm::Value * |
| 3015 | CodeGenFunction::FormX86ResolverCondition(const FMVResolverOption &RO) { |
| 3016 | llvm::Value *Condition = nullptr; |
| 3017 | |
| 3018 | if (RO.Architecture) { |
| 3019 | StringRef Arch = *RO.Architecture; |
| 3020 | // If arch= specifies an x86-64 micro-architecture level, test the feature |
| 3021 | // with __builtin_cpu_supports, otherwise use __builtin_cpu_is. |
| 3022 | if (Arch.starts_with(Prefix: "x86-64" )) |
| 3023 | Condition = EmitX86CpuSupports(FeatureStrs: {Arch}); |
| 3024 | else |
| 3025 | Condition = EmitX86CpuIs(CPUStr: Arch); |
| 3026 | } |
| 3027 | |
| 3028 | if (!RO.Features.empty()) { |
| 3029 | llvm::Value *FeatureCond = EmitX86CpuSupports(FeatureStrs: RO.Features); |
| 3030 | Condition = |
| 3031 | Condition ? Builder.CreateAnd(LHS: Condition, RHS: FeatureCond) : FeatureCond; |
| 3032 | } |
| 3033 | return Condition; |
| 3034 | } |
| 3035 | |
| 3036 | static void CreateMultiVersionResolverReturn(CodeGenModule &CGM, |
| 3037 | llvm::Function *Resolver, |
| 3038 | CGBuilderTy &Builder, |
| 3039 | llvm::Function *FuncToReturn, |
| 3040 | bool SupportsIFunc) { |
| 3041 | if (SupportsIFunc) { |
| 3042 | Builder.CreateRet(V: FuncToReturn); |
| 3043 | return; |
| 3044 | } |
| 3045 | |
| 3046 | llvm::SmallVector<llvm::Value *, 10> Args( |
| 3047 | llvm::make_pointer_range(Range: Resolver->args())); |
| 3048 | |
| 3049 | llvm::CallInst *Result = Builder.CreateCall(Callee: FuncToReturn, Args); |
| 3050 | Result->setTailCallKind(llvm::CallInst::TCK_MustTail); |
| 3051 | |
| 3052 | if (Resolver->getReturnType()->isVoidTy()) |
| 3053 | Builder.CreateRetVoid(); |
| 3054 | else |
| 3055 | Builder.CreateRet(V: Result); |
| 3056 | } |
| 3057 | |
| 3058 | void CodeGenFunction::EmitMultiVersionResolver( |
| 3059 | llvm::Function *Resolver, ArrayRef<FMVResolverOption> Options) { |
| 3060 | |
| 3061 | llvm::Triple::ArchType ArchType = |
| 3062 | getContext().getTargetInfo().getTriple().getArch(); |
| 3063 | |
| 3064 | switch (ArchType) { |
| 3065 | case llvm::Triple::x86: |
| 3066 | case llvm::Triple::x86_64: |
| 3067 | EmitX86MultiVersionResolver(Resolver, Options); |
| 3068 | return; |
| 3069 | case llvm::Triple::aarch64: |
| 3070 | EmitAArch64MultiVersionResolver(Resolver, Options); |
| 3071 | return; |
| 3072 | case llvm::Triple::riscv32: |
| 3073 | case llvm::Triple::riscv64: |
| 3074 | case llvm::Triple::riscv32be: |
| 3075 | case llvm::Triple::riscv64be: |
| 3076 | EmitRISCVMultiVersionResolver(Resolver, Options); |
| 3077 | return; |
| 3078 | |
| 3079 | default: |
| 3080 | assert(false && "Only implemented for x86, AArch64 and RISC-V targets" ); |
| 3081 | } |
| 3082 | } |
| 3083 | |
| 3084 | void CodeGenFunction::EmitRISCVMultiVersionResolver( |
| 3085 | llvm::Function *Resolver, ArrayRef<FMVResolverOption> Options) { |
| 3086 | |
| 3087 | if (getContext().getTargetInfo().getTriple().getOS() != |
| 3088 | llvm::Triple::OSType::Linux) { |
| 3089 | CGM.getDiags().Report(DiagID: diag::err_os_unsupport_riscv_fmv); |
| 3090 | return; |
| 3091 | } |
| 3092 | |
| 3093 | llvm::BasicBlock *CurBlock = createBasicBlock(name: "resolver_entry" , parent: Resolver); |
| 3094 | Builder.SetInsertPoint(CurBlock); |
| 3095 | EmitRISCVCpuInit(); |
| 3096 | |
| 3097 | bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc(); |
| 3098 | bool HasDefault = false; |
| 3099 | unsigned DefaultIndex = 0; |
| 3100 | |
| 3101 | // Check the each candidate function. |
| 3102 | for (unsigned Index = 0; Index < Options.size(); Index++) { |
| 3103 | |
| 3104 | if (Options[Index].Features.empty()) { |
| 3105 | HasDefault = true; |
| 3106 | DefaultIndex = Index; |
| 3107 | continue; |
| 3108 | } |
| 3109 | |
| 3110 | Builder.SetInsertPoint(CurBlock); |
| 3111 | |
| 3112 | // FeaturesCondition: The bitmask of the required extension has been |
| 3113 | // enabled by the runtime object. |
| 3114 | // (__riscv_feature_bits.features[i] & REQUIRED_BITMASK) == |
| 3115 | // REQUIRED_BITMASK |
| 3116 | // |
| 3117 | // When condition is met, return this version of the function. |
| 3118 | // Otherwise, try the next version. |
| 3119 | // |
| 3120 | // if (FeaturesConditionVersion1) |
| 3121 | // return Version1; |
| 3122 | // else if (FeaturesConditionVersion2) |
| 3123 | // return Version2; |
| 3124 | // else if (FeaturesConditionVersion3) |
| 3125 | // return Version3; |
| 3126 | // ... |
| 3127 | // else |
| 3128 | // return DefaultVersion; |
| 3129 | |
| 3130 | // TODO: Add a condition to check the length before accessing elements. |
| 3131 | // Without checking the length first, we may access an incorrect memory |
| 3132 | // address when using different versions. |
| 3133 | llvm::SmallVector<StringRef, 8> CurrTargetAttrFeats; |
| 3134 | llvm::SmallVector<std::string, 8> TargetAttrFeats; |
| 3135 | |
| 3136 | for (StringRef Feat : Options[Index].Features) { |
| 3137 | std::vector<std::string> FeatStr = |
| 3138 | getContext().getTargetInfo().parseTargetAttr(Str: Feat).Features; |
| 3139 | |
| 3140 | assert(FeatStr.size() == 1 && "Feature string not delimited" ); |
| 3141 | |
| 3142 | std::string &CurrFeat = FeatStr.front(); |
| 3143 | if (CurrFeat[0] == '+') |
| 3144 | TargetAttrFeats.push_back(Elt: CurrFeat.substr(pos: 1)); |
| 3145 | } |
| 3146 | |
| 3147 | if (TargetAttrFeats.empty()) |
| 3148 | continue; |
| 3149 | |
| 3150 | for (std::string &Feat : TargetAttrFeats) |
| 3151 | CurrTargetAttrFeats.push_back(Elt: Feat); |
| 3152 | |
| 3153 | Builder.SetInsertPoint(CurBlock); |
| 3154 | llvm::Value *FeatsCondition = EmitRISCVCpuSupports(FeaturesStrs: CurrTargetAttrFeats); |
| 3155 | |
| 3156 | llvm::BasicBlock *RetBlock = createBasicBlock(name: "resolver_return" , parent: Resolver); |
| 3157 | CGBuilderTy RetBuilder(CGM, RetBlock); |
| 3158 | CreateMultiVersionResolverReturn(CGM, Resolver, Builder&: RetBuilder, |
| 3159 | FuncToReturn: Options[Index].Function, SupportsIFunc); |
| 3160 | llvm::BasicBlock *ElseBlock = createBasicBlock(name: "resolver_else" , parent: Resolver); |
| 3161 | |
| 3162 | Builder.SetInsertPoint(CurBlock); |
| 3163 | Builder.CreateCondBr(Cond: FeatsCondition, True: RetBlock, False: ElseBlock); |
| 3164 | |
| 3165 | CurBlock = ElseBlock; |
| 3166 | } |
| 3167 | |
| 3168 | // Finally, emit the default one. |
| 3169 | if (HasDefault) { |
| 3170 | Builder.SetInsertPoint(CurBlock); |
| 3171 | CreateMultiVersionResolverReturn( |
| 3172 | CGM, Resolver, Builder, FuncToReturn: Options[DefaultIndex].Function, SupportsIFunc); |
| 3173 | return; |
| 3174 | } |
| 3175 | |
| 3176 | // If no generic/default, emit an unreachable. |
| 3177 | Builder.SetInsertPoint(CurBlock); |
| 3178 | llvm::CallInst *TrapCall = EmitTrapCall(IntrID: llvm::Intrinsic::trap); |
| 3179 | TrapCall->setDoesNotReturn(); |
| 3180 | TrapCall->setDoesNotThrow(); |
| 3181 | Builder.CreateUnreachable(); |
| 3182 | Builder.ClearInsertionPoint(); |
| 3183 | } |
| 3184 | |
| 3185 | void CodeGenFunction::EmitAArch64MultiVersionResolver( |
| 3186 | llvm::Function *Resolver, ArrayRef<FMVResolverOption> Options) { |
| 3187 | assert(!Options.empty() && "No multiversion resolver options found" ); |
| 3188 | assert(Options.back().Features.size() == 0 && "Default case must be last" ); |
| 3189 | bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc(); |
| 3190 | assert(SupportsIFunc && |
| 3191 | "Multiversion resolver requires target IFUNC support" ); |
| 3192 | bool AArch64CpuInitialized = false; |
| 3193 | llvm::BasicBlock *CurBlock = createBasicBlock(name: "resolver_entry" , parent: Resolver); |
| 3194 | |
| 3195 | for (const FMVResolverOption &RO : Options) { |
| 3196 | Builder.SetInsertPoint(CurBlock); |
| 3197 | llvm::Value *Condition = FormAArch64ResolverCondition(RO); |
| 3198 | |
| 3199 | // The 'default' or 'all features enabled' case. |
| 3200 | if (!Condition) { |
| 3201 | CreateMultiVersionResolverReturn(CGM, Resolver, Builder, FuncToReturn: RO.Function, |
| 3202 | SupportsIFunc); |
| 3203 | return; |
| 3204 | } |
| 3205 | |
| 3206 | if (!AArch64CpuInitialized) { |
| 3207 | Builder.SetInsertPoint(TheBB: CurBlock, IP: CurBlock->begin()); |
| 3208 | EmitAArch64CpuInit(); |
| 3209 | AArch64CpuInitialized = true; |
| 3210 | Builder.SetInsertPoint(CurBlock); |
| 3211 | } |
| 3212 | |
| 3213 | // Skip unreachable versions. |
| 3214 | if (RO.Function == nullptr) |
| 3215 | continue; |
| 3216 | |
| 3217 | llvm::BasicBlock *RetBlock = createBasicBlock(name: "resolver_return" , parent: Resolver); |
| 3218 | CGBuilderTy RetBuilder(CGM, RetBlock); |
| 3219 | CreateMultiVersionResolverReturn(CGM, Resolver, Builder&: RetBuilder, FuncToReturn: RO.Function, |
| 3220 | SupportsIFunc); |
| 3221 | CurBlock = createBasicBlock(name: "resolver_else" , parent: Resolver); |
| 3222 | Builder.CreateCondBr(Cond: Condition, True: RetBlock, False: CurBlock); |
| 3223 | } |
| 3224 | |
| 3225 | // If no default, emit an unreachable. |
| 3226 | Builder.SetInsertPoint(CurBlock); |
| 3227 | llvm::CallInst *TrapCall = EmitTrapCall(IntrID: llvm::Intrinsic::trap); |
| 3228 | TrapCall->setDoesNotReturn(); |
| 3229 | TrapCall->setDoesNotThrow(); |
| 3230 | Builder.CreateUnreachable(); |
| 3231 | Builder.ClearInsertionPoint(); |
| 3232 | } |
| 3233 | |
| 3234 | void CodeGenFunction::EmitX86MultiVersionResolver( |
| 3235 | llvm::Function *Resolver, ArrayRef<FMVResolverOption> Options) { |
| 3236 | |
| 3237 | bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc(); |
| 3238 | |
| 3239 | // Main function's basic block. |
| 3240 | llvm::BasicBlock *CurBlock = createBasicBlock(name: "resolver_entry" , parent: Resolver); |
| 3241 | Builder.SetInsertPoint(CurBlock); |
| 3242 | EmitX86CpuInit(); |
| 3243 | |
| 3244 | for (const FMVResolverOption &RO : Options) { |
| 3245 | Builder.SetInsertPoint(CurBlock); |
| 3246 | llvm::Value *Condition = FormX86ResolverCondition(RO); |
| 3247 | |
| 3248 | // The 'default' or 'generic' case. |
| 3249 | if (!Condition) { |
| 3250 | assert(&RO == Options.end() - 1 && |
| 3251 | "Default or Generic case must be last" ); |
| 3252 | CreateMultiVersionResolverReturn(CGM, Resolver, Builder, FuncToReturn: RO.Function, |
| 3253 | SupportsIFunc); |
| 3254 | return; |
| 3255 | } |
| 3256 | |
| 3257 | llvm::BasicBlock *RetBlock = createBasicBlock(name: "resolver_return" , parent: Resolver); |
| 3258 | CGBuilderTy RetBuilder(CGM, RetBlock); |
| 3259 | CreateMultiVersionResolverReturn(CGM, Resolver, Builder&: RetBuilder, FuncToReturn: RO.Function, |
| 3260 | SupportsIFunc); |
| 3261 | CurBlock = createBasicBlock(name: "resolver_else" , parent: Resolver); |
| 3262 | Builder.CreateCondBr(Cond: Condition, True: RetBlock, False: CurBlock); |
| 3263 | } |
| 3264 | |
| 3265 | // If no generic/default, emit an unreachable. |
| 3266 | Builder.SetInsertPoint(CurBlock); |
| 3267 | llvm::CallInst *TrapCall = EmitTrapCall(IntrID: llvm::Intrinsic::trap); |
| 3268 | TrapCall->setDoesNotReturn(); |
| 3269 | TrapCall->setDoesNotThrow(); |
| 3270 | Builder.CreateUnreachable(); |
| 3271 | Builder.ClearInsertionPoint(); |
| 3272 | } |
| 3273 | |
| 3274 | // Loc - where the diagnostic will point, where in the source code this |
| 3275 | // alignment has failed. |
| 3276 | // SecondaryLoc - if present (will be present if sufficiently different from |
| 3277 | // Loc), the diagnostic will additionally point a "Note:" to this location. |
| 3278 | // It should be the location where the __attribute__((assume_aligned)) |
| 3279 | // was written e.g. |
| 3280 | void CodeGenFunction::emitAlignmentAssumptionCheck( |
| 3281 | llvm::Value *Ptr, QualType Ty, SourceLocation Loc, |
| 3282 | SourceLocation SecondaryLoc, llvm::Value *Alignment, |
| 3283 | llvm::Value *OffsetValue, llvm::Value *TheCheck, |
| 3284 | llvm::Instruction *Assumption) { |
| 3285 | assert(isa_and_nonnull<llvm::CallInst>(Assumption) && |
| 3286 | cast<llvm::CallInst>(Assumption)->getCalledOperand() == |
| 3287 | llvm::Intrinsic::getOrInsertDeclaration( |
| 3288 | Builder.GetInsertBlock()->getParent()->getParent(), |
| 3289 | llvm::Intrinsic::assume) && |
| 3290 | "Assumption should be a call to llvm.assume()." ); |
| 3291 | assert(&(Builder.GetInsertBlock()->back()) == Assumption && |
| 3292 | "Assumption should be the last instruction of the basic block, " |
| 3293 | "since the basic block is still being generated." ); |
| 3294 | |
| 3295 | if (!SanOpts.has(K: SanitizerKind::Alignment)) |
| 3296 | return; |
| 3297 | |
| 3298 | // Don't check pointers to volatile data. The behavior here is implementation- |
| 3299 | // defined. |
| 3300 | if (Ty->getPointeeType().isVolatileQualified()) |
| 3301 | return; |
| 3302 | |
| 3303 | // We need to temorairly remove the assumption so we can insert the |
| 3304 | // sanitizer check before it, else the check will be dropped by optimizations. |
| 3305 | Assumption->removeFromParent(); |
| 3306 | |
| 3307 | { |
| 3308 | auto CheckOrdinal = SanitizerKind::SO_Alignment; |
| 3309 | auto CheckHandler = SanitizerHandler::AlignmentAssumption; |
| 3310 | SanitizerDebugLocation SanScope(this, {CheckOrdinal}, CheckHandler); |
| 3311 | |
| 3312 | if (!OffsetValue) |
| 3313 | OffsetValue = Builder.getInt1(V: false); // no offset. |
| 3314 | |
| 3315 | llvm::Constant *StaticData[] = {EmitCheckSourceLocation(Loc), |
| 3316 | EmitCheckSourceLocation(Loc: SecondaryLoc), |
| 3317 | EmitCheckTypeDescriptor(T: Ty)}; |
| 3318 | llvm::Value *DynamicData[] = {Ptr, Alignment, OffsetValue}; |
| 3319 | EmitCheck(Checked: {std::make_pair(x&: TheCheck, y&: CheckOrdinal)}, Check: CheckHandler, |
| 3320 | StaticArgs: StaticData, DynamicArgs: DynamicData); |
| 3321 | } |
| 3322 | |
| 3323 | // We are now in the (new, empty) "cont" basic block. |
| 3324 | // Reintroduce the assumption. |
| 3325 | Builder.Insert(I: Assumption); |
| 3326 | // FIXME: Assumption still has it's original basic block as it's Parent. |
| 3327 | } |
| 3328 | |
| 3329 | llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) { |
| 3330 | if (CGDebugInfo *DI = getDebugInfo()) |
| 3331 | return DI->SourceLocToDebugLoc(Loc: Location); |
| 3332 | |
| 3333 | return llvm::DebugLoc(); |
| 3334 | } |
| 3335 | |
| 3336 | llvm::Value * |
| 3337 | CodeGenFunction::emitCondLikelihoodViaExpectIntrinsic(llvm::Value *Cond, |
| 3338 | Stmt::Likelihood LH) { |
| 3339 | switch (LH) { |
| 3340 | case Stmt::LH_None: |
| 3341 | return Cond; |
| 3342 | case Stmt::LH_Likely: |
| 3343 | case Stmt::LH_Unlikely: |
| 3344 | // Don't generate llvm.expect on -O0 as the backend won't use it for |
| 3345 | // anything. |
| 3346 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
| 3347 | return Cond; |
| 3348 | llvm::Type *CondTy = Cond->getType(); |
| 3349 | assert(CondTy->isIntegerTy(1) && "expecting condition to be a boolean" ); |
| 3350 | llvm::Function *FnExpect = |
| 3351 | CGM.getIntrinsic(IID: llvm::Intrinsic::expect, Tys: CondTy); |
| 3352 | llvm::Value *ExpectedValueOfCond = |
| 3353 | llvm::ConstantInt::getBool(Ty: CondTy, V: LH == Stmt::LH_Likely); |
| 3354 | return Builder.CreateCall(Callee: FnExpect, Args: {Cond, ExpectedValueOfCond}, |
| 3355 | Name: Cond->getName() + ".expval" ); |
| 3356 | } |
| 3357 | llvm_unreachable("Unknown Likelihood" ); |
| 3358 | } |
| 3359 | |
| 3360 | llvm::Value *CodeGenFunction::emitBoolVecConversion(llvm::Value *SrcVec, |
| 3361 | unsigned NumElementsDst, |
| 3362 | const llvm::Twine &Name) { |
| 3363 | auto *SrcTy = cast<llvm::FixedVectorType>(Val: SrcVec->getType()); |
| 3364 | unsigned NumElementsSrc = SrcTy->getNumElements(); |
| 3365 | if (NumElementsSrc == NumElementsDst) |
| 3366 | return SrcVec; |
| 3367 | |
| 3368 | std::vector<int> ShuffleMask(NumElementsDst, -1); |
| 3369 | for (unsigned MaskIdx = 0; |
| 3370 | MaskIdx < std::min<>(a: NumElementsDst, b: NumElementsSrc); ++MaskIdx) |
| 3371 | ShuffleMask[MaskIdx] = MaskIdx; |
| 3372 | |
| 3373 | return Builder.CreateShuffleVector(V: SrcVec, Mask: ShuffleMask, Name); |
| 3374 | } |
| 3375 | |
| 3376 | void CodeGenFunction::EmitPointerAuthOperandBundle( |
| 3377 | const CGPointerAuthInfo &PointerAuth, |
| 3378 | SmallVectorImpl<llvm::OperandBundleDef> &Bundles) { |
| 3379 | if (!PointerAuth.isSigned()) |
| 3380 | return; |
| 3381 | |
| 3382 | auto *Key = Builder.getInt32(C: PointerAuth.getKey()); |
| 3383 | |
| 3384 | llvm::Value *Discriminator = PointerAuth.getDiscriminator(); |
| 3385 | if (!Discriminator) |
| 3386 | Discriminator = Builder.getSize(N: 0); |
| 3387 | |
| 3388 | llvm::Value *Args[] = {Key, Discriminator}; |
| 3389 | Bundles.emplace_back(Args: "ptrauth" , Args); |
| 3390 | } |
| 3391 | |
| 3392 | static llvm::Value *EmitPointerAuthCommon(CodeGenFunction &CGF, |
| 3393 | const CGPointerAuthInfo &PointerAuth, |
| 3394 | llvm::Value *Pointer, |
| 3395 | unsigned IntrinsicID) { |
| 3396 | if (!PointerAuth) |
| 3397 | return Pointer; |
| 3398 | |
| 3399 | auto Key = CGF.Builder.getInt32(C: PointerAuth.getKey()); |
| 3400 | |
| 3401 | llvm::Value *Discriminator = PointerAuth.getDiscriminator(); |
| 3402 | if (!Discriminator) { |
| 3403 | Discriminator = CGF.Builder.getSize(N: 0); |
| 3404 | } |
| 3405 | |
| 3406 | // Convert the pointer to intptr_t before signing it. |
| 3407 | auto OrigType = Pointer->getType(); |
| 3408 | Pointer = CGF.Builder.CreatePtrToInt(V: Pointer, DestTy: CGF.IntPtrTy); |
| 3409 | |
| 3410 | // call i64 @llvm.ptrauth.sign.i64(i64 %pointer, i32 %key, i64 %discriminator) |
| 3411 | auto Intrinsic = CGF.CGM.getIntrinsic(IID: IntrinsicID); |
| 3412 | Pointer = CGF.EmitRuntimeCall(callee: Intrinsic, args: {Pointer, Key, Discriminator}); |
| 3413 | |
| 3414 | // Convert back to the original type. |
| 3415 | Pointer = CGF.Builder.CreateIntToPtr(V: Pointer, DestTy: OrigType); |
| 3416 | return Pointer; |
| 3417 | } |
| 3418 | |
| 3419 | llvm::Value * |
| 3420 | CodeGenFunction::EmitPointerAuthSign(const CGPointerAuthInfo &PointerAuth, |
| 3421 | llvm::Value *Pointer) { |
| 3422 | if (!PointerAuth.shouldSign()) |
| 3423 | return Pointer; |
| 3424 | return EmitPointerAuthCommon(CGF&: *this, PointerAuth, Pointer, |
| 3425 | IntrinsicID: llvm::Intrinsic::ptrauth_sign); |
| 3426 | } |
| 3427 | |
| 3428 | static llvm::Value *EmitStrip(CodeGenFunction &CGF, |
| 3429 | const CGPointerAuthInfo &PointerAuth, |
| 3430 | llvm::Value *Pointer) { |
| 3431 | auto StripIntrinsic = CGF.CGM.getIntrinsic(IID: llvm::Intrinsic::ptrauth_strip); |
| 3432 | |
| 3433 | auto Key = CGF.Builder.getInt32(C: PointerAuth.getKey()); |
| 3434 | // Convert the pointer to intptr_t before signing it. |
| 3435 | auto OrigType = Pointer->getType(); |
| 3436 | Pointer = CGF.EmitRuntimeCall( |
| 3437 | callee: StripIntrinsic, args: {CGF.Builder.CreatePtrToInt(V: Pointer, DestTy: CGF.IntPtrTy), Key}); |
| 3438 | return CGF.Builder.CreateIntToPtr(V: Pointer, DestTy: OrigType); |
| 3439 | } |
| 3440 | |
| 3441 | llvm::Value * |
| 3442 | CodeGenFunction::EmitPointerAuthAuth(const CGPointerAuthInfo &PointerAuth, |
| 3443 | llvm::Value *Pointer) { |
| 3444 | if (PointerAuth.shouldStrip()) { |
| 3445 | return EmitStrip(CGF&: *this, PointerAuth, Pointer); |
| 3446 | } |
| 3447 | if (!PointerAuth.shouldAuth()) { |
| 3448 | return Pointer; |
| 3449 | } |
| 3450 | |
| 3451 | return EmitPointerAuthCommon(CGF&: *this, PointerAuth, Pointer, |
| 3452 | IntrinsicID: llvm::Intrinsic::ptrauth_auth); |
| 3453 | } |
| 3454 | |
| 3455 | void CodeGenFunction::addInstToCurrentSourceAtom( |
| 3456 | llvm::Instruction *KeyInstruction, llvm::Value *Backup) { |
| 3457 | if (CGDebugInfo *DI = getDebugInfo()) |
| 3458 | DI->addInstToCurrentSourceAtom(KeyInstruction, Backup); |
| 3459 | } |
| 3460 | |
| 3461 | void CodeGenFunction::addInstToSpecificSourceAtom( |
| 3462 | llvm::Instruction *KeyInstruction, llvm::Value *Backup, uint64_t Atom) { |
| 3463 | if (CGDebugInfo *DI = getDebugInfo()) |
| 3464 | DI->addInstToSpecificSourceAtom(KeyInstruction, Backup, Atom); |
| 3465 | } |
| 3466 | |
| 3467 | void CodeGenFunction::addInstToNewSourceAtom(llvm::Instruction *KeyInstruction, |
| 3468 | llvm::Value *Backup) { |
| 3469 | if (CGDebugInfo *DI = getDebugInfo()) { |
| 3470 | ApplyAtomGroup Grp(getDebugInfo()); |
| 3471 | DI->addInstToCurrentSourceAtom(KeyInstruction, Backup); |
| 3472 | } |
| 3473 | } |
| 3474 | |
| 3475 | void CodeGenFunction::emitPFPPostCopyUpdates(Address DestPtr, Address SrcPtr, |
| 3476 | QualType Ty) { |
| 3477 | for (auto &Field : getContext().findPFPFields(Ty)) { |
| 3478 | if (getContext().arePFPFieldsTriviallyCopyable(RD: Field.Field->getParent())) |
| 3479 | continue; |
| 3480 | auto DestFieldPtr = EmitAddressOfPFPField(RecordPtr: DestPtr, Field); |
| 3481 | auto SrcFieldPtr = EmitAddressOfPFPField(RecordPtr: SrcPtr, Field); |
| 3482 | Builder.CreateStore(Val: Builder.CreateLoad(Addr: SrcFieldPtr), Addr: DestFieldPtr); |
| 3483 | } |
| 3484 | } |
| 3485 | |