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