1 //===-- Lint.cpp - Check for common errors in LLVM IR ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass statically checks for common and easily-identified constructs
11 // which produce undefined or likely unintended behavior in LLVM IR.
13 // It is not a guarantee of correctness, in two ways. First, it isn't
14 // comprehensive. There are checks which could be done statically which are
15 // not yet implemented. Some of these are indicated by TODO comments, but
16 // those aren't comprehensive either. Second, many conditions cannot be
17 // checked statically. This pass does no dynamic instrumentation, so it
18 // can't check for all possible problems.
20 // Another limitation is that it assumes all code will be executed. A store
21 // through a null pointer in a basic block which is never reached is harmless,
22 // but this pass will warn about it anyway. This is the main reason why most
23 // of these checks live here instead of in the Verifier pass.
25 // Optimization passes may make conditions that this pass checks for more or
26 // less obvious. If an optimization pass appears to be introducing a warning,
27 // it may be that the optimization pass is merely exposing an existing
28 // condition in the code.
30 // This code may be run before instcombine. In many cases, instcombine checks
31 // for the same kinds of things and turns instructions with undefined behavior
32 // into unreachable (or equivalent). Because of this, this pass makes some
33 // effort to look through bitcasts and so on.
35 //===----------------------------------------------------------------------===//
37 #include "llvm/Analysis/Lint.h"
38 #include "llvm/ADT/STLExtras.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/Analysis/AliasAnalysis.h"
41 #include "llvm/Analysis/AssumptionCache.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/InstructionSimplify.h"
44 #include "llvm/Analysis/Loads.h"
45 #include "llvm/Analysis/Passes.h"
46 #include "llvm/Analysis/TargetLibraryInfo.h"
47 #include "llvm/Analysis/ValueTracking.h"
48 #include "llvm/IR/CallSite.h"
49 #include "llvm/IR/DataLayout.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/InstVisitor.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/LegacyPassManager.h"
55 #include "llvm/Pass.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
62 static const unsigned Read = 1;
63 static const unsigned Write = 2;
64 static const unsigned Callee = 4;
65 static const unsigned Branchee = 8;
68 class Lint : public FunctionPass, public InstVisitor<Lint> {
69 friend class InstVisitor<Lint>;
71 void visitFunction(Function &F);
73 void visitCallSite(CallSite CS);
74 void visitMemoryReference(Instruction &I, Value *Ptr,
75 uint64_t Size, unsigned Align,
76 Type *Ty, unsigned Flags);
77 void visitEHBeginCatch(IntrinsicInst *II);
78 void visitEHEndCatch(IntrinsicInst *II);
80 void visitCallInst(CallInst &I);
81 void visitInvokeInst(InvokeInst &I);
82 void visitReturnInst(ReturnInst &I);
83 void visitLoadInst(LoadInst &I);
84 void visitStoreInst(StoreInst &I);
85 void visitXor(BinaryOperator &I);
86 void visitSub(BinaryOperator &I);
87 void visitLShr(BinaryOperator &I);
88 void visitAShr(BinaryOperator &I);
89 void visitShl(BinaryOperator &I);
90 void visitSDiv(BinaryOperator &I);
91 void visitUDiv(BinaryOperator &I);
92 void visitSRem(BinaryOperator &I);
93 void visitURem(BinaryOperator &I);
94 void visitAllocaInst(AllocaInst &I);
95 void visitVAArgInst(VAArgInst &I);
96 void visitIndirectBrInst(IndirectBrInst &I);
97 void visitExtractElementInst(ExtractElementInst &I);
98 void visitInsertElementInst(InsertElementInst &I);
99 void visitUnreachableInst(UnreachableInst &I);
101 Value *findValue(Value *V, const DataLayout &DL, bool OffsetOk) const;
102 Value *findValueImpl(Value *V, const DataLayout &DL, bool OffsetOk,
103 SmallPtrSetImpl<Value *> &Visited) const;
110 TargetLibraryInfo *TLI;
112 std::string Messages;
113 raw_string_ostream MessagesStr;
115 static char ID; // Pass identification, replacement for typeid
116 Lint() : FunctionPass(ID), MessagesStr(Messages) {
117 initializeLintPass(*PassRegistry::getPassRegistry());
120 bool runOnFunction(Function &F) override;
122 void getAnalysisUsage(AnalysisUsage &AU) const override {
123 AU.setPreservesAll();
124 AU.addRequired<AliasAnalysis>();
125 AU.addRequired<AssumptionCacheTracker>();
126 AU.addRequired<TargetLibraryInfoWrapperPass>();
127 AU.addRequired<DominatorTreeWrapperPass>();
129 void print(raw_ostream &O, const Module *M) const override {}
131 void WriteValues(ArrayRef<const Value *> Vs) {
132 for (const Value *V : Vs) {
135 if (isa<Instruction>(V)) {
136 MessagesStr << *V << '\n';
138 V->printAsOperand(MessagesStr, true, Mod);
144 // CheckFailed - A check failed, so print out the condition and the message
145 // that failed. This provides a nice place to put a breakpoint if you want
146 // to see why something is not correct.
147 template <typename... Ts>
148 void CheckFailed(const Twine &Message, const Ts &...Vs) {
149 MessagesStr << Message << '\n';
150 WriteValues({Vs...});
156 INITIALIZE_PASS_BEGIN(Lint, "lint", "Statically lint-checks LLVM IR",
158 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
159 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
160 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
161 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
162 INITIALIZE_PASS_END(Lint, "lint", "Statically lint-checks LLVM IR",
165 // Assert - We know that cond should be true, if not print an error message.
166 #define Assert(C, ...) \
167 do { if (!(C)) { CheckFailed(__VA_ARGS__); return; } } while (0)
169 // Lint::run - This is the main Analysis entry point for a
172 bool Lint::runOnFunction(Function &F) {
174 AA = &getAnalysis<AliasAnalysis>();
175 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
176 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
177 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
179 dbgs() << MessagesStr.str();
184 void Lint::visitFunction(Function &F) {
185 // This isn't undefined behavior, it's just a little unusual, and it's a
186 // fairly common mistake to neglect to name a function.
187 Assert(F.hasName() || F.hasLocalLinkage(),
188 "Unusual: Unnamed function with non-local linkage", &F);
190 // TODO: Check for irreducible control flow.
193 void Lint::visitCallSite(CallSite CS) {
194 Instruction &I = *CS.getInstruction();
195 Value *Callee = CS.getCalledValue();
196 const DataLayout &DL = CS->getModule()->getDataLayout();
198 visitMemoryReference(I, Callee, AliasAnalysis::UnknownSize,
199 0, nullptr, MemRef::Callee);
201 if (Function *F = dyn_cast<Function>(findValue(Callee, DL,
202 /*OffsetOk=*/false))) {
203 Assert(CS.getCallingConv() == F->getCallingConv(),
204 "Undefined behavior: Caller and callee calling convention differ",
207 FunctionType *FT = F->getFunctionType();
208 unsigned NumActualArgs = CS.arg_size();
210 Assert(FT->isVarArg() ? FT->getNumParams() <= NumActualArgs
211 : FT->getNumParams() == NumActualArgs,
212 "Undefined behavior: Call argument count mismatches callee "
216 Assert(FT->getReturnType() == I.getType(),
217 "Undefined behavior: Call return type mismatches "
218 "callee return type",
221 // Check argument types (in case the callee was casted) and attributes.
222 // TODO: Verify that caller and callee attributes are compatible.
223 Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
224 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
225 for (; AI != AE; ++AI) {
228 Argument *Formal = PI++;
229 Assert(Formal->getType() == Actual->getType(),
230 "Undefined behavior: Call argument type mismatches "
231 "callee parameter type",
234 // Check that noalias arguments don't alias other arguments. This is
235 // not fully precise because we don't know the sizes of the dereferenced
237 if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
238 for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
239 if (AI != BI && (*BI)->getType()->isPointerTy()) {
240 AliasAnalysis::AliasResult Result = AA->alias(*AI, *BI);
241 Assert(Result != AliasAnalysis::MustAlias &&
242 Result != AliasAnalysis::PartialAlias,
243 "Unusual: noalias argument aliases another argument", &I);
246 // Check that an sret argument points to valid memory.
247 if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
249 cast<PointerType>(Formal->getType())->getElementType();
250 visitMemoryReference(I, Actual, AA->getTypeStoreSize(Ty),
251 DL.getABITypeAlignment(Ty), Ty,
252 MemRef::Read | MemRef::Write);
258 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
259 for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
261 Value *Obj = findValue(*AI, DL, /*OffsetOk=*/true);
262 Assert(!isa<AllocaInst>(Obj),
263 "Undefined behavior: Call with \"tail\" keyword references "
269 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
270 switch (II->getIntrinsicID()) {
273 // TODO: Check more intrinsics
275 case Intrinsic::memcpy: {
276 MemCpyInst *MCI = cast<MemCpyInst>(&I);
277 // TODO: If the size is known, use it.
278 visitMemoryReference(I, MCI->getDest(), AliasAnalysis::UnknownSize,
279 MCI->getAlignment(), nullptr,
281 visitMemoryReference(I, MCI->getSource(), AliasAnalysis::UnknownSize,
282 MCI->getAlignment(), nullptr,
285 // Check that the memcpy arguments don't overlap. The AliasAnalysis API
286 // isn't expressive enough for what we really want to do. Known partial
287 // overlap is not distinguished from the case where nothing is known.
289 if (const ConstantInt *Len =
290 dyn_cast<ConstantInt>(findValue(MCI->getLength(), DL,
291 /*OffsetOk=*/false)))
292 if (Len->getValue().isIntN(32))
293 Size = Len->getValue().getZExtValue();
294 Assert(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
295 AliasAnalysis::MustAlias,
296 "Undefined behavior: memcpy source and destination overlap", &I);
299 case Intrinsic::memmove: {
300 MemMoveInst *MMI = cast<MemMoveInst>(&I);
301 // TODO: If the size is known, use it.
302 visitMemoryReference(I, MMI->getDest(), AliasAnalysis::UnknownSize,
303 MMI->getAlignment(), nullptr,
305 visitMemoryReference(I, MMI->getSource(), AliasAnalysis::UnknownSize,
306 MMI->getAlignment(), nullptr,
310 case Intrinsic::memset: {
311 MemSetInst *MSI = cast<MemSetInst>(&I);
312 // TODO: If the size is known, use it.
313 visitMemoryReference(I, MSI->getDest(), AliasAnalysis::UnknownSize,
314 MSI->getAlignment(), nullptr,
319 case Intrinsic::vastart:
320 Assert(I.getParent()->getParent()->isVarArg(),
321 "Undefined behavior: va_start called in a non-varargs function",
324 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
325 0, nullptr, MemRef::Read | MemRef::Write);
327 case Intrinsic::vacopy:
328 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
329 0, nullptr, MemRef::Write);
330 visitMemoryReference(I, CS.getArgument(1), AliasAnalysis::UnknownSize,
331 0, nullptr, MemRef::Read);
333 case Intrinsic::vaend:
334 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
335 0, nullptr, MemRef::Read | MemRef::Write);
338 case Intrinsic::stackrestore:
339 // Stackrestore doesn't read or write memory, but it sets the
340 // stack pointer, which the compiler may read from or write to
341 // at any time, so check it for both readability and writeability.
342 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
343 0, nullptr, MemRef::Read | MemRef::Write);
346 case Intrinsic::eh_begincatch:
347 visitEHBeginCatch(II);
349 case Intrinsic::eh_endcatch:
355 void Lint::visitCallInst(CallInst &I) {
356 return visitCallSite(&I);
359 void Lint::visitInvokeInst(InvokeInst &I) {
360 return visitCallSite(&I);
363 void Lint::visitReturnInst(ReturnInst &I) {
364 Function *F = I.getParent()->getParent();
365 Assert(!F->doesNotReturn(),
366 "Unusual: Return statement in function with noreturn attribute", &I);
368 if (Value *V = I.getReturnValue()) {
370 findValue(V, F->getParent()->getDataLayout(), /*OffsetOk=*/true);
371 Assert(!isa<AllocaInst>(Obj), "Unusual: Returning alloca value", &I);
375 // TODO: Check that the reference is in bounds.
376 // TODO: Check readnone/readonly function attributes.
377 void Lint::visitMemoryReference(Instruction &I,
378 Value *Ptr, uint64_t Size, unsigned Align,
379 Type *Ty, unsigned Flags) {
380 // If no memory is being referenced, it doesn't matter if the pointer
385 Value *UnderlyingObject =
386 findValue(Ptr, I.getModule()->getDataLayout(), /*OffsetOk=*/true);
387 Assert(!isa<ConstantPointerNull>(UnderlyingObject),
388 "Undefined behavior: Null pointer dereference", &I);
389 Assert(!isa<UndefValue>(UnderlyingObject),
390 "Undefined behavior: Undef pointer dereference", &I);
391 Assert(!isa<ConstantInt>(UnderlyingObject) ||
392 !cast<ConstantInt>(UnderlyingObject)->isAllOnesValue(),
393 "Unusual: All-ones pointer dereference", &I);
394 Assert(!isa<ConstantInt>(UnderlyingObject) ||
395 !cast<ConstantInt>(UnderlyingObject)->isOne(),
396 "Unusual: Address one pointer dereference", &I);
398 if (Flags & MemRef::Write) {
399 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(UnderlyingObject))
400 Assert(!GV->isConstant(), "Undefined behavior: Write to read-only memory",
402 Assert(!isa<Function>(UnderlyingObject) &&
403 !isa<BlockAddress>(UnderlyingObject),
404 "Undefined behavior: Write to text section", &I);
406 if (Flags & MemRef::Read) {
407 Assert(!isa<Function>(UnderlyingObject), "Unusual: Load from function body",
409 Assert(!isa<BlockAddress>(UnderlyingObject),
410 "Undefined behavior: Load from block address", &I);
412 if (Flags & MemRef::Callee) {
413 Assert(!isa<BlockAddress>(UnderlyingObject),
414 "Undefined behavior: Call to block address", &I);
416 if (Flags & MemRef::Branchee) {
417 Assert(!isa<Constant>(UnderlyingObject) ||
418 isa<BlockAddress>(UnderlyingObject),
419 "Undefined behavior: Branch to non-blockaddress", &I);
422 // Check for buffer overflows and misalignment.
423 // Only handles memory references that read/write something simple like an
424 // alloca instruction or a global variable.
425 auto &DL = I.getModule()->getDataLayout();
427 if (Value *Base = GetPointerBaseWithConstantOffset(Ptr, Offset, DL)) {
428 // OK, so the access is to a constant offset from Ptr. Check that Ptr is
429 // something we can handle and if so extract the size of this base object
430 // along with its alignment.
431 uint64_t BaseSize = AliasAnalysis::UnknownSize;
432 unsigned BaseAlign = 0;
434 if (AllocaInst *AI = dyn_cast<AllocaInst>(Base)) {
435 Type *ATy = AI->getAllocatedType();
436 if (!AI->isArrayAllocation() && ATy->isSized())
437 BaseSize = DL.getTypeAllocSize(ATy);
438 BaseAlign = AI->getAlignment();
439 if (BaseAlign == 0 && ATy->isSized())
440 BaseAlign = DL.getABITypeAlignment(ATy);
441 } else if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) {
442 // If the global may be defined differently in another compilation unit
443 // then don't warn about funky memory accesses.
444 if (GV->hasDefinitiveInitializer()) {
445 Type *GTy = GV->getType()->getElementType();
447 BaseSize = DL.getTypeAllocSize(GTy);
448 BaseAlign = GV->getAlignment();
449 if (BaseAlign == 0 && GTy->isSized())
450 BaseAlign = DL.getABITypeAlignment(GTy);
454 // Accesses from before the start or after the end of the object are not
456 Assert(Size == AliasAnalysis::UnknownSize ||
457 BaseSize == AliasAnalysis::UnknownSize ||
458 (Offset >= 0 && Offset + Size <= BaseSize),
459 "Undefined behavior: Buffer overflow", &I);
461 // Accesses that say that the memory is more aligned than it is are not
463 if (Align == 0 && Ty && Ty->isSized())
464 Align = DL.getABITypeAlignment(Ty);
465 Assert(!BaseAlign || Align <= MinAlign(BaseAlign, Offset),
466 "Undefined behavior: Memory reference address is misaligned", &I);
470 void Lint::visitLoadInst(LoadInst &I) {
471 visitMemoryReference(I, I.getPointerOperand(),
472 AA->getTypeStoreSize(I.getType()), I.getAlignment(),
473 I.getType(), MemRef::Read);
476 void Lint::visitStoreInst(StoreInst &I) {
477 visitMemoryReference(I, I.getPointerOperand(),
478 AA->getTypeStoreSize(I.getOperand(0)->getType()),
480 I.getOperand(0)->getType(), MemRef::Write);
483 void Lint::visitXor(BinaryOperator &I) {
484 Assert(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
485 "Undefined result: xor(undef, undef)", &I);
488 void Lint::visitSub(BinaryOperator &I) {
489 Assert(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
490 "Undefined result: sub(undef, undef)", &I);
493 void Lint::visitLShr(BinaryOperator &I) {
494 if (ConstantInt *CI = dyn_cast<ConstantInt>(
495 findValue(I.getOperand(1), I.getModule()->getDataLayout(),
496 /*OffsetOk=*/false)))
497 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
498 "Undefined result: Shift count out of range", &I);
501 void Lint::visitAShr(BinaryOperator &I) {
502 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(
503 I.getOperand(1), I.getModule()->getDataLayout(), /*OffsetOk=*/false)))
504 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
505 "Undefined result: Shift count out of range", &I);
508 void Lint::visitShl(BinaryOperator &I) {
509 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(
510 I.getOperand(1), I.getModule()->getDataLayout(), /*OffsetOk=*/false)))
511 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
512 "Undefined result: Shift count out of range", &I);
516 allPredsCameFromLandingPad(BasicBlock *BB,
517 SmallSet<BasicBlock *, 4> &VisitedBlocks) {
518 VisitedBlocks.insert(BB);
519 if (BB->isLandingPad())
521 // If we find a block with no predecessors, the search failed.
524 for (BasicBlock *Pred : predecessors(BB)) {
525 if (VisitedBlocks.count(Pred))
527 if (!allPredsCameFromLandingPad(Pred, VisitedBlocks))
534 allSuccessorsReachEndCatch(BasicBlock *BB, BasicBlock::iterator InstBegin,
535 IntrinsicInst **SecondBeginCatch,
536 SmallSet<BasicBlock *, 4> &VisitedBlocks) {
537 VisitedBlocks.insert(BB);
538 for (BasicBlock::iterator I = InstBegin, E = BB->end(); I != E; ++I) {
539 IntrinsicInst *IC = dyn_cast<IntrinsicInst>(I);
540 if (IC && IC->getIntrinsicID() == Intrinsic::eh_endcatch)
542 // If we find another begincatch while looking for an endcatch,
543 // that's also an error.
544 if (IC && IC->getIntrinsicID() == Intrinsic::eh_begincatch) {
545 *SecondBeginCatch = IC;
550 // If we reach a block with no successors while searching, the
551 // search has failed.
554 // Otherwise, search all of the successors.
555 for (BasicBlock *Succ : successors(BB)) {
556 if (VisitedBlocks.count(Succ))
558 if (!allSuccessorsReachEndCatch(Succ, Succ->begin(), SecondBeginCatch,
565 void Lint::visitEHBeginCatch(IntrinsicInst *II) {
566 // The checks in this function make a potentially dubious assumption about
567 // the CFG, namely that any block involved in a catch is only used for the
568 // catch. This will very likely be true of IR generated by a front end,
569 // but it may cease to be true, for example, if the IR is run through a
570 // pass which combines similar blocks.
572 // In general, if we encounter a block the isn't dominated by the catch
573 // block while we are searching the catch block's successors for a call
574 // to end catch intrinsic, then it is possible that it will be legal for
575 // a path through this block to never reach a call to llvm.eh.endcatch.
576 // An analogous statement could be made about our search for a landing
577 // pad among the catch block's predecessors.
579 // What is actually required is that no path is possible at runtime that
580 // reaches a call to llvm.eh.begincatch without having previously visited
581 // a landingpad instruction and that no path is possible at runtime that
582 // calls llvm.eh.begincatch and does not subsequently call llvm.eh.endcatch
583 // (mentally adjusting for the fact that in reality these calls will be
584 // removed before code generation).
586 // Because this is a lint check, we take a pessimistic approach and warn if
587 // the control flow is potentially incorrect.
589 SmallSet<BasicBlock *, 4> VisitedBlocks;
590 BasicBlock *CatchBB = II->getParent();
592 // The begin catch must occur in a landing pad block or all paths
593 // to it must have come from a landing pad.
594 Assert(allPredsCameFromLandingPad(CatchBB, VisitedBlocks),
595 "llvm.eh.begincatch may be reachable without passing a landingpad",
598 // Reset the visited block list.
599 VisitedBlocks.clear();
601 IntrinsicInst *SecondBeginCatch = nullptr;
603 // This has to be called before it is asserted. Otherwise, the first assert
604 // below can never be hit.
605 bool EndCatchFound = allSuccessorsReachEndCatch(
606 CatchBB, std::next(static_cast<BasicBlock::iterator>(II)),
607 &SecondBeginCatch, VisitedBlocks);
609 SecondBeginCatch == nullptr,
610 "llvm.eh.begincatch may be called a second time before llvm.eh.endcatch",
611 II, SecondBeginCatch);
612 Assert(EndCatchFound,
613 "Some paths from llvm.eh.begincatch may not reach llvm.eh.endcatch",
617 static bool allPredCameFromBeginCatch(
618 BasicBlock *BB, BasicBlock::reverse_iterator InstRbegin,
619 IntrinsicInst **SecondEndCatch, SmallSet<BasicBlock *, 4> &VisitedBlocks) {
620 VisitedBlocks.insert(BB);
621 // Look for a begincatch in this block.
622 for (BasicBlock::reverse_iterator RI = InstRbegin, RE = BB->rend(); RI != RE;
624 IntrinsicInst *IC = dyn_cast<IntrinsicInst>(&*RI);
625 if (IC && IC->getIntrinsicID() == Intrinsic::eh_begincatch)
627 // If we find another end catch before we find a begin catch, that's
629 if (IC && IC->getIntrinsicID() == Intrinsic::eh_endcatch) {
630 *SecondEndCatch = IC;
633 // If we encounter a landingpad instruction, the search failed.
634 if (isa<LandingPadInst>(*RI))
637 // If while searching we find a block with no predeccesors,
638 // the search failed.
641 // Search any predecessors we haven't seen before.
642 for (BasicBlock *Pred : predecessors(BB)) {
643 if (VisitedBlocks.count(Pred))
645 if (!allPredCameFromBeginCatch(Pred, Pred->rbegin(), SecondEndCatch,
652 void Lint::visitEHEndCatch(IntrinsicInst *II) {
653 // The check in this function makes a potentially dubious assumption about
654 // the CFG, namely that any block involved in a catch is only used for the
655 // catch. This will very likely be true of IR generated by a front end,
656 // but it may cease to be true, for example, if the IR is run through a
657 // pass which combines similar blocks.
659 // In general, if we encounter a block the isn't post-dominated by the
660 // end catch block while we are searching the end catch block's predecessors
661 // for a call to the begin catch intrinsic, then it is possible that it will
662 // be legal for a path to reach the end catch block without ever having
663 // called llvm.eh.begincatch.
665 // What is actually required is that no path is possible at runtime that
666 // reaches a call to llvm.eh.endcatch without having previously visited
667 // a call to llvm.eh.begincatch (mentally adjusting for the fact that in
668 // reality these calls will be removed before code generation).
670 // Because this is a lint check, we take a pessimistic approach and warn if
671 // the control flow is potentially incorrect.
673 BasicBlock *EndCatchBB = II->getParent();
675 // Alls paths to the end catch call must pass through a begin catch call.
677 // If llvm.eh.begincatch wasn't called in the current block, we'll use this
678 // lambda to recursively look for it in predecessors.
679 SmallSet<BasicBlock *, 4> VisitedBlocks;
680 IntrinsicInst *SecondEndCatch = nullptr;
682 // This has to be called before it is asserted. Otherwise, the first assert
683 // below can never be hit.
684 bool BeginCatchFound =
685 allPredCameFromBeginCatch(EndCatchBB, BasicBlock::reverse_iterator(II),
686 &SecondEndCatch, VisitedBlocks);
688 SecondEndCatch == nullptr,
689 "llvm.eh.endcatch may be called a second time after llvm.eh.begincatch",
691 Assert(BeginCatchFound,
692 "llvm.eh.endcatch may be reachable without passing llvm.eh.begincatch",
696 static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT,
697 AssumptionCache *AC) {
698 // Assume undef could be zero.
699 if (isa<UndefValue>(V))
702 VectorType *VecTy = dyn_cast<VectorType>(V->getType());
704 unsigned BitWidth = V->getType()->getIntegerBitWidth();
705 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
706 computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC,
707 dyn_cast<Instruction>(V), DT);
708 return KnownZero.isAllOnesValue();
711 // Per-component check doesn't work with zeroinitializer
712 Constant *C = dyn_cast<Constant>(V);
716 if (C->isZeroValue())
719 // For a vector, KnownZero will only be true if all values are zero, so check
720 // this per component
721 unsigned BitWidth = VecTy->getElementType()->getIntegerBitWidth();
722 for (unsigned I = 0, N = VecTy->getNumElements(); I != N; ++I) {
723 Constant *Elem = C->getAggregateElement(I);
724 if (isa<UndefValue>(Elem))
727 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
728 computeKnownBits(Elem, KnownZero, KnownOne, DL);
729 if (KnownZero.isAllOnesValue())
736 void Lint::visitSDiv(BinaryOperator &I) {
737 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
738 "Undefined behavior: Division by zero", &I);
741 void Lint::visitUDiv(BinaryOperator &I) {
742 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
743 "Undefined behavior: Division by zero", &I);
746 void Lint::visitSRem(BinaryOperator &I) {
747 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
748 "Undefined behavior: Division by zero", &I);
751 void Lint::visitURem(BinaryOperator &I) {
752 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
753 "Undefined behavior: Division by zero", &I);
756 void Lint::visitAllocaInst(AllocaInst &I) {
757 if (isa<ConstantInt>(I.getArraySize()))
758 // This isn't undefined behavior, it's just an obvious pessimization.
759 Assert(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
760 "Pessimization: Static alloca outside of entry block", &I);
762 // TODO: Check for an unusual size (MSB set?)
765 void Lint::visitVAArgInst(VAArgInst &I) {
766 visitMemoryReference(I, I.getOperand(0), AliasAnalysis::UnknownSize, 0,
767 nullptr, MemRef::Read | MemRef::Write);
770 void Lint::visitIndirectBrInst(IndirectBrInst &I) {
771 visitMemoryReference(I, I.getAddress(), AliasAnalysis::UnknownSize, 0,
772 nullptr, MemRef::Branchee);
774 Assert(I.getNumDestinations() != 0,
775 "Undefined behavior: indirectbr with no destinations", &I);
778 void Lint::visitExtractElementInst(ExtractElementInst &I) {
779 if (ConstantInt *CI = dyn_cast<ConstantInt>(
780 findValue(I.getIndexOperand(), I.getModule()->getDataLayout(),
781 /*OffsetOk=*/false)))
782 Assert(CI->getValue().ult(I.getVectorOperandType()->getNumElements()),
783 "Undefined result: extractelement index out of range", &I);
786 void Lint::visitInsertElementInst(InsertElementInst &I) {
787 if (ConstantInt *CI = dyn_cast<ConstantInt>(
788 findValue(I.getOperand(2), I.getModule()->getDataLayout(),
789 /*OffsetOk=*/false)))
790 Assert(CI->getValue().ult(I.getType()->getNumElements()),
791 "Undefined result: insertelement index out of range", &I);
794 void Lint::visitUnreachableInst(UnreachableInst &I) {
795 // This isn't undefined behavior, it's merely suspicious.
796 Assert(&I == I.getParent()->begin() ||
797 std::prev(BasicBlock::iterator(&I))->mayHaveSideEffects(),
798 "Unusual: unreachable immediately preceded by instruction without "
803 /// findValue - Look through bitcasts and simple memory reference patterns
804 /// to identify an equivalent, but more informative, value. If OffsetOk
805 /// is true, look through getelementptrs with non-zero offsets too.
807 /// Most analysis passes don't require this logic, because instcombine
808 /// will simplify most of these kinds of things away. But it's a goal of
809 /// this Lint pass to be useful even on non-optimized IR.
810 Value *Lint::findValue(Value *V, const DataLayout &DL, bool OffsetOk) const {
811 SmallPtrSet<Value *, 4> Visited;
812 return findValueImpl(V, DL, OffsetOk, Visited);
815 /// findValueImpl - Implementation helper for findValue.
816 Value *Lint::findValueImpl(Value *V, const DataLayout &DL, bool OffsetOk,
817 SmallPtrSetImpl<Value *> &Visited) const {
818 // Detect self-referential values.
819 if (!Visited.insert(V).second)
820 return UndefValue::get(V->getType());
822 // TODO: Look through sext or zext cast, when the result is known to
823 // be interpreted as signed or unsigned, respectively.
824 // TODO: Look through eliminable cast pairs.
825 // TODO: Look through calls with unique return values.
826 // TODO: Look through vector insert/extract/shuffle.
827 V = OffsetOk ? GetUnderlyingObject(V, DL) : V->stripPointerCasts();
828 if (LoadInst *L = dyn_cast<LoadInst>(V)) {
829 BasicBlock::iterator BBI = L;
830 BasicBlock *BB = L->getParent();
831 SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
833 if (!VisitedBlocks.insert(BB).second)
835 if (Value *U = FindAvailableLoadedValue(L->getPointerOperand(),
837 return findValueImpl(U, DL, OffsetOk, Visited);
838 if (BBI != BB->begin()) break;
839 BB = BB->getUniquePredecessor();
843 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
844 if (Value *W = PN->hasConstantValue())
846 return findValueImpl(W, DL, OffsetOk, Visited);
847 } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
848 if (CI->isNoopCast(DL))
849 return findValueImpl(CI->getOperand(0), DL, OffsetOk, Visited);
850 } else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(V)) {
851 if (Value *W = FindInsertedValue(Ex->getAggregateOperand(),
854 return findValueImpl(W, DL, OffsetOk, Visited);
855 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
856 // Same as above, but for ConstantExpr instead of Instruction.
857 if (Instruction::isCast(CE->getOpcode())) {
858 if (CastInst::isNoopCast(Instruction::CastOps(CE->getOpcode()),
859 CE->getOperand(0)->getType(), CE->getType(),
860 DL.getIntPtrType(V->getType())))
861 return findValueImpl(CE->getOperand(0), DL, OffsetOk, Visited);
862 } else if (CE->getOpcode() == Instruction::ExtractValue) {
863 ArrayRef<unsigned> Indices = CE->getIndices();
864 if (Value *W = FindInsertedValue(CE->getOperand(0), Indices))
866 return findValueImpl(W, DL, OffsetOk, Visited);
870 // As a last resort, try SimplifyInstruction or constant folding.
871 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
872 if (Value *W = SimplifyInstruction(Inst, DL, TLI, DT, AC))
873 return findValueImpl(W, DL, OffsetOk, Visited);
874 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
875 if (Value *W = ConstantFoldConstantExpression(CE, DL, TLI))
877 return findValueImpl(W, DL, OffsetOk, Visited);
883 //===----------------------------------------------------------------------===//
884 // Implement the public interfaces to this file...
885 //===----------------------------------------------------------------------===//
887 FunctionPass *llvm::createLintPass() {
891 /// lintFunction - Check a function for errors, printing messages on stderr.
893 void llvm::lintFunction(const Function &f) {
894 Function &F = const_cast<Function&>(f);
895 assert(!F.isDeclaration() && "Cannot lint external functions");
897 legacy::FunctionPassManager FPM(F.getParent());
898 Lint *V = new Lint();
903 /// lintModule - Check a module for errors, printing messages on stderr.
905 void llvm::lintModule(const Module &M) {
906 legacy::PassManager PM;
907 Lint *V = new Lint();
909 PM.run(const_cast<Module&>(M));