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 /// \brief A check failed, so printout out the condition and the message.
146 /// This provides a nice place to put a breakpoint if you want to see why
147 /// something is not correct.
148 void CheckFailed(const Twine &Message) { MessagesStr << Message << '\n'; }
150 /// \brief A check failed (with values to print).
152 /// This calls the Message-only version so that the above is easier to set
154 template <typename T1, typename... Ts>
155 void CheckFailed(const Twine &Message, const T1 &V1, const Ts &...Vs) {
156 CheckFailed(Message);
157 WriteValues({V1, Vs...});
163 INITIALIZE_PASS_BEGIN(Lint, "lint", "Statically lint-checks LLVM IR",
165 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
166 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
167 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
168 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
169 INITIALIZE_PASS_END(Lint, "lint", "Statically lint-checks LLVM IR",
172 // Assert - We know that cond should be true, if not print an error message.
173 #define Assert(C, ...) \
174 do { if (!(C)) { CheckFailed(__VA_ARGS__); return; } } while (0)
176 // Lint::run - This is the main Analysis entry point for a
179 bool Lint::runOnFunction(Function &F) {
181 AA = &getAnalysis<AliasAnalysis>();
182 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
183 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
184 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
186 dbgs() << MessagesStr.str();
191 void Lint::visitFunction(Function &F) {
192 // This isn't undefined behavior, it's just a little unusual, and it's a
193 // fairly common mistake to neglect to name a function.
194 Assert(F.hasName() || F.hasLocalLinkage(),
195 "Unusual: Unnamed function with non-local linkage", &F);
197 // TODO: Check for irreducible control flow.
200 void Lint::visitCallSite(CallSite CS) {
201 Instruction &I = *CS.getInstruction();
202 Value *Callee = CS.getCalledValue();
203 const DataLayout &DL = CS->getModule()->getDataLayout();
205 visitMemoryReference(I, Callee, MemoryLocation::UnknownSize, 0, nullptr,
208 if (Function *F = dyn_cast<Function>(findValue(Callee, DL,
209 /*OffsetOk=*/false))) {
210 Assert(CS.getCallingConv() == F->getCallingConv(),
211 "Undefined behavior: Caller and callee calling convention differ",
214 FunctionType *FT = F->getFunctionType();
215 unsigned NumActualArgs = CS.arg_size();
217 Assert(FT->isVarArg() ? FT->getNumParams() <= NumActualArgs
218 : FT->getNumParams() == NumActualArgs,
219 "Undefined behavior: Call argument count mismatches callee "
223 Assert(FT->getReturnType() == I.getType(),
224 "Undefined behavior: Call return type mismatches "
225 "callee return type",
228 // Check argument types (in case the callee was casted) and attributes.
229 // TODO: Verify that caller and callee attributes are compatible.
230 Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
231 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
232 for (; AI != AE; ++AI) {
235 Argument *Formal = PI++;
236 Assert(Formal->getType() == Actual->getType(),
237 "Undefined behavior: Call argument type mismatches "
238 "callee parameter type",
241 // Check that noalias arguments don't alias other arguments. This is
242 // not fully precise because we don't know the sizes of the dereferenced
244 if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
245 for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
246 if (AI != BI && (*BI)->getType()->isPointerTy()) {
247 AliasAnalysis::AliasResult Result = AA->alias(*AI, *BI);
248 Assert(Result != AliasAnalysis::MustAlias &&
249 Result != AliasAnalysis::PartialAlias,
250 "Unusual: noalias argument aliases another argument", &I);
253 // Check that an sret argument points to valid memory.
254 if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
256 cast<PointerType>(Formal->getType())->getElementType();
257 visitMemoryReference(I, Actual, AA->getTypeStoreSize(Ty),
258 DL.getABITypeAlignment(Ty), Ty,
259 MemRef::Read | MemRef::Write);
265 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
266 for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
268 Value *Obj = findValue(*AI, DL, /*OffsetOk=*/true);
269 Assert(!isa<AllocaInst>(Obj),
270 "Undefined behavior: Call with \"tail\" keyword references "
276 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
277 switch (II->getIntrinsicID()) {
280 // TODO: Check more intrinsics
282 case Intrinsic::memcpy: {
283 MemCpyInst *MCI = cast<MemCpyInst>(&I);
284 // TODO: If the size is known, use it.
285 visitMemoryReference(I, MCI->getDest(), MemoryLocation::UnknownSize,
286 MCI->getAlignment(), nullptr, MemRef::Write);
287 visitMemoryReference(I, MCI->getSource(), MemoryLocation::UnknownSize,
288 MCI->getAlignment(), nullptr, MemRef::Read);
290 // Check that the memcpy arguments don't overlap. The AliasAnalysis API
291 // isn't expressive enough for what we really want to do. Known partial
292 // overlap is not distinguished from the case where nothing is known.
294 if (const ConstantInt *Len =
295 dyn_cast<ConstantInt>(findValue(MCI->getLength(), DL,
296 /*OffsetOk=*/false)))
297 if (Len->getValue().isIntN(32))
298 Size = Len->getValue().getZExtValue();
299 Assert(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
300 AliasAnalysis::MustAlias,
301 "Undefined behavior: memcpy source and destination overlap", &I);
304 case Intrinsic::memmove: {
305 MemMoveInst *MMI = cast<MemMoveInst>(&I);
306 // TODO: If the size is known, use it.
307 visitMemoryReference(I, MMI->getDest(), MemoryLocation::UnknownSize,
308 MMI->getAlignment(), nullptr, MemRef::Write);
309 visitMemoryReference(I, MMI->getSource(), MemoryLocation::UnknownSize,
310 MMI->getAlignment(), nullptr, MemRef::Read);
313 case Intrinsic::memset: {
314 MemSetInst *MSI = cast<MemSetInst>(&I);
315 // TODO: If the size is known, use it.
316 visitMemoryReference(I, MSI->getDest(), MemoryLocation::UnknownSize,
317 MSI->getAlignment(), nullptr, MemRef::Write);
321 case Intrinsic::vastart:
322 Assert(I.getParent()->getParent()->isVarArg(),
323 "Undefined behavior: va_start called in a non-varargs function",
326 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
327 nullptr, MemRef::Read | MemRef::Write);
329 case Intrinsic::vacopy:
330 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
331 nullptr, MemRef::Write);
332 visitMemoryReference(I, CS.getArgument(1), MemoryLocation::UnknownSize, 0,
333 nullptr, MemRef::Read);
335 case Intrinsic::vaend:
336 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
337 nullptr, MemRef::Read | MemRef::Write);
340 case Intrinsic::stackrestore:
341 // Stackrestore doesn't read or write memory, but it sets the
342 // stack pointer, which the compiler may read from or write to
343 // at any time, so check it for both readability and writeability.
344 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
345 nullptr, MemRef::Read | MemRef::Write);
348 case Intrinsic::eh_begincatch:
349 visitEHBeginCatch(II);
351 case Intrinsic::eh_endcatch:
357 void Lint::visitCallInst(CallInst &I) {
358 return visitCallSite(&I);
361 void Lint::visitInvokeInst(InvokeInst &I) {
362 return visitCallSite(&I);
365 void Lint::visitReturnInst(ReturnInst &I) {
366 Function *F = I.getParent()->getParent();
367 Assert(!F->doesNotReturn(),
368 "Unusual: Return statement in function with noreturn attribute", &I);
370 if (Value *V = I.getReturnValue()) {
372 findValue(V, F->getParent()->getDataLayout(), /*OffsetOk=*/true);
373 Assert(!isa<AllocaInst>(Obj), "Unusual: Returning alloca value", &I);
377 // TODO: Check that the reference is in bounds.
378 // TODO: Check readnone/readonly function attributes.
379 void Lint::visitMemoryReference(Instruction &I,
380 Value *Ptr, uint64_t Size, unsigned Align,
381 Type *Ty, unsigned Flags) {
382 // If no memory is being referenced, it doesn't matter if the pointer
387 Value *UnderlyingObject =
388 findValue(Ptr, I.getModule()->getDataLayout(), /*OffsetOk=*/true);
389 Assert(!isa<ConstantPointerNull>(UnderlyingObject),
390 "Undefined behavior: Null pointer dereference", &I);
391 Assert(!isa<UndefValue>(UnderlyingObject),
392 "Undefined behavior: Undef pointer dereference", &I);
393 Assert(!isa<ConstantInt>(UnderlyingObject) ||
394 !cast<ConstantInt>(UnderlyingObject)->isAllOnesValue(),
395 "Unusual: All-ones pointer dereference", &I);
396 Assert(!isa<ConstantInt>(UnderlyingObject) ||
397 !cast<ConstantInt>(UnderlyingObject)->isOne(),
398 "Unusual: Address one pointer dereference", &I);
400 if (Flags & MemRef::Write) {
401 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(UnderlyingObject))
402 Assert(!GV->isConstant(), "Undefined behavior: Write to read-only memory",
404 Assert(!isa<Function>(UnderlyingObject) &&
405 !isa<BlockAddress>(UnderlyingObject),
406 "Undefined behavior: Write to text section", &I);
408 if (Flags & MemRef::Read) {
409 Assert(!isa<Function>(UnderlyingObject), "Unusual: Load from function body",
411 Assert(!isa<BlockAddress>(UnderlyingObject),
412 "Undefined behavior: Load from block address", &I);
414 if (Flags & MemRef::Callee) {
415 Assert(!isa<BlockAddress>(UnderlyingObject),
416 "Undefined behavior: Call to block address", &I);
418 if (Flags & MemRef::Branchee) {
419 Assert(!isa<Constant>(UnderlyingObject) ||
420 isa<BlockAddress>(UnderlyingObject),
421 "Undefined behavior: Branch to non-blockaddress", &I);
424 // Check for buffer overflows and misalignment.
425 // Only handles memory references that read/write something simple like an
426 // alloca instruction or a global variable.
427 auto &DL = I.getModule()->getDataLayout();
429 if (Value *Base = GetPointerBaseWithConstantOffset(Ptr, Offset, DL)) {
430 // OK, so the access is to a constant offset from Ptr. Check that Ptr is
431 // something we can handle and if so extract the size of this base object
432 // along with its alignment.
433 uint64_t BaseSize = MemoryLocation::UnknownSize;
434 unsigned BaseAlign = 0;
436 if (AllocaInst *AI = dyn_cast<AllocaInst>(Base)) {
437 Type *ATy = AI->getAllocatedType();
438 if (!AI->isArrayAllocation() && ATy->isSized())
439 BaseSize = DL.getTypeAllocSize(ATy);
440 BaseAlign = AI->getAlignment();
441 if (BaseAlign == 0 && ATy->isSized())
442 BaseAlign = DL.getABITypeAlignment(ATy);
443 } else if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) {
444 // If the global may be defined differently in another compilation unit
445 // then don't warn about funky memory accesses.
446 if (GV->hasDefinitiveInitializer()) {
447 Type *GTy = GV->getType()->getElementType();
449 BaseSize = DL.getTypeAllocSize(GTy);
450 BaseAlign = GV->getAlignment();
451 if (BaseAlign == 0 && GTy->isSized())
452 BaseAlign = DL.getABITypeAlignment(GTy);
456 // Accesses from before the start or after the end of the object are not
458 Assert(Size == MemoryLocation::UnknownSize ||
459 BaseSize == MemoryLocation::UnknownSize ||
460 (Offset >= 0 && Offset + Size <= BaseSize),
461 "Undefined behavior: Buffer overflow", &I);
463 // Accesses that say that the memory is more aligned than it is are not
465 if (Align == 0 && Ty && Ty->isSized())
466 Align = DL.getABITypeAlignment(Ty);
467 Assert(!BaseAlign || Align <= MinAlign(BaseAlign, Offset),
468 "Undefined behavior: Memory reference address is misaligned", &I);
472 void Lint::visitLoadInst(LoadInst &I) {
473 visitMemoryReference(I, I.getPointerOperand(),
474 AA->getTypeStoreSize(I.getType()), I.getAlignment(),
475 I.getType(), MemRef::Read);
478 void Lint::visitStoreInst(StoreInst &I) {
479 visitMemoryReference(I, I.getPointerOperand(),
480 AA->getTypeStoreSize(I.getOperand(0)->getType()),
482 I.getOperand(0)->getType(), MemRef::Write);
485 void Lint::visitXor(BinaryOperator &I) {
486 Assert(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
487 "Undefined result: xor(undef, undef)", &I);
490 void Lint::visitSub(BinaryOperator &I) {
491 Assert(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
492 "Undefined result: sub(undef, undef)", &I);
495 void Lint::visitLShr(BinaryOperator &I) {
496 if (ConstantInt *CI = dyn_cast<ConstantInt>(
497 findValue(I.getOperand(1), I.getModule()->getDataLayout(),
498 /*OffsetOk=*/false)))
499 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
500 "Undefined result: Shift count out of range", &I);
503 void Lint::visitAShr(BinaryOperator &I) {
504 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(
505 I.getOperand(1), I.getModule()->getDataLayout(), /*OffsetOk=*/false)))
506 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
507 "Undefined result: Shift count out of range", &I);
510 void Lint::visitShl(BinaryOperator &I) {
511 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(
512 I.getOperand(1), I.getModule()->getDataLayout(), /*OffsetOk=*/false)))
513 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
514 "Undefined result: Shift count out of range", &I);
518 allPredsCameFromLandingPad(BasicBlock *BB,
519 SmallSet<BasicBlock *, 4> &VisitedBlocks) {
520 VisitedBlocks.insert(BB);
521 if (BB->isLandingPad())
523 // If we find a block with no predecessors, the search failed.
526 for (BasicBlock *Pred : predecessors(BB)) {
527 if (VisitedBlocks.count(Pred))
529 if (!allPredsCameFromLandingPad(Pred, VisitedBlocks))
536 allSuccessorsReachEndCatch(BasicBlock *BB, BasicBlock::iterator InstBegin,
537 IntrinsicInst **SecondBeginCatch,
538 SmallSet<BasicBlock *, 4> &VisitedBlocks) {
539 VisitedBlocks.insert(BB);
540 for (BasicBlock::iterator I = InstBegin, E = BB->end(); I != E; ++I) {
541 IntrinsicInst *IC = dyn_cast<IntrinsicInst>(I);
542 if (IC && IC->getIntrinsicID() == Intrinsic::eh_endcatch)
544 // If we find another begincatch while looking for an endcatch,
545 // that's also an error.
546 if (IC && IC->getIntrinsicID() == Intrinsic::eh_begincatch) {
547 *SecondBeginCatch = IC;
552 // If we reach a block with no successors while searching, the
553 // search has failed.
556 // Otherwise, search all of the successors.
557 for (BasicBlock *Succ : successors(BB)) {
558 if (VisitedBlocks.count(Succ))
560 if (!allSuccessorsReachEndCatch(Succ, Succ->begin(), SecondBeginCatch,
567 void Lint::visitEHBeginCatch(IntrinsicInst *II) {
568 // The checks in this function make a potentially dubious assumption about
569 // the CFG, namely that any block involved in a catch is only used for the
570 // catch. This will very likely be true of IR generated by a front end,
571 // but it may cease to be true, for example, if the IR is run through a
572 // pass which combines similar blocks.
574 // In general, if we encounter a block the isn't dominated by the catch
575 // block while we are searching the catch block's successors for a call
576 // to end catch intrinsic, then it is possible that it will be legal for
577 // a path through this block to never reach a call to llvm.eh.endcatch.
578 // An analogous statement could be made about our search for a landing
579 // pad among the catch block's predecessors.
581 // What is actually required is that no path is possible at runtime that
582 // reaches a call to llvm.eh.begincatch without having previously visited
583 // a landingpad instruction and that no path is possible at runtime that
584 // calls llvm.eh.begincatch and does not subsequently call llvm.eh.endcatch
585 // (mentally adjusting for the fact that in reality these calls will be
586 // removed before code generation).
588 // Because this is a lint check, we take a pessimistic approach and warn if
589 // the control flow is potentially incorrect.
591 SmallSet<BasicBlock *, 4> VisitedBlocks;
592 BasicBlock *CatchBB = II->getParent();
594 // The begin catch must occur in a landing pad block or all paths
595 // to it must have come from a landing pad.
596 Assert(allPredsCameFromLandingPad(CatchBB, VisitedBlocks),
597 "llvm.eh.begincatch may be reachable without passing a landingpad",
600 // Reset the visited block list.
601 VisitedBlocks.clear();
603 IntrinsicInst *SecondBeginCatch = nullptr;
605 // This has to be called before it is asserted. Otherwise, the first assert
606 // below can never be hit.
607 bool EndCatchFound = allSuccessorsReachEndCatch(
608 CatchBB, std::next(static_cast<BasicBlock::iterator>(II)),
609 &SecondBeginCatch, VisitedBlocks);
611 SecondBeginCatch == nullptr,
612 "llvm.eh.begincatch may be called a second time before llvm.eh.endcatch",
613 II, SecondBeginCatch);
614 Assert(EndCatchFound,
615 "Some paths from llvm.eh.begincatch may not reach llvm.eh.endcatch",
619 static bool allPredCameFromBeginCatch(
620 BasicBlock *BB, BasicBlock::reverse_iterator InstRbegin,
621 IntrinsicInst **SecondEndCatch, SmallSet<BasicBlock *, 4> &VisitedBlocks) {
622 VisitedBlocks.insert(BB);
623 // Look for a begincatch in this block.
624 for (BasicBlock::reverse_iterator RI = InstRbegin, RE = BB->rend(); RI != RE;
626 IntrinsicInst *IC = dyn_cast<IntrinsicInst>(&*RI);
627 if (IC && IC->getIntrinsicID() == Intrinsic::eh_begincatch)
629 // If we find another end catch before we find a begin catch, that's
631 if (IC && IC->getIntrinsicID() == Intrinsic::eh_endcatch) {
632 *SecondEndCatch = IC;
635 // If we encounter a landingpad instruction, the search failed.
636 if (isa<LandingPadInst>(*RI))
639 // If while searching we find a block with no predeccesors,
640 // the search failed.
643 // Search any predecessors we haven't seen before.
644 for (BasicBlock *Pred : predecessors(BB)) {
645 if (VisitedBlocks.count(Pred))
647 if (!allPredCameFromBeginCatch(Pred, Pred->rbegin(), SecondEndCatch,
654 void Lint::visitEHEndCatch(IntrinsicInst *II) {
655 // The check in this function makes a potentially dubious assumption about
656 // the CFG, namely that any block involved in a catch is only used for the
657 // catch. This will very likely be true of IR generated by a front end,
658 // but it may cease to be true, for example, if the IR is run through a
659 // pass which combines similar blocks.
661 // In general, if we encounter a block the isn't post-dominated by the
662 // end catch block while we are searching the end catch block's predecessors
663 // for a call to the begin catch intrinsic, then it is possible that it will
664 // be legal for a path to reach the end catch block without ever having
665 // called llvm.eh.begincatch.
667 // What is actually required is that no path is possible at runtime that
668 // reaches a call to llvm.eh.endcatch without having previously visited
669 // a call to llvm.eh.begincatch (mentally adjusting for the fact that in
670 // reality these calls will be removed before code generation).
672 // Because this is a lint check, we take a pessimistic approach and warn if
673 // the control flow is potentially incorrect.
675 BasicBlock *EndCatchBB = II->getParent();
677 // Alls paths to the end catch call must pass through a begin catch call.
679 // If llvm.eh.begincatch wasn't called in the current block, we'll use this
680 // lambda to recursively look for it in predecessors.
681 SmallSet<BasicBlock *, 4> VisitedBlocks;
682 IntrinsicInst *SecondEndCatch = nullptr;
684 // This has to be called before it is asserted. Otherwise, the first assert
685 // below can never be hit.
686 bool BeginCatchFound =
687 allPredCameFromBeginCatch(EndCatchBB, BasicBlock::reverse_iterator(II),
688 &SecondEndCatch, VisitedBlocks);
690 SecondEndCatch == nullptr,
691 "llvm.eh.endcatch may be called a second time after llvm.eh.begincatch",
693 Assert(BeginCatchFound,
694 "llvm.eh.endcatch may be reachable without passing llvm.eh.begincatch",
698 static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT,
699 AssumptionCache *AC) {
700 // Assume undef could be zero.
701 if (isa<UndefValue>(V))
704 VectorType *VecTy = dyn_cast<VectorType>(V->getType());
706 unsigned BitWidth = V->getType()->getIntegerBitWidth();
707 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
708 computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC,
709 dyn_cast<Instruction>(V), DT);
710 return KnownZero.isAllOnesValue();
713 // Per-component check doesn't work with zeroinitializer
714 Constant *C = dyn_cast<Constant>(V);
718 if (C->isZeroValue())
721 // For a vector, KnownZero will only be true if all values are zero, so check
722 // this per component
723 unsigned BitWidth = VecTy->getElementType()->getIntegerBitWidth();
724 for (unsigned I = 0, N = VecTy->getNumElements(); I != N; ++I) {
725 Constant *Elem = C->getAggregateElement(I);
726 if (isa<UndefValue>(Elem))
729 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
730 computeKnownBits(Elem, KnownZero, KnownOne, DL);
731 if (KnownZero.isAllOnesValue())
738 void Lint::visitSDiv(BinaryOperator &I) {
739 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
740 "Undefined behavior: Division by zero", &I);
743 void Lint::visitUDiv(BinaryOperator &I) {
744 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
745 "Undefined behavior: Division by zero", &I);
748 void Lint::visitSRem(BinaryOperator &I) {
749 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
750 "Undefined behavior: Division by zero", &I);
753 void Lint::visitURem(BinaryOperator &I) {
754 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
755 "Undefined behavior: Division by zero", &I);
758 void Lint::visitAllocaInst(AllocaInst &I) {
759 if (isa<ConstantInt>(I.getArraySize()))
760 // This isn't undefined behavior, it's just an obvious pessimization.
761 Assert(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
762 "Pessimization: Static alloca outside of entry block", &I);
764 // TODO: Check for an unusual size (MSB set?)
767 void Lint::visitVAArgInst(VAArgInst &I) {
768 visitMemoryReference(I, I.getOperand(0), MemoryLocation::UnknownSize, 0,
769 nullptr, MemRef::Read | MemRef::Write);
772 void Lint::visitIndirectBrInst(IndirectBrInst &I) {
773 visitMemoryReference(I, I.getAddress(), MemoryLocation::UnknownSize, 0,
774 nullptr, MemRef::Branchee);
776 Assert(I.getNumDestinations() != 0,
777 "Undefined behavior: indirectbr with no destinations", &I);
780 void Lint::visitExtractElementInst(ExtractElementInst &I) {
781 if (ConstantInt *CI = dyn_cast<ConstantInt>(
782 findValue(I.getIndexOperand(), I.getModule()->getDataLayout(),
783 /*OffsetOk=*/false)))
784 Assert(CI->getValue().ult(I.getVectorOperandType()->getNumElements()),
785 "Undefined result: extractelement index out of range", &I);
788 void Lint::visitInsertElementInst(InsertElementInst &I) {
789 if (ConstantInt *CI = dyn_cast<ConstantInt>(
790 findValue(I.getOperand(2), I.getModule()->getDataLayout(),
791 /*OffsetOk=*/false)))
792 Assert(CI->getValue().ult(I.getType()->getNumElements()),
793 "Undefined result: insertelement index out of range", &I);
796 void Lint::visitUnreachableInst(UnreachableInst &I) {
797 // This isn't undefined behavior, it's merely suspicious.
798 Assert(&I == I.getParent()->begin() ||
799 std::prev(BasicBlock::iterator(&I))->mayHaveSideEffects(),
800 "Unusual: unreachable immediately preceded by instruction without "
805 /// findValue - Look through bitcasts and simple memory reference patterns
806 /// to identify an equivalent, but more informative, value. If OffsetOk
807 /// is true, look through getelementptrs with non-zero offsets too.
809 /// Most analysis passes don't require this logic, because instcombine
810 /// will simplify most of these kinds of things away. But it's a goal of
811 /// this Lint pass to be useful even on non-optimized IR.
812 Value *Lint::findValue(Value *V, const DataLayout &DL, bool OffsetOk) const {
813 SmallPtrSet<Value *, 4> Visited;
814 return findValueImpl(V, DL, OffsetOk, Visited);
817 /// findValueImpl - Implementation helper for findValue.
818 Value *Lint::findValueImpl(Value *V, const DataLayout &DL, bool OffsetOk,
819 SmallPtrSetImpl<Value *> &Visited) const {
820 // Detect self-referential values.
821 if (!Visited.insert(V).second)
822 return UndefValue::get(V->getType());
824 // TODO: Look through sext or zext cast, when the result is known to
825 // be interpreted as signed or unsigned, respectively.
826 // TODO: Look through eliminable cast pairs.
827 // TODO: Look through calls with unique return values.
828 // TODO: Look through vector insert/extract/shuffle.
829 V = OffsetOk ? GetUnderlyingObject(V, DL) : V->stripPointerCasts();
830 if (LoadInst *L = dyn_cast<LoadInst>(V)) {
831 BasicBlock::iterator BBI = L;
832 BasicBlock *BB = L->getParent();
833 SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
835 if (!VisitedBlocks.insert(BB).second)
837 if (Value *U = FindAvailableLoadedValue(L->getPointerOperand(),
839 return findValueImpl(U, DL, OffsetOk, Visited);
840 if (BBI != BB->begin()) break;
841 BB = BB->getUniquePredecessor();
845 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
846 if (Value *W = PN->hasConstantValue())
848 return findValueImpl(W, DL, OffsetOk, Visited);
849 } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
850 if (CI->isNoopCast(DL))
851 return findValueImpl(CI->getOperand(0), DL, OffsetOk, Visited);
852 } else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(V)) {
853 if (Value *W = FindInsertedValue(Ex->getAggregateOperand(),
856 return findValueImpl(W, DL, OffsetOk, Visited);
857 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
858 // Same as above, but for ConstantExpr instead of Instruction.
859 if (Instruction::isCast(CE->getOpcode())) {
860 if (CastInst::isNoopCast(Instruction::CastOps(CE->getOpcode()),
861 CE->getOperand(0)->getType(), CE->getType(),
862 DL.getIntPtrType(V->getType())))
863 return findValueImpl(CE->getOperand(0), DL, OffsetOk, Visited);
864 } else if (CE->getOpcode() == Instruction::ExtractValue) {
865 ArrayRef<unsigned> Indices = CE->getIndices();
866 if (Value *W = FindInsertedValue(CE->getOperand(0), Indices))
868 return findValueImpl(W, DL, OffsetOk, Visited);
872 // As a last resort, try SimplifyInstruction or constant folding.
873 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
874 if (Value *W = SimplifyInstruction(Inst, DL, TLI, DT, AC))
875 return findValueImpl(W, DL, OffsetOk, Visited);
876 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
877 if (Value *W = ConstantFoldConstantExpression(CE, DL, TLI))
879 return findValueImpl(W, DL, OffsetOk, Visited);
885 //===----------------------------------------------------------------------===//
886 // Implement the public interfaces to this file...
887 //===----------------------------------------------------------------------===//
889 FunctionPass *llvm::createLintPass() {
893 /// lintFunction - Check a function for errors, printing messages on stderr.
895 void llvm::lintFunction(const Function &f) {
896 Function &F = const_cast<Function&>(f);
897 assert(!F.isDeclaration() && "Cannot lint external functions");
899 legacy::FunctionPassManager FPM(F.getParent());
900 Lint *V = new Lint();
905 /// lintModule - Check a module for errors, printing messages on stderr.
907 void llvm::lintModule(const Module &M) {
908 legacy::PassManager PM;
909 Lint *V = new Lint();
911 PM.run(const_cast<Module&>(M));