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/Module.h"
53 #include "llvm/IR/InstVisitor.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/LegacyPassManager.h"
56 #include "llvm/Pass.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/raw_ostream.h"
63 static const unsigned Read = 1;
64 static const unsigned Write = 2;
65 static const unsigned Callee = 4;
66 static const unsigned Branchee = 8;
69 class Lint : public FunctionPass, public InstVisitor<Lint> {
70 friend class InstVisitor<Lint>;
72 void visitFunction(Function &F);
74 void visitCallSite(CallSite CS);
75 void visitMemoryReference(Instruction &I, Value *Ptr,
76 uint64_t Size, unsigned Align,
77 Type *Ty, unsigned Flags);
78 void visitEHBeginCatch(IntrinsicInst *II);
79 void visitEHEndCatch(IntrinsicInst *II);
81 void visitCallInst(CallInst &I);
82 void visitInvokeInst(InvokeInst &I);
83 void visitReturnInst(ReturnInst &I);
84 void visitLoadInst(LoadInst &I);
85 void visitStoreInst(StoreInst &I);
86 void visitXor(BinaryOperator &I);
87 void visitSub(BinaryOperator &I);
88 void visitLShr(BinaryOperator &I);
89 void visitAShr(BinaryOperator &I);
90 void visitShl(BinaryOperator &I);
91 void visitSDiv(BinaryOperator &I);
92 void visitUDiv(BinaryOperator &I);
93 void visitSRem(BinaryOperator &I);
94 void visitURem(BinaryOperator &I);
95 void visitAllocaInst(AllocaInst &I);
96 void visitVAArgInst(VAArgInst &I);
97 void visitIndirectBrInst(IndirectBrInst &I);
98 void visitExtractElementInst(ExtractElementInst &I);
99 void visitInsertElementInst(InsertElementInst &I);
100 void visitUnreachableInst(UnreachableInst &I);
102 Value *findValue(Value *V, bool OffsetOk) const;
103 Value *findValueImpl(Value *V, bool OffsetOk,
104 SmallPtrSetImpl<Value *> &Visited) const;
108 const DataLayout *DL;
112 TargetLibraryInfo *TLI;
114 std::string Messages;
115 raw_string_ostream MessagesStr;
117 static char ID; // Pass identification, replacement for typeid
118 Lint() : FunctionPass(ID), MessagesStr(Messages) {
119 initializeLintPass(*PassRegistry::getPassRegistry());
122 bool runOnFunction(Function &F) override;
124 void getAnalysisUsage(AnalysisUsage &AU) const override {
125 AU.setPreservesAll();
126 AU.addRequired<AAResultsWrapperPass>();
127 AU.addRequired<AssumptionCacheTracker>();
128 AU.addRequired<TargetLibraryInfoWrapperPass>();
129 AU.addRequired<DominatorTreeWrapperPass>();
131 void print(raw_ostream &O, const Module *M) const override {}
133 void WriteValues(ArrayRef<const Value *> Vs) {
134 for (const Value *V : Vs) {
137 if (isa<Instruction>(V)) {
138 MessagesStr << *V << '\n';
140 V->printAsOperand(MessagesStr, true, Mod);
146 /// \brief A check failed, so printout out the condition and the message.
148 /// This provides a nice place to put a breakpoint if you want to see why
149 /// something is not correct.
150 void CheckFailed(const Twine &Message) { MessagesStr << Message << '\n'; }
152 /// \brief A check failed (with values to print).
154 /// This calls the Message-only version so that the above is easier to set
156 template <typename T1, typename... Ts>
157 void CheckFailed(const Twine &Message, const T1 &V1, const Ts &...Vs) {
158 CheckFailed(Message);
159 WriteValues({V1, Vs...});
165 INITIALIZE_PASS_BEGIN(Lint, "lint", "Statically lint-checks LLVM IR",
167 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
168 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
169 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
170 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
171 INITIALIZE_PASS_END(Lint, "lint", "Statically lint-checks LLVM IR",
174 // Assert - We know that cond should be true, if not print an error message.
175 #define Assert(C, ...) \
176 do { if (!(C)) { CheckFailed(__VA_ARGS__); return; } } while (0)
178 // Lint::run - This is the main Analysis entry point for a
181 bool Lint::runOnFunction(Function &F) {
183 DL = &F.getParent()->getDataLayout();
184 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
185 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
186 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
187 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
189 dbgs() << MessagesStr.str();
194 void Lint::visitFunction(Function &F) {
195 // This isn't undefined behavior, it's just a little unusual, and it's a
196 // fairly common mistake to neglect to name a function.
197 Assert(F.hasName() || F.hasLocalLinkage(),
198 "Unusual: Unnamed function with non-local linkage", &F);
200 // TODO: Check for irreducible control flow.
203 void Lint::visitCallSite(CallSite CS) {
204 Instruction &I = *CS.getInstruction();
205 Value *Callee = CS.getCalledValue();
207 visitMemoryReference(I, Callee, MemoryLocation::UnknownSize, 0, nullptr,
210 if (Function *F = dyn_cast<Function>(findValue(Callee,
211 /*OffsetOk=*/false))) {
212 Assert(CS.getCallingConv() == F->getCallingConv(),
213 "Undefined behavior: Caller and callee calling convention differ",
216 FunctionType *FT = F->getFunctionType();
217 unsigned NumActualArgs = CS.arg_size();
219 Assert(FT->isVarArg() ? FT->getNumParams() <= NumActualArgs
220 : FT->getNumParams() == NumActualArgs,
221 "Undefined behavior: Call argument count mismatches callee "
225 Assert(FT->getReturnType() == I.getType(),
226 "Undefined behavior: Call return type mismatches "
227 "callee return type",
230 // Check argument types (in case the callee was casted) and attributes.
231 // TODO: Verify that caller and callee attributes are compatible.
232 Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
233 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
234 for (; AI != AE; ++AI) {
237 Argument *Formal = PI++;
238 Assert(Formal->getType() == Actual->getType(),
239 "Undefined behavior: Call argument type mismatches "
240 "callee parameter type",
243 // Check that noalias arguments don't alias other arguments. This is
244 // not fully precise because we don't know the sizes of the dereferenced
246 if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
247 for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
248 if (AI != BI && (*BI)->getType()->isPointerTy()) {
249 AliasResult Result = AA->alias(*AI, *BI);
250 Assert(Result != MustAlias && Result != PartialAlias,
251 "Unusual: noalias argument aliases another argument", &I);
254 // Check that an sret argument points to valid memory.
255 if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
257 cast<PointerType>(Formal->getType())->getElementType();
258 visitMemoryReference(I, Actual, DL->getTypeStoreSize(Ty),
259 DL->getABITypeAlignment(Ty), Ty,
260 MemRef::Read | MemRef::Write);
266 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
267 for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
269 Value *Obj = findValue(*AI, /*OffsetOk=*/true);
270 Assert(!isa<AllocaInst>(Obj),
271 "Undefined behavior: Call with \"tail\" keyword references "
277 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
278 switch (II->getIntrinsicID()) {
281 // TODO: Check more intrinsics
283 case Intrinsic::memcpy: {
284 MemCpyInst *MCI = cast<MemCpyInst>(&I);
285 // TODO: If the size is known, use it.
286 visitMemoryReference(I, MCI->getDest(), MemoryLocation::UnknownSize,
287 MCI->getAlignment(), nullptr, MemRef::Write);
288 visitMemoryReference(I, MCI->getSource(), MemoryLocation::UnknownSize,
289 MCI->getAlignment(), nullptr, MemRef::Read);
291 // Check that the memcpy arguments don't overlap. The AliasAnalysis API
292 // isn't expressive enough for what we really want to do. Known partial
293 // overlap is not distinguished from the case where nothing is known.
295 if (const ConstantInt *Len =
296 dyn_cast<ConstantInt>(findValue(MCI->getLength(),
297 /*OffsetOk=*/false)))
298 if (Len->getValue().isIntN(32))
299 Size = Len->getValue().getZExtValue();
300 Assert(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
302 "Undefined behavior: memcpy source and destination overlap", &I);
305 case Intrinsic::memmove: {
306 MemMoveInst *MMI = cast<MemMoveInst>(&I);
307 // TODO: If the size is known, use it.
308 visitMemoryReference(I, MMI->getDest(), MemoryLocation::UnknownSize,
309 MMI->getAlignment(), nullptr, MemRef::Write);
310 visitMemoryReference(I, MMI->getSource(), MemoryLocation::UnknownSize,
311 MMI->getAlignment(), nullptr, MemRef::Read);
314 case Intrinsic::memset: {
315 MemSetInst *MSI = cast<MemSetInst>(&I);
316 // TODO: If the size is known, use it.
317 visitMemoryReference(I, MSI->getDest(), MemoryLocation::UnknownSize,
318 MSI->getAlignment(), nullptr, MemRef::Write);
322 case Intrinsic::vastart:
323 Assert(I.getParent()->getParent()->isVarArg(),
324 "Undefined behavior: va_start called in a non-varargs function",
327 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
328 nullptr, MemRef::Read | MemRef::Write);
330 case Intrinsic::vacopy:
331 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
332 nullptr, MemRef::Write);
333 visitMemoryReference(I, CS.getArgument(1), MemoryLocation::UnknownSize, 0,
334 nullptr, MemRef::Read);
336 case Intrinsic::vaend:
337 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
338 nullptr, MemRef::Read | MemRef::Write);
341 case Intrinsic::stackrestore:
342 // Stackrestore doesn't read or write memory, but it sets the
343 // stack pointer, which the compiler may read from or write to
344 // at any time, so check it for both readability and writeability.
345 visitMemoryReference(I, CS.getArgument(0), MemoryLocation::UnknownSize, 0,
346 nullptr, MemRef::Read | MemRef::Write);
349 case Intrinsic::eh_begincatch:
350 visitEHBeginCatch(II);
352 case Intrinsic::eh_endcatch:
358 void Lint::visitCallInst(CallInst &I) {
359 return visitCallSite(&I);
362 void Lint::visitInvokeInst(InvokeInst &I) {
363 return visitCallSite(&I);
366 void Lint::visitReturnInst(ReturnInst &I) {
367 Function *F = I.getParent()->getParent();
368 Assert(!F->doesNotReturn(),
369 "Unusual: Return statement in function with noreturn attribute", &I);
371 if (Value *V = I.getReturnValue()) {
372 Value *Obj = findValue(V, /*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 = findValue(Ptr, /*OffsetOk=*/true);
388 Assert(!isa<ConstantPointerNull>(UnderlyingObject),
389 "Undefined behavior: Null pointer dereference", &I);
390 Assert(!isa<UndefValue>(UnderlyingObject),
391 "Undefined behavior: Undef pointer dereference", &I);
392 Assert(!isa<ConstantInt>(UnderlyingObject) ||
393 !cast<ConstantInt>(UnderlyingObject)->isAllOnesValue(),
394 "Unusual: All-ones pointer dereference", &I);
395 Assert(!isa<ConstantInt>(UnderlyingObject) ||
396 !cast<ConstantInt>(UnderlyingObject)->isOne(),
397 "Unusual: Address one pointer dereference", &I);
399 if (Flags & MemRef::Write) {
400 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(UnderlyingObject))
401 Assert(!GV->isConstant(), "Undefined behavior: Write to read-only memory",
403 Assert(!isa<Function>(UnderlyingObject) &&
404 !isa<BlockAddress>(UnderlyingObject),
405 "Undefined behavior: Write to text section", &I);
407 if (Flags & MemRef::Read) {
408 Assert(!isa<Function>(UnderlyingObject), "Unusual: Load from function body",
410 Assert(!isa<BlockAddress>(UnderlyingObject),
411 "Undefined behavior: Load from block address", &I);
413 if (Flags & MemRef::Callee) {
414 Assert(!isa<BlockAddress>(UnderlyingObject),
415 "Undefined behavior: Call to block address", &I);
417 if (Flags & MemRef::Branchee) {
418 Assert(!isa<Constant>(UnderlyingObject) ||
419 isa<BlockAddress>(UnderlyingObject),
420 "Undefined behavior: Branch to non-blockaddress", &I);
423 // Check for buffer overflows and misalignment.
424 // Only handles memory references that read/write something simple like an
425 // alloca instruction or a global variable.
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 = MemoryLocation::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 == MemoryLocation::UnknownSize ||
457 BaseSize == MemoryLocation::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 DL->getTypeStoreSize(I.getType()), I.getAlignment(),
473 I.getType(), MemRef::Read);
476 void Lint::visitStoreInst(StoreInst &I) {
477 visitMemoryReference(I, I.getPointerOperand(),
478 DL->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>(findValue(I.getOperand(1),
495 /*OffsetOk=*/false)))
496 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
497 "Undefined result: Shift count out of range", &I);
500 void Lint::visitAShr(BinaryOperator &I) {
501 if (ConstantInt *CI =
502 dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
503 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
504 "Undefined result: Shift count out of range", &I);
507 void Lint::visitShl(BinaryOperator &I) {
508 if (ConstantInt *CI =
509 dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
510 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
511 "Undefined result: Shift count out of range", &I);
515 allPredsCameFromLandingPad(BasicBlock *BB,
516 SmallSet<BasicBlock *, 4> &VisitedBlocks) {
517 VisitedBlocks.insert(BB);
518 if (BB->isLandingPad())
520 // If we find a block with no predecessors, the search failed.
523 for (BasicBlock *Pred : predecessors(BB)) {
524 if (VisitedBlocks.count(Pred))
526 if (!allPredsCameFromLandingPad(Pred, VisitedBlocks))
533 allSuccessorsReachEndCatch(BasicBlock *BB, BasicBlock::iterator InstBegin,
534 IntrinsicInst **SecondBeginCatch,
535 SmallSet<BasicBlock *, 4> &VisitedBlocks) {
536 VisitedBlocks.insert(BB);
537 for (BasicBlock::iterator I = InstBegin, E = BB->end(); I != E; ++I) {
538 IntrinsicInst *IC = dyn_cast<IntrinsicInst>(I);
539 if (IC && IC->getIntrinsicID() == Intrinsic::eh_endcatch)
541 // If we find another begincatch while looking for an endcatch,
542 // that's also an error.
543 if (IC && IC->getIntrinsicID() == Intrinsic::eh_begincatch) {
544 *SecondBeginCatch = IC;
549 // If we reach a block with no successors while searching, the
550 // search has failed.
553 // Otherwise, search all of the successors.
554 for (BasicBlock *Succ : successors(BB)) {
555 if (VisitedBlocks.count(Succ))
557 if (!allSuccessorsReachEndCatch(Succ, Succ->begin(), SecondBeginCatch,
564 void Lint::visitEHBeginCatch(IntrinsicInst *II) {
565 // The checks in this function make a potentially dubious assumption about
566 // the CFG, namely that any block involved in a catch is only used for the
567 // catch. This will very likely be true of IR generated by a front end,
568 // but it may cease to be true, for example, if the IR is run through a
569 // pass which combines similar blocks.
571 // In general, if we encounter a block the isn't dominated by the catch
572 // block while we are searching the catch block's successors for a call
573 // to end catch intrinsic, then it is possible that it will be legal for
574 // a path through this block to never reach a call to llvm.eh.endcatch.
575 // An analogous statement could be made about our search for a landing
576 // pad among the catch block's predecessors.
578 // What is actually required is that no path is possible at runtime that
579 // reaches a call to llvm.eh.begincatch without having previously visited
580 // a landingpad instruction and that no path is possible at runtime that
581 // calls llvm.eh.begincatch and does not subsequently call llvm.eh.endcatch
582 // (mentally adjusting for the fact that in reality these calls will be
583 // removed before code generation).
585 // Because this is a lint check, we take a pessimistic approach and warn if
586 // the control flow is potentially incorrect.
588 SmallSet<BasicBlock *, 4> VisitedBlocks;
589 BasicBlock *CatchBB = II->getParent();
591 // The begin catch must occur in a landing pad block or all paths
592 // to it must have come from a landing pad.
593 Assert(allPredsCameFromLandingPad(CatchBB, VisitedBlocks),
594 "llvm.eh.begincatch may be reachable without passing a landingpad",
597 // Reset the visited block list.
598 VisitedBlocks.clear();
600 IntrinsicInst *SecondBeginCatch = nullptr;
602 // This has to be called before it is asserted. Otherwise, the first assert
603 // below can never be hit.
604 bool EndCatchFound = allSuccessorsReachEndCatch(
605 CatchBB, std::next(static_cast<BasicBlock::iterator>(II)),
606 &SecondBeginCatch, VisitedBlocks);
608 SecondBeginCatch == nullptr,
609 "llvm.eh.begincatch may be called a second time before llvm.eh.endcatch",
610 II, SecondBeginCatch);
611 Assert(EndCatchFound,
612 "Some paths from llvm.eh.begincatch may not reach llvm.eh.endcatch",
616 static bool allPredCameFromBeginCatch(
617 BasicBlock *BB, BasicBlock::reverse_iterator InstRbegin,
618 IntrinsicInst **SecondEndCatch, SmallSet<BasicBlock *, 4> &VisitedBlocks) {
619 VisitedBlocks.insert(BB);
620 // Look for a begincatch in this block.
621 for (BasicBlock::reverse_iterator RI = InstRbegin, RE = BB->rend(); RI != RE;
623 IntrinsicInst *IC = dyn_cast<IntrinsicInst>(&*RI);
624 if (IC && IC->getIntrinsicID() == Intrinsic::eh_begincatch)
626 // If we find another end catch before we find a begin catch, that's
628 if (IC && IC->getIntrinsicID() == Intrinsic::eh_endcatch) {
629 *SecondEndCatch = IC;
632 // If we encounter a landingpad instruction, the search failed.
633 if (isa<LandingPadInst>(*RI))
636 // If while searching we find a block with no predeccesors,
637 // the search failed.
640 // Search any predecessors we haven't seen before.
641 for (BasicBlock *Pred : predecessors(BB)) {
642 if (VisitedBlocks.count(Pred))
644 if (!allPredCameFromBeginCatch(Pred, Pred->rbegin(), SecondEndCatch,
651 void Lint::visitEHEndCatch(IntrinsicInst *II) {
652 // The check in this function makes a potentially dubious assumption about
653 // the CFG, namely that any block involved in a catch is only used for the
654 // catch. This will very likely be true of IR generated by a front end,
655 // but it may cease to be true, for example, if the IR is run through a
656 // pass which combines similar blocks.
658 // In general, if we encounter a block the isn't post-dominated by the
659 // end catch block while we are searching the end catch block's predecessors
660 // for a call to the begin catch intrinsic, then it is possible that it will
661 // be legal for a path to reach the end catch block without ever having
662 // called llvm.eh.begincatch.
664 // What is actually required is that no path is possible at runtime that
665 // reaches a call to llvm.eh.endcatch without having previously visited
666 // a call to llvm.eh.begincatch (mentally adjusting for the fact that in
667 // reality these calls will be removed before code generation).
669 // Because this is a lint check, we take a pessimistic approach and warn if
670 // the control flow is potentially incorrect.
672 BasicBlock *EndCatchBB = II->getParent();
674 // Alls paths to the end catch call must pass through a begin catch call.
676 // If llvm.eh.begincatch wasn't called in the current block, we'll use this
677 // lambda to recursively look for it in predecessors.
678 SmallSet<BasicBlock *, 4> VisitedBlocks;
679 IntrinsicInst *SecondEndCatch = nullptr;
681 // This has to be called before it is asserted. Otherwise, the first assert
682 // below can never be hit.
683 bool BeginCatchFound =
684 allPredCameFromBeginCatch(EndCatchBB, BasicBlock::reverse_iterator(II),
685 &SecondEndCatch, VisitedBlocks);
687 SecondEndCatch == nullptr,
688 "llvm.eh.endcatch may be called a second time after llvm.eh.begincatch",
690 Assert(BeginCatchFound,
691 "llvm.eh.endcatch may be reachable without passing llvm.eh.begincatch",
695 static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT,
696 AssumptionCache *AC) {
697 // Assume undef could be zero.
698 if (isa<UndefValue>(V))
701 VectorType *VecTy = dyn_cast<VectorType>(V->getType());
703 unsigned BitWidth = V->getType()->getIntegerBitWidth();
704 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
705 computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC,
706 dyn_cast<Instruction>(V), DT);
707 return KnownZero.isAllOnesValue();
710 // Per-component check doesn't work with zeroinitializer
711 Constant *C = dyn_cast<Constant>(V);
715 if (C->isZeroValue())
718 // For a vector, KnownZero will only be true if all values are zero, so check
719 // this per component
720 unsigned BitWidth = VecTy->getElementType()->getIntegerBitWidth();
721 for (unsigned I = 0, N = VecTy->getNumElements(); I != N; ++I) {
722 Constant *Elem = C->getAggregateElement(I);
723 if (isa<UndefValue>(Elem))
726 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
727 computeKnownBits(Elem, KnownZero, KnownOne, DL);
728 if (KnownZero.isAllOnesValue())
735 void Lint::visitSDiv(BinaryOperator &I) {
736 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
737 "Undefined behavior: Division by zero", &I);
740 void Lint::visitUDiv(BinaryOperator &I) {
741 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
742 "Undefined behavior: Division by zero", &I);
745 void Lint::visitSRem(BinaryOperator &I) {
746 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
747 "Undefined behavior: Division by zero", &I);
750 void Lint::visitURem(BinaryOperator &I) {
751 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
752 "Undefined behavior: Division by zero", &I);
755 void Lint::visitAllocaInst(AllocaInst &I) {
756 if (isa<ConstantInt>(I.getArraySize()))
757 // This isn't undefined behavior, it's just an obvious pessimization.
758 Assert(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
759 "Pessimization: Static alloca outside of entry block", &I);
761 // TODO: Check for an unusual size (MSB set?)
764 void Lint::visitVAArgInst(VAArgInst &I) {
765 visitMemoryReference(I, I.getOperand(0), MemoryLocation::UnknownSize, 0,
766 nullptr, MemRef::Read | MemRef::Write);
769 void Lint::visitIndirectBrInst(IndirectBrInst &I) {
770 visitMemoryReference(I, I.getAddress(), MemoryLocation::UnknownSize, 0,
771 nullptr, MemRef::Branchee);
773 Assert(I.getNumDestinations() != 0,
774 "Undefined behavior: indirectbr with no destinations", &I);
777 void Lint::visitExtractElementInst(ExtractElementInst &I) {
778 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(I.getIndexOperand(),
779 /*OffsetOk=*/false)))
780 Assert(CI->getValue().ult(I.getVectorOperandType()->getNumElements()),
781 "Undefined result: extractelement index out of range", &I);
784 void Lint::visitInsertElementInst(InsertElementInst &I) {
785 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(I.getOperand(2),
786 /*OffsetOk=*/false)))
787 Assert(CI->getValue().ult(I.getType()->getNumElements()),
788 "Undefined result: insertelement index out of range", &I);
791 void Lint::visitUnreachableInst(UnreachableInst &I) {
792 // This isn't undefined behavior, it's merely suspicious.
793 Assert(&I == I.getParent()->begin() ||
794 std::prev(BasicBlock::iterator(&I))->mayHaveSideEffects(),
795 "Unusual: unreachable immediately preceded by instruction without "
800 /// findValue - Look through bitcasts and simple memory reference patterns
801 /// to identify an equivalent, but more informative, value. If OffsetOk
802 /// is true, look through getelementptrs with non-zero offsets too.
804 /// Most analysis passes don't require this logic, because instcombine
805 /// will simplify most of these kinds of things away. But it's a goal of
806 /// this Lint pass to be useful even on non-optimized IR.
807 Value *Lint::findValue(Value *V, bool OffsetOk) const {
808 SmallPtrSet<Value *, 4> Visited;
809 return findValueImpl(V, OffsetOk, Visited);
812 /// findValueImpl - Implementation helper for findValue.
813 Value *Lint::findValueImpl(Value *V, bool OffsetOk,
814 SmallPtrSetImpl<Value *> &Visited) const {
815 // Detect self-referential values.
816 if (!Visited.insert(V).second)
817 return UndefValue::get(V->getType());
819 // TODO: Look through sext or zext cast, when the result is known to
820 // be interpreted as signed or unsigned, respectively.
821 // TODO: Look through eliminable cast pairs.
822 // TODO: Look through calls with unique return values.
823 // TODO: Look through vector insert/extract/shuffle.
824 V = OffsetOk ? GetUnderlyingObject(V, *DL) : V->stripPointerCasts();
825 if (LoadInst *L = dyn_cast<LoadInst>(V)) {
826 BasicBlock::iterator BBI = L;
827 BasicBlock *BB = L->getParent();
828 SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
830 if (!VisitedBlocks.insert(BB).second)
832 if (Value *U = FindAvailableLoadedValue(L->getPointerOperand(),
834 return findValueImpl(U, OffsetOk, Visited);
835 if (BBI != BB->begin()) break;
836 BB = BB->getUniquePredecessor();
840 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
841 if (Value *W = PN->hasConstantValue())
843 return findValueImpl(W, OffsetOk, Visited);
844 } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
845 if (CI->isNoopCast(*DL))
846 return findValueImpl(CI->getOperand(0), OffsetOk, Visited);
847 } else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(V)) {
848 if (Value *W = FindInsertedValue(Ex->getAggregateOperand(),
851 return findValueImpl(W, OffsetOk, Visited);
852 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
853 // Same as above, but for ConstantExpr instead of Instruction.
854 if (Instruction::isCast(CE->getOpcode())) {
855 if (CastInst::isNoopCast(Instruction::CastOps(CE->getOpcode()),
856 CE->getOperand(0)->getType(), CE->getType(),
857 DL->getIntPtrType(V->getType())))
858 return findValueImpl(CE->getOperand(0), OffsetOk, Visited);
859 } else if (CE->getOpcode() == Instruction::ExtractValue) {
860 ArrayRef<unsigned> Indices = CE->getIndices();
861 if (Value *W = FindInsertedValue(CE->getOperand(0), Indices))
863 return findValueImpl(W, OffsetOk, Visited);
867 // As a last resort, try SimplifyInstruction or constant folding.
868 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
869 if (Value *W = SimplifyInstruction(Inst, *DL, TLI, DT, AC))
870 return findValueImpl(W, OffsetOk, Visited);
871 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
872 if (Value *W = ConstantFoldConstantExpression(CE, *DL, TLI))
874 return findValueImpl(W, OffsetOk, Visited);
880 //===----------------------------------------------------------------------===//
881 // Implement the public interfaces to this file...
882 //===----------------------------------------------------------------------===//
884 FunctionPass *llvm::createLintPass() {
888 /// lintFunction - Check a function for errors, printing messages on stderr.
890 void llvm::lintFunction(const Function &f) {
891 Function &F = const_cast<Function&>(f);
892 assert(!F.isDeclaration() && "Cannot lint external functions");
894 legacy::FunctionPassManager FPM(F.getParent());
895 Lint *V = new Lint();
900 /// lintModule - Check a module for errors, printing messages on stderr.
902 void llvm::lintModule(const Module &M) {
903 legacy::PassManager PM;
904 Lint *V = new Lint();
906 PM.run(const_cast<Module&>(M));