• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===-- Lint.cpp - Check for common errors in LLVM IR ---------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass statically checks for common and easily-identified constructs
11 // which produce undefined or likely unintended behavior in LLVM IR.
12 //
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.
19 //
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.
24 //
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.
29 //
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.
34 //
35 //===----------------------------------------------------------------------===//
36 
37 #include "llvm/Analysis/Passes.h"
38 #include "llvm/Analysis/AliasAnalysis.h"
39 #include "llvm/Analysis/InstructionSimplify.h"
40 #include "llvm/Analysis/ConstantFolding.h"
41 #include "llvm/Analysis/Dominators.h"
42 #include "llvm/Analysis/Lint.h"
43 #include "llvm/Analysis/Loads.h"
44 #include "llvm/Analysis/ValueTracking.h"
45 #include "llvm/Assembly/Writer.h"
46 #include "llvm/Target/TargetData.h"
47 #include "llvm/Pass.h"
48 #include "llvm/PassManager.h"
49 #include "llvm/IntrinsicInst.h"
50 #include "llvm/Function.h"
51 #include "llvm/Support/CallSite.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/InstVisitor.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/ADT/STLExtras.h"
56 using namespace llvm;
57 
58 namespace {
59   namespace MemRef {
60     static unsigned Read     = 1;
61     static unsigned Write    = 2;
62     static unsigned Callee   = 4;
63     static unsigned Branchee = 8;
64   }
65 
66   class Lint : public FunctionPass, public InstVisitor<Lint> {
67     friend class InstVisitor<Lint>;
68 
69     void visitFunction(Function &F);
70 
71     void visitCallSite(CallSite CS);
72     void visitMemoryReference(Instruction &I, Value *Ptr,
73                               uint64_t Size, unsigned Align,
74                               Type *Ty, unsigned Flags);
75 
76     void visitCallInst(CallInst &I);
77     void visitInvokeInst(InvokeInst &I);
78     void visitReturnInst(ReturnInst &I);
79     void visitLoadInst(LoadInst &I);
80     void visitStoreInst(StoreInst &I);
81     void visitXor(BinaryOperator &I);
82     void visitSub(BinaryOperator &I);
83     void visitLShr(BinaryOperator &I);
84     void visitAShr(BinaryOperator &I);
85     void visitShl(BinaryOperator &I);
86     void visitSDiv(BinaryOperator &I);
87     void visitUDiv(BinaryOperator &I);
88     void visitSRem(BinaryOperator &I);
89     void visitURem(BinaryOperator &I);
90     void visitAllocaInst(AllocaInst &I);
91     void visitVAArgInst(VAArgInst &I);
92     void visitIndirectBrInst(IndirectBrInst &I);
93     void visitExtractElementInst(ExtractElementInst &I);
94     void visitInsertElementInst(InsertElementInst &I);
95     void visitUnreachableInst(UnreachableInst &I);
96 
97     Value *findValue(Value *V, bool OffsetOk) const;
98     Value *findValueImpl(Value *V, bool OffsetOk,
99                          SmallPtrSet<Value *, 4> &Visited) const;
100 
101   public:
102     Module *Mod;
103     AliasAnalysis *AA;
104     DominatorTree *DT;
105     TargetData *TD;
106 
107     std::string Messages;
108     raw_string_ostream MessagesStr;
109 
110     static char ID; // Pass identification, replacement for typeid
Lint()111     Lint() : FunctionPass(ID), MessagesStr(Messages) {
112       initializeLintPass(*PassRegistry::getPassRegistry());
113     }
114 
115     virtual bool runOnFunction(Function &F);
116 
getAnalysisUsage(AnalysisUsage & AU) const117     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
118       AU.setPreservesAll();
119       AU.addRequired<AliasAnalysis>();
120       AU.addRequired<DominatorTree>();
121     }
print(raw_ostream & O,const Module * M) const122     virtual void print(raw_ostream &O, const Module *M) const {}
123 
WriteValue(const Value * V)124     void WriteValue(const Value *V) {
125       if (!V) return;
126       if (isa<Instruction>(V)) {
127         MessagesStr << *V << '\n';
128       } else {
129         WriteAsOperand(MessagesStr, V, true, Mod);
130         MessagesStr << '\n';
131       }
132     }
133 
134     // CheckFailed - A check failed, so print out the condition and the message
135     // that failed.  This provides a nice place to put a breakpoint if you want
136     // to see why something is not correct.
CheckFailed(const Twine & Message,const Value * V1=0,const Value * V2=0,const Value * V3=0,const Value * V4=0)137     void CheckFailed(const Twine &Message,
138                      const Value *V1 = 0, const Value *V2 = 0,
139                      const Value *V3 = 0, const Value *V4 = 0) {
140       MessagesStr << Message.str() << "\n";
141       WriteValue(V1);
142       WriteValue(V2);
143       WriteValue(V3);
144       WriteValue(V4);
145     }
146   };
147 }
148 
149 char Lint::ID = 0;
150 INITIALIZE_PASS_BEGIN(Lint, "lint", "Statically lint-checks LLVM IR",
151                       false, true)
INITIALIZE_PASS_DEPENDENCY(DominatorTree)152 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
153 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
154 INITIALIZE_PASS_END(Lint, "lint", "Statically lint-checks LLVM IR",
155                     false, true)
156 
157 // Assert - We know that cond should be true, if not print an error message.
158 #define Assert(C, M) \
159     do { if (!(C)) { CheckFailed(M); return; } } while (0)
160 #define Assert1(C, M, V1) \
161     do { if (!(C)) { CheckFailed(M, V1); return; } } while (0)
162 #define Assert2(C, M, V1, V2) \
163     do { if (!(C)) { CheckFailed(M, V1, V2); return; } } while (0)
164 #define Assert3(C, M, V1, V2, V3) \
165     do { if (!(C)) { CheckFailed(M, V1, V2, V3); return; } } while (0)
166 #define Assert4(C, M, V1, V2, V3, V4) \
167     do { if (!(C)) { CheckFailed(M, V1, V2, V3, V4); return; } } while (0)
168 
169 // Lint::run - This is the main Analysis entry point for a
170 // function.
171 //
172 bool Lint::runOnFunction(Function &F) {
173   Mod = F.getParent();
174   AA = &getAnalysis<AliasAnalysis>();
175   DT = &getAnalysis<DominatorTree>();
176   TD = getAnalysisIfAvailable<TargetData>();
177   visit(F);
178   dbgs() << MessagesStr.str();
179   Messages.clear();
180   return false;
181 }
182 
visitFunction(Function & F)183 void Lint::visitFunction(Function &F) {
184   // This isn't undefined behavior, it's just a little unusual, and it's a
185   // fairly common mistake to neglect to name a function.
186   Assert1(F.hasName() || F.hasLocalLinkage(),
187           "Unusual: Unnamed function with non-local linkage", &F);
188 
189   // TODO: Check for irreducible control flow.
190 }
191 
visitCallSite(CallSite CS)192 void Lint::visitCallSite(CallSite CS) {
193   Instruction &I = *CS.getInstruction();
194   Value *Callee = CS.getCalledValue();
195 
196   visitMemoryReference(I, Callee, AliasAnalysis::UnknownSize,
197                        0, 0, MemRef::Callee);
198 
199   if (Function *F = dyn_cast<Function>(findValue(Callee, /*OffsetOk=*/false))) {
200     Assert1(CS.getCallingConv() == F->getCallingConv(),
201             "Undefined behavior: Caller and callee calling convention differ",
202             &I);
203 
204     FunctionType *FT = F->getFunctionType();
205     unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
206 
207     Assert1(FT->isVarArg() ?
208               FT->getNumParams() <= NumActualArgs :
209               FT->getNumParams() == NumActualArgs,
210             "Undefined behavior: Call argument count mismatches callee "
211             "argument count", &I);
212 
213     Assert1(FT->getReturnType() == I.getType(),
214             "Undefined behavior: Call return type mismatches "
215             "callee return type", &I);
216 
217     // Check argument types (in case the callee was casted) and attributes.
218     // TODO: Verify that caller and callee attributes are compatible.
219     Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
220     CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
221     for (; AI != AE; ++AI) {
222       Value *Actual = *AI;
223       if (PI != PE) {
224         Argument *Formal = PI++;
225         Assert1(Formal->getType() == Actual->getType(),
226                 "Undefined behavior: Call argument type mismatches "
227                 "callee parameter type", &I);
228 
229         // Check that noalias arguments don't alias other arguments. This is
230         // not fully precise because we don't know the sizes of the dereferenced
231         // memory regions.
232         if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
233           for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
234             if (AI != BI && (*BI)->getType()->isPointerTy()) {
235               AliasAnalysis::AliasResult Result = AA->alias(*AI, *BI);
236               Assert1(Result != AliasAnalysis::MustAlias &&
237                       Result != AliasAnalysis::PartialAlias,
238                       "Unusual: noalias argument aliases another argument", &I);
239             }
240 
241         // Check that an sret argument points to valid memory.
242         if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
243           Type *Ty =
244             cast<PointerType>(Formal->getType())->getElementType();
245           visitMemoryReference(I, Actual, AA->getTypeStoreSize(Ty),
246                                TD ? TD->getABITypeAlignment(Ty) : 0,
247                                Ty, MemRef::Read | MemRef::Write);
248         }
249       }
250     }
251   }
252 
253   if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
254     for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
255          AI != AE; ++AI) {
256       Value *Obj = findValue(*AI, /*OffsetOk=*/true);
257       Assert1(!isa<AllocaInst>(Obj),
258               "Undefined behavior: Call with \"tail\" keyword references "
259               "alloca", &I);
260     }
261 
262 
263   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
264     switch (II->getIntrinsicID()) {
265     default: break;
266 
267     // TODO: Check more intrinsics
268 
269     case Intrinsic::memcpy: {
270       MemCpyInst *MCI = cast<MemCpyInst>(&I);
271       // TODO: If the size is known, use it.
272       visitMemoryReference(I, MCI->getDest(), AliasAnalysis::UnknownSize,
273                            MCI->getAlignment(), 0,
274                            MemRef::Write);
275       visitMemoryReference(I, MCI->getSource(), AliasAnalysis::UnknownSize,
276                            MCI->getAlignment(), 0,
277                            MemRef::Read);
278 
279       // Check that the memcpy arguments don't overlap. The AliasAnalysis API
280       // isn't expressive enough for what we really want to do. Known partial
281       // overlap is not distinguished from the case where nothing is known.
282       uint64_t Size = 0;
283       if (const ConstantInt *Len =
284             dyn_cast<ConstantInt>(findValue(MCI->getLength(),
285                                             /*OffsetOk=*/false)))
286         if (Len->getValue().isIntN(32))
287           Size = Len->getValue().getZExtValue();
288       Assert1(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
289               AliasAnalysis::MustAlias,
290               "Undefined behavior: memcpy source and destination overlap", &I);
291       break;
292     }
293     case Intrinsic::memmove: {
294       MemMoveInst *MMI = cast<MemMoveInst>(&I);
295       // TODO: If the size is known, use it.
296       visitMemoryReference(I, MMI->getDest(), AliasAnalysis::UnknownSize,
297                            MMI->getAlignment(), 0,
298                            MemRef::Write);
299       visitMemoryReference(I, MMI->getSource(), AliasAnalysis::UnknownSize,
300                            MMI->getAlignment(), 0,
301                            MemRef::Read);
302       break;
303     }
304     case Intrinsic::memset: {
305       MemSetInst *MSI = cast<MemSetInst>(&I);
306       // TODO: If the size is known, use it.
307       visitMemoryReference(I, MSI->getDest(), AliasAnalysis::UnknownSize,
308                            MSI->getAlignment(), 0,
309                            MemRef::Write);
310       break;
311     }
312 
313     case Intrinsic::vastart:
314       Assert1(I.getParent()->getParent()->isVarArg(),
315               "Undefined behavior: va_start called in a non-varargs function",
316               &I);
317 
318       visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
319                            0, 0, MemRef::Read | MemRef::Write);
320       break;
321     case Intrinsic::vacopy:
322       visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
323                            0, 0, MemRef::Write);
324       visitMemoryReference(I, CS.getArgument(1), AliasAnalysis::UnknownSize,
325                            0, 0, MemRef::Read);
326       break;
327     case Intrinsic::vaend:
328       visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
329                            0, 0, MemRef::Read | MemRef::Write);
330       break;
331 
332     case Intrinsic::stackrestore:
333       // Stackrestore doesn't read or write memory, but it sets the
334       // stack pointer, which the compiler may read from or write to
335       // at any time, so check it for both readability and writeability.
336       visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
337                            0, 0, MemRef::Read | MemRef::Write);
338       break;
339     }
340 }
341 
visitCallInst(CallInst & I)342 void Lint::visitCallInst(CallInst &I) {
343   return visitCallSite(&I);
344 }
345 
visitInvokeInst(InvokeInst & I)346 void Lint::visitInvokeInst(InvokeInst &I) {
347   return visitCallSite(&I);
348 }
349 
visitReturnInst(ReturnInst & I)350 void Lint::visitReturnInst(ReturnInst &I) {
351   Function *F = I.getParent()->getParent();
352   Assert1(!F->doesNotReturn(),
353           "Unusual: Return statement in function with noreturn attribute",
354           &I);
355 
356   if (Value *V = I.getReturnValue()) {
357     Value *Obj = findValue(V, /*OffsetOk=*/true);
358     Assert1(!isa<AllocaInst>(Obj),
359             "Unusual: Returning alloca value", &I);
360   }
361 }
362 
363 // TODO: Check that the reference is in bounds.
364 // TODO: Check readnone/readonly function attributes.
visitMemoryReference(Instruction & I,Value * Ptr,uint64_t Size,unsigned Align,Type * Ty,unsigned Flags)365 void Lint::visitMemoryReference(Instruction &I,
366                                 Value *Ptr, uint64_t Size, unsigned Align,
367                                 Type *Ty, unsigned Flags) {
368   // If no memory is being referenced, it doesn't matter if the pointer
369   // is valid.
370   if (Size == 0)
371     return;
372 
373   Value *UnderlyingObject = findValue(Ptr, /*OffsetOk=*/true);
374   Assert1(!isa<ConstantPointerNull>(UnderlyingObject),
375           "Undefined behavior: Null pointer dereference", &I);
376   Assert1(!isa<UndefValue>(UnderlyingObject),
377           "Undefined behavior: Undef pointer dereference", &I);
378   Assert1(!isa<ConstantInt>(UnderlyingObject) ||
379           !cast<ConstantInt>(UnderlyingObject)->isAllOnesValue(),
380           "Unusual: All-ones pointer dereference", &I);
381   Assert1(!isa<ConstantInt>(UnderlyingObject) ||
382           !cast<ConstantInt>(UnderlyingObject)->isOne(),
383           "Unusual: Address one pointer dereference", &I);
384 
385   if (Flags & MemRef::Write) {
386     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(UnderlyingObject))
387       Assert1(!GV->isConstant(),
388               "Undefined behavior: Write to read-only memory", &I);
389     Assert1(!isa<Function>(UnderlyingObject) &&
390             !isa<BlockAddress>(UnderlyingObject),
391             "Undefined behavior: Write to text section", &I);
392   }
393   if (Flags & MemRef::Read) {
394     Assert1(!isa<Function>(UnderlyingObject),
395             "Unusual: Load from function body", &I);
396     Assert1(!isa<BlockAddress>(UnderlyingObject),
397             "Undefined behavior: Load from block address", &I);
398   }
399   if (Flags & MemRef::Callee) {
400     Assert1(!isa<BlockAddress>(UnderlyingObject),
401             "Undefined behavior: Call to block address", &I);
402   }
403   if (Flags & MemRef::Branchee) {
404     Assert1(!isa<Constant>(UnderlyingObject) ||
405             isa<BlockAddress>(UnderlyingObject),
406             "Undefined behavior: Branch to non-blockaddress", &I);
407   }
408 
409   if (TD) {
410     if (Align == 0 && Ty) Align = TD->getABITypeAlignment(Ty);
411 
412     if (Align != 0) {
413       unsigned BitWidth = TD->getTypeSizeInBits(Ptr->getType());
414       APInt Mask = APInt::getAllOnesValue(BitWidth),
415                    KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
416       ComputeMaskedBits(Ptr, Mask, KnownZero, KnownOne, TD);
417       Assert1(!(KnownOne & APInt::getLowBitsSet(BitWidth, Log2_32(Align))),
418               "Undefined behavior: Memory reference address is misaligned", &I);
419     }
420   }
421 }
422 
visitLoadInst(LoadInst & I)423 void Lint::visitLoadInst(LoadInst &I) {
424   visitMemoryReference(I, I.getPointerOperand(),
425                        AA->getTypeStoreSize(I.getType()), I.getAlignment(),
426                        I.getType(), MemRef::Read);
427 }
428 
visitStoreInst(StoreInst & I)429 void Lint::visitStoreInst(StoreInst &I) {
430   visitMemoryReference(I, I.getPointerOperand(),
431                        AA->getTypeStoreSize(I.getOperand(0)->getType()),
432                        I.getAlignment(),
433                        I.getOperand(0)->getType(), MemRef::Write);
434 }
435 
visitXor(BinaryOperator & I)436 void Lint::visitXor(BinaryOperator &I) {
437   Assert1(!isa<UndefValue>(I.getOperand(0)) ||
438           !isa<UndefValue>(I.getOperand(1)),
439           "Undefined result: xor(undef, undef)", &I);
440 }
441 
visitSub(BinaryOperator & I)442 void Lint::visitSub(BinaryOperator &I) {
443   Assert1(!isa<UndefValue>(I.getOperand(0)) ||
444           !isa<UndefValue>(I.getOperand(1)),
445           "Undefined result: sub(undef, undef)", &I);
446 }
447 
visitLShr(BinaryOperator & I)448 void Lint::visitLShr(BinaryOperator &I) {
449   if (ConstantInt *CI =
450         dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
451     Assert1(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
452             "Undefined result: Shift count out of range", &I);
453 }
454 
visitAShr(BinaryOperator & I)455 void Lint::visitAShr(BinaryOperator &I) {
456   if (ConstantInt *CI =
457         dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
458     Assert1(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
459             "Undefined result: Shift count out of range", &I);
460 }
461 
visitShl(BinaryOperator & I)462 void Lint::visitShl(BinaryOperator &I) {
463   if (ConstantInt *CI =
464         dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
465     Assert1(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
466             "Undefined result: Shift count out of range", &I);
467 }
468 
isZero(Value * V,TargetData * TD)469 static bool isZero(Value *V, TargetData *TD) {
470   // Assume undef could be zero.
471   if (isa<UndefValue>(V)) return true;
472 
473   unsigned BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
474   APInt Mask = APInt::getAllOnesValue(BitWidth),
475                KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
476   ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD);
477   return KnownZero.isAllOnesValue();
478 }
479 
visitSDiv(BinaryOperator & I)480 void Lint::visitSDiv(BinaryOperator &I) {
481   Assert1(!isZero(I.getOperand(1), TD),
482           "Undefined behavior: Division by zero", &I);
483 }
484 
visitUDiv(BinaryOperator & I)485 void Lint::visitUDiv(BinaryOperator &I) {
486   Assert1(!isZero(I.getOperand(1), TD),
487           "Undefined behavior: Division by zero", &I);
488 }
489 
visitSRem(BinaryOperator & I)490 void Lint::visitSRem(BinaryOperator &I) {
491   Assert1(!isZero(I.getOperand(1), TD),
492           "Undefined behavior: Division by zero", &I);
493 }
494 
visitURem(BinaryOperator & I)495 void Lint::visitURem(BinaryOperator &I) {
496   Assert1(!isZero(I.getOperand(1), TD),
497           "Undefined behavior: Division by zero", &I);
498 }
499 
visitAllocaInst(AllocaInst & I)500 void Lint::visitAllocaInst(AllocaInst &I) {
501   if (isa<ConstantInt>(I.getArraySize()))
502     // This isn't undefined behavior, it's just an obvious pessimization.
503     Assert1(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
504             "Pessimization: Static alloca outside of entry block", &I);
505 
506   // TODO: Check for an unusual size (MSB set?)
507 }
508 
visitVAArgInst(VAArgInst & I)509 void Lint::visitVAArgInst(VAArgInst &I) {
510   visitMemoryReference(I, I.getOperand(0), AliasAnalysis::UnknownSize, 0, 0,
511                        MemRef::Read | MemRef::Write);
512 }
513 
visitIndirectBrInst(IndirectBrInst & I)514 void Lint::visitIndirectBrInst(IndirectBrInst &I) {
515   visitMemoryReference(I, I.getAddress(), AliasAnalysis::UnknownSize, 0, 0,
516                        MemRef::Branchee);
517 
518   Assert1(I.getNumDestinations() != 0,
519           "Undefined behavior: indirectbr with no destinations", &I);
520 }
521 
visitExtractElementInst(ExtractElementInst & I)522 void Lint::visitExtractElementInst(ExtractElementInst &I) {
523   if (ConstantInt *CI =
524         dyn_cast<ConstantInt>(findValue(I.getIndexOperand(),
525                                         /*OffsetOk=*/false)))
526     Assert1(CI->getValue().ult(I.getVectorOperandType()->getNumElements()),
527             "Undefined result: extractelement index out of range", &I);
528 }
529 
visitInsertElementInst(InsertElementInst & I)530 void Lint::visitInsertElementInst(InsertElementInst &I) {
531   if (ConstantInt *CI =
532         dyn_cast<ConstantInt>(findValue(I.getOperand(2),
533                                         /*OffsetOk=*/false)))
534     Assert1(CI->getValue().ult(I.getType()->getNumElements()),
535             "Undefined result: insertelement index out of range", &I);
536 }
537 
visitUnreachableInst(UnreachableInst & I)538 void Lint::visitUnreachableInst(UnreachableInst &I) {
539   // This isn't undefined behavior, it's merely suspicious.
540   Assert1(&I == I.getParent()->begin() ||
541           prior(BasicBlock::iterator(&I))->mayHaveSideEffects(),
542           "Unusual: unreachable immediately preceded by instruction without "
543           "side effects", &I);
544 }
545 
546 /// findValue - Look through bitcasts and simple memory reference patterns
547 /// to identify an equivalent, but more informative, value.  If OffsetOk
548 /// is true, look through getelementptrs with non-zero offsets too.
549 ///
550 /// Most analysis passes don't require this logic, because instcombine
551 /// will simplify most of these kinds of things away. But it's a goal of
552 /// this Lint pass to be useful even on non-optimized IR.
findValue(Value * V,bool OffsetOk) const553 Value *Lint::findValue(Value *V, bool OffsetOk) const {
554   SmallPtrSet<Value *, 4> Visited;
555   return findValueImpl(V, OffsetOk, Visited);
556 }
557 
558 /// findValueImpl - Implementation helper for findValue.
findValueImpl(Value * V,bool OffsetOk,SmallPtrSet<Value *,4> & Visited) const559 Value *Lint::findValueImpl(Value *V, bool OffsetOk,
560                            SmallPtrSet<Value *, 4> &Visited) const {
561   // Detect self-referential values.
562   if (!Visited.insert(V))
563     return UndefValue::get(V->getType());
564 
565   // TODO: Look through sext or zext cast, when the result is known to
566   // be interpreted as signed or unsigned, respectively.
567   // TODO: Look through eliminable cast pairs.
568   // TODO: Look through calls with unique return values.
569   // TODO: Look through vector insert/extract/shuffle.
570   V = OffsetOk ? GetUnderlyingObject(V, TD) : V->stripPointerCasts();
571   if (LoadInst *L = dyn_cast<LoadInst>(V)) {
572     BasicBlock::iterator BBI = L;
573     BasicBlock *BB = L->getParent();
574     SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
575     for (;;) {
576       if (!VisitedBlocks.insert(BB)) break;
577       if (Value *U = FindAvailableLoadedValue(L->getPointerOperand(),
578                                               BB, BBI, 6, AA))
579         return findValueImpl(U, OffsetOk, Visited);
580       if (BBI != BB->begin()) break;
581       BB = BB->getUniquePredecessor();
582       if (!BB) break;
583       BBI = BB->end();
584     }
585   } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
586     if (Value *W = PN->hasConstantValue())
587       if (W != V)
588         return findValueImpl(W, OffsetOk, Visited);
589   } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
590     if (CI->isNoopCast(TD ? TD->getIntPtrType(V->getContext()) :
591                             Type::getInt64Ty(V->getContext())))
592       return findValueImpl(CI->getOperand(0), OffsetOk, Visited);
593   } else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(V)) {
594     if (Value *W = FindInsertedValue(Ex->getAggregateOperand(),
595                                      Ex->getIndices()))
596       if (W != V)
597         return findValueImpl(W, OffsetOk, Visited);
598   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
599     // Same as above, but for ConstantExpr instead of Instruction.
600     if (Instruction::isCast(CE->getOpcode())) {
601       if (CastInst::isNoopCast(Instruction::CastOps(CE->getOpcode()),
602                                CE->getOperand(0)->getType(),
603                                CE->getType(),
604                                TD ? TD->getIntPtrType(V->getContext()) :
605                                     Type::getInt64Ty(V->getContext())))
606         return findValueImpl(CE->getOperand(0), OffsetOk, Visited);
607     } else if (CE->getOpcode() == Instruction::ExtractValue) {
608       ArrayRef<unsigned> Indices = CE->getIndices();
609       if (Value *W = FindInsertedValue(CE->getOperand(0), Indices))
610         if (W != V)
611           return findValueImpl(W, OffsetOk, Visited);
612     }
613   }
614 
615   // As a last resort, try SimplifyInstruction or constant folding.
616   if (Instruction *Inst = dyn_cast<Instruction>(V)) {
617     if (Value *W = SimplifyInstruction(Inst, TD, DT))
618       return findValueImpl(W, OffsetOk, Visited);
619   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
620     if (Value *W = ConstantFoldConstantExpression(CE, TD))
621       if (W != V)
622         return findValueImpl(W, OffsetOk, Visited);
623   }
624 
625   return V;
626 }
627 
628 //===----------------------------------------------------------------------===//
629 //  Implement the public interfaces to this file...
630 //===----------------------------------------------------------------------===//
631 
createLintPass()632 FunctionPass *llvm::createLintPass() {
633   return new Lint();
634 }
635 
636 /// lintFunction - Check a function for errors, printing messages on stderr.
637 ///
lintFunction(const Function & f)638 void llvm::lintFunction(const Function &f) {
639   Function &F = const_cast<Function&>(f);
640   assert(!F.isDeclaration() && "Cannot lint external functions");
641 
642   FunctionPassManager FPM(F.getParent());
643   Lint *V = new Lint();
644   FPM.add(V);
645   FPM.run(F);
646 }
647 
648 /// lintModule - Check a module for errors, printing messages on stderr.
649 ///
lintModule(const Module & M)650 void llvm::lintModule(const Module &M) {
651   PassManager PM;
652   Lint *V = new Lint();
653   PM.add(V);
654   PM.run(const_cast<Module&>(M));
655 }
656