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/Lint.h"
38 #include "llvm/ADT/STLExtras.h"
39 #include "llvm/Analysis/AliasAnalysis.h"
40 #include "llvm/Analysis/ConstantFolding.h"
41 #include "llvm/Analysis/InstructionSimplify.h"
42 #include "llvm/Analysis/Loads.h"
43 #include "llvm/Analysis/Passes.h"
44 #include "llvm/Analysis/ValueTracking.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/DataLayout.h"
47 #include "llvm/IR/Dominators.h"
48 #include "llvm/IR/Function.h"
49 #include "llvm/IR/InstVisitor.h"
50 #include "llvm/IR/IntrinsicInst.h"
51 #include "llvm/Pass.h"
52 #include "llvm/PassManager.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Target/TargetLibraryInfo.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 const DataLayout *DL;
106 TargetLibraryInfo *TLI;
107
108 std::string Messages;
109 raw_string_ostream MessagesStr;
110
111 static char ID; // Pass identification, replacement for typeid
Lint()112 Lint() : FunctionPass(ID), MessagesStr(Messages) {
113 initializeLintPass(*PassRegistry::getPassRegistry());
114 }
115
116 bool runOnFunction(Function &F) override;
117
getAnalysisUsage(AnalysisUsage & AU) const118 void getAnalysisUsage(AnalysisUsage &AU) const override {
119 AU.setPreservesAll();
120 AU.addRequired<AliasAnalysis>();
121 AU.addRequired<TargetLibraryInfo>();
122 AU.addRequired<DominatorTreeWrapperPass>();
123 }
print(raw_ostream & O,const Module * M) const124 void print(raw_ostream &O, const Module *M) const override {}
125
WriteValue(const Value * V)126 void WriteValue(const Value *V) {
127 if (!V) return;
128 if (isa<Instruction>(V)) {
129 MessagesStr << *V << '\n';
130 } else {
131 V->printAsOperand(MessagesStr, true, Mod);
132 MessagesStr << '\n';
133 }
134 }
135
136 // CheckFailed - A check failed, so print out the condition and the message
137 // that failed. This provides a nice place to put a breakpoint if you want
138 // to see why something is not correct.
CheckFailed(const Twine & Message,const Value * V1=nullptr,const Value * V2=nullptr,const Value * V3=nullptr,const Value * V4=nullptr)139 void CheckFailed(const Twine &Message,
140 const Value *V1 = nullptr, const Value *V2 = nullptr,
141 const Value *V3 = nullptr, const Value *V4 = nullptr) {
142 MessagesStr << Message.str() << "\n";
143 WriteValue(V1);
144 WriteValue(V2);
145 WriteValue(V3);
146 WriteValue(V4);
147 }
148 };
149 }
150
151 char Lint::ID = 0;
152 INITIALIZE_PASS_BEGIN(Lint, "lint", "Statically lint-checks LLVM IR",
153 false, true)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)154 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
155 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
156 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
157 INITIALIZE_PASS_END(Lint, "lint", "Statically lint-checks LLVM IR",
158 false, true)
159
160 // Assert - We know that cond should be true, if not print an error message.
161 #define Assert(C, M) \
162 do { if (!(C)) { CheckFailed(M); return; } } while (0)
163 #define Assert1(C, M, V1) \
164 do { if (!(C)) { CheckFailed(M, V1); return; } } while (0)
165 #define Assert2(C, M, V1, V2) \
166 do { if (!(C)) { CheckFailed(M, V1, V2); return; } } while (0)
167 #define Assert3(C, M, V1, V2, V3) \
168 do { if (!(C)) { CheckFailed(M, V1, V2, V3); return; } } while (0)
169 #define Assert4(C, M, V1, V2, V3, V4) \
170 do { if (!(C)) { CheckFailed(M, V1, V2, V3, V4); return; } } while (0)
171
172 // Lint::run - This is the main Analysis entry point for a
173 // function.
174 //
175 bool Lint::runOnFunction(Function &F) {
176 Mod = F.getParent();
177 AA = &getAnalysis<AliasAnalysis>();
178 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
179 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
180 DL = DLP ? &DLP->getDataLayout() : nullptr;
181 TLI = &getAnalysis<TargetLibraryInfo>();
182 visit(F);
183 dbgs() << MessagesStr.str();
184 Messages.clear();
185 return false;
186 }
187
visitFunction(Function & F)188 void Lint::visitFunction(Function &F) {
189 // This isn't undefined behavior, it's just a little unusual, and it's a
190 // fairly common mistake to neglect to name a function.
191 Assert1(F.hasName() || F.hasLocalLinkage(),
192 "Unusual: Unnamed function with non-local linkage", &F);
193
194 // TODO: Check for irreducible control flow.
195 }
196
visitCallSite(CallSite CS)197 void Lint::visitCallSite(CallSite CS) {
198 Instruction &I = *CS.getInstruction();
199 Value *Callee = CS.getCalledValue();
200
201 visitMemoryReference(I, Callee, AliasAnalysis::UnknownSize,
202 0, nullptr, MemRef::Callee);
203
204 if (Function *F = dyn_cast<Function>(findValue(Callee, /*OffsetOk=*/false))) {
205 Assert1(CS.getCallingConv() == F->getCallingConv(),
206 "Undefined behavior: Caller and callee calling convention differ",
207 &I);
208
209 FunctionType *FT = F->getFunctionType();
210 unsigned NumActualArgs = CS.arg_size();
211
212 Assert1(FT->isVarArg() ?
213 FT->getNumParams() <= NumActualArgs :
214 FT->getNumParams() == NumActualArgs,
215 "Undefined behavior: Call argument count mismatches callee "
216 "argument count", &I);
217
218 Assert1(FT->getReturnType() == I.getType(),
219 "Undefined behavior: Call return type mismatches "
220 "callee return type", &I);
221
222 // Check argument types (in case the callee was casted) and attributes.
223 // TODO: Verify that caller and callee attributes are compatible.
224 Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
225 CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
226 for (; AI != AE; ++AI) {
227 Value *Actual = *AI;
228 if (PI != PE) {
229 Argument *Formal = PI++;
230 Assert1(Formal->getType() == Actual->getType(),
231 "Undefined behavior: Call argument type mismatches "
232 "callee parameter type", &I);
233
234 // Check that noalias arguments don't alias other arguments. This is
235 // not fully precise because we don't know the sizes of the dereferenced
236 // memory regions.
237 if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy())
238 for (CallSite::arg_iterator BI = CS.arg_begin(); BI != AE; ++BI)
239 if (AI != BI && (*BI)->getType()->isPointerTy()) {
240 AliasAnalysis::AliasResult Result = AA->alias(*AI, *BI);
241 Assert1(Result != AliasAnalysis::MustAlias &&
242 Result != AliasAnalysis::PartialAlias,
243 "Unusual: noalias argument aliases another argument", &I);
244 }
245
246 // Check that an sret argument points to valid memory.
247 if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
248 Type *Ty =
249 cast<PointerType>(Formal->getType())->getElementType();
250 visitMemoryReference(I, Actual, AA->getTypeStoreSize(Ty),
251 DL ? DL->getABITypeAlignment(Ty) : 0,
252 Ty, MemRef::Read | MemRef::Write);
253 }
254 }
255 }
256 }
257
258 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->isTailCall())
259 for (CallSite::arg_iterator AI = CS.arg_begin(), AE = CS.arg_end();
260 AI != AE; ++AI) {
261 Value *Obj = findValue(*AI, /*OffsetOk=*/true);
262 Assert1(!isa<AllocaInst>(Obj),
263 "Undefined behavior: Call with \"tail\" keyword references "
264 "alloca", &I);
265 }
266
267
268 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
269 switch (II->getIntrinsicID()) {
270 default: break;
271
272 // TODO: Check more intrinsics
273
274 case Intrinsic::memcpy: {
275 MemCpyInst *MCI = cast<MemCpyInst>(&I);
276 // TODO: If the size is known, use it.
277 visitMemoryReference(I, MCI->getDest(), AliasAnalysis::UnknownSize,
278 MCI->getAlignment(), nullptr,
279 MemRef::Write);
280 visitMemoryReference(I, MCI->getSource(), AliasAnalysis::UnknownSize,
281 MCI->getAlignment(), nullptr,
282 MemRef::Read);
283
284 // Check that the memcpy arguments don't overlap. The AliasAnalysis API
285 // isn't expressive enough for what we really want to do. Known partial
286 // overlap is not distinguished from the case where nothing is known.
287 uint64_t Size = 0;
288 if (const ConstantInt *Len =
289 dyn_cast<ConstantInt>(findValue(MCI->getLength(),
290 /*OffsetOk=*/false)))
291 if (Len->getValue().isIntN(32))
292 Size = Len->getValue().getZExtValue();
293 Assert1(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
294 AliasAnalysis::MustAlias,
295 "Undefined behavior: memcpy source and destination overlap", &I);
296 break;
297 }
298 case Intrinsic::memmove: {
299 MemMoveInst *MMI = cast<MemMoveInst>(&I);
300 // TODO: If the size is known, use it.
301 visitMemoryReference(I, MMI->getDest(), AliasAnalysis::UnknownSize,
302 MMI->getAlignment(), nullptr,
303 MemRef::Write);
304 visitMemoryReference(I, MMI->getSource(), AliasAnalysis::UnknownSize,
305 MMI->getAlignment(), nullptr,
306 MemRef::Read);
307 break;
308 }
309 case Intrinsic::memset: {
310 MemSetInst *MSI = cast<MemSetInst>(&I);
311 // TODO: If the size is known, use it.
312 visitMemoryReference(I, MSI->getDest(), AliasAnalysis::UnknownSize,
313 MSI->getAlignment(), nullptr,
314 MemRef::Write);
315 break;
316 }
317
318 case Intrinsic::vastart:
319 Assert1(I.getParent()->getParent()->isVarArg(),
320 "Undefined behavior: va_start called in a non-varargs function",
321 &I);
322
323 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
324 0, nullptr, MemRef::Read | MemRef::Write);
325 break;
326 case Intrinsic::vacopy:
327 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
328 0, nullptr, MemRef::Write);
329 visitMemoryReference(I, CS.getArgument(1), AliasAnalysis::UnknownSize,
330 0, nullptr, MemRef::Read);
331 break;
332 case Intrinsic::vaend:
333 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
334 0, nullptr, MemRef::Read | MemRef::Write);
335 break;
336
337 case Intrinsic::stackrestore:
338 // Stackrestore doesn't read or write memory, but it sets the
339 // stack pointer, which the compiler may read from or write to
340 // at any time, so check it for both readability and writeability.
341 visitMemoryReference(I, CS.getArgument(0), AliasAnalysis::UnknownSize,
342 0, nullptr, MemRef::Read | MemRef::Write);
343 break;
344 }
345 }
346
visitCallInst(CallInst & I)347 void Lint::visitCallInst(CallInst &I) {
348 return visitCallSite(&I);
349 }
350
visitInvokeInst(InvokeInst & I)351 void Lint::visitInvokeInst(InvokeInst &I) {
352 return visitCallSite(&I);
353 }
354
visitReturnInst(ReturnInst & I)355 void Lint::visitReturnInst(ReturnInst &I) {
356 Function *F = I.getParent()->getParent();
357 Assert1(!F->doesNotReturn(),
358 "Unusual: Return statement in function with noreturn attribute",
359 &I);
360
361 if (Value *V = I.getReturnValue()) {
362 Value *Obj = findValue(V, /*OffsetOk=*/true);
363 Assert1(!isa<AllocaInst>(Obj),
364 "Unusual: Returning alloca value", &I);
365 }
366 }
367
368 // TODO: Check that the reference is in bounds.
369 // TODO: Check readnone/readonly function attributes.
visitMemoryReference(Instruction & I,Value * Ptr,uint64_t Size,unsigned Align,Type * Ty,unsigned Flags)370 void Lint::visitMemoryReference(Instruction &I,
371 Value *Ptr, uint64_t Size, unsigned Align,
372 Type *Ty, unsigned Flags) {
373 // If no memory is being referenced, it doesn't matter if the pointer
374 // is valid.
375 if (Size == 0)
376 return;
377
378 Value *UnderlyingObject = findValue(Ptr, /*OffsetOk=*/true);
379 Assert1(!isa<ConstantPointerNull>(UnderlyingObject),
380 "Undefined behavior: Null pointer dereference", &I);
381 Assert1(!isa<UndefValue>(UnderlyingObject),
382 "Undefined behavior: Undef pointer dereference", &I);
383 Assert1(!isa<ConstantInt>(UnderlyingObject) ||
384 !cast<ConstantInt>(UnderlyingObject)->isAllOnesValue(),
385 "Unusual: All-ones pointer dereference", &I);
386 Assert1(!isa<ConstantInt>(UnderlyingObject) ||
387 !cast<ConstantInt>(UnderlyingObject)->isOne(),
388 "Unusual: Address one pointer dereference", &I);
389
390 if (Flags & MemRef::Write) {
391 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(UnderlyingObject))
392 Assert1(!GV->isConstant(),
393 "Undefined behavior: Write to read-only memory", &I);
394 Assert1(!isa<Function>(UnderlyingObject) &&
395 !isa<BlockAddress>(UnderlyingObject),
396 "Undefined behavior: Write to text section", &I);
397 }
398 if (Flags & MemRef::Read) {
399 Assert1(!isa<Function>(UnderlyingObject),
400 "Unusual: Load from function body", &I);
401 Assert1(!isa<BlockAddress>(UnderlyingObject),
402 "Undefined behavior: Load from block address", &I);
403 }
404 if (Flags & MemRef::Callee) {
405 Assert1(!isa<BlockAddress>(UnderlyingObject),
406 "Undefined behavior: Call to block address", &I);
407 }
408 if (Flags & MemRef::Branchee) {
409 Assert1(!isa<Constant>(UnderlyingObject) ||
410 isa<BlockAddress>(UnderlyingObject),
411 "Undefined behavior: Branch to non-blockaddress", &I);
412 }
413
414 // Check for buffer overflows and misalignment.
415 // Only handles memory references that read/write something simple like an
416 // alloca instruction or a global variable.
417 int64_t Offset = 0;
418 if (Value *Base = GetPointerBaseWithConstantOffset(Ptr, Offset, DL)) {
419 // OK, so the access is to a constant offset from Ptr. Check that Ptr is
420 // something we can handle and if so extract the size of this base object
421 // along with its alignment.
422 uint64_t BaseSize = AliasAnalysis::UnknownSize;
423 unsigned BaseAlign = 0;
424
425 if (AllocaInst *AI = dyn_cast<AllocaInst>(Base)) {
426 Type *ATy = AI->getAllocatedType();
427 if (DL && !AI->isArrayAllocation() && ATy->isSized())
428 BaseSize = DL->getTypeAllocSize(ATy);
429 BaseAlign = AI->getAlignment();
430 if (DL && BaseAlign == 0 && ATy->isSized())
431 BaseAlign = DL->getABITypeAlignment(ATy);
432 } else if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) {
433 // If the global may be defined differently in another compilation unit
434 // then don't warn about funky memory accesses.
435 if (GV->hasDefinitiveInitializer()) {
436 Type *GTy = GV->getType()->getElementType();
437 if (DL && GTy->isSized())
438 BaseSize = DL->getTypeAllocSize(GTy);
439 BaseAlign = GV->getAlignment();
440 if (DL && BaseAlign == 0 && GTy->isSized())
441 BaseAlign = DL->getABITypeAlignment(GTy);
442 }
443 }
444
445 // Accesses from before the start or after the end of the object are not
446 // defined.
447 Assert1(Size == AliasAnalysis::UnknownSize ||
448 BaseSize == AliasAnalysis::UnknownSize ||
449 (Offset >= 0 && Offset + Size <= BaseSize),
450 "Undefined behavior: Buffer overflow", &I);
451
452 // Accesses that say that the memory is more aligned than it is are not
453 // defined.
454 if (DL && Align == 0 && Ty && Ty->isSized())
455 Align = DL->getABITypeAlignment(Ty);
456 Assert1(!BaseAlign || Align <= MinAlign(BaseAlign, Offset),
457 "Undefined behavior: Memory reference address is misaligned", &I);
458 }
459 }
460
visitLoadInst(LoadInst & I)461 void Lint::visitLoadInst(LoadInst &I) {
462 visitMemoryReference(I, I.getPointerOperand(),
463 AA->getTypeStoreSize(I.getType()), I.getAlignment(),
464 I.getType(), MemRef::Read);
465 }
466
visitStoreInst(StoreInst & I)467 void Lint::visitStoreInst(StoreInst &I) {
468 visitMemoryReference(I, I.getPointerOperand(),
469 AA->getTypeStoreSize(I.getOperand(0)->getType()),
470 I.getAlignment(),
471 I.getOperand(0)->getType(), MemRef::Write);
472 }
473
visitXor(BinaryOperator & I)474 void Lint::visitXor(BinaryOperator &I) {
475 Assert1(!isa<UndefValue>(I.getOperand(0)) ||
476 !isa<UndefValue>(I.getOperand(1)),
477 "Undefined result: xor(undef, undef)", &I);
478 }
479
visitSub(BinaryOperator & I)480 void Lint::visitSub(BinaryOperator &I) {
481 Assert1(!isa<UndefValue>(I.getOperand(0)) ||
482 !isa<UndefValue>(I.getOperand(1)),
483 "Undefined result: sub(undef, undef)", &I);
484 }
485
visitLShr(BinaryOperator & I)486 void Lint::visitLShr(BinaryOperator &I) {
487 if (ConstantInt *CI =
488 dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
489 Assert1(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
490 "Undefined result: Shift count out of range", &I);
491 }
492
visitAShr(BinaryOperator & I)493 void Lint::visitAShr(BinaryOperator &I) {
494 if (ConstantInt *CI =
495 dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
496 Assert1(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
497 "Undefined result: Shift count out of range", &I);
498 }
499
visitShl(BinaryOperator & I)500 void Lint::visitShl(BinaryOperator &I) {
501 if (ConstantInt *CI =
502 dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
503 Assert1(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
504 "Undefined result: Shift count out of range", &I);
505 }
506
isZero(Value * V,const DataLayout * DL)507 static bool isZero(Value *V, const DataLayout *DL) {
508 // Assume undef could be zero.
509 if (isa<UndefValue>(V))
510 return true;
511
512 VectorType *VecTy = dyn_cast<VectorType>(V->getType());
513 if (!VecTy) {
514 unsigned BitWidth = V->getType()->getIntegerBitWidth();
515 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
516 computeKnownBits(V, KnownZero, KnownOne, DL);
517 return KnownZero.isAllOnesValue();
518 }
519
520 // Per-component check doesn't work with zeroinitializer
521 Constant *C = dyn_cast<Constant>(V);
522 if (!C)
523 return false;
524
525 if (C->isZeroValue())
526 return true;
527
528 // For a vector, KnownZero will only be true if all values are zero, so check
529 // this per component
530 unsigned BitWidth = VecTy->getElementType()->getIntegerBitWidth();
531 for (unsigned I = 0, N = VecTy->getNumElements(); I != N; ++I) {
532 Constant *Elem = C->getAggregateElement(I);
533 if (isa<UndefValue>(Elem))
534 return true;
535
536 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
537 computeKnownBits(Elem, KnownZero, KnownOne, DL);
538 if (KnownZero.isAllOnesValue())
539 return true;
540 }
541
542 return false;
543 }
544
visitSDiv(BinaryOperator & I)545 void Lint::visitSDiv(BinaryOperator &I) {
546 Assert1(!isZero(I.getOperand(1), DL),
547 "Undefined behavior: Division by zero", &I);
548 }
549
visitUDiv(BinaryOperator & I)550 void Lint::visitUDiv(BinaryOperator &I) {
551 Assert1(!isZero(I.getOperand(1), DL),
552 "Undefined behavior: Division by zero", &I);
553 }
554
visitSRem(BinaryOperator & I)555 void Lint::visitSRem(BinaryOperator &I) {
556 Assert1(!isZero(I.getOperand(1), DL),
557 "Undefined behavior: Division by zero", &I);
558 }
559
visitURem(BinaryOperator & I)560 void Lint::visitURem(BinaryOperator &I) {
561 Assert1(!isZero(I.getOperand(1), DL),
562 "Undefined behavior: Division by zero", &I);
563 }
564
visitAllocaInst(AllocaInst & I)565 void Lint::visitAllocaInst(AllocaInst &I) {
566 if (isa<ConstantInt>(I.getArraySize()))
567 // This isn't undefined behavior, it's just an obvious pessimization.
568 Assert1(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
569 "Pessimization: Static alloca outside of entry block", &I);
570
571 // TODO: Check for an unusual size (MSB set?)
572 }
573
visitVAArgInst(VAArgInst & I)574 void Lint::visitVAArgInst(VAArgInst &I) {
575 visitMemoryReference(I, I.getOperand(0), AliasAnalysis::UnknownSize, 0,
576 nullptr, MemRef::Read | MemRef::Write);
577 }
578
visitIndirectBrInst(IndirectBrInst & I)579 void Lint::visitIndirectBrInst(IndirectBrInst &I) {
580 visitMemoryReference(I, I.getAddress(), AliasAnalysis::UnknownSize, 0,
581 nullptr, MemRef::Branchee);
582
583 Assert1(I.getNumDestinations() != 0,
584 "Undefined behavior: indirectbr with no destinations", &I);
585 }
586
visitExtractElementInst(ExtractElementInst & I)587 void Lint::visitExtractElementInst(ExtractElementInst &I) {
588 if (ConstantInt *CI =
589 dyn_cast<ConstantInt>(findValue(I.getIndexOperand(),
590 /*OffsetOk=*/false)))
591 Assert1(CI->getValue().ult(I.getVectorOperandType()->getNumElements()),
592 "Undefined result: extractelement index out of range", &I);
593 }
594
visitInsertElementInst(InsertElementInst & I)595 void Lint::visitInsertElementInst(InsertElementInst &I) {
596 if (ConstantInt *CI =
597 dyn_cast<ConstantInt>(findValue(I.getOperand(2),
598 /*OffsetOk=*/false)))
599 Assert1(CI->getValue().ult(I.getType()->getNumElements()),
600 "Undefined result: insertelement index out of range", &I);
601 }
602
visitUnreachableInst(UnreachableInst & I)603 void Lint::visitUnreachableInst(UnreachableInst &I) {
604 // This isn't undefined behavior, it's merely suspicious.
605 Assert1(&I == I.getParent()->begin() ||
606 std::prev(BasicBlock::iterator(&I))->mayHaveSideEffects(),
607 "Unusual: unreachable immediately preceded by instruction without "
608 "side effects", &I);
609 }
610
611 /// findValue - Look through bitcasts and simple memory reference patterns
612 /// to identify an equivalent, but more informative, value. If OffsetOk
613 /// is true, look through getelementptrs with non-zero offsets too.
614 ///
615 /// Most analysis passes don't require this logic, because instcombine
616 /// will simplify most of these kinds of things away. But it's a goal of
617 /// this Lint pass to be useful even on non-optimized IR.
findValue(Value * V,bool OffsetOk) const618 Value *Lint::findValue(Value *V, bool OffsetOk) const {
619 SmallPtrSet<Value *, 4> Visited;
620 return findValueImpl(V, OffsetOk, Visited);
621 }
622
623 /// findValueImpl - Implementation helper for findValue.
findValueImpl(Value * V,bool OffsetOk,SmallPtrSet<Value *,4> & Visited) const624 Value *Lint::findValueImpl(Value *V, bool OffsetOk,
625 SmallPtrSet<Value *, 4> &Visited) const {
626 // Detect self-referential values.
627 if (!Visited.insert(V))
628 return UndefValue::get(V->getType());
629
630 // TODO: Look through sext or zext cast, when the result is known to
631 // be interpreted as signed or unsigned, respectively.
632 // TODO: Look through eliminable cast pairs.
633 // TODO: Look through calls with unique return values.
634 // TODO: Look through vector insert/extract/shuffle.
635 V = OffsetOk ? GetUnderlyingObject(V, DL) : V->stripPointerCasts();
636 if (LoadInst *L = dyn_cast<LoadInst>(V)) {
637 BasicBlock::iterator BBI = L;
638 BasicBlock *BB = L->getParent();
639 SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
640 for (;;) {
641 if (!VisitedBlocks.insert(BB)) break;
642 if (Value *U = FindAvailableLoadedValue(L->getPointerOperand(),
643 BB, BBI, 6, AA))
644 return findValueImpl(U, OffsetOk, Visited);
645 if (BBI != BB->begin()) break;
646 BB = BB->getUniquePredecessor();
647 if (!BB) break;
648 BBI = BB->end();
649 }
650 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
651 if (Value *W = PN->hasConstantValue())
652 if (W != V)
653 return findValueImpl(W, OffsetOk, Visited);
654 } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
655 if (CI->isNoopCast(DL))
656 return findValueImpl(CI->getOperand(0), OffsetOk, Visited);
657 } else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(V)) {
658 if (Value *W = FindInsertedValue(Ex->getAggregateOperand(),
659 Ex->getIndices()))
660 if (W != V)
661 return findValueImpl(W, OffsetOk, Visited);
662 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
663 // Same as above, but for ConstantExpr instead of Instruction.
664 if (Instruction::isCast(CE->getOpcode())) {
665 if (CastInst::isNoopCast(Instruction::CastOps(CE->getOpcode()),
666 CE->getOperand(0)->getType(),
667 CE->getType(),
668 DL ? DL->getIntPtrType(V->getType()) :
669 Type::getInt64Ty(V->getContext())))
670 return findValueImpl(CE->getOperand(0), OffsetOk, Visited);
671 } else if (CE->getOpcode() == Instruction::ExtractValue) {
672 ArrayRef<unsigned> Indices = CE->getIndices();
673 if (Value *W = FindInsertedValue(CE->getOperand(0), Indices))
674 if (W != V)
675 return findValueImpl(W, OffsetOk, Visited);
676 }
677 }
678
679 // As a last resort, try SimplifyInstruction or constant folding.
680 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
681 if (Value *W = SimplifyInstruction(Inst, DL, TLI, DT))
682 return findValueImpl(W, OffsetOk, Visited);
683 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
684 if (Value *W = ConstantFoldConstantExpression(CE, DL, TLI))
685 if (W != V)
686 return findValueImpl(W, OffsetOk, Visited);
687 }
688
689 return V;
690 }
691
692 //===----------------------------------------------------------------------===//
693 // Implement the public interfaces to this file...
694 //===----------------------------------------------------------------------===//
695
createLintPass()696 FunctionPass *llvm::createLintPass() {
697 return new Lint();
698 }
699
700 /// lintFunction - Check a function for errors, printing messages on stderr.
701 ///
lintFunction(const Function & f)702 void llvm::lintFunction(const Function &f) {
703 Function &F = const_cast<Function&>(f);
704 assert(!F.isDeclaration() && "Cannot lint external functions");
705
706 FunctionPassManager FPM(F.getParent());
707 Lint *V = new Lint();
708 FPM.add(V);
709 FPM.run(F);
710 }
711
712 /// lintModule - Check a module for errors, printing messages on stderr.
713 ///
lintModule(const Module & M)714 void llvm::lintModule(const Module &M) {
715 PassManager PM;
716 Lint *V = new Lint();
717 PM.add(V);
718 PM.run(const_cast<Module&>(M));
719 }
720