1 //===-- Lint.cpp - Check for common errors in LLVM IR ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass statically checks for common and easily-identified constructs
10 // which produce undefined or likely unintended behavior in LLVM IR.
11 //
12 // It is not a guarantee of correctness, in two ways. First, it isn't
13 // comprehensive. There are checks which could be done statically which are
14 // not yet implemented. Some of these are indicated by TODO comments, but
15 // those aren't comprehensive either. Second, many conditions cannot be
16 // checked statically. This pass does no dynamic instrumentation, so it
17 // can't check for all possible problems.
18 //
19 // Another limitation is that it assumes all code will be executed. A store
20 // through a null pointer in a basic block which is never reached is harmless,
21 // but this pass will warn about it anyway. This is the main reason why most
22 // of these checks live here instead of in the Verifier pass.
23 //
24 // Optimization passes may make conditions that this pass checks for more or
25 // less obvious. If an optimization pass appears to be introducing a warning,
26 // it may be that the optimization pass is merely exposing an existing
27 // condition in the code.
28 //
29 // This code may be run before instcombine. In many cases, instcombine checks
30 // for the same kinds of things and turns instructions with undefined behavior
31 // into unreachable (or equivalent). Because of this, this pass makes some
32 // effort to look through bitcasts and so on.
33 //
34 //===----------------------------------------------------------------------===//
35
36 #include "llvm/Analysis/Lint.h"
37 #include "llvm/ADT/APInt.h"
38 #include "llvm/ADT/ArrayRef.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/Twine.h"
41 #include "llvm/Analysis/AliasAnalysis.h"
42 #include "llvm/Analysis/AssumptionCache.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Analysis/InstructionSimplify.h"
45 #include "llvm/Analysis/Loads.h"
46 #include "llvm/Analysis/MemoryLocation.h"
47 #include "llvm/Analysis/Passes.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Analysis/ValueTracking.h"
50 #include "llvm/IR/Argument.h"
51 #include "llvm/IR/BasicBlock.h"
52 #include "llvm/IR/Constant.h"
53 #include "llvm/IR/Constants.h"
54 #include "llvm/IR/DataLayout.h"
55 #include "llvm/IR/DerivedTypes.h"
56 #include "llvm/IR/Dominators.h"
57 #include "llvm/IR/Function.h"
58 #include "llvm/IR/GlobalVariable.h"
59 #include "llvm/IR/InstVisitor.h"
60 #include "llvm/IR/InstrTypes.h"
61 #include "llvm/IR/Instruction.h"
62 #include "llvm/IR/Instructions.h"
63 #include "llvm/IR/IntrinsicInst.h"
64 #include "llvm/IR/LegacyPassManager.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/IR/PassManager.h"
67 #include "llvm/IR/Type.h"
68 #include "llvm/IR/Value.h"
69 #include "llvm/InitializePasses.h"
70 #include "llvm/Pass.h"
71 #include "llvm/Support/Casting.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/KnownBits.h"
74 #include "llvm/Support/MathExtras.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include <cassert>
77 #include <cstdint>
78 #include <iterator>
79 #include <string>
80
81 using namespace llvm;
82
83 namespace {
84 namespace MemRef {
85 static const unsigned Read = 1;
86 static const unsigned Write = 2;
87 static const unsigned Callee = 4;
88 static const unsigned Branchee = 8;
89 } // end namespace MemRef
90
91 class Lint : public InstVisitor<Lint> {
92 friend class InstVisitor<Lint>;
93
94 void visitFunction(Function &F);
95
96 void visitCallBase(CallBase &CB);
97 void visitMemoryReference(Instruction &I, const MemoryLocation &Loc,
98 MaybeAlign Alignment, Type *Ty, unsigned Flags);
99 void visitEHBeginCatch(IntrinsicInst *II);
100 void visitEHEndCatch(IntrinsicInst *II);
101
102 void visitReturnInst(ReturnInst &I);
103 void visitLoadInst(LoadInst &I);
104 void visitStoreInst(StoreInst &I);
105 void visitXor(BinaryOperator &I);
106 void visitSub(BinaryOperator &I);
107 void visitLShr(BinaryOperator &I);
108 void visitAShr(BinaryOperator &I);
109 void visitShl(BinaryOperator &I);
110 void visitSDiv(BinaryOperator &I);
111 void visitUDiv(BinaryOperator &I);
112 void visitSRem(BinaryOperator &I);
113 void visitURem(BinaryOperator &I);
114 void visitAllocaInst(AllocaInst &I);
115 void visitVAArgInst(VAArgInst &I);
116 void visitIndirectBrInst(IndirectBrInst &I);
117 void visitExtractElementInst(ExtractElementInst &I);
118 void visitInsertElementInst(InsertElementInst &I);
119 void visitUnreachableInst(UnreachableInst &I);
120
121 Value *findValue(Value *V, bool OffsetOk) const;
122 Value *findValueImpl(Value *V, bool OffsetOk,
123 SmallPtrSetImpl<Value *> &Visited) const;
124
125 public:
126 Module *Mod;
127 const DataLayout *DL;
128 AliasAnalysis *AA;
129 AssumptionCache *AC;
130 DominatorTree *DT;
131 TargetLibraryInfo *TLI;
132
133 std::string Messages;
134 raw_string_ostream MessagesStr;
135
Lint(Module * Mod,const DataLayout * DL,AliasAnalysis * AA,AssumptionCache * AC,DominatorTree * DT,TargetLibraryInfo * TLI)136 Lint(Module *Mod, const DataLayout *DL, AliasAnalysis *AA,
137 AssumptionCache *AC, DominatorTree *DT, TargetLibraryInfo *TLI)
138 : Mod(Mod), DL(DL), AA(AA), AC(AC), DT(DT), TLI(TLI),
139 MessagesStr(Messages) {}
140
WriteValues(ArrayRef<const Value * > Vs)141 void WriteValues(ArrayRef<const Value *> Vs) {
142 for (const Value *V : Vs) {
143 if (!V)
144 continue;
145 if (isa<Instruction>(V)) {
146 MessagesStr << *V << '\n';
147 } else {
148 V->printAsOperand(MessagesStr, true, Mod);
149 MessagesStr << '\n';
150 }
151 }
152 }
153
154 /// A check failed, so printout out the condition and the message.
155 ///
156 /// This provides a nice place to put a breakpoint if you want to see why
157 /// something is not correct.
CheckFailed(const Twine & Message)158 void CheckFailed(const Twine &Message) { MessagesStr << Message << '\n'; }
159
160 /// A check failed (with values to print).
161 ///
162 /// This calls the Message-only version so that the above is easier to set
163 /// a breakpoint on.
164 template <typename T1, typename... Ts>
CheckFailed(const Twine & Message,const T1 & V1,const Ts &...Vs)165 void CheckFailed(const Twine &Message, const T1 &V1, const Ts &... Vs) {
166 CheckFailed(Message);
167 WriteValues({V1, Vs...});
168 }
169 };
170 } // end anonymous namespace
171
172 // Assert - We know that cond should be true, if not print an error message.
173 #define Assert(C, ...) \
174 do { \
175 if (!(C)) { \
176 CheckFailed(__VA_ARGS__); \
177 return; \
178 } \
179 } while (false)
180
visitFunction(Function & F)181 void Lint::visitFunction(Function &F) {
182 // This isn't undefined behavior, it's just a little unusual, and it's a
183 // fairly common mistake to neglect to name a function.
184 Assert(F.hasName() || F.hasLocalLinkage(),
185 "Unusual: Unnamed function with non-local linkage", &F);
186
187 // TODO: Check for irreducible control flow.
188 }
189
visitCallBase(CallBase & I)190 void Lint::visitCallBase(CallBase &I) {
191 Value *Callee = I.getCalledOperand();
192
193 visitMemoryReference(I, MemoryLocation::getAfter(Callee), None, nullptr,
194 MemRef::Callee);
195
196 if (Function *F = dyn_cast<Function>(findValue(Callee,
197 /*OffsetOk=*/false))) {
198 Assert(I.getCallingConv() == F->getCallingConv(),
199 "Undefined behavior: Caller and callee calling convention differ",
200 &I);
201
202 FunctionType *FT = F->getFunctionType();
203 unsigned NumActualArgs = I.arg_size();
204
205 Assert(FT->isVarArg() ? FT->getNumParams() <= NumActualArgs
206 : FT->getNumParams() == NumActualArgs,
207 "Undefined behavior: Call argument count mismatches callee "
208 "argument count",
209 &I);
210
211 Assert(FT->getReturnType() == I.getType(),
212 "Undefined behavior: Call return type mismatches "
213 "callee return type",
214 &I);
215
216 // Check argument types (in case the callee was casted) and attributes.
217 // TODO: Verify that caller and callee attributes are compatible.
218 Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
219 auto AI = I.arg_begin(), AE = I.arg_end();
220 for (; AI != AE; ++AI) {
221 Value *Actual = *AI;
222 if (PI != PE) {
223 Argument *Formal = &*PI++;
224 Assert(Formal->getType() == Actual->getType(),
225 "Undefined behavior: Call argument type mismatches "
226 "callee parameter type",
227 &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 AttributeList PAL = I.getAttributes();
234 unsigned ArgNo = 0;
235 for (auto BI = I.arg_begin(); BI != AE; ++BI, ++ArgNo) {
236 // Skip ByVal arguments since they will be memcpy'd to the callee's
237 // stack so we're not really passing the pointer anyway.
238 if (PAL.hasParamAttribute(ArgNo, Attribute::ByVal))
239 continue;
240 // If both arguments are readonly, they have no dependence.
241 if (Formal->onlyReadsMemory() && I.onlyReadsMemory(ArgNo))
242 continue;
243 if (AI != BI && (*BI)->getType()->isPointerTy()) {
244 AliasResult Result = AA->alias(*AI, *BI);
245 Assert(Result != AliasResult::MustAlias &&
246 Result != AliasResult::PartialAlias,
247 "Unusual: noalias argument aliases another argument", &I);
248 }
249 }
250 }
251
252 // Check that an sret argument points to valid memory.
253 if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
254 Type *Ty = Formal->getParamStructRetType();
255 MemoryLocation Loc(
256 Actual, LocationSize::precise(DL->getTypeStoreSize(Ty)));
257 visitMemoryReference(I, Loc, DL->getABITypeAlign(Ty), Ty,
258 MemRef::Read | MemRef::Write);
259 }
260 }
261 }
262 }
263
264 if (const auto *CI = dyn_cast<CallInst>(&I)) {
265 if (CI->isTailCall()) {
266 const AttributeList &PAL = CI->getAttributes();
267 unsigned ArgNo = 0;
268 for (Value *Arg : I.args()) {
269 // Skip ByVal arguments since they will be memcpy'd to the callee's
270 // stack anyway.
271 if (PAL.hasParamAttribute(ArgNo++, Attribute::ByVal))
272 continue;
273 Value *Obj = findValue(Arg, /*OffsetOk=*/true);
274 Assert(!isa<AllocaInst>(Obj),
275 "Undefined behavior: Call with \"tail\" keyword references "
276 "alloca",
277 &I);
278 }
279 }
280 }
281
282 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I))
283 switch (II->getIntrinsicID()) {
284 default:
285 break;
286
287 // TODO: Check more intrinsics
288
289 case Intrinsic::memcpy: {
290 MemCpyInst *MCI = cast<MemCpyInst>(&I);
291 visitMemoryReference(I, MemoryLocation::getForDest(MCI),
292 MCI->getDestAlign(), nullptr, MemRef::Write);
293 visitMemoryReference(I, MemoryLocation::getForSource(MCI),
294 MCI->getSourceAlign(), nullptr, MemRef::Read);
295
296 // Check that the memcpy arguments don't overlap. The AliasAnalysis API
297 // isn't expressive enough for what we really want to do. Known partial
298 // overlap is not distinguished from the case where nothing is known.
299 auto Size = LocationSize::afterPointer();
300 if (const ConstantInt *Len =
301 dyn_cast<ConstantInt>(findValue(MCI->getLength(),
302 /*OffsetOk=*/false)))
303 if (Len->getValue().isIntN(32))
304 Size = LocationSize::precise(Len->getValue().getZExtValue());
305 Assert(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
306 AliasResult::MustAlias,
307 "Undefined behavior: memcpy source and destination overlap", &I);
308 break;
309 }
310 case Intrinsic::memcpy_inline: {
311 MemCpyInlineInst *MCII = cast<MemCpyInlineInst>(&I);
312 const uint64_t Size = MCII->getLength()->getValue().getLimitedValue();
313 visitMemoryReference(I, MemoryLocation::getForDest(MCII),
314 MCII->getDestAlign(), nullptr, MemRef::Write);
315 visitMemoryReference(I, MemoryLocation::getForSource(MCII),
316 MCII->getSourceAlign(), nullptr, MemRef::Read);
317
318 // Check that the memcpy arguments don't overlap. The AliasAnalysis API
319 // isn't expressive enough for what we really want to do. Known partial
320 // overlap is not distinguished from the case where nothing is known.
321 const LocationSize LS = LocationSize::precise(Size);
322 Assert(AA->alias(MCII->getSource(), LS, MCII->getDest(), LS) !=
323 AliasResult::MustAlias,
324 "Undefined behavior: memcpy source and destination overlap", &I);
325 break;
326 }
327 case Intrinsic::memmove: {
328 MemMoveInst *MMI = cast<MemMoveInst>(&I);
329 visitMemoryReference(I, MemoryLocation::getForDest(MMI),
330 MMI->getDestAlign(), nullptr, MemRef::Write);
331 visitMemoryReference(I, MemoryLocation::getForSource(MMI),
332 MMI->getSourceAlign(), nullptr, MemRef::Read);
333 break;
334 }
335 case Intrinsic::memset: {
336 MemSetInst *MSI = cast<MemSetInst>(&I);
337 visitMemoryReference(I, MemoryLocation::getForDest(MSI),
338 MSI->getDestAlign(), nullptr, MemRef::Write);
339 break;
340 }
341
342 case Intrinsic::vastart:
343 Assert(I.getParent()->getParent()->isVarArg(),
344 "Undefined behavior: va_start called in a non-varargs function",
345 &I);
346
347 visitMemoryReference(I, MemoryLocation::getForArgument(&I, 0, TLI), None,
348 nullptr, MemRef::Read | MemRef::Write);
349 break;
350 case Intrinsic::vacopy:
351 visitMemoryReference(I, MemoryLocation::getForArgument(&I, 0, TLI), None,
352 nullptr, MemRef::Write);
353 visitMemoryReference(I, MemoryLocation::getForArgument(&I, 1, TLI), None,
354 nullptr, MemRef::Read);
355 break;
356 case Intrinsic::vaend:
357 visitMemoryReference(I, MemoryLocation::getForArgument(&I, 0, TLI), None,
358 nullptr, MemRef::Read | MemRef::Write);
359 break;
360
361 case Intrinsic::stackrestore:
362 // Stackrestore doesn't read or write memory, but it sets the
363 // stack pointer, which the compiler may read from or write to
364 // at any time, so check it for both readability and writeability.
365 visitMemoryReference(I, MemoryLocation::getForArgument(&I, 0, TLI), None,
366 nullptr, MemRef::Read | MemRef::Write);
367 break;
368 case Intrinsic::get_active_lane_mask:
369 if (auto *TripCount = dyn_cast<ConstantInt>(I.getArgOperand(1)))
370 Assert(!TripCount->isZero(), "get_active_lane_mask: operand #2 "
371 "must be greater than 0", &I);
372 break;
373 }
374 }
375
visitReturnInst(ReturnInst & I)376 void Lint::visitReturnInst(ReturnInst &I) {
377 Function *F = I.getParent()->getParent();
378 Assert(!F->doesNotReturn(),
379 "Unusual: Return statement in function with noreturn attribute", &I);
380
381 if (Value *V = I.getReturnValue()) {
382 Value *Obj = findValue(V, /*OffsetOk=*/true);
383 Assert(!isa<AllocaInst>(Obj), "Unusual: Returning alloca value", &I);
384 }
385 }
386
387 // TODO: Check that the reference is in bounds.
388 // TODO: Check readnone/readonly function attributes.
visitMemoryReference(Instruction & I,const MemoryLocation & Loc,MaybeAlign Align,Type * Ty,unsigned Flags)389 void Lint::visitMemoryReference(Instruction &I, const MemoryLocation &Loc,
390 MaybeAlign Align, Type *Ty, unsigned Flags) {
391 // If no memory is being referenced, it doesn't matter if the pointer
392 // is valid.
393 if (Loc.Size.isZero())
394 return;
395
396 Value *Ptr = const_cast<Value *>(Loc.Ptr);
397 Value *UnderlyingObject = findValue(Ptr, /*OffsetOk=*/true);
398 Assert(!isa<ConstantPointerNull>(UnderlyingObject),
399 "Undefined behavior: Null pointer dereference", &I);
400 Assert(!isa<UndefValue>(UnderlyingObject),
401 "Undefined behavior: Undef pointer dereference", &I);
402 Assert(!isa<ConstantInt>(UnderlyingObject) ||
403 !cast<ConstantInt>(UnderlyingObject)->isMinusOne(),
404 "Unusual: All-ones pointer dereference", &I);
405 Assert(!isa<ConstantInt>(UnderlyingObject) ||
406 !cast<ConstantInt>(UnderlyingObject)->isOne(),
407 "Unusual: Address one pointer dereference", &I);
408
409 if (Flags & MemRef::Write) {
410 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(UnderlyingObject))
411 Assert(!GV->isConstant(), "Undefined behavior: Write to read-only memory",
412 &I);
413 Assert(!isa<Function>(UnderlyingObject) &&
414 !isa<BlockAddress>(UnderlyingObject),
415 "Undefined behavior: Write to text section", &I);
416 }
417 if (Flags & MemRef::Read) {
418 Assert(!isa<Function>(UnderlyingObject), "Unusual: Load from function body",
419 &I);
420 Assert(!isa<BlockAddress>(UnderlyingObject),
421 "Undefined behavior: Load from block address", &I);
422 }
423 if (Flags & MemRef::Callee) {
424 Assert(!isa<BlockAddress>(UnderlyingObject),
425 "Undefined behavior: Call to block address", &I);
426 }
427 if (Flags & MemRef::Branchee) {
428 Assert(!isa<Constant>(UnderlyingObject) ||
429 isa<BlockAddress>(UnderlyingObject),
430 "Undefined behavior: Branch to non-blockaddress", &I);
431 }
432
433 // Check for buffer overflows and misalignment.
434 // Only handles memory references that read/write something simple like an
435 // alloca instruction or a global variable.
436 int64_t Offset = 0;
437 if (Value *Base = GetPointerBaseWithConstantOffset(Ptr, Offset, *DL)) {
438 // OK, so the access is to a constant offset from Ptr. Check that Ptr is
439 // something we can handle and if so extract the size of this base object
440 // along with its alignment.
441 uint64_t BaseSize = MemoryLocation::UnknownSize;
442 MaybeAlign BaseAlign;
443
444 if (AllocaInst *AI = dyn_cast<AllocaInst>(Base)) {
445 Type *ATy = AI->getAllocatedType();
446 if (!AI->isArrayAllocation() && ATy->isSized())
447 BaseSize = DL->getTypeAllocSize(ATy);
448 BaseAlign = AI->getAlign();
449 } else if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) {
450 // If the global may be defined differently in another compilation unit
451 // then don't warn about funky memory accesses.
452 if (GV->hasDefinitiveInitializer()) {
453 Type *GTy = GV->getValueType();
454 if (GTy->isSized())
455 BaseSize = DL->getTypeAllocSize(GTy);
456 BaseAlign = GV->getAlign();
457 if (!BaseAlign && GTy->isSized())
458 BaseAlign = DL->getABITypeAlign(GTy);
459 }
460 }
461
462 // Accesses from before the start or after the end of the object are not
463 // defined.
464 Assert(!Loc.Size.hasValue() || BaseSize == MemoryLocation::UnknownSize ||
465 (Offset >= 0 && Offset + Loc.Size.getValue() <= BaseSize),
466 "Undefined behavior: Buffer overflow", &I);
467
468 // Accesses that say that the memory is more aligned than it is are not
469 // defined.
470 if (!Align && Ty && Ty->isSized())
471 Align = DL->getABITypeAlign(Ty);
472 if (BaseAlign && Align)
473 Assert(*Align <= commonAlignment(*BaseAlign, Offset),
474 "Undefined behavior: Memory reference address is misaligned", &I);
475 }
476 }
477
visitLoadInst(LoadInst & I)478 void Lint::visitLoadInst(LoadInst &I) {
479 visitMemoryReference(I, MemoryLocation::get(&I), I.getAlign(), I.getType(),
480 MemRef::Read);
481 }
482
visitStoreInst(StoreInst & I)483 void Lint::visitStoreInst(StoreInst &I) {
484 visitMemoryReference(I, MemoryLocation::get(&I), I.getAlign(),
485 I.getOperand(0)->getType(), MemRef::Write);
486 }
487
visitXor(BinaryOperator & I)488 void Lint::visitXor(BinaryOperator &I) {
489 Assert(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
490 "Undefined result: xor(undef, undef)", &I);
491 }
492
visitSub(BinaryOperator & I)493 void Lint::visitSub(BinaryOperator &I) {
494 Assert(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
495 "Undefined result: sub(undef, undef)", &I);
496 }
497
visitLShr(BinaryOperator & I)498 void Lint::visitLShr(BinaryOperator &I) {
499 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(I.getOperand(1),
500 /*OffsetOk=*/false)))
501 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
502 "Undefined result: Shift count out of range", &I);
503 }
504
visitAShr(BinaryOperator & I)505 void Lint::visitAShr(BinaryOperator &I) {
506 if (ConstantInt *CI =
507 dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
508 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
509 "Undefined result: Shift count out of range", &I);
510 }
511
visitShl(BinaryOperator & I)512 void Lint::visitShl(BinaryOperator &I) {
513 if (ConstantInt *CI =
514 dyn_cast<ConstantInt>(findValue(I.getOperand(1), /*OffsetOk=*/false)))
515 Assert(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
516 "Undefined result: Shift count out of range", &I);
517 }
518
isZero(Value * V,const DataLayout & DL,DominatorTree * DT,AssumptionCache * AC)519 static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT,
520 AssumptionCache *AC) {
521 // Assume undef could be zero.
522 if (isa<UndefValue>(V))
523 return true;
524
525 VectorType *VecTy = dyn_cast<VectorType>(V->getType());
526 if (!VecTy) {
527 KnownBits Known =
528 computeKnownBits(V, DL, 0, AC, dyn_cast<Instruction>(V), DT);
529 return Known.isZero();
530 }
531
532 // Per-component check doesn't work with zeroinitializer
533 Constant *C = dyn_cast<Constant>(V);
534 if (!C)
535 return false;
536
537 if (C->isZeroValue())
538 return true;
539
540 // For a vector, KnownZero will only be true if all values are zero, so check
541 // this per component
542 for (unsigned I = 0, N = cast<FixedVectorType>(VecTy)->getNumElements();
543 I != N; ++I) {
544 Constant *Elem = C->getAggregateElement(I);
545 if (isa<UndefValue>(Elem))
546 return true;
547
548 KnownBits Known = computeKnownBits(Elem, DL);
549 if (Known.isZero())
550 return true;
551 }
552
553 return false;
554 }
555
visitSDiv(BinaryOperator & I)556 void Lint::visitSDiv(BinaryOperator &I) {
557 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
558 "Undefined behavior: Division by zero", &I);
559 }
560
visitUDiv(BinaryOperator & I)561 void Lint::visitUDiv(BinaryOperator &I) {
562 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
563 "Undefined behavior: Division by zero", &I);
564 }
565
visitSRem(BinaryOperator & I)566 void Lint::visitSRem(BinaryOperator &I) {
567 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
568 "Undefined behavior: Division by zero", &I);
569 }
570
visitURem(BinaryOperator & I)571 void Lint::visitURem(BinaryOperator &I) {
572 Assert(!isZero(I.getOperand(1), I.getModule()->getDataLayout(), DT, AC),
573 "Undefined behavior: Division by zero", &I);
574 }
575
visitAllocaInst(AllocaInst & I)576 void Lint::visitAllocaInst(AllocaInst &I) {
577 if (isa<ConstantInt>(I.getArraySize()))
578 // This isn't undefined behavior, it's just an obvious pessimization.
579 Assert(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
580 "Pessimization: Static alloca outside of entry block", &I);
581
582 // TODO: Check for an unusual size (MSB set?)
583 }
584
visitVAArgInst(VAArgInst & I)585 void Lint::visitVAArgInst(VAArgInst &I) {
586 visitMemoryReference(I, MemoryLocation::get(&I), None, nullptr,
587 MemRef::Read | MemRef::Write);
588 }
589
visitIndirectBrInst(IndirectBrInst & I)590 void Lint::visitIndirectBrInst(IndirectBrInst &I) {
591 visitMemoryReference(I, MemoryLocation::getAfter(I.getAddress()), None,
592 nullptr, MemRef::Branchee);
593
594 Assert(I.getNumDestinations() != 0,
595 "Undefined behavior: indirectbr with no destinations", &I);
596 }
597
visitExtractElementInst(ExtractElementInst & I)598 void Lint::visitExtractElementInst(ExtractElementInst &I) {
599 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(I.getIndexOperand(),
600 /*OffsetOk=*/false)))
601 Assert(
602 CI->getValue().ult(
603 cast<FixedVectorType>(I.getVectorOperandType())->getNumElements()),
604 "Undefined result: extractelement index out of range", &I);
605 }
606
visitInsertElementInst(InsertElementInst & I)607 void Lint::visitInsertElementInst(InsertElementInst &I) {
608 if (ConstantInt *CI = dyn_cast<ConstantInt>(findValue(I.getOperand(2),
609 /*OffsetOk=*/false)))
610 Assert(CI->getValue().ult(
611 cast<FixedVectorType>(I.getType())->getNumElements()),
612 "Undefined result: insertelement index out of range", &I);
613 }
614
visitUnreachableInst(UnreachableInst & I)615 void Lint::visitUnreachableInst(UnreachableInst &I) {
616 // This isn't undefined behavior, it's merely suspicious.
617 Assert(&I == &I.getParent()->front() ||
618 std::prev(I.getIterator())->mayHaveSideEffects(),
619 "Unusual: unreachable immediately preceded by instruction without "
620 "side effects",
621 &I);
622 }
623
624 /// findValue - Look through bitcasts and simple memory reference patterns
625 /// to identify an equivalent, but more informative, value. If OffsetOk
626 /// is true, look through getelementptrs with non-zero offsets too.
627 ///
628 /// Most analysis passes don't require this logic, because instcombine
629 /// will simplify most of these kinds of things away. But it's a goal of
630 /// this Lint pass to be useful even on non-optimized IR.
findValue(Value * V,bool OffsetOk) const631 Value *Lint::findValue(Value *V, bool OffsetOk) const {
632 SmallPtrSet<Value *, 4> Visited;
633 return findValueImpl(V, OffsetOk, Visited);
634 }
635
636 /// findValueImpl - Implementation helper for findValue.
findValueImpl(Value * V,bool OffsetOk,SmallPtrSetImpl<Value * > & Visited) const637 Value *Lint::findValueImpl(Value *V, bool OffsetOk,
638 SmallPtrSetImpl<Value *> &Visited) const {
639 // Detect self-referential values.
640 if (!Visited.insert(V).second)
641 return UndefValue::get(V->getType());
642
643 // TODO: Look through sext or zext cast, when the result is known to
644 // be interpreted as signed or unsigned, respectively.
645 // TODO: Look through eliminable cast pairs.
646 // TODO: Look through calls with unique return values.
647 // TODO: Look through vector insert/extract/shuffle.
648 V = OffsetOk ? getUnderlyingObject(V) : V->stripPointerCasts();
649 if (LoadInst *L = dyn_cast<LoadInst>(V)) {
650 BasicBlock::iterator BBI = L->getIterator();
651 BasicBlock *BB = L->getParent();
652 SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
653 for (;;) {
654 if (!VisitedBlocks.insert(BB).second)
655 break;
656 if (Value *U =
657 FindAvailableLoadedValue(L, BB, BBI, DefMaxInstsToScan, AA))
658 return findValueImpl(U, OffsetOk, Visited);
659 if (BBI != BB->begin())
660 break;
661 BB = BB->getUniquePredecessor();
662 if (!BB)
663 break;
664 BBI = BB->end();
665 }
666 } else if (PHINode *PN = dyn_cast<PHINode>(V)) {
667 if (Value *W = PN->hasConstantValue())
668 return findValueImpl(W, OffsetOk, Visited);
669 } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
670 if (CI->isNoopCast(*DL))
671 return findValueImpl(CI->getOperand(0), OffsetOk, Visited);
672 } else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(V)) {
673 if (Value *W =
674 FindInsertedValue(Ex->getAggregateOperand(), Ex->getIndices()))
675 if (W != V)
676 return findValueImpl(W, OffsetOk, Visited);
677 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
678 // Same as above, but for ConstantExpr instead of Instruction.
679 if (Instruction::isCast(CE->getOpcode())) {
680 if (CastInst::isNoopCast(Instruction::CastOps(CE->getOpcode()),
681 CE->getOperand(0)->getType(), CE->getType(),
682 *DL))
683 return findValueImpl(CE->getOperand(0), OffsetOk, Visited);
684 } else if (CE->getOpcode() == Instruction::ExtractValue) {
685 ArrayRef<unsigned> Indices = CE->getIndices();
686 if (Value *W = FindInsertedValue(CE->getOperand(0), Indices))
687 if (W != V)
688 return findValueImpl(W, OffsetOk, Visited);
689 }
690 }
691
692 // As a last resort, try SimplifyInstruction or constant folding.
693 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
694 if (Value *W = SimplifyInstruction(Inst, {*DL, TLI, DT, AC}))
695 return findValueImpl(W, OffsetOk, Visited);
696 } else if (auto *C = dyn_cast<Constant>(V)) {
697 Value *W = ConstantFoldConstant(C, *DL, TLI);
698 if (W != V)
699 return findValueImpl(W, OffsetOk, Visited);
700 }
701
702 return V;
703 }
704
run(Function & F,FunctionAnalysisManager & AM)705 PreservedAnalyses LintPass::run(Function &F, FunctionAnalysisManager &AM) {
706 auto *Mod = F.getParent();
707 auto *DL = &F.getParent()->getDataLayout();
708 auto *AA = &AM.getResult<AAManager>(F);
709 auto *AC = &AM.getResult<AssumptionAnalysis>(F);
710 auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
711 auto *TLI = &AM.getResult<TargetLibraryAnalysis>(F);
712 Lint L(Mod, DL, AA, AC, DT, TLI);
713 L.visit(F);
714 dbgs() << L.MessagesStr.str();
715 return PreservedAnalyses::all();
716 }
717
718 class LintLegacyPass : public FunctionPass {
719 public:
720 static char ID; // Pass identification, replacement for typeid
LintLegacyPass()721 LintLegacyPass() : FunctionPass(ID) {
722 initializeLintLegacyPassPass(*PassRegistry::getPassRegistry());
723 }
724
725 bool runOnFunction(Function &F) override;
726
getAnalysisUsage(AnalysisUsage & AU) const727 void getAnalysisUsage(AnalysisUsage &AU) const override {
728 AU.setPreservesAll();
729 AU.addRequired<AAResultsWrapperPass>();
730 AU.addRequired<AssumptionCacheTracker>();
731 AU.addRequired<TargetLibraryInfoWrapperPass>();
732 AU.addRequired<DominatorTreeWrapperPass>();
733 }
print(raw_ostream & O,const Module * M) const734 void print(raw_ostream &O, const Module *M) const override {}
735 };
736
737 char LintLegacyPass::ID = 0;
738 INITIALIZE_PASS_BEGIN(LintLegacyPass, "lint", "Statically lint-checks LLVM IR",
739 false, true)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)740 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
741 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
742 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
743 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
744 INITIALIZE_PASS_END(LintLegacyPass, "lint", "Statically lint-checks LLVM IR",
745 false, true)
746
747 bool LintLegacyPass::runOnFunction(Function &F) {
748 auto *Mod = F.getParent();
749 auto *DL = &F.getParent()->getDataLayout();
750 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
751 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
752 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
753 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
754 Lint L(Mod, DL, AA, AC, DT, TLI);
755 L.visit(F);
756 dbgs() << L.MessagesStr.str();
757 return false;
758 }
759
760 //===----------------------------------------------------------------------===//
761 // Implement the public interfaces to this file...
762 //===----------------------------------------------------------------------===//
763
createLintLegacyPassPass()764 FunctionPass *llvm::createLintLegacyPassPass() { return new LintLegacyPass(); }
765
766 /// lintFunction - Check a function for errors, printing messages on stderr.
767 ///
lintFunction(const Function & f)768 void llvm::lintFunction(const Function &f) {
769 Function &F = const_cast<Function &>(f);
770 assert(!F.isDeclaration() && "Cannot lint external functions");
771
772 legacy::FunctionPassManager FPM(F.getParent());
773 auto *V = new LintLegacyPass();
774 FPM.add(V);
775 FPM.run(F);
776 }
777
778 /// lintModule - Check a module for errors, printing messages on stderr.
779 ///
lintModule(const Module & M)780 void llvm::lintModule(const Module &M) {
781 legacy::PassManager PM;
782 auto *V = new LintLegacyPass();
783 PM.add(V);
784 PM.run(const_cast<Module &>(M));
785 }
786