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/BasicAliasAnalysis.h"
44#include "llvm/Analysis/ConstantFolding.h"
45#include "llvm/Analysis/InstructionSimplify.h"
46#include "llvm/Analysis/Loads.h"
47#include "llvm/Analysis/MemoryLocation.h"
48#include "llvm/Analysis/ScopedNoAliasAA.h"
49#include "llvm/Analysis/TargetLibraryInfo.h"
50#include "llvm/Analysis/TypeBasedAliasAnalysis.h"
51#include "llvm/Analysis/ValueTracking.h"
52#include "llvm/IR/Argument.h"
53#include "llvm/IR/BasicBlock.h"
54#include "llvm/IR/Constant.h"
55#include "llvm/IR/Constants.h"
56#include "llvm/IR/DataLayout.h"
57#include "llvm/IR/DerivedTypes.h"
58#include "llvm/IR/Dominators.h"
59#include "llvm/IR/Function.h"
60#include "llvm/IR/GlobalVariable.h"
61#include "llvm/IR/InstVisitor.h"
62#include "llvm/IR/InstrTypes.h"
63#include "llvm/IR/Instruction.h"
64#include "llvm/IR/Instructions.h"
65#include "llvm/IR/IntrinsicInst.h"
66#include "llvm/IR/Module.h"
67#include "llvm/IR/PassManager.h"
68#include "llvm/IR/Type.h"
69#include "llvm/IR/Value.h"
70#include "llvm/Support/AMDGPUAddrSpace.h"
71#include "llvm/Support/Casting.h"
72#include "llvm/Support/KnownBits.h"
73#include "llvm/Support/raw_ostream.h"
74#include <cassert>
75#include <cstdint>
76#include <iterator>
77#include <string>
78
79using namespace llvm;
80
81namespace {
82namespace MemRef {
83static const unsigned Read = 1;
84static const unsigned Write = 2;
85static const unsigned Callee = 4;
86static const unsigned Branchee = 8;
87} // end namespace MemRef
88
89class Lint : public InstVisitor<Lint> {
90 friend class InstVisitor<Lint>;
91
92 void visitFunction(Function &F);
93
94 void visitCallBase(CallBase &CB);
95 void visitMemoryReference(Instruction &I, const MemoryLocation &Loc,
96 MaybeAlign Alignment, Type *Ty, unsigned Flags);
97
98 void visitReturnInst(ReturnInst &I);
99 void visitLoadInst(LoadInst &I);
100 void visitStoreInst(StoreInst &I);
101 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I);
102 void visitAtomicRMWInst(AtomicRMWInst &I);
103 void visitXor(BinaryOperator &I);
104 void visitSub(BinaryOperator &I);
105 void visitLShr(BinaryOperator &I);
106 void visitAShr(BinaryOperator &I);
107 void visitShl(BinaryOperator &I);
108 void visitSDiv(BinaryOperator &I);
109 void visitUDiv(BinaryOperator &I);
110 void visitSRem(BinaryOperator &I);
111 void visitURem(BinaryOperator &I);
112 void visitAllocaInst(AllocaInst &I);
113 void visitVAArgInst(VAArgInst &I);
114 void visitIndirectBrInst(IndirectBrInst &I);
115 void visitExtractElementInst(ExtractElementInst &I);
116 void visitInsertElementInst(InsertElementInst &I);
117 void visitUnreachableInst(UnreachableInst &I);
118
119 Value *findValue(Value *V, bool OffsetOk) const;
120 Value *findValueImpl(Value *V, bool OffsetOk,
121 SmallPtrSetImpl<Value *> &Visited) const;
122
123public:
124 Module *Mod;
125 const Triple &TT;
126 const DataLayout *DL;
127 AliasAnalysis *AA;
128 AssumptionCache *AC;
129 DominatorTree *DT;
130 TargetLibraryInfo *TLI;
131
132 std::string Messages;
133 raw_string_ostream MessagesStr;
134
135 Lint(Module *Mod, const DataLayout *DL, AliasAnalysis *AA,
136 AssumptionCache *AC, DominatorTree *DT, TargetLibraryInfo *TLI)
137 : Mod(Mod), TT(Mod->getTargetTriple()), DL(DL), AA(AA), AC(AC), DT(DT),
138 TLI(TLI), MessagesStr(Messages) {}
139
140 void WriteValues(ArrayRef<const Value *> Vs) {
141 for (const Value *V : Vs) {
142 if (!V)
143 continue;
144 if (isa<Instruction>(Val: V)) {
145 MessagesStr << *V << '\n';
146 } else {
147 V->printAsOperand(O&: MessagesStr, PrintType: true, M: Mod);
148 MessagesStr << '\n';
149 }
150 }
151 }
152
153 /// A check failed, so printout out the condition and the message.
154 ///
155 /// This provides a nice place to put a breakpoint if you want to see why
156 /// something is not correct.
157 void CheckFailed(const Twine &Message) { MessagesStr << Message << '\n'; }
158
159 /// A check failed (with values to print).
160 ///
161 /// This calls the Message-only version so that the above is easier to set
162 /// a breakpoint on.
163 template <typename T1, typename... Ts>
164 void CheckFailed(const Twine &Message, const T1 &V1, const Ts &... Vs) {
165 CheckFailed(Message);
166 WriteValues(Vs: {V1, Vs...});
167 }
168};
169} // end anonymous namespace
170
171// Check - We know that cond should be true, if not print an error message.
172#define Check(C, ...) \
173 do { \
174 if (!(C)) { \
175 CheckFailed(__VA_ARGS__); \
176 return; \
177 } \
178 } while (false)
179
180void Lint::visitFunction(Function &F) {
181 // This isn't undefined behavior, it's just a little unusual, and it's a
182 // fairly common mistake to neglect to name a function.
183 Check(F.hasName() || F.hasLocalLinkage(),
184 "Unusual: Unnamed function with non-local linkage", &F);
185
186 // TODO: Check for irreducible control flow.
187}
188
189void Lint::visitCallBase(CallBase &I) {
190 Value *Callee = I.getCalledOperand();
191
192 visitMemoryReference(I, Loc: MemoryLocation::getAfter(Ptr: Callee), Alignment: std::nullopt,
193 Ty: nullptr, Flags: MemRef::Callee);
194
195 if (Function *F = dyn_cast<Function>(Val: findValue(V: Callee,
196 /*OffsetOk=*/false))) {
197 Check(I.getCallingConv() == F->getCallingConv(),
198 "Undefined behavior: Caller and callee calling convention differ",
199 &I);
200
201 FunctionType *FT = F->getFunctionType();
202 unsigned NumActualArgs = I.arg_size();
203
204 Check(FT->isVarArg() ? FT->getNumParams() <= NumActualArgs
205 : FT->getNumParams() == NumActualArgs,
206 "Undefined behavior: Call argument count mismatches callee "
207 "argument count",
208 &I);
209
210 Check(FT->getReturnType() == I.getType(),
211 "Undefined behavior: Call return type mismatches "
212 "callee return type",
213 &I);
214
215 // Check argument types (in case the callee was casted) and attributes.
216 // TODO: Verify that caller and callee attributes are compatible.
217 Function::arg_iterator PI = F->arg_begin(), PE = F->arg_end();
218 auto AI = I.arg_begin(), AE = I.arg_end();
219 for (; AI != AE; ++AI) {
220 Value *Actual = *AI;
221 if (PI != PE) {
222 Argument *Formal = &*PI++;
223 Check(Formal->getType() == Actual->getType(),
224 "Undefined behavior: Call argument type mismatches "
225 "callee parameter type",
226 &I);
227
228 // Check that noalias arguments don't alias other arguments. This is
229 // not fully precise because we don't know the sizes of the dereferenced
230 // memory regions.
231 if (Formal->hasNoAliasAttr() && Actual->getType()->isPointerTy()) {
232 AttributeList PAL = I.getAttributes();
233 unsigned ArgNo = 0;
234 for (auto *BI = I.arg_begin(); BI != AE; ++BI, ++ArgNo) {
235 // Skip ByVal arguments since they will be memcpy'd to the callee's
236 // stack so we're not really passing the pointer anyway.
237 if (PAL.hasParamAttr(ArgNo, Kind: Attribute::ByVal))
238 continue;
239 // If both arguments are readonly, they have no dependence.
240 if (Formal->onlyReadsMemory() && I.onlyReadsMemory(OpNo: ArgNo))
241 continue;
242 // Skip readnone arguments since those are guaranteed not to be
243 // dereferenced anyway.
244 if (I.doesNotAccessMemory(OpNo: ArgNo))
245 continue;
246 if (AI != BI && (*BI)->getType()->isPointerTy() &&
247 !isa<ConstantPointerNull>(Val: *BI)) {
248 AliasResult Result = AA->alias(V1: *AI, V2: *BI);
249 Check(Result != AliasResult::MustAlias &&
250 Result != AliasResult::PartialAlias,
251 "Unusual: noalias argument aliases another argument", &I);
252 }
253 }
254 }
255
256 // Check that an sret argument points to valid memory.
257 if (Formal->hasStructRetAttr() && Actual->getType()->isPointerTy()) {
258 Type *Ty = Formal->getParamStructRetType();
259 MemoryLocation Loc(
260 Actual, LocationSize::precise(Value: DL->getTypeStoreSize(Ty)));
261 visitMemoryReference(I, Loc, Alignment: DL->getABITypeAlign(Ty), Ty,
262 Flags: MemRef::Read | MemRef::Write);
263 }
264
265 // Check that ABI attributes for the function and call-site match.
266 unsigned ArgNo = AI->getOperandNo();
267 AttributeList CallAttrs = I.getAttributes();
268 for (Attribute::AttrKind Attr :
269 drop_begin(RangeOrContainer: enum_seq(Begin: Attribute::None, End: Attribute::EndAttrKinds,
270 force_iteration_on_noniterable_enum))) {
271 if (!Attribute::isABIAttr(Kind: Attr))
272 continue;
273
274 Attribute CallAttr = CallAttrs.getParamAttr(ArgNo, Kind: Attr);
275 Attribute FnAttr = F->getParamAttribute(ArgNo, Kind: Attr);
276
277 if ((Attr == Attribute::ZExt || Attr == Attribute::SExt) &&
278 CallAttr.isValid() && !FnAttr.isValid()) {
279 // As an exception, allow zeroext/signext at the call-site, while
280 // missing on the function. This means that the caller extends,
281 // while the callee does not assume extension. The call-site
282 // extension is redundant in that case, but perfectly well-defined.
283 continue;
284 }
285
286 Check(CallAttr.isValid() == FnAttr.isValid(),
287 Twine("Undefined behavior: ABI attribute ") +
288 Attribute::getNameFromAttrKind(Attr) +
289 " not present on both function and call-site",
290 &I);
291 if (CallAttr.isValid() && FnAttr.isValid()) {
292 Check(CallAttr == FnAttr,
293 Twine("Undefined behavior: ABI attribute ") +
294 Attribute::getNameFromAttrKind(Attr) +
295 " does not have same argument for function and call-site",
296 &I);
297 }
298 }
299 }
300 }
301 }
302
303 if (const auto *CI = dyn_cast<CallInst>(Val: &I)) {
304 if (CI->isTailCall()) {
305 const AttributeList &PAL = CI->getAttributes();
306 unsigned ArgNo = 0;
307 for (Value *Arg : I.args()) {
308 // Skip ByVal arguments since they will be memcpy'd to the callee's
309 // stack anyway.
310 if (PAL.hasParamAttr(ArgNo: ArgNo++, Kind: Attribute::ByVal))
311 continue;
312 Value *Obj = findValue(V: Arg, /*OffsetOk=*/true);
313 Check(!isa<AllocaInst>(Obj),
314 "Undefined behavior: Call with \"tail\" keyword references "
315 "alloca",
316 &I);
317 }
318 }
319 }
320
321 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &I))
322 switch (II->getIntrinsicID()) {
323 default:
324 break;
325
326 // TODO: Check more intrinsics
327
328 case Intrinsic::memcpy:
329 case Intrinsic::memcpy_inline: {
330 MemCpyInst *MCI = cast<MemCpyInst>(Val: &I);
331 visitMemoryReference(I, Loc: MemoryLocation::getForDest(MI: MCI),
332 Alignment: MCI->getDestAlign(), Ty: nullptr, Flags: MemRef::Write);
333 visitMemoryReference(I, Loc: MemoryLocation::getForSource(MTI: MCI),
334 Alignment: MCI->getSourceAlign(), Ty: nullptr, Flags: MemRef::Read);
335
336 // Check that the memcpy arguments don't overlap. The AliasAnalysis API
337 // isn't expressive enough for what we really want to do. Known partial
338 // overlap is not distinguished from the case where nothing is known.
339 auto Size = LocationSize::afterPointer();
340 if (const ConstantInt *Len =
341 dyn_cast<ConstantInt>(Val: findValue(V: MCI->getLength(),
342 /*OffsetOk=*/false)))
343 if (Len->getValue().isIntN(N: 32))
344 Size = LocationSize::precise(Value: Len->getValue().getZExtValue());
345 Check(AA->alias(MCI->getSource(), Size, MCI->getDest(), Size) !=
346 AliasResult::MustAlias,
347 "Undefined behavior: memcpy source and destination overlap", &I);
348 break;
349 }
350 case Intrinsic::memmove: {
351 MemMoveInst *MMI = cast<MemMoveInst>(Val: &I);
352 visitMemoryReference(I, Loc: MemoryLocation::getForDest(MI: MMI),
353 Alignment: MMI->getDestAlign(), Ty: nullptr, Flags: MemRef::Write);
354 visitMemoryReference(I, Loc: MemoryLocation::getForSource(MTI: MMI),
355 Alignment: MMI->getSourceAlign(), Ty: nullptr, Flags: MemRef::Read);
356 break;
357 }
358 case Intrinsic::memset:
359 case Intrinsic::memset_inline: {
360 MemSetInst *MSI = cast<MemSetInst>(Val: &I);
361 visitMemoryReference(I, Loc: MemoryLocation::getForDest(MI: MSI),
362 Alignment: MSI->getDestAlign(), Ty: nullptr, Flags: MemRef::Write);
363 break;
364 }
365 case Intrinsic::vastart:
366 // vastart in non-varargs function is rejected by the verifier
367 visitMemoryReference(I, Loc: MemoryLocation::getForArgument(Call: &I, ArgIdx: 0, TLI),
368 Alignment: std::nullopt, Ty: nullptr, Flags: MemRef::Read | MemRef::Write);
369 break;
370 case Intrinsic::vacopy:
371 visitMemoryReference(I, Loc: MemoryLocation::getForArgument(Call: &I, ArgIdx: 0, TLI),
372 Alignment: std::nullopt, Ty: nullptr, Flags: MemRef::Write);
373 visitMemoryReference(I, Loc: MemoryLocation::getForArgument(Call: &I, ArgIdx: 1, TLI),
374 Alignment: std::nullopt, Ty: nullptr, Flags: MemRef::Read);
375 break;
376 case Intrinsic::vaend:
377 visitMemoryReference(I, Loc: MemoryLocation::getForArgument(Call: &I, ArgIdx: 0, TLI),
378 Alignment: std::nullopt, Ty: nullptr, Flags: MemRef::Read | MemRef::Write);
379 break;
380
381 case Intrinsic::stackrestore:
382 // Stackrestore doesn't read or write memory, but it sets the
383 // stack pointer, which the compiler may read from or write to
384 // at any time, so check it for both readability and writeability.
385 visitMemoryReference(I, Loc: MemoryLocation::getForArgument(Call: &I, ArgIdx: 0, TLI),
386 Alignment: std::nullopt, Ty: nullptr, Flags: MemRef::Read | MemRef::Write);
387 break;
388 }
389}
390
391void Lint::visitReturnInst(ReturnInst &I) {
392 Function *F = I.getParent()->getParent();
393 Check(!F->doesNotReturn(),
394 "Unusual: Return statement in function with noreturn attribute", &I);
395
396 if (Value *V = I.getReturnValue()) {
397 Value *Obj = findValue(V, /*OffsetOk=*/true);
398 Check(!isa<AllocaInst>(Obj), "Unusual: Returning alloca value", &I);
399 }
400}
401
402// TODO: Check that the reference is in bounds.
403// TODO: Check readnone/readonly function attributes.
404void Lint::visitMemoryReference(Instruction &I, const MemoryLocation &Loc,
405 MaybeAlign Align, Type *Ty, unsigned Flags) {
406 // If no memory is being referenced, it doesn't matter if the pointer
407 // is valid.
408 if (Loc.Size.isZero())
409 return;
410
411 Value *Ptr = const_cast<Value *>(Loc.Ptr);
412 Value *UnderlyingObject = findValue(V: Ptr, /*OffsetOk=*/true);
413 Check(!isa<ConstantPointerNull>(UnderlyingObject),
414 "Undefined behavior: Null pointer dereference", &I);
415 Check(!isa<UndefValue>(UnderlyingObject),
416 "Undefined behavior: Undef pointer dereference", &I);
417 Check(!isa<ConstantInt>(UnderlyingObject) ||
418 !cast<ConstantInt>(UnderlyingObject)->isMinusOne(),
419 "Unusual: All-ones pointer dereference", &I);
420 Check(!isa<ConstantInt>(UnderlyingObject) ||
421 !cast<ConstantInt>(UnderlyingObject)->isOne(),
422 "Unusual: Address one pointer dereference", &I);
423
424 if (Flags & MemRef::Write) {
425 if (TT.isAMDGPU())
426 Check(!AMDGPU::isConstantAddressSpace(
427 UnderlyingObject->getType()->getPointerAddressSpace()),
428 "Undefined behavior: Write to memory in const addrspace", &I);
429
430 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: UnderlyingObject))
431 Check(!GV->isConstant(), "Undefined behavior: Write to read-only memory",
432 &I);
433 Check(!isa<Function>(UnderlyingObject) &&
434 !isa<BlockAddress>(UnderlyingObject),
435 "Undefined behavior: Write to text section", &I);
436 }
437 if (Flags & MemRef::Read) {
438 Check(!isa<Function>(UnderlyingObject), "Unusual: Load from function body",
439 &I);
440 Check(!isa<BlockAddress>(UnderlyingObject),
441 "Undefined behavior: Load from block address", &I);
442 }
443 if (Flags & MemRef::Callee) {
444 Check(!isa<BlockAddress>(UnderlyingObject),
445 "Undefined behavior: Call to block address", &I);
446 }
447 if (Flags & MemRef::Branchee) {
448 Check(!isa<Constant>(UnderlyingObject) ||
449 isa<BlockAddress>(UnderlyingObject),
450 "Undefined behavior: Branch to non-blockaddress", &I);
451 }
452
453 // Check for buffer overflows and misalignment.
454 // Only handles memory references that read/write something simple like an
455 // alloca instruction or a global variable.
456 int64_t Offset = 0;
457 if (Value *Base = GetPointerBaseWithConstantOffset(Ptr, Offset, DL: *DL)) {
458 // OK, so the access is to a constant offset from Ptr. Check that Ptr is
459 // something we can handle and if so extract the size of this base object
460 // along with its alignment.
461 uint64_t BaseSize = MemoryLocation::UnknownSize;
462 MaybeAlign BaseAlign;
463
464 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: Base)) {
465 std::optional<TypeSize> ATy = AI->getAllocationSize(DL: *DL);
466 if (ATy && !ATy->isScalable())
467 BaseSize = ATy->getFixedValue();
468 BaseAlign = AI->getAlign();
469 } else if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: Base)) {
470 // If the global may be defined differently in another compilation unit
471 // then don't warn about funky memory accesses.
472 if (GV->hasDefinitiveInitializer()) {
473 Type *GTy = GV->getValueType();
474 if (GTy->isSized())
475 BaseSize = DL->getTypeAllocSize(Ty: GTy);
476 BaseAlign = GV->getAlign();
477 if (!BaseAlign && GTy->isSized())
478 BaseAlign = DL->getABITypeAlign(Ty: GTy);
479 }
480 }
481
482 // Accesses from before the start or after the end of the object are not
483 // defined.
484 Check(!Loc.Size.hasValue() || Loc.Size.isScalable() ||
485 BaseSize == MemoryLocation::UnknownSize ||
486 (Offset >= 0 && Offset + Loc.Size.getValue() <= BaseSize),
487 "Undefined behavior: Buffer overflow", &I);
488
489 // Accesses that say that the memory is more aligned than it is are not
490 // defined.
491 if (!Align && Ty && Ty->isSized())
492 Align = DL->getABITypeAlign(Ty);
493 if (BaseAlign && Align)
494 Check(*Align <= commonAlignment(*BaseAlign, Offset),
495 "Undefined behavior: Memory reference address is misaligned", &I);
496 }
497}
498
499void Lint::visitLoadInst(LoadInst &I) {
500 visitMemoryReference(I, Loc: MemoryLocation::get(LI: &I), Align: I.getAlign(), Ty: I.getType(),
501 Flags: MemRef::Read);
502}
503
504void Lint::visitStoreInst(StoreInst &I) {
505 visitMemoryReference(I, Loc: MemoryLocation::get(SI: &I), Align: I.getAlign(),
506 Ty: I.getOperand(i_nocapture: 0)->getType(), Flags: MemRef::Write);
507}
508
509void Lint::visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
510 visitMemoryReference(I, Loc: MemoryLocation::get(CXI: &I), Align: I.getAlign(),
511 Ty: I.getOperand(i_nocapture: 0)->getType(), Flags: MemRef::Write);
512}
513
514void Lint::visitAtomicRMWInst(AtomicRMWInst &I) {
515 visitMemoryReference(I, Loc: MemoryLocation::get(RMWI: &I), Align: I.getAlign(),
516 Ty: I.getOperand(i_nocapture: 0)->getType(), Flags: MemRef::Write);
517}
518
519void Lint::visitXor(BinaryOperator &I) {
520 Check(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
521 "Undefined result: xor(undef, undef)", &I);
522}
523
524void Lint::visitSub(BinaryOperator &I) {
525 Check(!isa<UndefValue>(I.getOperand(0)) || !isa<UndefValue>(I.getOperand(1)),
526 "Undefined result: sub(undef, undef)", &I);
527}
528
529void Lint::visitLShr(BinaryOperator &I) {
530 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: findValue(V: I.getOperand(i_nocapture: 1),
531 /*OffsetOk=*/false)))
532 Check(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
533 "Undefined result: Shift count out of range", &I);
534}
535
536void Lint::visitAShr(BinaryOperator &I) {
537 if (ConstantInt *CI =
538 dyn_cast<ConstantInt>(Val: findValue(V: I.getOperand(i_nocapture: 1), /*OffsetOk=*/false)))
539 Check(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
540 "Undefined result: Shift count out of range", &I);
541}
542
543void Lint::visitShl(BinaryOperator &I) {
544 if (ConstantInt *CI =
545 dyn_cast<ConstantInt>(Val: findValue(V: I.getOperand(i_nocapture: 1), /*OffsetOk=*/false)))
546 Check(CI->getValue().ult(cast<IntegerType>(I.getType())->getBitWidth()),
547 "Undefined result: Shift count out of range", &I);
548}
549
550static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT,
551 AssumptionCache *AC) {
552 // Assume undef could be zero.
553 if (isa<UndefValue>(Val: V))
554 return true;
555
556 VectorType *VecTy = dyn_cast<VectorType>(Val: V->getType());
557 if (!VecTy) {
558 KnownBits Known = computeKnownBits(V, DL, AC, CtxI: dyn_cast<Instruction>(Val: V), DT);
559 return Known.isZero();
560 }
561
562 // Per-component check doesn't work with zeroinitializer
563 Constant *C = dyn_cast<Constant>(Val: V);
564 if (!C)
565 return false;
566
567 if (C->isNullValue())
568 return true;
569
570 // For a vector, KnownZero will only be true if all values are zero, so check
571 // this per component
572 for (unsigned I = 0, N = cast<FixedVectorType>(Val: VecTy)->getNumElements();
573 I != N; ++I) {
574 Constant *Elem = C->getAggregateElement(Elt: I);
575 if (isa<UndefValue>(Val: Elem))
576 return true;
577
578 KnownBits Known = computeKnownBits(V: Elem, DL);
579 if (Known.isZero())
580 return true;
581 }
582
583 return false;
584}
585
586void Lint::visitSDiv(BinaryOperator &I) {
587 Check(!isZero(I.getOperand(1), I.getDataLayout(), DT, AC),
588 "Undefined behavior: Division by zero", &I);
589}
590
591void Lint::visitUDiv(BinaryOperator &I) {
592 Check(!isZero(I.getOperand(1), I.getDataLayout(), DT, AC),
593 "Undefined behavior: Division by zero", &I);
594}
595
596void Lint::visitSRem(BinaryOperator &I) {
597 Check(!isZero(I.getOperand(1), I.getDataLayout(), DT, AC),
598 "Undefined behavior: Division by zero", &I);
599}
600
601void Lint::visitURem(BinaryOperator &I) {
602 Check(!isZero(I.getOperand(1), I.getDataLayout(), DT, AC),
603 "Undefined behavior: Division by zero", &I);
604}
605
606void Lint::visitAllocaInst(AllocaInst &I) {
607 if (isa<ConstantInt>(Val: I.getArraySize()))
608 // This isn't undefined behavior, it's just an obvious pessimization.
609 Check(&I.getParent()->getParent()->getEntryBlock() == I.getParent(),
610 "Pessimization: Static alloca outside of entry block", &I);
611
612 // TODO: Check for an unusual size (MSB set?)
613}
614
615void Lint::visitVAArgInst(VAArgInst &I) {
616 visitMemoryReference(I, Loc: MemoryLocation::get(VI: &I), Align: std::nullopt, Ty: nullptr,
617 Flags: MemRef::Read | MemRef::Write);
618}
619
620void Lint::visitIndirectBrInst(IndirectBrInst &I) {
621 visitMemoryReference(I, Loc: MemoryLocation::getAfter(Ptr: I.getAddress()),
622 Align: std::nullopt, Ty: nullptr, Flags: MemRef::Branchee);
623
624 Check(I.getNumDestinations() != 0,
625 "Undefined behavior: indirectbr with no destinations", &I);
626}
627
628void Lint::visitExtractElementInst(ExtractElementInst &I) {
629 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: findValue(V: I.getIndexOperand(),
630 /*OffsetOk=*/false))) {
631 ElementCount EC = I.getVectorOperandType()->getElementCount();
632 Check(EC.isScalable() || CI->getValue().ult(EC.getFixedValue()),
633 "Undefined result: extractelement index out of range", &I);
634 }
635}
636
637void Lint::visitInsertElementInst(InsertElementInst &I) {
638 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: findValue(V: I.getOperand(i_nocapture: 2),
639 /*OffsetOk=*/false))) {
640 ElementCount EC = I.getType()->getElementCount();
641 Check(EC.isScalable() || CI->getValue().ult(EC.getFixedValue()),
642 "Undefined result: insertelement index out of range", &I);
643 }
644}
645
646void Lint::visitUnreachableInst(UnreachableInst &I) {
647 // This isn't undefined behavior, it's merely suspicious.
648 Check(&I == &I.getParent()->front() ||
649 std::prev(I.getIterator())->mayHaveSideEffects(),
650 "Unusual: unreachable immediately preceded by instruction without "
651 "side effects",
652 &I);
653}
654
655/// findValue - Look through bitcasts and simple memory reference patterns
656/// to identify an equivalent, but more informative, value. If OffsetOk
657/// is true, look through getelementptrs with non-zero offsets too.
658///
659/// Most analysis passes don't require this logic, because instcombine
660/// will simplify most of these kinds of things away. But it's a goal of
661/// this Lint pass to be useful even on non-optimized IR.
662Value *Lint::findValue(Value *V, bool OffsetOk) const {
663 SmallPtrSet<Value *, 4> Visited;
664 return findValueImpl(V, OffsetOk, Visited);
665}
666
667/// findValueImpl - Implementation helper for findValue.
668Value *Lint::findValueImpl(Value *V, bool OffsetOk,
669 SmallPtrSetImpl<Value *> &Visited) const {
670 // Detect self-referential values.
671 if (!Visited.insert(Ptr: V).second)
672 return PoisonValue::get(T: V->getType());
673
674 // TODO: Look through sext or zext cast, when the result is known to
675 // be interpreted as signed or unsigned, respectively.
676 // TODO: Look through eliminable cast pairs.
677 // TODO: Look through calls with unique return values.
678 // TODO: Look through vector insert/extract/shuffle.
679 V = OffsetOk ? getUnderlyingObject(V) : V->stripPointerCasts();
680 if (LoadInst *L = dyn_cast<LoadInst>(Val: V)) {
681 BasicBlock::iterator BBI = L->getIterator();
682 BasicBlock *BB = L->getParent();
683 SmallPtrSet<BasicBlock *, 4> VisitedBlocks;
684 BatchAAResults BatchAA(*AA);
685 for (;;) {
686 if (!VisitedBlocks.insert(Ptr: BB).second)
687 break;
688 if (Value *U =
689 FindAvailableLoadedValue(Load: L, ScanBB: BB, ScanFrom&: BBI, MaxInstsToScan: DefMaxInstsToScan, AA: &BatchAA))
690 return findValueImpl(V: U, OffsetOk, Visited);
691 if (BBI != BB->begin())
692 break;
693 BB = BB->getUniquePredecessor();
694 if (!BB)
695 break;
696 BBI = BB->end();
697 }
698 } else if (PHINode *PN = dyn_cast<PHINode>(Val: V)) {
699 if (Value *W = PN->hasConstantValue())
700 return findValueImpl(V: W, OffsetOk, Visited);
701 } else if (CastInst *CI = dyn_cast<CastInst>(Val: V)) {
702 if (CI->isNoopCast(DL: *DL))
703 return findValueImpl(V: CI->getOperand(i_nocapture: 0), OffsetOk, Visited);
704 } else if (ExtractValueInst *Ex = dyn_cast<ExtractValueInst>(Val: V)) {
705 if (Value *W =
706 FindInsertedValue(V: Ex->getAggregateOperand(), idx_range: Ex->getIndices()))
707 if (W != V)
708 return findValueImpl(V: W, OffsetOk, Visited);
709 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: V)) {
710 // Same as above, but for ConstantExpr instead of Instruction.
711 if (Instruction::isCast(Opcode: CE->getOpcode())) {
712 if (CastInst::isNoopCast(Opcode: Instruction::CastOps(CE->getOpcode()),
713 SrcTy: CE->getOperand(i_nocapture: 0)->getType(), DstTy: CE->getType(),
714 DL: *DL))
715 return findValueImpl(V: CE->getOperand(i_nocapture: 0), OffsetOk, Visited);
716 }
717 }
718
719 // As a last resort, try SimplifyInstruction or constant folding.
720 if (Instruction *Inst = dyn_cast<Instruction>(Val: V)) {
721 if (Value *W = simplifyInstruction(I: Inst, Q: {*DL, TLI, DT, AC}))
722 return findValueImpl(V: W, OffsetOk, Visited);
723 } else if (auto *C = dyn_cast<Constant>(Val: V)) {
724 Value *W = ConstantFoldConstant(C, DL: *DL, TLI);
725 if (W != V)
726 return findValueImpl(V: W, OffsetOk, Visited);
727 }
728
729 return V;
730}
731
732PreservedAnalyses LintPass::run(Function &F, FunctionAnalysisManager &AM) {
733 auto *Mod = F.getParent();
734 auto *DL = &F.getDataLayout();
735 auto *AA = &AM.getResult<AAManager>(IR&: F);
736 auto *AC = &AM.getResult<AssumptionAnalysis>(IR&: F);
737 auto *DT = &AM.getResult<DominatorTreeAnalysis>(IR&: F);
738 auto *TLI = &AM.getResult<TargetLibraryAnalysis>(IR&: F);
739 Lint L(Mod, DL, AA, AC, DT, TLI);
740 L.visit(F);
741 dbgs() << L.MessagesStr.str();
742 if (AbortOnError && !L.MessagesStr.str().empty())
743 report_fatal_error(
744 reason: "linter found errors, aborting. (enabled by abort-on-error)", gen_crash_diag: false);
745 return PreservedAnalyses::all();
746}
747
748void LintPass::printPipeline(
749 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
750 PassInfoMixin<LintPass>::printPipeline(OS, MapClassName2PassName);
751 if (AbortOnError)
752 OS << "<abort-on-error>";
753}
754
755//===----------------------------------------------------------------------===//
756// Implement the public interfaces to this file...
757//===----------------------------------------------------------------------===//
758
759/// lintFunction - Check a function for errors, printing messages on stderr.
760///
761void llvm::lintFunction(const Function &f, bool AbortOnError) {
762 Function &F = const_cast<Function &>(f);
763 assert(!F.isDeclaration() && "Cannot lint external functions");
764
765 FunctionAnalysisManager FAM;
766 FAM.registerPass(PassBuilder: [&] { return TargetLibraryAnalysis(); });
767 FAM.registerPass(PassBuilder: [&] { return DominatorTreeAnalysis(); });
768 FAM.registerPass(PassBuilder: [&] { return AssumptionAnalysis(); });
769 FAM.registerPass(PassBuilder: [&] {
770 AAManager AA;
771 AA.registerFunctionAnalysis<BasicAA>();
772 AA.registerFunctionAnalysis<ScopedNoAliasAA>();
773 AA.registerFunctionAnalysis<TypeBasedAA>();
774 return AA;
775 });
776 LintPass(AbortOnError).run(F, AM&: FAM);
777}
778
779/// lintModule - Check a module for errors, printing messages on stderr.
780///
781void llvm::lintModule(const Module &M, bool AbortOnError) {
782 for (const Function &F : M) {
783 if (!F.isDeclaration())
784 lintFunction(f: F, AbortOnError);
785 }
786}
787