1 //===- CrashDebugger.cpp - Debug compilation crashes ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the bugpoint internals that narrow down compilation crashes
12 //===----------------------------------------------------------------------===//
14 #include "BugDriver.h"
15 #include "ListReducer.h"
16 #include "ToolRunner.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/IR/CFG.h"
19 #include "llvm/IR/Constants.h"
20 #include "llvm/IR/DerivedTypes.h"
21 #include "llvm/IR/Instructions.h"
22 #include "llvm/IR/LegacyPassManager.h"
23 #include "llvm/IR/Module.h"
24 #include "llvm/IR/ValueSymbolTable.h"
25 #include "llvm/IR/Verifier.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/FileUtilities.h"
29 #include "llvm/Transforms/Scalar.h"
30 #include "llvm/Transforms/Utils/Cloning.h"
37 cl::desc("Force function reduction to keep main"),
40 NoGlobalRM ("disable-global-remove",
41 cl::desc("Do not remove global variables"),
45 ReplaceFuncsWithNull("replace-funcs-with-null",
46 cl::desc("When stubbing functions, replace all uses will null"),
49 DontReducePassList("disable-pass-list-reduction",
50 cl::desc("Skip pass list reduction steps"),
55 class ReducePassList : public ListReducer<std::string> {
58 ReducePassList(BugDriver &bd) : BD(bd) {}
60 // doTest - Return true iff running the "removed" passes succeeds, and
61 // running the "Kept" passes fail when run on the output of the "removed"
62 // passes. If we return true, we update the current module of bugpoint.
64 TestResult doTest(std::vector<std::string> &Removed,
65 std::vector<std::string> &Kept,
66 std::string &Error) override;
70 ReducePassList::TestResult
71 ReducePassList::doTest(std::vector<std::string> &Prefix,
72 std::vector<std::string> &Suffix,
74 std::string PrefixOutput;
75 Module *OrigProgram = nullptr;
76 if (!Prefix.empty()) {
77 outs() << "Checking to see if these passes crash: "
78 << getPassesString(Prefix) << ": ";
79 if (BD.runPasses(BD.getProgram(), Prefix, PrefixOutput))
82 OrigProgram = BD.Program;
84 BD.Program = parseInputFile(PrefixOutput, BD.getContext()).release();
85 if (BD.Program == nullptr) {
86 errs() << BD.getToolName() << ": Error reading bitcode file '"
87 << PrefixOutput << "'!\n";
90 sys::fs::remove(PrefixOutput);
93 outs() << "Checking to see if these passes crash: "
94 << getPassesString(Suffix) << ": ";
96 if (BD.runPasses(BD.getProgram(), Suffix)) {
97 delete OrigProgram; // The suffix crashes alone...
101 // Nothing failed, restore state...
104 BD.Program = OrigProgram;
110 /// ReduceCrashingGlobalVariables - This works by removing the global
111 /// variable's initializer and seeing if the program still crashes. If it
112 /// does, then we keep that program and try again.
114 class ReduceCrashingGlobalVariables : public ListReducer<GlobalVariable*> {
116 bool (*TestFn)(const BugDriver &, Module *);
118 ReduceCrashingGlobalVariables(BugDriver &bd,
119 bool (*testFn)(const BugDriver &, Module *))
120 : BD(bd), TestFn(testFn) {}
122 TestResult doTest(std::vector<GlobalVariable*> &Prefix,
123 std::vector<GlobalVariable*> &Kept,
124 std::string &Error) override {
125 if (!Kept.empty() && TestGlobalVariables(Kept))
127 if (!Prefix.empty() && TestGlobalVariables(Prefix))
132 bool TestGlobalVariables(std::vector<GlobalVariable*> &GVs);
137 ReduceCrashingGlobalVariables::TestGlobalVariables(
138 std::vector<GlobalVariable*> &GVs) {
139 // Clone the program to try hacking it apart...
140 ValueToValueMapTy VMap;
141 Module *M = CloneModule(BD.getProgram(), VMap);
143 // Convert list to set for fast lookup...
144 std::set<GlobalVariable*> GVSet;
146 for (unsigned i = 0, e = GVs.size(); i != e; ++i) {
147 GlobalVariable* CMGV = cast<GlobalVariable>(VMap[GVs[i]]);
148 assert(CMGV && "Global Variable not in module?!");
152 outs() << "Checking for crash with only these global variables: ";
153 PrintGlobalVariableList(GVs);
156 // Loop over and delete any global variables which we aren't supposed to be
158 for (GlobalVariable &I : M->globals())
159 if (I.hasInitializer() && !GVSet.count(&I)) {
160 I.setInitializer(nullptr);
161 I.setLinkage(GlobalValue::ExternalLinkage);
164 // Try running the hacked up program...
166 BD.setNewProgram(M); // It crashed, keep the trimmed version...
168 // Make sure to use global variable pointers that point into the now-current
170 GVs.assign(GVSet.begin(), GVSet.end());
179 /// ReduceCrashingFunctions reducer - This works by removing functions and
180 /// seeing if the program still crashes. If it does, then keep the newer,
183 class ReduceCrashingFunctions : public ListReducer<Function*> {
185 bool (*TestFn)(const BugDriver &, Module *);
187 ReduceCrashingFunctions(BugDriver &bd,
188 bool (*testFn)(const BugDriver &, Module *))
189 : BD(bd), TestFn(testFn) {}
191 TestResult doTest(std::vector<Function*> &Prefix,
192 std::vector<Function*> &Kept,
193 std::string &Error) override {
194 if (!Kept.empty() && TestFuncs(Kept))
196 if (!Prefix.empty() && TestFuncs(Prefix))
201 bool TestFuncs(std::vector<Function*> &Prefix);
205 static void RemoveFunctionReferences(Module *M, const char* Name) {
206 auto *UsedVar = M->getGlobalVariable(Name, true);
207 if (!UsedVar || !UsedVar->hasInitializer()) return;
208 if (isa<ConstantAggregateZero>(UsedVar->getInitializer())) {
209 assert(UsedVar->use_empty());
210 UsedVar->eraseFromParent();
213 auto *OldUsedVal = cast<ConstantArray>(UsedVar->getInitializer());
214 std::vector<Constant*> Used;
215 for(Value *V : OldUsedVal->operand_values()) {
216 Constant *Op = cast<Constant>(V->stripPointerCasts());
217 if(!Op->isNullValue()) {
218 Used.push_back(cast<Constant>(V));
221 auto *NewValElemTy = OldUsedVal->getType()->getElementType();
222 auto *NewValTy = ArrayType::get(NewValElemTy, Used.size());
223 auto *NewUsedVal = ConstantArray::get(NewValTy, Used);
224 UsedVar->mutateType(NewUsedVal->getType()->getPointerTo());
225 UsedVar->setInitializer(NewUsedVal);
228 bool ReduceCrashingFunctions::TestFuncs(std::vector<Function*> &Funcs) {
229 // If main isn't present, claim there is no problem.
230 if (KeepMain && std::find(Funcs.begin(), Funcs.end(),
231 BD.getProgram()->getFunction("main")) ==
235 // Clone the program to try hacking it apart...
236 ValueToValueMapTy VMap;
237 Module *M = CloneModule(BD.getProgram(), VMap);
239 // Convert list to set for fast lookup...
240 std::set<Function*> Functions;
241 for (unsigned i = 0, e = Funcs.size(); i != e; ++i) {
242 Function *CMF = cast<Function>(VMap[Funcs[i]]);
243 assert(CMF && "Function not in module?!");
244 assert(CMF->getFunctionType() == Funcs[i]->getFunctionType() && "wrong ty");
245 assert(CMF->getName() == Funcs[i]->getName() && "wrong name");
246 Functions.insert(CMF);
249 outs() << "Checking for crash with only these functions: ";
250 PrintFunctionList(Funcs);
252 if (!ReplaceFuncsWithNull) {
253 // Loop over and delete any functions which we aren't supposed to be playing
255 for (Function &I : *M)
256 if (!I.isDeclaration() && !Functions.count(&I))
257 DeleteFunctionBody(&I);
259 std::vector<GlobalValue*> ToRemove;
260 // First, remove aliases to functions we're about to purge.
261 for (GlobalAlias &Alias : M->aliases()) {
262 Constant *Root = Alias.getAliasee()->stripPointerCasts();
263 Function *F = dyn_cast<Function>(Root);
265 if (Functions.count(F))
266 // We're keeping this function.
268 } else if (Root->isNullValue()) {
269 // This referenced a globalalias that we've already replaced,
270 // so we still need to replace this alias.
272 // Not a function, therefore not something we mess with.
276 PointerType *Ty = cast<PointerType>(Alias.getType());
277 Constant *Replacement = ConstantPointerNull::get(Ty);
278 Alias.replaceAllUsesWith(Replacement);
279 ToRemove.push_back(&Alias);
282 for (Function &I : *M) {
283 if (!I.isDeclaration() && !Functions.count(&I)) {
284 PointerType *Ty = cast<PointerType>(I.getType());
285 Constant *Replacement = ConstantPointerNull::get(Ty);
286 I.replaceAllUsesWith(Replacement);
287 ToRemove.push_back(&I);
291 for (auto *F : ToRemove) {
292 F->eraseFromParent();
295 // Finally, remove any null members from any global intrinsic.
296 RemoveFunctionReferences(M, "llvm.used");
297 RemoveFunctionReferences(M, "llvm.compiler.used");
299 // Try running the hacked up program...
301 BD.setNewProgram(M); // It crashed, keep the trimmed version...
303 // Make sure to use function pointers that point into the now-current
305 Funcs.assign(Functions.begin(), Functions.end());
314 /// ReduceCrashingBlocks reducer - This works by setting the terminators of
315 /// all terminators except the specified basic blocks to a 'ret' instruction,
316 /// then running the simplify-cfg pass. This has the effect of chopping up
317 /// the CFG really fast which can reduce large functions quickly.
319 class ReduceCrashingBlocks : public ListReducer<const BasicBlock*> {
321 bool (*TestFn)(const BugDriver &, Module *);
323 ReduceCrashingBlocks(BugDriver &bd,
324 bool (*testFn)(const BugDriver &, Module *))
325 : BD(bd), TestFn(testFn) {}
327 TestResult doTest(std::vector<const BasicBlock*> &Prefix,
328 std::vector<const BasicBlock*> &Kept,
329 std::string &Error) override {
330 if (!Kept.empty() && TestBlocks(Kept))
332 if (!Prefix.empty() && TestBlocks(Prefix))
337 bool TestBlocks(std::vector<const BasicBlock*> &Prefix);
341 bool ReduceCrashingBlocks::TestBlocks(std::vector<const BasicBlock*> &BBs) {
342 // Clone the program to try hacking it apart...
343 ValueToValueMapTy VMap;
344 Module *M = CloneModule(BD.getProgram(), VMap);
346 // Convert list to set for fast lookup...
347 SmallPtrSet<BasicBlock*, 8> Blocks;
348 for (unsigned i = 0, e = BBs.size(); i != e; ++i)
349 Blocks.insert(cast<BasicBlock>(VMap[BBs[i]]));
351 outs() << "Checking for crash with only these blocks:";
352 unsigned NumPrint = Blocks.size();
353 if (NumPrint > 10) NumPrint = 10;
354 for (unsigned i = 0, e = NumPrint; i != e; ++i)
355 outs() << " " << BBs[i]->getName();
356 if (NumPrint < Blocks.size())
357 outs() << "... <" << Blocks.size() << " total>";
360 // Loop over and delete any hack up any blocks that are not listed...
361 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
362 for (Function::iterator BB = I->begin(), E = I->end(); BB != E; ++BB)
363 if (!Blocks.count(&*BB) && BB->getTerminator()->getNumSuccessors()) {
364 // Loop over all of the successors of this block, deleting any PHI nodes
365 // that might include it.
366 for (succ_iterator SI = succ_begin(&*BB), E = succ_end(&*BB); SI != E;
368 (*SI)->removePredecessor(&*BB);
370 TerminatorInst *BBTerm = BB->getTerminator();
372 if (!BB->getTerminator()->getType()->isVoidTy())
373 BBTerm->replaceAllUsesWith(Constant::getNullValue(BBTerm->getType()));
375 // Replace the old terminator instruction.
376 BB->getInstList().pop_back();
377 new UnreachableInst(BB->getContext(), &*BB);
380 // The CFG Simplifier pass may delete one of the basic blocks we are
381 // interested in. If it does we need to take the block out of the list. Make
382 // a "persistent mapping" by turning basic blocks into <function, name> pairs.
383 // This won't work well if blocks are unnamed, but that is just the risk we
385 std::vector<std::pair<std::string, std::string> > BlockInfo;
387 for (BasicBlock *BB : Blocks)
388 BlockInfo.emplace_back(BB->getParent()->getName(), BB->getName());
390 // Now run the CFG simplify pass on the function...
391 std::vector<std::string> Passes;
392 Passes.push_back("simplifycfg");
393 Passes.push_back("verify");
394 std::unique_ptr<Module> New = BD.runPassesOn(M, Passes);
397 errs() << "simplifycfg failed!\n";
402 // Try running on the hacked up program...
404 BD.setNewProgram(M); // It crashed, keep the trimmed version...
406 // Make sure to use basic block pointers that point into the now-current
407 // module, and that they don't include any deleted blocks.
409 const ValueSymbolTable &GST = M->getValueSymbolTable();
410 for (unsigned i = 0, e = BlockInfo.size(); i != e; ++i) {
411 Function *F = cast<Function>(GST.lookup(BlockInfo[i].first));
412 ValueSymbolTable &ST = F->getValueSymbolTable();
413 Value* V = ST.lookup(BlockInfo[i].second);
414 if (V && V->getType() == Type::getLabelTy(V->getContext()))
415 BBs.push_back(cast<BasicBlock>(V));
419 delete M; // It didn't crash, try something else.
424 /// ReduceCrashingInstructions reducer - This works by removing the specified
425 /// non-terminator instructions and replacing them with undef.
427 class ReduceCrashingInstructions : public ListReducer<const Instruction*> {
429 bool (*TestFn)(const BugDriver &, Module *);
431 ReduceCrashingInstructions(BugDriver &bd,
432 bool (*testFn)(const BugDriver &, Module *))
433 : BD(bd), TestFn(testFn) {}
435 TestResult doTest(std::vector<const Instruction*> &Prefix,
436 std::vector<const Instruction*> &Kept,
437 std::string &Error) override {
438 if (!Kept.empty() && TestInsts(Kept))
440 if (!Prefix.empty() && TestInsts(Prefix))
445 bool TestInsts(std::vector<const Instruction*> &Prefix);
449 bool ReduceCrashingInstructions::TestInsts(std::vector<const Instruction*>
451 // Clone the program to try hacking it apart...
452 ValueToValueMapTy VMap;
453 Module *M = CloneModule(BD.getProgram(), VMap);
455 // Convert list to set for fast lookup...
456 SmallPtrSet<Instruction*, 64> Instructions;
457 for (unsigned i = 0, e = Insts.size(); i != e; ++i) {
458 assert(!isa<TerminatorInst>(Insts[i]));
459 Instructions.insert(cast<Instruction>(VMap[Insts[i]]));
462 outs() << "Checking for crash with only " << Instructions.size();
463 if (Instructions.size() == 1)
464 outs() << " instruction: ";
466 outs() << " instructions: ";
468 for (Module::iterator MI = M->begin(), ME = M->end(); MI != ME; ++MI)
469 for (Function::iterator FI = MI->begin(), FE = MI->end(); FI != FE; ++FI)
470 for (BasicBlock::iterator I = FI->begin(), E = FI->end(); I != E;) {
471 Instruction *Inst = &*I++;
472 if (!Instructions.count(Inst) && !isa<TerminatorInst>(Inst) &&
474 if (!Inst->getType()->isVoidTy())
475 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
476 Inst->eraseFromParent();
480 // Verify that this is still valid.
481 legacy::PassManager Passes;
482 Passes.add(createVerifierPass());
485 // Try running on the hacked up program...
487 BD.setNewProgram(M); // It crashed, keep the trimmed version...
489 // Make sure to use instruction pointers that point into the now-current
490 // module, and that they don't include any deleted blocks.
492 for (Instruction *Inst : Instructions)
493 Insts.push_back(Inst);
496 delete M; // It didn't crash, try something else.
500 /// DebugACrash - Given a predicate that determines whether a component crashes
501 /// on a program, try to destructively reduce the program while still keeping
502 /// the predicate true.
503 static bool DebugACrash(BugDriver &BD,
504 bool (*TestFn)(const BugDriver &, Module *),
505 std::string &Error) {
506 // See if we can get away with nuking some of the global variable initializers
509 BD.getProgram()->global_begin() != BD.getProgram()->global_end()) {
510 // Now try to reduce the number of global variable initializers in the
511 // module to something small.
512 Module *M = CloneModule(BD.getProgram());
513 bool DeletedInit = false;
515 for (Module::global_iterator I = M->global_begin(), E = M->global_end();
517 if (I->hasInitializer()) {
518 I->setInitializer(nullptr);
519 I->setLinkage(GlobalValue::ExternalLinkage);
524 delete M; // No change made...
526 // See if the program still causes a crash...
527 outs() << "\nChecking to see if we can delete global inits: ";
529 if (TestFn(BD, M)) { // Still crashes?
531 outs() << "\n*** Able to remove all global initializers!\n";
532 } else { // No longer crashes?
533 outs() << " - Removing all global inits hides problem!\n";
536 std::vector<GlobalVariable*> GVs;
538 for (Module::global_iterator I = BD.getProgram()->global_begin(),
539 E = BD.getProgram()->global_end(); I != E; ++I)
540 if (I->hasInitializer())
543 if (GVs.size() > 1 && !BugpointIsInterrupted) {
544 outs() << "\n*** Attempting to reduce the number of global "
545 << "variables in the testcase\n";
547 unsigned OldSize = GVs.size();
548 ReduceCrashingGlobalVariables(BD, TestFn).reduceList(GVs, Error);
552 if (GVs.size() < OldSize)
553 BD.EmitProgressBitcode(BD.getProgram(), "reduced-global-variables");
559 // Now try to reduce the number of functions in the module to something small.
560 std::vector<Function*> Functions;
561 for (Function &F : *BD.getProgram())
562 if (!F.isDeclaration())
563 Functions.push_back(&F);
565 if (Functions.size() > 1 && !BugpointIsInterrupted) {
566 outs() << "\n*** Attempting to reduce the number of functions "
569 unsigned OldSize = Functions.size();
570 ReduceCrashingFunctions(BD, TestFn).reduceList(Functions, Error);
572 if (Functions.size() < OldSize)
573 BD.EmitProgressBitcode(BD.getProgram(), "reduced-function");
576 // Attempt to delete entire basic blocks at a time to speed up
577 // convergence... this actually works by setting the terminator of the blocks
578 // to a return instruction then running simplifycfg, which can potentially
579 // shrinks the code dramatically quickly
581 if (!DisableSimplifyCFG && !BugpointIsInterrupted) {
582 std::vector<const BasicBlock*> Blocks;
583 for (Function &F : *BD.getProgram())
584 for (BasicBlock &BB : F)
585 Blocks.push_back(&BB);
586 unsigned OldSize = Blocks.size();
587 ReduceCrashingBlocks(BD, TestFn).reduceList(Blocks, Error);
588 if (Blocks.size() < OldSize)
589 BD.EmitProgressBitcode(BD.getProgram(), "reduced-blocks");
592 // Attempt to delete instructions using bisection. This should help out nasty
593 // cases with large basic blocks where the problem is at one end.
594 if (!BugpointIsInterrupted) {
595 std::vector<const Instruction*> Insts;
596 for (const Function &F : *BD.getProgram())
597 for (const BasicBlock &BB : F)
598 for (const Instruction &I : BB)
599 if (!isa<TerminatorInst>(&I))
602 ReduceCrashingInstructions(BD, TestFn).reduceList(Insts, Error);
605 // FIXME: This should use the list reducer to converge faster by deleting
606 // larger chunks of instructions at a time!
607 unsigned Simplification = 2;
609 if (BugpointIsInterrupted) break;
611 outs() << "\n*** Attempting to reduce testcase by deleting instruc"
612 << "tions: Simplification Level #" << Simplification << '\n';
614 // Now that we have deleted the functions that are unnecessary for the
615 // program, try to remove instructions that are not necessary to cause the
616 // crash. To do this, we loop through all of the instructions in the
617 // remaining functions, deleting them (replacing any values produced with
618 // nulls), and then running ADCE and SimplifyCFG. If the transformed input
619 // still triggers failure, keep deleting until we cannot trigger failure
622 unsigned InstructionsToSkipBeforeDeleting = 0;
625 // Loop over all of the (non-terminator) instructions remaining in the
626 // function, attempting to delete them.
627 unsigned CurInstructionNum = 0;
628 for (Module::const_iterator FI = BD.getProgram()->begin(),
629 E = BD.getProgram()->end(); FI != E; ++FI)
630 if (!FI->isDeclaration())
631 for (Function::const_iterator BI = FI->begin(), E = FI->end(); BI != E;
633 for (BasicBlock::const_iterator I = BI->begin(), E = --BI->end();
634 I != E; ++I, ++CurInstructionNum) {
635 if (InstructionsToSkipBeforeDeleting) {
636 --InstructionsToSkipBeforeDeleting;
638 if (BugpointIsInterrupted) goto ExitLoops;
640 if (isa<LandingPadInst>(I))
643 outs() << "Checking instruction: " << *I;
644 std::unique_ptr<Module> M =
645 BD.deleteInstructionFromProgram(&*I, Simplification);
647 // Find out if the pass still crashes on this pass...
648 if (TestFn(BD, M.get())) {
649 // Yup, it does, we delete the old module, and continue trying
650 // to reduce the testcase...
651 BD.setNewProgram(M.release());
652 InstructionsToSkipBeforeDeleting = CurInstructionNum;
653 goto TryAgain; // I wish I had a multi-level break here!
658 if (InstructionsToSkipBeforeDeleting) {
659 InstructionsToSkipBeforeDeleting = 0;
663 } while (Simplification);
666 // Try to clean up the testcase by running funcresolve and globaldce...
667 if (!BugpointIsInterrupted) {
668 outs() << "\n*** Attempting to perform final cleanups: ";
669 Module *M = CloneModule(BD.getProgram());
670 M = BD.performFinalCleanups(M, true).release();
672 // Find out if the pass still crashes on the cleaned up program...
674 BD.setNewProgram(M); // Yup, it does, keep the reduced version...
680 BD.EmitProgressBitcode(BD.getProgram(), "reduced-simplified");
685 static bool TestForOptimizerCrash(const BugDriver &BD, Module *M) {
686 return BD.runPasses(M);
689 /// debugOptimizerCrash - This method is called when some pass crashes on input.
690 /// It attempts to prune down the testcase to something reasonable, and figure
691 /// out exactly which pass is crashing.
693 bool BugDriver::debugOptimizerCrash(const std::string &ID) {
694 outs() << "\n*** Debugging optimizer crash!\n";
697 // Reduce the list of passes which causes the optimizer to crash...
698 if (!BugpointIsInterrupted && !DontReducePassList)
699 ReducePassList(*this).reduceList(PassesToRun, Error);
700 assert(Error.empty());
702 outs() << "\n*** Found crashing pass"
703 << (PassesToRun.size() == 1 ? ": " : "es: ")
704 << getPassesString(PassesToRun) << '\n';
706 EmitProgressBitcode(Program, ID);
708 bool Success = DebugACrash(*this, TestForOptimizerCrash, Error);
709 assert(Error.empty());
713 static bool TestForCodeGenCrash(const BugDriver &BD, Module *M) {
715 BD.compileProgram(M, &Error);
716 if (!Error.empty()) {
717 errs() << "<crash>\n";
718 return true; // Tool is still crashing.
724 /// debugCodeGeneratorCrash - This method is called when the code generator
725 /// crashes on an input. It attempts to reduce the input as much as possible
726 /// while still causing the code generator to crash.
727 bool BugDriver::debugCodeGeneratorCrash(std::string &Error) {
728 errs() << "*** Debugging code generator crash!\n";
730 return DebugACrash(*this, TestForCodeGenCrash, Error);