xref: /aosp_15_r20/external/llvm/tools/bugpoint/CrashDebugger.cpp (revision 9880d6810fe72a1726cb53787c6711e909410d58)
1 //===- CrashDebugger.cpp - Debug compilation crashes ----------------------===//
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 file defines the bugpoint internals that narrow down compilation crashes
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "BugDriver.h"
15 #include "ListReducer.h"
16 #include "ToolRunner.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/StringSet.h"
19 #include "llvm/IR/CFG.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/DerivedTypes.h"
22 #include "llvm/IR/Instructions.h"
23 #include "llvm/IR/LegacyPassManager.h"
24 #include "llvm/IR/Module.h"
25 #include "llvm/IR/ValueSymbolTable.h"
26 #include "llvm/IR/Verifier.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/FileUtilities.h"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/Transforms/Utils/Cloning.h"
32 #include <set>
33 using namespace llvm;
34 
35 namespace {
36   cl::opt<bool>
37   KeepMain("keep-main",
38            cl::desc("Force function reduction to keep main"),
39            cl::init(false));
40   cl::opt<bool>
41   NoGlobalRM ("disable-global-remove",
42          cl::desc("Do not remove global variables"),
43          cl::init(false));
44 
45   cl::opt<bool>
46   ReplaceFuncsWithNull("replace-funcs-with-null",
47          cl::desc("When stubbing functions, replace all uses will null"),
48          cl::init(false));
49   cl::opt<bool>
50   DontReducePassList("disable-pass-list-reduction",
51                      cl::desc("Skip pass list reduction steps"),
52                      cl::init(false));
53 
54   cl::opt<bool> NoNamedMDRM("disable-namedmd-remove",
55                             cl::desc("Do not remove global named metadata"),
56                             cl::init(false));
57 }
58 
59 namespace llvm {
60   class ReducePassList : public ListReducer<std::string> {
61     BugDriver &BD;
62   public:
ReducePassList(BugDriver & bd)63     ReducePassList(BugDriver &bd) : BD(bd) {}
64 
65     // doTest - Return true iff running the "removed" passes succeeds, and
66     // running the "Kept" passes fail when run on the output of the "removed"
67     // passes.  If we return true, we update the current module of bugpoint.
68     //
69     TestResult doTest(std::vector<std::string> &Removed,
70                       std::vector<std::string> &Kept,
71                       std::string &Error) override;
72   };
73 }
74 
75 ReducePassList::TestResult
doTest(std::vector<std::string> & Prefix,std::vector<std::string> & Suffix,std::string & Error)76 ReducePassList::doTest(std::vector<std::string> &Prefix,
77                        std::vector<std::string> &Suffix,
78                        std::string &Error) {
79   std::string PrefixOutput;
80   Module *OrigProgram = nullptr;
81   if (!Prefix.empty()) {
82     outs() << "Checking to see if these passes crash: "
83            << getPassesString(Prefix) << ": ";
84     if (BD.runPasses(BD.getProgram(), Prefix, PrefixOutput))
85       return KeepPrefix;
86 
87     OrigProgram = BD.Program;
88 
89     BD.Program = parseInputFile(PrefixOutput, BD.getContext()).release();
90     if (BD.Program == nullptr) {
91       errs() << BD.getToolName() << ": Error reading bitcode file '"
92              << PrefixOutput << "'!\n";
93       exit(1);
94     }
95     sys::fs::remove(PrefixOutput);
96   }
97 
98   outs() << "Checking to see if these passes crash: "
99          << getPassesString(Suffix) << ": ";
100 
101   if (BD.runPasses(BD.getProgram(), Suffix)) {
102     delete OrigProgram;            // The suffix crashes alone...
103     return KeepSuffix;
104   }
105 
106   // Nothing failed, restore state...
107   if (OrigProgram) {
108     delete BD.Program;
109     BD.Program = OrigProgram;
110   }
111   return NoFailure;
112 }
113 
114 namespace {
115   /// ReduceCrashingGlobalVariables - This works by removing the global
116   /// variable's initializer and seeing if the program still crashes. If it
117   /// does, then we keep that program and try again.
118   ///
119   class ReduceCrashingGlobalVariables : public ListReducer<GlobalVariable*> {
120     BugDriver &BD;
121     bool (*TestFn)(const BugDriver &, Module *);
122   public:
ReduceCrashingGlobalVariables(BugDriver & bd,bool (* testFn)(const BugDriver &,Module *))123     ReduceCrashingGlobalVariables(BugDriver &bd,
124                                   bool (*testFn)(const BugDriver &, Module *))
125       : BD(bd), TestFn(testFn) {}
126 
doTest(std::vector<GlobalVariable * > & Prefix,std::vector<GlobalVariable * > & Kept,std::string & Error)127     TestResult doTest(std::vector<GlobalVariable*> &Prefix,
128                       std::vector<GlobalVariable*> &Kept,
129                       std::string &Error) override {
130       if (!Kept.empty() && TestGlobalVariables(Kept))
131         return KeepSuffix;
132       if (!Prefix.empty() && TestGlobalVariables(Prefix))
133         return KeepPrefix;
134       return NoFailure;
135     }
136 
137     bool TestGlobalVariables(std::vector<GlobalVariable*> &GVs);
138   };
139 }
140 
141 bool
TestGlobalVariables(std::vector<GlobalVariable * > & GVs)142 ReduceCrashingGlobalVariables::TestGlobalVariables(
143                               std::vector<GlobalVariable*> &GVs) {
144   // Clone the program to try hacking it apart...
145   ValueToValueMapTy VMap;
146   Module *M = CloneModule(BD.getProgram(), VMap).release();
147 
148   // Convert list to set for fast lookup...
149   std::set<GlobalVariable*> GVSet;
150 
151   for (unsigned i = 0, e = GVs.size(); i != e; ++i) {
152     GlobalVariable* CMGV = cast<GlobalVariable>(VMap[GVs[i]]);
153     assert(CMGV && "Global Variable not in module?!");
154     GVSet.insert(CMGV);
155   }
156 
157   outs() << "Checking for crash with only these global variables: ";
158   PrintGlobalVariableList(GVs);
159   outs() << ": ";
160 
161   // Loop over and delete any global variables which we aren't supposed to be
162   // playing with...
163   for (GlobalVariable &I : M->globals())
164     if (I.hasInitializer() && !GVSet.count(&I)) {
165       DeleteGlobalInitializer(&I);
166       I.setLinkage(GlobalValue::ExternalLinkage);
167       I.setComdat(nullptr);
168     }
169 
170   // Try running the hacked up program...
171   if (TestFn(BD, M)) {
172     BD.setNewProgram(M);        // It crashed, keep the trimmed version...
173 
174     // Make sure to use global variable pointers that point into the now-current
175     // module.
176     GVs.assign(GVSet.begin(), GVSet.end());
177     return true;
178   }
179 
180   delete M;
181   return false;
182 }
183 
184 namespace {
185   /// ReduceCrashingFunctions reducer - This works by removing functions and
186   /// seeing if the program still crashes. If it does, then keep the newer,
187   /// smaller program.
188   ///
189   class ReduceCrashingFunctions : public ListReducer<Function*> {
190     BugDriver &BD;
191     bool (*TestFn)(const BugDriver &, Module *);
192   public:
ReduceCrashingFunctions(BugDriver & bd,bool (* testFn)(const BugDriver &,Module *))193     ReduceCrashingFunctions(BugDriver &bd,
194                             bool (*testFn)(const BugDriver &, Module *))
195       : BD(bd), TestFn(testFn) {}
196 
doTest(std::vector<Function * > & Prefix,std::vector<Function * > & Kept,std::string & Error)197     TestResult doTest(std::vector<Function*> &Prefix,
198                       std::vector<Function*> &Kept,
199                       std::string &Error) override {
200       if (!Kept.empty() && TestFuncs(Kept))
201         return KeepSuffix;
202       if (!Prefix.empty() && TestFuncs(Prefix))
203         return KeepPrefix;
204       return NoFailure;
205     }
206 
207     bool TestFuncs(std::vector<Function*> &Prefix);
208   };
209 }
210 
RemoveFunctionReferences(Module * M,const char * Name)211 static void RemoveFunctionReferences(Module *M, const char* Name) {
212   auto *UsedVar = M->getGlobalVariable(Name, true);
213   if (!UsedVar || !UsedVar->hasInitializer()) return;
214   if (isa<ConstantAggregateZero>(UsedVar->getInitializer())) {
215     assert(UsedVar->use_empty());
216     UsedVar->eraseFromParent();
217     return;
218   }
219   auto *OldUsedVal = cast<ConstantArray>(UsedVar->getInitializer());
220   std::vector<Constant*> Used;
221   for(Value *V : OldUsedVal->operand_values()) {
222     Constant *Op = cast<Constant>(V->stripPointerCasts());
223     if(!Op->isNullValue()) {
224       Used.push_back(cast<Constant>(V));
225     }
226   }
227   auto *NewValElemTy = OldUsedVal->getType()->getElementType();
228   auto *NewValTy = ArrayType::get(NewValElemTy, Used.size());
229   auto *NewUsedVal = ConstantArray::get(NewValTy, Used);
230   UsedVar->mutateType(NewUsedVal->getType()->getPointerTo());
231   UsedVar->setInitializer(NewUsedVal);
232 }
233 
TestFuncs(std::vector<Function * > & Funcs)234 bool ReduceCrashingFunctions::TestFuncs(std::vector<Function*> &Funcs) {
235   // If main isn't present, claim there is no problem.
236   if (KeepMain && std::find(Funcs.begin(), Funcs.end(),
237                             BD.getProgram()->getFunction("main")) ==
238                       Funcs.end())
239     return false;
240 
241   // Clone the program to try hacking it apart...
242   ValueToValueMapTy VMap;
243   Module *M = CloneModule(BD.getProgram(), VMap).release();
244 
245   // Convert list to set for fast lookup...
246   std::set<Function*> Functions;
247   for (unsigned i = 0, e = Funcs.size(); i != e; ++i) {
248     Function *CMF = cast<Function>(VMap[Funcs[i]]);
249     assert(CMF && "Function not in module?!");
250     assert(CMF->getFunctionType() == Funcs[i]->getFunctionType() && "wrong ty");
251     assert(CMF->getName() == Funcs[i]->getName() && "wrong name");
252     Functions.insert(CMF);
253   }
254 
255   outs() << "Checking for crash with only these functions: ";
256   PrintFunctionList(Funcs);
257   outs() << ": ";
258   if (!ReplaceFuncsWithNull) {
259     // Loop over and delete any functions which we aren't supposed to be playing
260     // with...
261     for (Function &I : *M)
262       if (!I.isDeclaration() && !Functions.count(&I))
263         DeleteFunctionBody(&I);
264   } else {
265     std::vector<GlobalValue*> ToRemove;
266     // First, remove aliases to functions we're about to purge.
267     for (GlobalAlias &Alias : M->aliases()) {
268       GlobalObject *Root = Alias.getBaseObject();
269       Function *F = dyn_cast_or_null<Function>(Root);
270       if (F) {
271         if (Functions.count(F))
272           // We're keeping this function.
273           continue;
274       } else if (Root->isNullValue()) {
275         // This referenced a globalalias that we've already replaced,
276         // so we still need to replace this alias.
277       } else if (!F) {
278         // Not a function, therefore not something we mess with.
279         continue;
280       }
281 
282       PointerType *Ty = cast<PointerType>(Alias.getType());
283       Constant *Replacement = ConstantPointerNull::get(Ty);
284       Alias.replaceAllUsesWith(Replacement);
285       ToRemove.push_back(&Alias);
286     }
287 
288     for (Function &I : *M) {
289       if (!I.isDeclaration() && !Functions.count(&I)) {
290         PointerType *Ty = cast<PointerType>(I.getType());
291         Constant *Replacement = ConstantPointerNull::get(Ty);
292         I.replaceAllUsesWith(Replacement);
293         ToRemove.push_back(&I);
294       }
295     }
296 
297     for (auto *F : ToRemove) {
298       F->eraseFromParent();
299     }
300 
301     // Finally, remove any null members from any global intrinsic.
302     RemoveFunctionReferences(M, "llvm.used");
303     RemoveFunctionReferences(M, "llvm.compiler.used");
304   }
305   // Try running the hacked up program...
306   if (TestFn(BD, M)) {
307     BD.setNewProgram(M);        // It crashed, keep the trimmed version...
308 
309     // Make sure to use function pointers that point into the now-current
310     // module.
311     Funcs.assign(Functions.begin(), Functions.end());
312     return true;
313   }
314   delete M;
315   return false;
316 }
317 
318 
319 namespace {
320   /// ReduceCrashingBlocks reducer - This works by setting the terminators of
321   /// all terminators except the specified basic blocks to a 'ret' instruction,
322   /// then running the simplify-cfg pass.  This has the effect of chopping up
323   /// the CFG really fast which can reduce large functions quickly.
324   ///
325   class ReduceCrashingBlocks : public ListReducer<const BasicBlock*> {
326     BugDriver &BD;
327     bool (*TestFn)(const BugDriver &, Module *);
328   public:
ReduceCrashingBlocks(BugDriver & bd,bool (* testFn)(const BugDriver &,Module *))329     ReduceCrashingBlocks(BugDriver &bd,
330                          bool (*testFn)(const BugDriver &, Module *))
331       : BD(bd), TestFn(testFn) {}
332 
doTest(std::vector<const BasicBlock * > & Prefix,std::vector<const BasicBlock * > & Kept,std::string & Error)333     TestResult doTest(std::vector<const BasicBlock*> &Prefix,
334                       std::vector<const BasicBlock*> &Kept,
335                       std::string &Error) override {
336       if (!Kept.empty() && TestBlocks(Kept))
337         return KeepSuffix;
338       if (!Prefix.empty() && TestBlocks(Prefix))
339         return KeepPrefix;
340       return NoFailure;
341     }
342 
343     bool TestBlocks(std::vector<const BasicBlock*> &Prefix);
344   };
345 }
346 
TestBlocks(std::vector<const BasicBlock * > & BBs)347 bool ReduceCrashingBlocks::TestBlocks(std::vector<const BasicBlock*> &BBs) {
348   // Clone the program to try hacking it apart...
349   ValueToValueMapTy VMap;
350   Module *M = CloneModule(BD.getProgram(), VMap).release();
351 
352   // Convert list to set for fast lookup...
353   SmallPtrSet<BasicBlock*, 8> Blocks;
354   for (unsigned i = 0, e = BBs.size(); i != e; ++i)
355     Blocks.insert(cast<BasicBlock>(VMap[BBs[i]]));
356 
357   outs() << "Checking for crash with only these blocks:";
358   unsigned NumPrint = Blocks.size();
359   if (NumPrint > 10) NumPrint = 10;
360   for (unsigned i = 0, e = NumPrint; i != e; ++i)
361     outs() << " " << BBs[i]->getName();
362   if (NumPrint < Blocks.size())
363     outs() << "... <" << Blocks.size() << " total>";
364   outs() << ": ";
365 
366   // Loop over and delete any hack up any blocks that are not listed...
367   for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
368     for (Function::iterator BB = I->begin(), E = I->end(); BB != E; ++BB)
369       if (!Blocks.count(&*BB) && BB->getTerminator()->getNumSuccessors()) {
370         // Loop over all of the successors of this block, deleting any PHI nodes
371         // that might include it.
372         for (succ_iterator SI = succ_begin(&*BB), E = succ_end(&*BB); SI != E;
373              ++SI)
374           (*SI)->removePredecessor(&*BB);
375 
376         TerminatorInst *BBTerm = BB->getTerminator();
377         if (BBTerm->isEHPad() || BBTerm->getType()->isTokenTy())
378           continue;
379         if (!BBTerm->getType()->isVoidTy())
380           BBTerm->replaceAllUsesWith(Constant::getNullValue(BBTerm->getType()));
381 
382         // Replace the old terminator instruction.
383         BB->getInstList().pop_back();
384         new UnreachableInst(BB->getContext(), &*BB);
385       }
386 
387   // The CFG Simplifier pass may delete one of the basic blocks we are
388   // interested in.  If it does we need to take the block out of the list.  Make
389   // a "persistent mapping" by turning basic blocks into <function, name> pairs.
390   // This won't work well if blocks are unnamed, but that is just the risk we
391   // have to take.
392   std::vector<std::pair<std::string, std::string> > BlockInfo;
393 
394   for (BasicBlock *BB : Blocks)
395     BlockInfo.emplace_back(BB->getParent()->getName(), BB->getName());
396 
397   // Now run the CFG simplify pass on the function...
398   std::vector<std::string> Passes;
399   Passes.push_back("simplifycfg");
400   Passes.push_back("verify");
401   std::unique_ptr<Module> New = BD.runPassesOn(M, Passes);
402   delete M;
403   if (!New) {
404     errs() << "simplifycfg failed!\n";
405     exit(1);
406   }
407   M = New.release();
408 
409   // Try running on the hacked up program...
410   if (TestFn(BD, M)) {
411     BD.setNewProgram(M);      // It crashed, keep the trimmed version...
412 
413     // Make sure to use basic block pointers that point into the now-current
414     // module, and that they don't include any deleted blocks.
415     BBs.clear();
416     const ValueSymbolTable &GST = M->getValueSymbolTable();
417     for (unsigned i = 0, e = BlockInfo.size(); i != e; ++i) {
418       Function *F = cast<Function>(GST.lookup(BlockInfo[i].first));
419       ValueSymbolTable &ST = F->getValueSymbolTable();
420       Value* V = ST.lookup(BlockInfo[i].second);
421       if (V && V->getType() == Type::getLabelTy(V->getContext()))
422         BBs.push_back(cast<BasicBlock>(V));
423     }
424     return true;
425   }
426   delete M;  // It didn't crash, try something else.
427   return false;
428 }
429 
430 namespace {
431   /// ReduceCrashingInstructions reducer - This works by removing the specified
432   /// non-terminator instructions and replacing them with undef.
433   ///
434   class ReduceCrashingInstructions : public ListReducer<const Instruction*> {
435     BugDriver &BD;
436     bool (*TestFn)(const BugDriver &, Module *);
437   public:
ReduceCrashingInstructions(BugDriver & bd,bool (* testFn)(const BugDriver &,Module *))438     ReduceCrashingInstructions(BugDriver &bd,
439                                bool (*testFn)(const BugDriver &, Module *))
440       : BD(bd), TestFn(testFn) {}
441 
doTest(std::vector<const Instruction * > & Prefix,std::vector<const Instruction * > & Kept,std::string & Error)442     TestResult doTest(std::vector<const Instruction*> &Prefix,
443                       std::vector<const Instruction*> &Kept,
444                       std::string &Error) override {
445       if (!Kept.empty() && TestInsts(Kept))
446         return KeepSuffix;
447       if (!Prefix.empty() && TestInsts(Prefix))
448         return KeepPrefix;
449       return NoFailure;
450     }
451 
452     bool TestInsts(std::vector<const Instruction*> &Prefix);
453   };
454 }
455 
TestInsts(std::vector<const Instruction * > & Insts)456 bool ReduceCrashingInstructions::TestInsts(std::vector<const Instruction*>
457                                            &Insts) {
458   // Clone the program to try hacking it apart...
459   ValueToValueMapTy VMap;
460   Module *M = CloneModule(BD.getProgram(), VMap).release();
461 
462   // Convert list to set for fast lookup...
463   SmallPtrSet<Instruction*, 32> Instructions;
464   for (unsigned i = 0, e = Insts.size(); i != e; ++i) {
465     assert(!isa<TerminatorInst>(Insts[i]));
466     Instructions.insert(cast<Instruction>(VMap[Insts[i]]));
467   }
468 
469   outs() << "Checking for crash with only " << Instructions.size();
470   if (Instructions.size() == 1)
471     outs() << " instruction: ";
472   else
473     outs() << " instructions: ";
474 
475   for (Module::iterator MI = M->begin(), ME = M->end(); MI != ME; ++MI)
476     for (Function::iterator FI = MI->begin(), FE = MI->end(); FI != FE; ++FI)
477       for (BasicBlock::iterator I = FI->begin(), E = FI->end(); I != E;) {
478         Instruction *Inst = &*I++;
479         if (!Instructions.count(Inst) && !isa<TerminatorInst>(Inst) &&
480             !Inst->isEHPad() && !Inst->getType()->isTokenTy()) {
481           if (!Inst->getType()->isVoidTy())
482             Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
483           Inst->eraseFromParent();
484         }
485       }
486 
487   // Verify that this is still valid.
488   legacy::PassManager Passes;
489   Passes.add(createVerifierPass());
490   Passes.run(*M);
491 
492   // Try running on the hacked up program...
493   if (TestFn(BD, M)) {
494     BD.setNewProgram(M);      // It crashed, keep the trimmed version...
495 
496     // Make sure to use instruction pointers that point into the now-current
497     // module, and that they don't include any deleted blocks.
498     Insts.clear();
499     for (Instruction *Inst : Instructions)
500       Insts.push_back(Inst);
501     return true;
502   }
503   delete M;  // It didn't crash, try something else.
504   return false;
505 }
506 
507 namespace {
508 // Reduce the list of Named Metadata nodes. We keep this as a list of
509 // names to avoid having to convert back and forth every time.
510 class ReduceCrashingNamedMD : public ListReducer<std::string> {
511   BugDriver &BD;
512   bool (*TestFn)(const BugDriver &, Module *);
513 
514 public:
ReduceCrashingNamedMD(BugDriver & bd,bool (* testFn)(const BugDriver &,Module *))515   ReduceCrashingNamedMD(BugDriver &bd,
516                         bool (*testFn)(const BugDriver &, Module *))
517       : BD(bd), TestFn(testFn) {}
518 
doTest(std::vector<std::string> & Prefix,std::vector<std::string> & Kept,std::string & Error)519   TestResult doTest(std::vector<std::string> &Prefix,
520                     std::vector<std::string> &Kept,
521                     std::string &Error) override {
522     if (!Kept.empty() && TestNamedMDs(Kept))
523       return KeepSuffix;
524     if (!Prefix.empty() && TestNamedMDs(Prefix))
525       return KeepPrefix;
526     return NoFailure;
527   }
528 
529   bool TestNamedMDs(std::vector<std::string> &NamedMDs);
530 };
531 }
532 
TestNamedMDs(std::vector<std::string> & NamedMDs)533 bool ReduceCrashingNamedMD::TestNamedMDs(std::vector<std::string> &NamedMDs) {
534 
535   ValueToValueMapTy VMap;
536   Module *M = CloneModule(BD.getProgram(), VMap).release();
537 
538   outs() << "Checking for crash with only these named metadata nodes:";
539   unsigned NumPrint = std::min<size_t>(NamedMDs.size(), 10);
540   for (unsigned i = 0, e = NumPrint; i != e; ++i)
541     outs() << " " << NamedMDs[i];
542   if (NumPrint < NamedMDs.size())
543     outs() << "... <" << NamedMDs.size() << " total>";
544   outs() << ": ";
545 
546   // Make a StringMap for faster lookup
547   StringSet<> Names;
548   for (const std::string &Name : NamedMDs)
549     Names.insert(Name);
550 
551   // First collect all the metadata to delete in a vector, then
552   // delete them all at once to avoid invalidating the iterator
553   std::vector<NamedMDNode *> ToDelete;
554   ToDelete.reserve(M->named_metadata_size() - Names.size());
555   for (auto &NamedMD : M->named_metadata())
556     // Always keep a nonempty llvm.dbg.cu because the Verifier would complain.
557     if (!Names.count(NamedMD.getName()) &&
558         (!(NamedMD.getName() == "llvm.dbg.cu" && NamedMD.getNumOperands() > 0)))
559       ToDelete.push_back(&NamedMD);
560 
561   for (auto *NamedMD : ToDelete)
562     NamedMD->eraseFromParent();
563 
564   // Verify that this is still valid.
565   legacy::PassManager Passes;
566   Passes.add(createVerifierPass());
567   Passes.run(*M);
568 
569   // Try running on the hacked up program...
570   if (TestFn(BD, M)) {
571     BD.setNewProgram(M); // It crashed, keep the trimmed version...
572     return true;
573   }
574   delete M; // It didn't crash, try something else.
575   return false;
576 }
577 
578 namespace {
579 // Reduce the list of operands to named metadata nodes
580 class ReduceCrashingNamedMDOps : public ListReducer<const MDNode *> {
581   BugDriver &BD;
582   bool (*TestFn)(const BugDriver &, Module *);
583 
584 public:
ReduceCrashingNamedMDOps(BugDriver & bd,bool (* testFn)(const BugDriver &,Module *))585   ReduceCrashingNamedMDOps(BugDriver &bd,
586                            bool (*testFn)(const BugDriver &, Module *))
587       : BD(bd), TestFn(testFn) {}
588 
doTest(std::vector<const MDNode * > & Prefix,std::vector<const MDNode * > & Kept,std::string & Error)589   TestResult doTest(std::vector<const MDNode *> &Prefix,
590                     std::vector<const MDNode *> &Kept,
591                     std::string &Error) override {
592     if (!Kept.empty() && TestNamedMDOps(Kept))
593       return KeepSuffix;
594     if (!Prefix.empty() && TestNamedMDOps(Prefix))
595       return KeepPrefix;
596     return NoFailure;
597   }
598 
599   bool TestNamedMDOps(std::vector<const MDNode *> &NamedMDOps);
600 };
601 }
602 
TestNamedMDOps(std::vector<const MDNode * > & NamedMDOps)603 bool ReduceCrashingNamedMDOps::TestNamedMDOps(
604     std::vector<const MDNode *> &NamedMDOps) {
605   // Convert list to set for fast lookup...
606   SmallPtrSet<const MDNode *, 32> OldMDNodeOps;
607   for (unsigned i = 0, e = NamedMDOps.size(); i != e; ++i) {
608     OldMDNodeOps.insert(NamedMDOps[i]);
609   }
610 
611   outs() << "Checking for crash with only " << OldMDNodeOps.size();
612   if (OldMDNodeOps.size() == 1)
613     outs() << " named metadata operand: ";
614   else
615     outs() << " named metadata operands: ";
616 
617   ValueToValueMapTy VMap;
618   Module *M = CloneModule(BD.getProgram(), VMap).release();
619 
620   // This is a little wasteful. In the future it might be good if we could have
621   // these dropped during cloning.
622   for (auto &NamedMD : BD.getProgram()->named_metadata()) {
623     // Drop the old one and create a new one
624     M->eraseNamedMetadata(M->getNamedMetadata(NamedMD.getName()));
625     NamedMDNode *NewNamedMDNode =
626         M->getOrInsertNamedMetadata(NamedMD.getName());
627     for (MDNode *op : NamedMD.operands())
628       if (OldMDNodeOps.count(op))
629         NewNamedMDNode->addOperand(cast<MDNode>(MapMetadata(op, VMap)));
630   }
631 
632   // Verify that this is still valid.
633   legacy::PassManager Passes;
634   Passes.add(createVerifierPass());
635   Passes.run(*M);
636 
637   // Try running on the hacked up program...
638   if (TestFn(BD, M)) {
639     // Make sure to use instruction pointers that point into the now-current
640     // module, and that they don't include any deleted blocks.
641     NamedMDOps.clear();
642     for (const MDNode *Node : OldMDNodeOps)
643       NamedMDOps.push_back(cast<MDNode>(*VMap.getMappedMD(Node)));
644 
645     BD.setNewProgram(M); // It crashed, keep the trimmed version...
646     return true;
647   }
648   delete M; // It didn't crash, try something else.
649   return false;
650 }
651 
ReduceGlobalInitializers(BugDriver & BD,bool (* TestFn)(const BugDriver &,Module *),std::string & Error)652 static void ReduceGlobalInitializers(BugDriver &BD,
653                                      bool (*TestFn)(const BugDriver &, Module *),
654                                      std::string &Error) {
655   if (BD.getProgram()->global_begin() != BD.getProgram()->global_end()) {
656     // Now try to reduce the number of global variable initializers in the
657     // module to something small.
658     Module *M = CloneModule(BD.getProgram()).release();
659     bool DeletedInit = false;
660 
661     for (Module::global_iterator I = M->global_begin(), E = M->global_end();
662          I != E; ++I)
663       if (I->hasInitializer()) {
664         DeleteGlobalInitializer(&*I);
665         I->setLinkage(GlobalValue::ExternalLinkage);
666         I->setComdat(nullptr);
667         DeletedInit = true;
668       }
669 
670     if (!DeletedInit) {
671       delete M;  // No change made...
672     } else {
673       // See if the program still causes a crash...
674       outs() << "\nChecking to see if we can delete global inits: ";
675 
676       if (TestFn(BD, M)) {      // Still crashes?
677         BD.setNewProgram(M);
678         outs() << "\n*** Able to remove all global initializers!\n";
679       } else {                  // No longer crashes?
680         outs() << "  - Removing all global inits hides problem!\n";
681         delete M;
682 
683         std::vector<GlobalVariable*> GVs;
684 
685         for (Module::global_iterator I = BD.getProgram()->global_begin(),
686                E = BD.getProgram()->global_end(); I != E; ++I)
687           if (I->hasInitializer())
688             GVs.push_back(&*I);
689 
690         if (GVs.size() > 1 && !BugpointIsInterrupted) {
691           outs() << "\n*** Attempting to reduce the number of global "
692                     << "variables in the testcase\n";
693 
694           unsigned OldSize = GVs.size();
695           ReduceCrashingGlobalVariables(BD, TestFn).reduceList(GVs, Error);
696           assert(!Error.empty());
697 
698           if (GVs.size() < OldSize)
699             BD.EmitProgressBitcode(BD.getProgram(), "reduced-global-variables");
700         }
701       }
702     }
703   }
704 }
705 
ReduceInsts(BugDriver & BD,bool (* TestFn)(const BugDriver &,Module *),std::string & Error)706 static void ReduceInsts(BugDriver &BD,
707                         bool (*TestFn)(const BugDriver &, Module *),
708                         std::string &Error) {
709   // Attempt to delete instructions using bisection. This should help out nasty
710   // cases with large basic blocks where the problem is at one end.
711   if (!BugpointIsInterrupted) {
712     std::vector<const Instruction*> Insts;
713     for (const Function &F : *BD.getProgram())
714       for (const BasicBlock &BB : F)
715         for (const Instruction &I : BB)
716           if (!isa<TerminatorInst>(&I))
717             Insts.push_back(&I);
718 
719     ReduceCrashingInstructions(BD, TestFn).reduceList(Insts, Error);
720   }
721 
722   unsigned Simplification = 2;
723   do {
724     if (BugpointIsInterrupted)
725       return;
726     --Simplification;
727     outs() << "\n*** Attempting to reduce testcase by deleting instruc"
728            << "tions: Simplification Level #" << Simplification << '\n';
729 
730     // Now that we have deleted the functions that are unnecessary for the
731     // program, try to remove instructions that are not necessary to cause the
732     // crash.  To do this, we loop through all of the instructions in the
733     // remaining functions, deleting them (replacing any values produced with
734     // nulls), and then running ADCE and SimplifyCFG.  If the transformed input
735     // still triggers failure, keep deleting until we cannot trigger failure
736     // anymore.
737     //
738     unsigned InstructionsToSkipBeforeDeleting = 0;
739   TryAgain:
740 
741     // Loop over all of the (non-terminator) instructions remaining in the
742     // function, attempting to delete them.
743     unsigned CurInstructionNum = 0;
744     for (Module::const_iterator FI = BD.getProgram()->begin(),
745            E = BD.getProgram()->end(); FI != E; ++FI)
746       if (!FI->isDeclaration())
747         for (Function::const_iterator BI = FI->begin(), E = FI->end(); BI != E;
748              ++BI)
749           for (BasicBlock::const_iterator I = BI->begin(), E = --BI->end();
750                I != E; ++I, ++CurInstructionNum) {
751             if (InstructionsToSkipBeforeDeleting) {
752               --InstructionsToSkipBeforeDeleting;
753             } else {
754               if (BugpointIsInterrupted)
755                 return;
756 
757               if (I->isEHPad() || I->getType()->isTokenTy())
758                 continue;
759 
760               outs() << "Checking instruction: " << *I;
761               std::unique_ptr<Module> M =
762                   BD.deleteInstructionFromProgram(&*I, Simplification);
763 
764               // Find out if the pass still crashes on this pass...
765               if (TestFn(BD, M.get())) {
766                 // Yup, it does, we delete the old module, and continue trying
767                 // to reduce the testcase...
768                 BD.setNewProgram(M.release());
769                 InstructionsToSkipBeforeDeleting = CurInstructionNum;
770                 goto TryAgain;  // I wish I had a multi-level break here!
771               }
772             }
773           }
774 
775     if (InstructionsToSkipBeforeDeleting) {
776       InstructionsToSkipBeforeDeleting = 0;
777       goto TryAgain;
778     }
779 
780   } while (Simplification);
781   BD.EmitProgressBitcode(BD.getProgram(), "reduced-instructions");
782 }
783 
784 
785 /// DebugACrash - Given a predicate that determines whether a component crashes
786 /// on a program, try to destructively reduce the program while still keeping
787 /// the predicate true.
DebugACrash(BugDriver & BD,bool (* TestFn)(const BugDriver &,Module *),std::string & Error)788 static bool DebugACrash(BugDriver &BD,
789                         bool (*TestFn)(const BugDriver &, Module *),
790                         std::string &Error) {
791   // See if we can get away with nuking some of the global variable initializers
792   // in the program...
793   if (!NoGlobalRM)
794     ReduceGlobalInitializers(BD, TestFn, Error);
795 
796   // Now try to reduce the number of functions in the module to something small.
797   std::vector<Function*> Functions;
798   for (Function &F : *BD.getProgram())
799     if (!F.isDeclaration())
800       Functions.push_back(&F);
801 
802   if (Functions.size() > 1 && !BugpointIsInterrupted) {
803     outs() << "\n*** Attempting to reduce the number of functions "
804       "in the testcase\n";
805 
806     unsigned OldSize = Functions.size();
807     ReduceCrashingFunctions(BD, TestFn).reduceList(Functions, Error);
808 
809     if (Functions.size() < OldSize)
810       BD.EmitProgressBitcode(BD.getProgram(), "reduced-function");
811   }
812 
813   // Attempt to delete entire basic blocks at a time to speed up
814   // convergence... this actually works by setting the terminator of the blocks
815   // to a return instruction then running simplifycfg, which can potentially
816   // shrinks the code dramatically quickly
817   //
818   if (!DisableSimplifyCFG && !BugpointIsInterrupted) {
819     std::vector<const BasicBlock*> Blocks;
820     for (Function &F : *BD.getProgram())
821       for (BasicBlock &BB : F)
822         Blocks.push_back(&BB);
823     unsigned OldSize = Blocks.size();
824     ReduceCrashingBlocks(BD, TestFn).reduceList(Blocks, Error);
825     if (Blocks.size() < OldSize)
826       BD.EmitProgressBitcode(BD.getProgram(), "reduced-blocks");
827   }
828 
829   // Attempt to delete instructions using bisection. This should help out nasty
830   // cases with large basic blocks where the problem is at one end.
831   if (!BugpointIsInterrupted)
832     ReduceInsts(BD, TestFn, Error);
833 
834   if (!NoNamedMDRM) {
835     if (!BugpointIsInterrupted) {
836       // Try to reduce the amount of global metadata (particularly debug info),
837       // by dropping global named metadata that anchors them
838       outs() << "\n*** Attempting to remove named metadata: ";
839       std::vector<std::string> NamedMDNames;
840       for (auto &NamedMD : BD.getProgram()->named_metadata())
841         NamedMDNames.push_back(NamedMD.getName().str());
842       ReduceCrashingNamedMD(BD, TestFn).reduceList(NamedMDNames, Error);
843     }
844 
845     if (!BugpointIsInterrupted) {
846       // Now that we quickly dropped all the named metadata that doesn't
847       // contribute to the crash, bisect the operands of the remaining ones
848       std::vector<const MDNode *> NamedMDOps;
849       for (auto &NamedMD : BD.getProgram()->named_metadata())
850         for (auto op : NamedMD.operands())
851           NamedMDOps.push_back(op);
852       ReduceCrashingNamedMDOps(BD, TestFn).reduceList(NamedMDOps, Error);
853     }
854     BD.EmitProgressBitcode(BD.getProgram(), "reduced-named-md");
855   }
856 
857   // Try to clean up the testcase by running funcresolve and globaldce...
858   if (!BugpointIsInterrupted) {
859     outs() << "\n*** Attempting to perform final cleanups: ";
860     Module *M = CloneModule(BD.getProgram()).release();
861     M = BD.performFinalCleanups(M, true).release();
862 
863     // Find out if the pass still crashes on the cleaned up program...
864     if (TestFn(BD, M)) {
865       BD.setNewProgram(M);     // Yup, it does, keep the reduced version...
866     } else {
867       delete M;
868     }
869   }
870 
871   BD.EmitProgressBitcode(BD.getProgram(), "reduced-simplified");
872 
873   return false;
874 }
875 
TestForOptimizerCrash(const BugDriver & BD,Module * M)876 static bool TestForOptimizerCrash(const BugDriver &BD, Module *M) {
877   return BD.runPasses(M, BD.getPassesToRun());
878 }
879 
880 /// debugOptimizerCrash - This method is called when some pass crashes on input.
881 /// It attempts to prune down the testcase to something reasonable, and figure
882 /// out exactly which pass is crashing.
883 ///
debugOptimizerCrash(const std::string & ID)884 bool BugDriver::debugOptimizerCrash(const std::string &ID) {
885   outs() << "\n*** Debugging optimizer crash!\n";
886 
887   std::string Error;
888   // Reduce the list of passes which causes the optimizer to crash...
889   if (!BugpointIsInterrupted && !DontReducePassList)
890     ReducePassList(*this).reduceList(PassesToRun, Error);
891   assert(Error.empty());
892 
893   outs() << "\n*** Found crashing pass"
894          << (PassesToRun.size() == 1 ? ": " : "es: ")
895          << getPassesString(PassesToRun) << '\n';
896 
897   EmitProgressBitcode(Program, ID);
898 
899   bool Success = DebugACrash(*this, TestForOptimizerCrash, Error);
900   assert(Error.empty());
901   return Success;
902 }
903 
TestForCodeGenCrash(const BugDriver & BD,Module * M)904 static bool TestForCodeGenCrash(const BugDriver &BD, Module *M) {
905   std::string Error;
906   BD.compileProgram(M, &Error);
907   if (!Error.empty()) {
908     errs() << "<crash>\n";
909     return true;  // Tool is still crashing.
910   }
911   errs() << '\n';
912   return false;
913 }
914 
915 /// debugCodeGeneratorCrash - This method is called when the code generator
916 /// crashes on an input.  It attempts to reduce the input as much as possible
917 /// while still causing the code generator to crash.
debugCodeGeneratorCrash(std::string & Error)918 bool BugDriver::debugCodeGeneratorCrash(std::string &Error) {
919   errs() << "*** Debugging code generator crash!\n";
920 
921   return DebugACrash(*this, TestForCodeGenCrash, Error);
922 }
923