[llvm-exegesis] Fix wrong index type.
[llvm-complete.git] / tools / bugpoint / CrashDebugger.cpp
blobe973bfef4dc60eed372949860bc3ef78e0462263
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
12 //===----------------------------------------------------------------------===//
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/Analysis/TargetTransformInfo.h"
20 #include "llvm/Transforms/Utils/Local.h"
21 #include "llvm/IR/CFG.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DebugInfo.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/LegacyPassManager.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/IR/ValueSymbolTable.h"
29 #include "llvm/IR/Verifier.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/FileUtilities.h"
33 #include "llvm/Transforms/Scalar.h"
34 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
35 #include "llvm/Transforms/Utils/Cloning.h"
36 #include <set>
37 using namespace llvm;
39 namespace {
40 cl::opt<bool> KeepMain("keep-main",
41 cl::desc("Force function reduction to keep main"),
42 cl::init(false));
43 cl::opt<bool> NoGlobalRM("disable-global-remove",
44 cl::desc("Do not remove global variables"),
45 cl::init(false));
47 cl::opt<bool> ReplaceFuncsWithNull(
48 "replace-funcs-with-null",
49 cl::desc("When stubbing functions, replace all uses will null"),
50 cl::init(false));
51 cl::opt<bool> DontReducePassList("disable-pass-list-reduction",
52 cl::desc("Skip pass list reduction steps"),
53 cl::init(false));
55 cl::opt<bool> NoNamedMDRM("disable-namedmd-remove",
56 cl::desc("Do not remove global named metadata"),
57 cl::init(false));
58 cl::opt<bool> NoStripDebugInfo("disable-strip-debuginfo",
59 cl::desc("Do not strip debug info metadata"),
60 cl::init(false));
61 cl::opt<bool> NoStripDebugTypeInfo("disable-strip-debug-types",
62 cl::desc("Do not strip debug type info metadata"),
63 cl::init(false));
64 cl::opt<bool> VerboseErrors("verbose-errors",
65 cl::desc("Print the output of crashing program"),
66 cl::init(false));
69 namespace llvm {
70 class ReducePassList : public ListReducer<std::string> {
71 BugDriver &BD;
73 public:
74 ReducePassList(BugDriver &bd) : BD(bd) {}
76 // Return true iff running the "removed" passes succeeds, and running the
77 // "Kept" passes fail when run on the output of the "removed" passes. If we
78 // return true, we update the current module of bugpoint.
79 Expected<TestResult> doTest(std::vector<std::string> &Removed,
80 std::vector<std::string> &Kept) override;
84 Expected<ReducePassList::TestResult>
85 ReducePassList::doTest(std::vector<std::string> &Prefix,
86 std::vector<std::string> &Suffix) {
87 std::string PrefixOutput;
88 std::unique_ptr<Module> OrigProgram;
89 if (!Prefix.empty()) {
90 outs() << "Checking to see if these passes crash: "
91 << getPassesString(Prefix) << ": ";
92 if (BD.runPasses(BD.getProgram(), Prefix, PrefixOutput))
93 return KeepPrefix;
95 OrigProgram = std::move(BD.Program);
97 BD.Program = parseInputFile(PrefixOutput, BD.getContext());
98 if (BD.Program == nullptr) {
99 errs() << BD.getToolName() << ": Error reading bitcode file '"
100 << PrefixOutput << "'!\n";
101 exit(1);
103 sys::fs::remove(PrefixOutput);
106 outs() << "Checking to see if these passes crash: " << getPassesString(Suffix)
107 << ": ";
109 if (BD.runPasses(BD.getProgram(), Suffix))
110 return KeepSuffix; // The suffix crashes alone...
112 // Nothing failed, restore state...
113 if (OrigProgram)
114 BD.Program = std::move(OrigProgram);
115 return NoFailure;
118 using BugTester = bool (*)(const BugDriver &, Module *);
120 namespace {
121 /// ReduceCrashingGlobalInitializers - This works by removing global variable
122 /// initializers and seeing if the program still crashes. If it does, then we
123 /// keep that program and try again.
124 class ReduceCrashingGlobalInitializers : public ListReducer<GlobalVariable *> {
125 BugDriver &BD;
126 BugTester TestFn;
128 public:
129 ReduceCrashingGlobalInitializers(BugDriver &bd, BugTester testFn)
130 : BD(bd), TestFn(testFn) {}
132 Expected<TestResult> doTest(std::vector<GlobalVariable *> &Prefix,
133 std::vector<GlobalVariable *> &Kept) override {
134 if (!Kept.empty() && TestGlobalVariables(Kept))
135 return KeepSuffix;
136 if (!Prefix.empty() && TestGlobalVariables(Prefix))
137 return KeepPrefix;
138 return NoFailure;
141 bool TestGlobalVariables(std::vector<GlobalVariable *> &GVs);
145 bool ReduceCrashingGlobalInitializers::TestGlobalVariables(
146 std::vector<GlobalVariable *> &GVs) {
147 // Clone the program to try hacking it apart...
148 ValueToValueMapTy VMap;
149 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
151 // Convert list to set for fast lookup...
152 std::set<GlobalVariable *> GVSet;
154 for (unsigned i = 0, e = GVs.size(); i != e; ++i) {
155 GlobalVariable *CMGV = cast<GlobalVariable>(VMap[GVs[i]]);
156 assert(CMGV && "Global Variable not in module?!");
157 GVSet.insert(CMGV);
160 outs() << "Checking for crash with only these global variables: ";
161 PrintGlobalVariableList(GVs);
162 outs() << ": ";
164 // Loop over and delete any global variables which we aren't supposed to be
165 // playing with...
166 for (GlobalVariable &I : M->globals())
167 if (I.hasInitializer() && !GVSet.count(&I)) {
168 DeleteGlobalInitializer(&I);
169 I.setLinkage(GlobalValue::ExternalLinkage);
170 I.setComdat(nullptr);
173 // Try running the hacked up program...
174 if (TestFn(BD, M.get())) {
175 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
177 // Make sure to use global variable pointers that point into the now-current
178 // module.
179 GVs.assign(GVSet.begin(), GVSet.end());
180 return true;
183 return false;
186 namespace {
187 /// ReduceCrashingFunctions reducer - This works by removing functions and
188 /// seeing if the program still crashes. If it does, then keep the newer,
189 /// smaller program.
191 class ReduceCrashingFunctions : public ListReducer<Function *> {
192 BugDriver &BD;
193 BugTester TestFn;
195 public:
196 ReduceCrashingFunctions(BugDriver &bd, BugTester testFn)
197 : BD(bd), TestFn(testFn) {}
199 Expected<TestResult> doTest(std::vector<Function *> &Prefix,
200 std::vector<Function *> &Kept) override {
201 if (!Kept.empty() && TestFuncs(Kept))
202 return KeepSuffix;
203 if (!Prefix.empty() && TestFuncs(Prefix))
204 return KeepPrefix;
205 return NoFailure;
208 bool TestFuncs(std::vector<Function *> &Prefix);
212 static void RemoveFunctionReferences(Module *M, const char *Name) {
213 auto *UsedVar = M->getGlobalVariable(Name, true);
214 if (!UsedVar || !UsedVar->hasInitializer())
215 return;
216 if (isa<ConstantAggregateZero>(UsedVar->getInitializer())) {
217 assert(UsedVar->use_empty());
218 UsedVar->eraseFromParent();
219 return;
221 auto *OldUsedVal = cast<ConstantArray>(UsedVar->getInitializer());
222 std::vector<Constant *> Used;
223 for (Value *V : OldUsedVal->operand_values()) {
224 Constant *Op = cast<Constant>(V->stripPointerCasts());
225 if (!Op->isNullValue()) {
226 Used.push_back(cast<Constant>(V));
229 auto *NewValElemTy = OldUsedVal->getType()->getElementType();
230 auto *NewValTy = ArrayType::get(NewValElemTy, Used.size());
231 auto *NewUsedVal = ConstantArray::get(NewValTy, Used);
232 UsedVar->mutateType(NewUsedVal->getType()->getPointerTo());
233 UsedVar->setInitializer(NewUsedVal);
236 bool ReduceCrashingFunctions::TestFuncs(std::vector<Function *> &Funcs) {
237 // If main isn't present, claim there is no problem.
238 if (KeepMain && !is_contained(Funcs, BD.getProgram().getFunction("main")))
239 return false;
241 // Clone the program to try hacking it apart...
242 ValueToValueMapTy VMap;
243 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
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);
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;
282 PointerType *Ty = cast<PointerType>(Alias.getType());
283 Constant *Replacement = ConstantPointerNull::get(Ty);
284 Alias.replaceAllUsesWith(Replacement);
285 ToRemove.push_back(&Alias);
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);
297 for (auto *F : ToRemove) {
298 F->eraseFromParent();
301 // Finally, remove any null members from any global intrinsic.
302 RemoveFunctionReferences(M.get(), "llvm.used");
303 RemoveFunctionReferences(M.get(), "llvm.compiler.used");
305 // Try running the hacked up program...
306 if (TestFn(BD, M.get())) {
307 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
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;
314 return false;
317 namespace {
318 /// Simplify the CFG without completely destroying it.
319 /// This is not well defined, but basically comes down to "try to eliminate
320 /// unreachable blocks and constant fold terminators without deciding that
321 /// certain undefined behavior cuts off the program at the legs".
322 void simpleSimplifyCfg(Function &F, SmallVectorImpl<BasicBlock *> &BBs) {
323 if (F.empty())
324 return;
326 for (auto *BB : BBs) {
327 ConstantFoldTerminator(BB);
328 MergeBlockIntoPredecessor(BB);
331 // Remove unreachable blocks
332 // removeUnreachableBlocks can't be used here, it will turn various
333 // undefined behavior into unreachables, but bugpoint was the thing that
334 // generated the undefined behavior, and we don't want it to kill the entire
335 // program.
336 SmallPtrSet<BasicBlock *, 16> Visited;
337 for (auto *BB : depth_first(&F.getEntryBlock()))
338 Visited.insert(BB);
340 SmallVector<BasicBlock *, 16> Unreachable;
341 for (auto &BB : F)
342 if (!Visited.count(&BB))
343 Unreachable.push_back(&BB);
345 // The dead BB's may be in a dead cycle or otherwise have references to each
346 // other. Because of this, we have to drop all references first, then delete
347 // them all at once.
348 for (auto *BB : Unreachable) {
349 for (BasicBlock *Successor : successors(&*BB))
350 if (Visited.count(Successor))
351 Successor->removePredecessor(&*BB);
352 BB->dropAllReferences();
354 for (auto *BB : Unreachable)
355 BB->eraseFromParent();
357 /// ReduceCrashingBlocks reducer - This works by setting the terminators of
358 /// all terminators except the specified basic blocks to a 'ret' instruction,
359 /// then running the simplify-cfg pass. This has the effect of chopping up
360 /// the CFG really fast which can reduce large functions quickly.
362 class ReduceCrashingBlocks : public ListReducer<const BasicBlock *> {
363 BugDriver &BD;
364 BugTester TestFn;
366 public:
367 ReduceCrashingBlocks(BugDriver &BD, BugTester testFn)
368 : BD(BD), TestFn(testFn) {}
370 Expected<TestResult> doTest(std::vector<const BasicBlock *> &Prefix,
371 std::vector<const BasicBlock *> &Kept) override {
372 if (!Kept.empty() && TestBlocks(Kept))
373 return KeepSuffix;
374 if (!Prefix.empty() && TestBlocks(Prefix))
375 return KeepPrefix;
376 return NoFailure;
379 bool TestBlocks(std::vector<const BasicBlock *> &Prefix);
383 bool ReduceCrashingBlocks::TestBlocks(std::vector<const BasicBlock *> &BBs) {
384 // Clone the program to try hacking it apart...
385 ValueToValueMapTy VMap;
386 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
388 // Convert list to set for fast lookup...
389 SmallPtrSet<BasicBlock *, 8> Blocks;
390 for (unsigned i = 0, e = BBs.size(); i != e; ++i)
391 Blocks.insert(cast<BasicBlock>(VMap[BBs[i]]));
393 outs() << "Checking for crash with only these blocks:";
394 unsigned NumPrint = Blocks.size();
395 if (NumPrint > 10)
396 NumPrint = 10;
397 for (unsigned i = 0, e = NumPrint; i != e; ++i)
398 outs() << " " << BBs[i]->getName();
399 if (NumPrint < Blocks.size())
400 outs() << "... <" << Blocks.size() << " total>";
401 outs() << ": ";
403 // Loop over and delete any hack up any blocks that are not listed...
404 for (Function &F : M->functions()) {
405 for (BasicBlock &BB : F) {
406 if (!Blocks.count(&BB) && BB.getTerminator()->getNumSuccessors()) {
407 // Loop over all of the successors of this block, deleting any PHI nodes
408 // that might include it.
409 for (BasicBlock *Succ : successors(&BB))
410 Succ->removePredecessor(&BB);
412 TerminatorInst *BBTerm = BB.getTerminator();
413 if (BBTerm->isEHPad() || BBTerm->getType()->isTokenTy())
414 continue;
415 if (!BBTerm->getType()->isVoidTy())
416 BBTerm->replaceAllUsesWith(Constant::getNullValue(BBTerm->getType()));
418 // Replace the old terminator instruction.
419 BB.getInstList().pop_back();
420 new UnreachableInst(BB.getContext(), &BB);
425 // The CFG Simplifier pass may delete one of the basic blocks we are
426 // interested in. If it does we need to take the block out of the list. Make
427 // a "persistent mapping" by turning basic blocks into <function, name> pairs.
428 // This won't work well if blocks are unnamed, but that is just the risk we
429 // have to take. FIXME: Can we just name the blocks?
430 std::vector<std::pair<std::string, std::string>> BlockInfo;
432 for (BasicBlock *BB : Blocks)
433 BlockInfo.emplace_back(BB->getParent()->getName(), BB->getName());
435 SmallVector<BasicBlock *, 16> ToProcess;
436 for (auto &F : *M) {
437 for (auto &BB : F)
438 if (!Blocks.count(&BB))
439 ToProcess.push_back(&BB);
440 simpleSimplifyCfg(F, ToProcess);
441 ToProcess.clear();
443 // Verify we didn't break anything
444 std::vector<std::string> Passes;
445 Passes.push_back("verify");
446 std::unique_ptr<Module> New = BD.runPassesOn(M.get(), Passes);
447 if (!New) {
448 errs() << "verify failed!\n";
449 exit(1);
451 M = std::move(New);
453 // Try running on the hacked up program...
454 if (TestFn(BD, M.get())) {
455 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
457 // Make sure to use basic block pointers that point into the now-current
458 // module, and that they don't include any deleted blocks.
459 BBs.clear();
460 const ValueSymbolTable &GST = BD.getProgram().getValueSymbolTable();
461 for (const auto &BI : BlockInfo) {
462 Function *F = cast<Function>(GST.lookup(BI.first));
463 Value *V = F->getValueSymbolTable()->lookup(BI.second);
464 if (V && V->getType() == Type::getLabelTy(V->getContext()))
465 BBs.push_back(cast<BasicBlock>(V));
467 return true;
469 // It didn't crash, try something else.
470 return false;
473 namespace {
474 /// ReduceCrashingConditionals reducer - This works by changing
475 /// conditional branches to unconditional ones, then simplifying the CFG
476 /// This has the effect of chopping up the CFG really fast which can reduce
477 /// large functions quickly.
479 class ReduceCrashingConditionals : public ListReducer<const BasicBlock *> {
480 BugDriver &BD;
481 BugTester TestFn;
482 bool Direction;
484 public:
485 ReduceCrashingConditionals(BugDriver &bd, BugTester testFn, bool Direction)
486 : BD(bd), TestFn(testFn), Direction(Direction) {}
488 Expected<TestResult> doTest(std::vector<const BasicBlock *> &Prefix,
489 std::vector<const BasicBlock *> &Kept) override {
490 if (!Kept.empty() && TestBlocks(Kept))
491 return KeepSuffix;
492 if (!Prefix.empty() && TestBlocks(Prefix))
493 return KeepPrefix;
494 return NoFailure;
497 bool TestBlocks(std::vector<const BasicBlock *> &Prefix);
501 bool ReduceCrashingConditionals::TestBlocks(
502 std::vector<const BasicBlock *> &BBs) {
503 // Clone the program to try hacking it apart...
504 ValueToValueMapTy VMap;
505 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
507 // Convert list to set for fast lookup...
508 SmallPtrSet<const BasicBlock *, 8> Blocks;
509 for (const auto *BB : BBs)
510 Blocks.insert(cast<BasicBlock>(VMap[BB]));
512 outs() << "Checking for crash with changing conditionals to always jump to "
513 << (Direction ? "true" : "false") << ":";
514 unsigned NumPrint = Blocks.size();
515 if (NumPrint > 10)
516 NumPrint = 10;
517 for (unsigned i = 0, e = NumPrint; i != e; ++i)
518 outs() << " " << BBs[i]->getName();
519 if (NumPrint < Blocks.size())
520 outs() << "... <" << Blocks.size() << " total>";
521 outs() << ": ";
523 // Loop over and delete any hack up any blocks that are not listed...
524 for (auto &F : *M)
525 for (auto &BB : F)
526 if (!Blocks.count(&BB)) {
527 auto *BR = dyn_cast<BranchInst>(BB.getTerminator());
528 if (!BR || !BR->isConditional())
529 continue;
530 if (Direction)
531 BR->setCondition(ConstantInt::getTrue(BR->getContext()));
532 else
533 BR->setCondition(ConstantInt::getFalse(BR->getContext()));
536 // The following may destroy some blocks, so we save them first
537 std::vector<std::pair<std::string, std::string>> BlockInfo;
539 for (const BasicBlock *BB : Blocks)
540 BlockInfo.emplace_back(BB->getParent()->getName(), BB->getName());
542 SmallVector<BasicBlock *, 16> ToProcess;
543 for (auto &F : *M) {
544 for (auto &BB : F)
545 if (!Blocks.count(&BB))
546 ToProcess.push_back(&BB);
547 simpleSimplifyCfg(F, ToProcess);
548 ToProcess.clear();
550 // Verify we didn't break anything
551 std::vector<std::string> Passes;
552 Passes.push_back("verify");
553 std::unique_ptr<Module> New = BD.runPassesOn(M.get(), Passes);
554 if (!New) {
555 errs() << "verify failed!\n";
556 exit(1);
558 M = std::move(New);
560 // Try running on the hacked up program...
561 if (TestFn(BD, M.get())) {
562 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
564 // Make sure to use basic block pointers that point into the now-current
565 // module, and that they don't include any deleted blocks.
566 BBs.clear();
567 const ValueSymbolTable &GST = BD.getProgram().getValueSymbolTable();
568 for (auto &BI : BlockInfo) {
569 auto *F = cast<Function>(GST.lookup(BI.first));
570 Value *V = F->getValueSymbolTable()->lookup(BI.second);
571 if (V && V->getType() == Type::getLabelTy(V->getContext()))
572 BBs.push_back(cast<BasicBlock>(V));
574 return true;
576 // It didn't crash, try something else.
577 return false;
580 namespace {
581 /// SimplifyCFG reducer - This works by calling SimplifyCFG on each basic block
582 /// in the program.
584 class ReduceSimplifyCFG : public ListReducer<const BasicBlock *> {
585 BugDriver &BD;
586 BugTester TestFn;
587 TargetTransformInfo TTI;
589 public:
590 ReduceSimplifyCFG(BugDriver &bd, BugTester testFn)
591 : BD(bd), TestFn(testFn), TTI(bd.getProgram().getDataLayout()) {}
593 Expected<TestResult> doTest(std::vector<const BasicBlock *> &Prefix,
594 std::vector<const BasicBlock *> &Kept) override {
595 if (!Kept.empty() && TestBlocks(Kept))
596 return KeepSuffix;
597 if (!Prefix.empty() && TestBlocks(Prefix))
598 return KeepPrefix;
599 return NoFailure;
602 bool TestBlocks(std::vector<const BasicBlock *> &Prefix);
606 bool ReduceSimplifyCFG::TestBlocks(std::vector<const BasicBlock *> &BBs) {
607 // Clone the program to try hacking it apart...
608 ValueToValueMapTy VMap;
609 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
611 // Convert list to set for fast lookup...
612 SmallPtrSet<const BasicBlock *, 8> Blocks;
613 for (const auto *BB : BBs)
614 Blocks.insert(cast<BasicBlock>(VMap[BB]));
616 outs() << "Checking for crash with CFG simplifying:";
617 unsigned NumPrint = Blocks.size();
618 if (NumPrint > 10)
619 NumPrint = 10;
620 for (unsigned i = 0, e = NumPrint; i != e; ++i)
621 outs() << " " << BBs[i]->getName();
622 if (NumPrint < Blocks.size())
623 outs() << "... <" << Blocks.size() << " total>";
624 outs() << ": ";
626 // The following may destroy some blocks, so we save them first
627 std::vector<std::pair<std::string, std::string>> BlockInfo;
629 for (const BasicBlock *BB : Blocks)
630 BlockInfo.emplace_back(BB->getParent()->getName(), BB->getName());
632 // Loop over and delete any hack up any blocks that are not listed...
633 for (auto &F : *M)
634 // Loop over all of the basic blocks and remove them if they are unneeded.
635 for (Function::iterator BBIt = F.begin(); BBIt != F.end();) {
636 if (!Blocks.count(&*BBIt)) {
637 ++BBIt;
638 continue;
640 simplifyCFG(&*BBIt++, TTI);
642 // Verify we didn't break anything
643 std::vector<std::string> Passes;
644 Passes.push_back("verify");
645 std::unique_ptr<Module> New = BD.runPassesOn(M.get(), Passes);
646 if (!New) {
647 errs() << "verify failed!\n";
648 exit(1);
650 M = std::move(New);
652 // Try running on the hacked up program...
653 if (TestFn(BD, M.get())) {
654 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
656 // Make sure to use basic block pointers that point into the now-current
657 // module, and that they don't include any deleted blocks.
658 BBs.clear();
659 const ValueSymbolTable &GST = BD.getProgram().getValueSymbolTable();
660 for (auto &BI : BlockInfo) {
661 auto *F = cast<Function>(GST.lookup(BI.first));
662 Value *V = F->getValueSymbolTable()->lookup(BI.second);
663 if (V && V->getType() == Type::getLabelTy(V->getContext()))
664 BBs.push_back(cast<BasicBlock>(V));
666 return true;
668 // It didn't crash, try something else.
669 return false;
672 namespace {
673 /// ReduceCrashingInstructions reducer - This works by removing the specified
674 /// non-terminator instructions and replacing them with undef.
676 class ReduceCrashingInstructions : public ListReducer<const Instruction *> {
677 BugDriver &BD;
678 BugTester TestFn;
680 public:
681 ReduceCrashingInstructions(BugDriver &bd, BugTester testFn)
682 : BD(bd), TestFn(testFn) {}
684 Expected<TestResult> doTest(std::vector<const Instruction *> &Prefix,
685 std::vector<const Instruction *> &Kept) override {
686 if (!Kept.empty() && TestInsts(Kept))
687 return KeepSuffix;
688 if (!Prefix.empty() && TestInsts(Prefix))
689 return KeepPrefix;
690 return NoFailure;
693 bool TestInsts(std::vector<const Instruction *> &Prefix);
697 bool ReduceCrashingInstructions::TestInsts(
698 std::vector<const Instruction *> &Insts) {
699 // Clone the program to try hacking it apart...
700 ValueToValueMapTy VMap;
701 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
703 // Convert list to set for fast lookup...
704 SmallPtrSet<Instruction *, 32> Instructions;
705 for (unsigned i = 0, e = Insts.size(); i != e; ++i) {
706 assert(!Insts[i]->isTerminator());
707 Instructions.insert(cast<Instruction>(VMap[Insts[i]]));
710 outs() << "Checking for crash with only " << Instructions.size();
711 if (Instructions.size() == 1)
712 outs() << " instruction: ";
713 else
714 outs() << " instructions: ";
716 for (Module::iterator MI = M->begin(), ME = M->end(); MI != ME; ++MI)
717 for (Function::iterator FI = MI->begin(), FE = MI->end(); FI != FE; ++FI)
718 for (BasicBlock::iterator I = FI->begin(), E = FI->end(); I != E;) {
719 Instruction *Inst = &*I++;
720 if (!Instructions.count(Inst) && !Inst->isTerminator() &&
721 !Inst->isEHPad() && !Inst->getType()->isTokenTy() &&
722 !Inst->isSwiftError()) {
723 if (!Inst->getType()->isVoidTy())
724 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
725 Inst->eraseFromParent();
729 // Verify that this is still valid.
730 legacy::PassManager Passes;
731 Passes.add(createVerifierPass(/*FatalErrors=*/false));
732 Passes.run(*M);
734 // Try running on the hacked up program...
735 if (TestFn(BD, M.get())) {
736 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
738 // Make sure to use instruction pointers that point into the now-current
739 // module, and that they don't include any deleted blocks.
740 Insts.clear();
741 for (Instruction *Inst : Instructions)
742 Insts.push_back(Inst);
743 return true;
745 // It didn't crash, try something else.
746 return false;
749 namespace {
750 // Reduce the list of Named Metadata nodes. We keep this as a list of
751 // names to avoid having to convert back and forth every time.
752 class ReduceCrashingNamedMD : public ListReducer<std::string> {
753 BugDriver &BD;
754 BugTester TestFn;
756 public:
757 ReduceCrashingNamedMD(BugDriver &bd, BugTester testFn)
758 : BD(bd), TestFn(testFn) {}
760 Expected<TestResult> doTest(std::vector<std::string> &Prefix,
761 std::vector<std::string> &Kept) override {
762 if (!Kept.empty() && TestNamedMDs(Kept))
763 return KeepSuffix;
764 if (!Prefix.empty() && TestNamedMDs(Prefix))
765 return KeepPrefix;
766 return NoFailure;
769 bool TestNamedMDs(std::vector<std::string> &NamedMDs);
773 bool ReduceCrashingNamedMD::TestNamedMDs(std::vector<std::string> &NamedMDs) {
775 ValueToValueMapTy VMap;
776 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
778 outs() << "Checking for crash with only these named metadata nodes:";
779 unsigned NumPrint = std::min<size_t>(NamedMDs.size(), 10);
780 for (unsigned i = 0, e = NumPrint; i != e; ++i)
781 outs() << " " << NamedMDs[i];
782 if (NumPrint < NamedMDs.size())
783 outs() << "... <" << NamedMDs.size() << " total>";
784 outs() << ": ";
786 // Make a StringMap for faster lookup
787 StringSet<> Names;
788 for (const std::string &Name : NamedMDs)
789 Names.insert(Name);
791 // First collect all the metadata to delete in a vector, then
792 // delete them all at once to avoid invalidating the iterator
793 std::vector<NamedMDNode *> ToDelete;
794 ToDelete.reserve(M->named_metadata_size() - Names.size());
795 for (auto &NamedMD : M->named_metadata())
796 // Always keep a nonempty llvm.dbg.cu because the Verifier would complain.
797 if (!Names.count(NamedMD.getName()) &&
798 (!(NamedMD.getName() == "llvm.dbg.cu" && NamedMD.getNumOperands() > 0)))
799 ToDelete.push_back(&NamedMD);
801 for (auto *NamedMD : ToDelete)
802 NamedMD->eraseFromParent();
804 // Verify that this is still valid.
805 legacy::PassManager Passes;
806 Passes.add(createVerifierPass(/*FatalErrors=*/false));
807 Passes.run(*M);
809 // Try running on the hacked up program...
810 if (TestFn(BD, M.get())) {
811 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
812 return true;
814 return false;
817 namespace {
818 // Reduce the list of operands to named metadata nodes
819 class ReduceCrashingNamedMDOps : public ListReducer<const MDNode *> {
820 BugDriver &BD;
821 BugTester TestFn;
823 public:
824 ReduceCrashingNamedMDOps(BugDriver &bd, BugTester testFn)
825 : BD(bd), TestFn(testFn) {}
827 Expected<TestResult> doTest(std::vector<const MDNode *> &Prefix,
828 std::vector<const MDNode *> &Kept) override {
829 if (!Kept.empty() && TestNamedMDOps(Kept))
830 return KeepSuffix;
831 if (!Prefix.empty() && TestNamedMDOps(Prefix))
832 return KeepPrefix;
833 return NoFailure;
836 bool TestNamedMDOps(std::vector<const MDNode *> &NamedMDOps);
840 bool ReduceCrashingNamedMDOps::TestNamedMDOps(
841 std::vector<const MDNode *> &NamedMDOps) {
842 // Convert list to set for fast lookup...
843 SmallPtrSet<const MDNode *, 32> OldMDNodeOps;
844 for (unsigned i = 0, e = NamedMDOps.size(); i != e; ++i) {
845 OldMDNodeOps.insert(NamedMDOps[i]);
848 outs() << "Checking for crash with only " << OldMDNodeOps.size();
849 if (OldMDNodeOps.size() == 1)
850 outs() << " named metadata operand: ";
851 else
852 outs() << " named metadata operands: ";
854 ValueToValueMapTy VMap;
855 std::unique_ptr<Module> M = CloneModule(BD.getProgram(), VMap);
857 // This is a little wasteful. In the future it might be good if we could have
858 // these dropped during cloning.
859 for (auto &NamedMD : BD.getProgram().named_metadata()) {
860 // Drop the old one and create a new one
861 M->eraseNamedMetadata(M->getNamedMetadata(NamedMD.getName()));
862 NamedMDNode *NewNamedMDNode =
863 M->getOrInsertNamedMetadata(NamedMD.getName());
864 for (MDNode *op : NamedMD.operands())
865 if (OldMDNodeOps.count(op))
866 NewNamedMDNode->addOperand(cast<MDNode>(MapMetadata(op, VMap)));
869 // Verify that this is still valid.
870 legacy::PassManager Passes;
871 Passes.add(createVerifierPass(/*FatalErrors=*/false));
872 Passes.run(*M);
874 // Try running on the hacked up program...
875 if (TestFn(BD, M.get())) {
876 // Make sure to use instruction pointers that point into the now-current
877 // module, and that they don't include any deleted blocks.
878 NamedMDOps.clear();
879 for (const MDNode *Node : OldMDNodeOps)
880 NamedMDOps.push_back(cast<MDNode>(*VMap.getMappedMD(Node)));
882 BD.setNewProgram(std::move(M)); // It crashed, keep the trimmed version...
883 return true;
885 // It didn't crash, try something else.
886 return false;
889 /// Attempt to eliminate as many global initializers as possible.
890 static Error ReduceGlobalInitializers(BugDriver &BD, BugTester TestFn) {
891 Module &OrigM = BD.getProgram();
892 if (OrigM.global_empty())
893 return Error::success();
895 // Now try to reduce the number of global variable initializers in the
896 // module to something small.
897 std::unique_ptr<Module> M = CloneModule(OrigM);
898 bool DeletedInit = false;
900 for (GlobalVariable &GV : M->globals()) {
901 if (GV.hasInitializer()) {
902 DeleteGlobalInitializer(&GV);
903 GV.setLinkage(GlobalValue::ExternalLinkage);
904 GV.setComdat(nullptr);
905 DeletedInit = true;
909 if (!DeletedInit)
910 return Error::success();
912 // See if the program still causes a crash...
913 outs() << "\nChecking to see if we can delete global inits: ";
915 if (TestFn(BD, M.get())) { // Still crashes?
916 BD.setNewProgram(std::move(M));
917 outs() << "\n*** Able to remove all global initializers!\n";
918 return Error::success();
921 // No longer crashes.
922 outs() << " - Removing all global inits hides problem!\n";
924 std::vector<GlobalVariable *> GVs;
925 for (GlobalVariable &GV : OrigM.globals())
926 if (GV.hasInitializer())
927 GVs.push_back(&GV);
929 if (GVs.size() > 1 && !BugpointIsInterrupted) {
930 outs() << "\n*** Attempting to reduce the number of global initializers "
931 << "in the testcase\n";
933 unsigned OldSize = GVs.size();
934 Expected<bool> Result =
935 ReduceCrashingGlobalInitializers(BD, TestFn).reduceList(GVs);
936 if (Error E = Result.takeError())
937 return E;
939 if (GVs.size() < OldSize)
940 BD.EmitProgressBitcode(BD.getProgram(), "reduced-global-variables");
942 return Error::success();
945 static Error ReduceInsts(BugDriver &BD, BugTester TestFn) {
946 // Attempt to delete instructions using bisection. This should help out nasty
947 // cases with large basic blocks where the problem is at one end.
948 if (!BugpointIsInterrupted) {
949 std::vector<const Instruction *> Insts;
950 for (const Function &F : BD.getProgram())
951 for (const BasicBlock &BB : F)
952 for (const Instruction &I : BB)
953 if (!I.isTerminator())
954 Insts.push_back(&I);
956 Expected<bool> Result =
957 ReduceCrashingInstructions(BD, TestFn).reduceList(Insts);
958 if (Error E = Result.takeError())
959 return E;
962 unsigned Simplification = 2;
963 do {
964 if (BugpointIsInterrupted)
965 // TODO: Should we distinguish this with an "interrupted error"?
966 return Error::success();
967 --Simplification;
968 outs() << "\n*** Attempting to reduce testcase by deleting instruc"
969 << "tions: Simplification Level #" << Simplification << '\n';
971 // Now that we have deleted the functions that are unnecessary for the
972 // program, try to remove instructions that are not necessary to cause the
973 // crash. To do this, we loop through all of the instructions in the
974 // remaining functions, deleting them (replacing any values produced with
975 // nulls), and then running ADCE and SimplifyCFG. If the transformed input
976 // still triggers failure, keep deleting until we cannot trigger failure
977 // anymore.
979 unsigned InstructionsToSkipBeforeDeleting = 0;
980 TryAgain:
982 // Loop over all of the (non-terminator) instructions remaining in the
983 // function, attempting to delete them.
984 unsigned CurInstructionNum = 0;
985 for (Module::const_iterator FI = BD.getProgram().begin(),
986 E = BD.getProgram().end();
987 FI != E; ++FI)
988 if (!FI->isDeclaration())
989 for (Function::const_iterator BI = FI->begin(), E = FI->end(); BI != E;
990 ++BI)
991 for (BasicBlock::const_iterator I = BI->begin(), E = --BI->end();
992 I != E; ++I, ++CurInstructionNum) {
993 if (InstructionsToSkipBeforeDeleting) {
994 --InstructionsToSkipBeforeDeleting;
995 } else {
996 if (BugpointIsInterrupted)
997 // TODO: Should this be some kind of interrupted error?
998 return Error::success();
1000 if (I->isEHPad() || I->getType()->isTokenTy() ||
1001 I->isSwiftError())
1002 continue;
1004 outs() << "Checking instruction: " << *I;
1005 std::unique_ptr<Module> M =
1006 BD.deleteInstructionFromProgram(&*I, Simplification);
1008 // Find out if the pass still crashes on this pass...
1009 if (TestFn(BD, M.get())) {
1010 // Yup, it does, we delete the old module, and continue trying
1011 // to reduce the testcase...
1012 BD.setNewProgram(std::move(M));
1013 InstructionsToSkipBeforeDeleting = CurInstructionNum;
1014 goto TryAgain; // I wish I had a multi-level break here!
1019 if (InstructionsToSkipBeforeDeleting) {
1020 InstructionsToSkipBeforeDeleting = 0;
1021 goto TryAgain;
1024 } while (Simplification);
1025 BD.EmitProgressBitcode(BD.getProgram(), "reduced-instructions");
1026 return Error::success();
1029 /// DebugACrash - Given a predicate that determines whether a component crashes
1030 /// on a program, try to destructively reduce the program while still keeping
1031 /// the predicate true.
1032 static Error DebugACrash(BugDriver &BD, BugTester TestFn) {
1033 // See if we can get away with nuking some of the global variable initializers
1034 // in the program...
1035 if (!NoGlobalRM)
1036 if (Error E = ReduceGlobalInitializers(BD, TestFn))
1037 return E;
1039 // Now try to reduce the number of functions in the module to something small.
1040 std::vector<Function *> Functions;
1041 for (Function &F : BD.getProgram())
1042 if (!F.isDeclaration())
1043 Functions.push_back(&F);
1045 if (Functions.size() > 1 && !BugpointIsInterrupted) {
1046 outs() << "\n*** Attempting to reduce the number of functions "
1047 "in the testcase\n";
1049 unsigned OldSize = Functions.size();
1050 Expected<bool> Result =
1051 ReduceCrashingFunctions(BD, TestFn).reduceList(Functions);
1052 if (Error E = Result.takeError())
1053 return E;
1055 if (Functions.size() < OldSize)
1056 BD.EmitProgressBitcode(BD.getProgram(), "reduced-function");
1059 // Attempt to change conditional branches into unconditional branches to
1060 // eliminate blocks.
1061 if (!DisableSimplifyCFG && !BugpointIsInterrupted) {
1062 std::vector<const BasicBlock *> Blocks;
1063 for (Function &F : BD.getProgram())
1064 for (BasicBlock &BB : F)
1065 Blocks.push_back(&BB);
1066 unsigned OldSize = Blocks.size();
1067 Expected<bool> Result =
1068 ReduceCrashingConditionals(BD, TestFn, true).reduceList(Blocks);
1069 if (Error E = Result.takeError())
1070 return E;
1071 Result = ReduceCrashingConditionals(BD, TestFn, false).reduceList(Blocks);
1072 if (Error E = Result.takeError())
1073 return E;
1074 if (Blocks.size() < OldSize)
1075 BD.EmitProgressBitcode(BD.getProgram(), "reduced-conditionals");
1078 // Attempt to delete entire basic blocks at a time to speed up
1079 // convergence... this actually works by setting the terminator of the blocks
1080 // to a return instruction then running simplifycfg, which can potentially
1081 // shrinks the code dramatically quickly
1083 if (!DisableSimplifyCFG && !BugpointIsInterrupted) {
1084 std::vector<const BasicBlock *> Blocks;
1085 for (Function &F : BD.getProgram())
1086 for (BasicBlock &BB : F)
1087 Blocks.push_back(&BB);
1088 unsigned OldSize = Blocks.size();
1089 Expected<bool> Result = ReduceCrashingBlocks(BD, TestFn).reduceList(Blocks);
1090 if (Error E = Result.takeError())
1091 return E;
1092 if (Blocks.size() < OldSize)
1093 BD.EmitProgressBitcode(BD.getProgram(), "reduced-blocks");
1096 if (!DisableSimplifyCFG && !BugpointIsInterrupted) {
1097 std::vector<const BasicBlock *> Blocks;
1098 for (Function &F : BD.getProgram())
1099 for (BasicBlock &BB : F)
1100 Blocks.push_back(&BB);
1101 unsigned OldSize = Blocks.size();
1102 Expected<bool> Result = ReduceSimplifyCFG(BD, TestFn).reduceList(Blocks);
1103 if (Error E = Result.takeError())
1104 return E;
1105 if (Blocks.size() < OldSize)
1106 BD.EmitProgressBitcode(BD.getProgram(), "reduced-simplifycfg");
1109 // Attempt to delete instructions using bisection. This should help out nasty
1110 // cases with large basic blocks where the problem is at one end.
1111 if (!BugpointIsInterrupted)
1112 if (Error E = ReduceInsts(BD, TestFn))
1113 return E;
1115 // Attempt to strip debug info metadata.
1116 auto stripMetadata = [&](std::function<bool(Module &)> strip) {
1117 std::unique_ptr<Module> M = CloneModule(BD.getProgram());
1118 strip(*M);
1119 if (TestFn(BD, M.get()))
1120 BD.setNewProgram(std::move(M));
1122 if (!NoStripDebugInfo && !BugpointIsInterrupted) {
1123 outs() << "\n*** Attempting to strip the debug info: ";
1124 stripMetadata(StripDebugInfo);
1126 if (!NoStripDebugTypeInfo && !BugpointIsInterrupted) {
1127 outs() << "\n*** Attempting to strip the debug type info: ";
1128 stripMetadata(stripNonLineTableDebugInfo);
1131 if (!NoNamedMDRM) {
1132 if (!BugpointIsInterrupted) {
1133 // Try to reduce the amount of global metadata (particularly debug info),
1134 // by dropping global named metadata that anchors them
1135 outs() << "\n*** Attempting to remove named metadata: ";
1136 std::vector<std::string> NamedMDNames;
1137 for (auto &NamedMD : BD.getProgram().named_metadata())
1138 NamedMDNames.push_back(NamedMD.getName().str());
1139 Expected<bool> Result =
1140 ReduceCrashingNamedMD(BD, TestFn).reduceList(NamedMDNames);
1141 if (Error E = Result.takeError())
1142 return E;
1145 if (!BugpointIsInterrupted) {
1146 // Now that we quickly dropped all the named metadata that doesn't
1147 // contribute to the crash, bisect the operands of the remaining ones
1148 std::vector<const MDNode *> NamedMDOps;
1149 for (auto &NamedMD : BD.getProgram().named_metadata())
1150 for (auto op : NamedMD.operands())
1151 NamedMDOps.push_back(op);
1152 Expected<bool> Result =
1153 ReduceCrashingNamedMDOps(BD, TestFn).reduceList(NamedMDOps);
1154 if (Error E = Result.takeError())
1155 return E;
1157 BD.EmitProgressBitcode(BD.getProgram(), "reduced-named-md");
1160 // Try to clean up the testcase by running funcresolve and globaldce...
1161 if (!BugpointIsInterrupted) {
1162 outs() << "\n*** Attempting to perform final cleanups: ";
1163 std::unique_ptr<Module> M = CloneModule(BD.getProgram());
1164 M = BD.performFinalCleanups(std::move(M), true);
1166 // Find out if the pass still crashes on the cleaned up program...
1167 if (M && TestFn(BD, M.get()))
1168 BD.setNewProgram(
1169 std::move(M)); // Yup, it does, keep the reduced version...
1172 BD.EmitProgressBitcode(BD.getProgram(), "reduced-simplified");
1174 return Error::success();
1177 static bool TestForOptimizerCrash(const BugDriver &BD, Module *M) {
1178 return BD.runPasses(*M, BD.getPassesToRun());
1181 /// debugOptimizerCrash - This method is called when some pass crashes on input.
1182 /// It attempts to prune down the testcase to something reasonable, and figure
1183 /// out exactly which pass is crashing.
1185 Error BugDriver::debugOptimizerCrash(const std::string &ID) {
1186 outs() << "\n*** Debugging optimizer crash!\n";
1188 // Reduce the list of passes which causes the optimizer to crash...
1189 if (!BugpointIsInterrupted && !DontReducePassList) {
1190 Expected<bool> Result = ReducePassList(*this).reduceList(PassesToRun);
1191 if (Error E = Result.takeError())
1192 return E;
1195 outs() << "\n*** Found crashing pass"
1196 << (PassesToRun.size() == 1 ? ": " : "es: ")
1197 << getPassesString(PassesToRun) << '\n';
1199 EmitProgressBitcode(*Program, ID);
1201 return DebugACrash(*this, TestForOptimizerCrash);
1204 static bool TestForCodeGenCrash(const BugDriver &BD, Module *M) {
1205 if (Error E = BD.compileProgram(*M)) {
1206 if (VerboseErrors)
1207 errs() << toString(std::move(E)) << "\n";
1208 else {
1209 consumeError(std::move(E));
1210 errs() << "<crash>\n";
1212 return true; // Tool is still crashing.
1214 errs() << '\n';
1215 return false;
1218 /// debugCodeGeneratorCrash - This method is called when the code generator
1219 /// crashes on an input. It attempts to reduce the input as much as possible
1220 /// while still causing the code generator to crash.
1221 Error BugDriver::debugCodeGeneratorCrash() {
1222 errs() << "*** Debugging code generator crash!\n";
1224 return DebugACrash(*this, TestForCodeGenCrash);