1 //===- Miscompilation.cpp - Debug program miscompilations -----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements optimizer and code generation miscompilation debugging
13 //===----------------------------------------------------------------------===//
15 #include "BugDriver.h"
16 #include "ListReducer.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/Module.h"
21 #include "llvm/Pass.h"
22 #include "llvm/Analysis/Verifier.h"
23 #include "llvm/Support/Mangler.h"
24 #include "llvm/Transforms/Utils/Cloning.h"
25 #include "llvm/Transforms/Utils/Linker.h"
26 #include "Support/CommandLine.h"
27 #include "Support/FileUtilities.h"
31 extern cl::list<std::string> InputArgv;
35 class ReduceMiscompilingPasses : public ListReducer<const PassInfo*> {
38 ReduceMiscompilingPasses(BugDriver &bd) : BD(bd) {}
40 virtual TestResult doTest(std::vector<const PassInfo*> &Prefix,
41 std::vector<const PassInfo*> &Suffix);
45 ReduceMiscompilingPasses::TestResult
46 ReduceMiscompilingPasses::doTest(std::vector<const PassInfo*> &Prefix,
47 std::vector<const PassInfo*> &Suffix) {
48 // First, run the program with just the Suffix passes. If it is still broken
49 // with JUST the kept passes, discard the prefix passes.
50 std::cout << "Checking to see if '" << getPassesString(Suffix)
51 << "' compile correctly: ";
53 std::string BytecodeResult;
54 if (BD.runPasses(Suffix, BytecodeResult, false/*delete*/, true/*quiet*/)) {
55 std::cerr << " Error running this sequence of passes"
56 << " on the input program!\n";
57 BD.setPassesToRun(Suffix);
58 BD.EmitProgressBytecode("pass-error", false);
59 exit(BD.debugOptimizerCrash());
62 // Check to see if the finished program matches the reference output...
63 if (BD.diffProgram(BytecodeResult, "", true /*delete bytecode*/)) {
64 std::cout << "nope.\n";
65 return KeepSuffix; // Miscompilation detected!
67 std::cout << "yup.\n"; // No miscompilation!
69 if (Prefix.empty()) return NoFailure;
71 // Next, see if the program is broken if we run the "prefix" passes first,
72 // then separately run the "kept" passes.
73 std::cout << "Checking to see if '" << getPassesString(Prefix)
74 << "' compile correctly: ";
76 // If it is not broken with the kept passes, it's possible that the prefix
77 // passes must be run before the kept passes to break it. If the program
78 // WORKS after the prefix passes, but then fails if running the prefix AND
79 // kept passes, we can update our bytecode file to include the result of the
80 // prefix passes, then discard the prefix passes.
82 if (BD.runPasses(Prefix, BytecodeResult, false/*delete*/, true/*quiet*/)) {
83 std::cerr << " Error running this sequence of passes"
84 << " on the input program!\n";
85 BD.setPassesToRun(Prefix);
86 BD.EmitProgressBytecode("pass-error", false);
87 exit(BD.debugOptimizerCrash());
90 // If the prefix maintains the predicate by itself, only keep the prefix!
91 if (BD.diffProgram(BytecodeResult)) {
92 std::cout << "nope.\n";
93 removeFile(BytecodeResult);
96 std::cout << "yup.\n"; // No miscompilation!
98 // Ok, so now we know that the prefix passes work, try running the suffix
99 // passes on the result of the prefix passes.
101 Module *PrefixOutput = ParseInputFile(BytecodeResult);
102 if (PrefixOutput == 0) {
103 std::cerr << BD.getToolName() << ": Error reading bytecode file '"
104 << BytecodeResult << "'!\n";
107 removeFile(BytecodeResult); // No longer need the file on disk
109 std::cout << "Checking to see if '" << getPassesString(Suffix)
110 << "' passes compile correctly after the '"
111 << getPassesString(Prefix) << "' passes: ";
113 Module *OriginalInput = BD.swapProgramIn(PrefixOutput);
114 if (BD.runPasses(Suffix, BytecodeResult, false/*delete*/, true/*quiet*/)) {
115 std::cerr << " Error running this sequence of passes"
116 << " on the input program!\n";
117 BD.setPassesToRun(Suffix);
118 BD.EmitProgressBytecode("pass-error", false);
119 exit(BD.debugOptimizerCrash());
123 if (BD.diffProgram(BytecodeResult, "", true/*delete bytecode*/)) {
124 std::cout << "nope.\n";
125 delete OriginalInput; // We pruned down the original input...
129 // Otherwise, we must not be running the bad pass anymore.
130 std::cout << "yup.\n"; // No miscompilation!
131 delete BD.swapProgramIn(OriginalInput); // Restore orig program & free test
136 class ReduceMiscompilingFunctions : public ListReducer<Function*> {
138 bool (*TestFn)(BugDriver &, Module *, Module *);
140 ReduceMiscompilingFunctions(BugDriver &bd,
141 bool (*F)(BugDriver &, Module *, Module *))
142 : BD(bd), TestFn(F) {}
144 virtual TestResult doTest(std::vector<Function*> &Prefix,
145 std::vector<Function*> &Suffix) {
146 if (!Suffix.empty() && TestFuncs(Suffix))
148 if (!Prefix.empty() && TestFuncs(Prefix))
153 bool TestFuncs(const std::vector<Function*> &Prefix);
157 /// TestMergedProgram - Given two modules, link them together and run the
158 /// program, checking to see if the program matches the diff. If the diff
159 /// matches, return false, otherwise return true. If the DeleteInputs argument
160 /// is set to true then this function deletes both input modules before it
162 static bool TestMergedProgram(BugDriver &BD, Module *M1, Module *M2,
164 // Link the two portions of the program back to together.
165 std::string ErrorMsg;
166 if (!DeleteInputs) M1 = CloneModule(M1);
167 if (LinkModules(M1, M2, &ErrorMsg)) {
168 std::cerr << BD.getToolName() << ": Error linking modules together:"
172 if (DeleteInputs) delete M2; // We are done with this module...
174 Module *OldProgram = BD.swapProgramIn(M1);
176 // Execute the program. If it does not match the expected output, we must
178 bool Broken = BD.diffProgram();
180 // Delete the linked module & restore the original
181 BD.swapProgramIn(OldProgram);
186 bool ReduceMiscompilingFunctions::TestFuncs(const std::vector<Function*>&Funcs){
187 // Test to see if the function is misoptimized if we ONLY run it on the
188 // functions listed in Funcs.
189 std::cout << "Checking to see if the program is misoptimized when "
190 << (Funcs.size()==1 ? "this function is" : "these functions are")
191 << " run through the pass"
192 << (BD.getPassesToRun().size() == 1 ? "" : "es") << ":";
193 PrintFunctionList(Funcs);
196 // Split the module into the two halves of the program we want.
197 Module *ToNotOptimize = CloneModule(BD.getProgram());
198 Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize, Funcs);
200 // Run the predicate, not that the predicate will delete both input modules.
201 return TestFn(BD, ToOptimize, ToNotOptimize);
204 static void DisambiguateGlobalSymbols(Module *M) {
205 // Try not to cause collisions by minimizing chances of renaming an
206 // already-external symbol, so take in external globals and functions as-is.
207 // The code should work correctly without disambiguation (assuming the same
208 // mangler is used by the two code generators), but having symbols with the
209 // same name causes warnings to be emitted by the code generator.
211 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I)
212 I->setName(Mang.getValueName(I));
213 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
214 I->setName(Mang.getValueName(I));
217 /// ExtractLoops - Given a reduced list of functions that still exposed the bug,
218 /// check to see if we can extract the loops in the region without obscuring the
219 /// bug. If so, it reduces the amount of code identified.
220 static bool ExtractLoops(BugDriver &BD,
221 bool (*TestFn)(BugDriver &, Module *, Module *),
222 std::vector<Function*> &MiscompiledFunctions) {
223 bool MadeChange = false;
225 Module *ToNotOptimize = CloneModule(BD.getProgram());
226 Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize,
227 MiscompiledFunctions);
228 Module *ToOptimizeLoopExtracted = BD.ExtractLoop(ToOptimize);
229 if (!ToOptimizeLoopExtracted) {
230 // If the loop extractor crashed or if there were no extractible loops,
231 // then this chapter of our odyssey is over with.
232 delete ToNotOptimize;
237 std::cerr << "Extracted a loop from the breaking portion of the program.\n";
240 // Bugpoint is intentionally not very trusting of LLVM transformations. In
241 // particular, we're not going to assume that the loop extractor works, so
242 // we're going to test the newly loop extracted program to make sure nothing
243 // has broken. If something broke, then we'll inform the user and stop
245 AbstractInterpreter *AI = BD.switchToCBE();
246 if (TestMergedProgram(BD, ToOptimizeLoopExtracted, ToNotOptimize, false)) {
247 BD.switchToInterpreter(AI);
249 // Merged program doesn't work anymore!
250 std::cerr << " *** ERROR: Loop extraction broke the program. :("
251 << " Please report a bug!\n";
252 std::cerr << " Continuing on with un-loop-extracted version.\n";
253 delete ToNotOptimize;
254 delete ToOptimizeLoopExtracted;
257 BD.switchToInterpreter(AI);
259 std::cout << " Testing after loop extraction:\n";
260 // Clone modules, the tester function will free them.
261 Module *TOLEBackup = CloneModule(ToOptimizeLoopExtracted);
262 Module *TNOBackup = CloneModule(ToNotOptimize);
263 if (!TestFn(BD, ToOptimizeLoopExtracted, ToNotOptimize)) {
264 std::cout << "*** Loop extraction masked the problem. Undoing.\n";
265 // If the program is not still broken, then loop extraction did something
266 // that masked the error. Stop loop extraction now.
271 ToOptimizeLoopExtracted = TOLEBackup;
272 ToNotOptimize = TNOBackup;
274 std::cout << "*** Loop extraction successful!\n";
276 // Okay, great! Now we know that we extracted a loop and that loop
277 // extraction both didn't break the program, and didn't mask the problem.
278 // Replace the current program with the loop extracted version, and try to
279 // extract another loop.
280 std::string ErrorMsg;
281 if (LinkModules(ToNotOptimize, ToOptimizeLoopExtracted, &ErrorMsg)) {
282 std::cerr << BD.getToolName() << ": Error linking modules together:"
287 // All of the Function*'s in the MiscompiledFunctions list are in the old
288 // module. Update this list to include all of the functions in the
289 // optimized and loop extracted module.
290 MiscompiledFunctions.clear();
291 for (Module::iterator I = ToOptimizeLoopExtracted->begin(),
292 E = ToOptimizeLoopExtracted->end(); I != E; ++I) {
293 if (!I->isExternal()) {
294 Function *NewF = ToNotOptimize->getFunction(I->getName(),
295 I->getFunctionType());
296 assert(NewF && "Function not found??");
297 MiscompiledFunctions.push_back(NewF);
300 delete ToOptimizeLoopExtracted;
302 BD.setNewProgram(ToNotOptimize);
307 /// DebugAMiscompilation - This is a generic driver to narrow down
308 /// miscompilations, either in an optimization or a code generator.
309 static std::vector<Function*>
310 DebugAMiscompilation(BugDriver &BD,
311 bool (*TestFn)(BugDriver &, Module *, Module *)) {
312 // Okay, now that we have reduced the list of passes which are causing the
313 // failure, see if we can pin down which functions are being
314 // miscompiled... first build a list of all of the non-external functions in
316 std::vector<Function*> MiscompiledFunctions;
317 Module *Prog = BD.getProgram();
318 for (Module::iterator I = Prog->begin(), E = Prog->end(); I != E; ++I)
319 if (!I->isExternal())
320 MiscompiledFunctions.push_back(I);
322 // Do the reduction...
323 ReduceMiscompilingFunctions(BD, TestFn).reduceList(MiscompiledFunctions);
325 std::cout << "\n*** The following function"
326 << (MiscompiledFunctions.size() == 1 ? " is" : "s are")
327 << " being miscompiled: ";
328 PrintFunctionList(MiscompiledFunctions);
331 // See if we can rip any loops out of the miscompiled functions and still
332 // trigger the problem.
333 if (ExtractLoops(BD, TestFn, MiscompiledFunctions)) {
334 // Okay, we extracted some loops and the problem still appears. See if we
335 // can eliminate some of the created functions from being candidates.
337 // Loop extraction can introduce functions with the same name (foo_code).
338 // Make sure to disambiguate the symbols so that when the program is split
339 // apart that we can link it back together again.
340 DisambiguateGlobalSymbols(BD.getProgram());
342 // Do the reduction...
343 ReduceMiscompilingFunctions(BD, TestFn).reduceList(MiscompiledFunctions);
345 std::cout << "\n*** The following function"
346 << (MiscompiledFunctions.size() == 1 ? " is" : "s are")
347 << " being miscompiled: ";
348 PrintFunctionList(MiscompiledFunctions);
352 return MiscompiledFunctions;
355 /// TestOptimizer - This is the predicate function used to check to see if the
356 /// "Test" portion of the program is misoptimized. If so, return true. In any
357 /// case, both module arguments are deleted.
358 static bool TestOptimizer(BugDriver &BD, Module *Test, Module *Safe) {
359 // Run the optimization passes on ToOptimize, producing a transformed version
360 // of the functions being tested.
361 std::cout << " Optimizing functions being tested: ";
362 Module *Optimized = BD.runPassesOn(Test, BD.getPassesToRun(),
363 /*AutoDebugCrashes*/true);
364 std::cout << "done.\n";
367 std::cout << " Checking to see if the merged program executes correctly: ";
368 bool Broken = TestMergedProgram(BD, Optimized, Safe, true);
369 std::cout << (Broken ? " nope.\n" : " yup.\n");
374 /// debugMiscompilation - This method is used when the passes selected are not
375 /// crashing, but the generated output is semantically different from the
378 bool BugDriver::debugMiscompilation() {
379 // Make sure something was miscompiled...
380 if (!ReduceMiscompilingPasses(*this).reduceList(PassesToRun)) {
381 std::cerr << "*** Optimized program matches reference output! No problem "
382 << "detected...\nbugpoint can't help you with your problem!\n";
386 std::cout << "\n*** Found miscompiling pass"
387 << (getPassesToRun().size() == 1 ? "" : "es") << ": "
388 << getPassesString(getPassesToRun()) << "\n";
389 EmitProgressBytecode("passinput");
391 std::vector<Function*> MiscompiledFunctions =
392 DebugAMiscompilation(*this, TestOptimizer);
394 // Output a bunch of bytecode files for the user...
395 std::cout << "Outputting reduced bytecode files which expose the problem:\n";
396 Module *ToNotOptimize = CloneModule(getProgram());
397 Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize,
398 MiscompiledFunctions);
400 std::cout << " Non-optimized portion: ";
401 ToNotOptimize = swapProgramIn(ToNotOptimize);
402 EmitProgressBytecode("tonotoptimize", true);
403 setNewProgram(ToNotOptimize); // Delete hacked module.
405 std::cout << " Portion that is input to optimizer: ";
406 ToOptimize = swapProgramIn(ToOptimize);
407 EmitProgressBytecode("tooptimize");
408 setNewProgram(ToOptimize); // Delete hacked module.
413 /// CleanupAndPrepareModules - Get the specified modules ready for code
414 /// generator testing.
415 static void CleanupAndPrepareModules(BugDriver &BD, Module *&Test,
417 // Clean up the modules, removing extra cruft that we don't need anymore...
418 Test = BD.performFinalCleanups(Test);
420 // If we are executing the JIT, we have several nasty issues to take care of.
421 if (!BD.isExecutingJIT()) return;
423 // First, if the main function is in the Safe module, we must add a stub to
424 // the Test module to call into it. Thus, we create a new function `main'
425 // which just calls the old one.
426 if (Function *oldMain = Safe->getNamedFunction("main"))
427 if (!oldMain->isExternal()) {
429 oldMain->setName("llvm_bugpoint_old_main");
430 // Create a NEW `main' function with same type in the test module.
431 Function *newMain = new Function(oldMain->getFunctionType(),
432 GlobalValue::ExternalLinkage,
434 // Create an `oldmain' prototype in the test module, which will
435 // corresponds to the real main function in the same module.
436 Function *oldMainProto = new Function(oldMain->getFunctionType(),
437 GlobalValue::ExternalLinkage,
438 oldMain->getName(), Test);
439 // Set up and remember the argument list for the main function.
440 std::vector<Value*> args;
441 for (Function::aiterator I = newMain->abegin(), E = newMain->aend(),
442 OI = oldMain->abegin(); I != E; ++I, ++OI) {
443 I->setName(OI->getName()); // Copy argument names from oldMain
447 // Call the old main function and return its result
448 BasicBlock *BB = new BasicBlock("entry", newMain);
449 CallInst *call = new CallInst(oldMainProto, args);
450 BB->getInstList().push_back(call);
452 // If the type of old function wasn't void, return value of call
453 new ReturnInst(oldMain->getReturnType() != Type::VoidTy ? call : 0, BB);
456 // The second nasty issue we must deal with in the JIT is that the Safe
457 // module cannot directly reference any functions defined in the test
458 // module. Instead, we use a JIT API call to dynamically resolve the
461 // Add the resolver to the Safe module.
462 // Prototype: void *getPointerToNamedFunction(const char* Name)
463 Function *resolverFunc =
464 Safe->getOrInsertFunction("getPointerToNamedFunction",
465 PointerType::get(Type::SByteTy),
466 PointerType::get(Type::SByteTy), 0);
468 // Use the function we just added to get addresses of functions we need.
469 for (Module::iterator F = Safe->begin(), E = Safe->end(); F != E; ++F){
470 if (F->isExternal() && !F->use_empty() && &*F != resolverFunc &&
471 F->getIntrinsicID() == 0 /* ignore intrinsics */) {
472 Function *TestFn =Test->getFunction(F->getName(), F->getFunctionType());
474 // Don't forward functions which are external in the test module too.
475 if (TestFn && !TestFn->isExternal()) {
476 // 1. Add a string constant with its name to the global file
477 Constant *InitArray = ConstantArray::get(F->getName());
478 GlobalVariable *funcName =
479 new GlobalVariable(InitArray->getType(), true /*isConstant*/,
480 GlobalValue::InternalLinkage, InitArray,
481 F->getName() + "_name", Safe);
483 // 2. Use `GetElementPtr *funcName, 0, 0' to convert the string to an
484 // sbyte* so it matches the signature of the resolver function.
486 // GetElementPtr *funcName, ulong 0, ulong 0
487 std::vector<Constant*> GEPargs(2,Constant::getNullValue(Type::IntTy));
489 ConstantExpr::getGetElementPtr(ConstantPointerRef::get(funcName),
491 std::vector<Value*> ResolverArgs;
492 ResolverArgs.push_back(GEP);
494 // 3. Replace all uses of `func' with calls to resolver by:
495 // (a) Iterating through the list of uses of this function
496 // (b) Insert a cast instruction in front of each use
497 // (c) Replace use of old call with new call
499 // Insert code at the beginning of the function
500 while (!F->use_empty())
501 if (Instruction *Inst = dyn_cast<Instruction>(F->use_back())) {
502 // call resolver(GetElementPtr...)
503 CallInst *resolve = new CallInst(resolverFunc, ResolverArgs,
505 // cast the result from the resolver to correctly-typed function
506 CastInst *castResolver =
507 new CastInst(resolve, PointerType::get(F->getFunctionType()),
508 "resolverCast", Inst);
509 // actually use the resolved function
510 Inst->replaceUsesOfWith(F, castResolver);
512 // FIXME: need to take care of cases where a function is used by
513 // something other than an instruction; e.g., global variable
514 // initializers and constant expressions.
515 std::cerr << "UNSUPPORTED: Non-instruction is using an external "
516 << "function, " << F->getName() << "().\n";
523 if (verifyModule(*Test) || verifyModule(*Safe)) {
524 std::cerr << "Bugpoint has a bug, which corrupted a module!!\n";
531 /// TestCodeGenerator - This is the predicate function used to check to see if
532 /// the "Test" portion of the program is miscompiled by the code generator under
533 /// test. If so, return true. In any case, both module arguments are deleted.
534 static bool TestCodeGenerator(BugDriver &BD, Module *Test, Module *Safe) {
535 CleanupAndPrepareModules(BD, Test, Safe);
537 std::string TestModuleBC = getUniqueFilename("bugpoint.test.bc");
538 if (BD.writeProgramToFile(TestModuleBC, Test)) {
539 std::cerr << "Error writing bytecode to `" << TestModuleBC << "'\nExiting.";
544 // Make the shared library
545 std::string SafeModuleBC = getUniqueFilename("bugpoint.safe.bc");
547 if (BD.writeProgramToFile(SafeModuleBC, Safe)) {
548 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
551 std::string SharedObject = BD.compileSharedObject(SafeModuleBC);
554 // Run the code generator on the `Test' code, loading the shared library.
555 // The function returns whether or not the new output differs from reference.
556 int Result = BD.diffProgram(TestModuleBC, SharedObject, false);
559 std::cerr << ": still failing!\n";
561 std::cerr << ": didn't fail.\n";
562 removeFile(TestModuleBC);
563 removeFile(SafeModuleBC);
564 removeFile(SharedObject);
570 bool BugDriver::debugCodeGenerator() {
571 if ((void*)cbe == (void*)Interpreter) {
572 std::string Result = executeProgramWithCBE("bugpoint.cbe.out");
573 std::cout << "\n*** The C backend cannot match the reference diff, but it "
574 << "is used as the 'known good'\n code generator, so I can't"
575 << " debug it. Perhaps you have a front-end problem?\n As a"
576 << " sanity check, I left the result of executing the program "
577 << "with the C backend\n in this file for you: '"
582 DisambiguateGlobalSymbols(Program);
584 std::vector<Function*> Funcs = DebugAMiscompilation(*this, TestCodeGenerator);
586 // Split the module into the two halves of the program we want.
587 Module *ToNotCodeGen = CloneModule(getProgram());
588 Module *ToCodeGen = SplitFunctionsOutOfModule(ToNotCodeGen, Funcs);
590 // Condition the modules
591 CleanupAndPrepareModules(*this, ToCodeGen, ToNotCodeGen);
593 std::string TestModuleBC = getUniqueFilename("bugpoint.test.bc");
594 if (writeProgramToFile(TestModuleBC, ToCodeGen)) {
595 std::cerr << "Error writing bytecode to `" << TestModuleBC << "'\nExiting.";
600 // Make the shared library
601 std::string SafeModuleBC = getUniqueFilename("bugpoint.safe.bc");
602 if (writeProgramToFile(SafeModuleBC, ToNotCodeGen)) {
603 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
606 std::string SharedObject = compileSharedObject(SafeModuleBC);
609 std::cout << "You can reproduce the problem with the command line: \n";
610 if (isExecutingJIT()) {
611 std::cout << " lli -load " << SharedObject << " " << TestModuleBC;
613 std::cout << " llc " << TestModuleBC << " -o " << TestModuleBC << ".s\n";
614 std::cout << " gcc " << SharedObject << " " << TestModuleBC
615 << ".s -o " << TestModuleBC << ".exe -Wl,-R.\n";
616 std::cout << " " << TestModuleBC << ".exe";
618 for (unsigned i=0, e = InputArgv.size(); i != e; ++i)
619 std::cout << " " << InputArgv[i];
621 std::cout << "The shared object was created with:\n llc -march=c "
622 << SafeModuleBC << " -o temporary.c\n"
623 << " gcc -xc temporary.c -O2 -o " << SharedObject
624 #if defined(sparc) || defined(__sparc__) || defined(__sparcv9)
625 << " -G" // Compile a shared library, `-G' for Sparc
627 << " -shared" // `-shared' for Linux/X86, maybe others
629 << " -fno-strict-aliasing\n";