X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=tools%2Fbugpoint%2FMiscompilation.cpp;h=b3e616111e824a14938c710bf014b03004316d0e;hp=b88f87c9b1e2de3c9f609ce5adbb015adc63485d;hb=66c5fd6c537269eaef0f630fa14360dcaff6a295;hpb=3d9cafa003a114bf9974bc80d5b69b0ed1d29290 diff --git a/tools/bugpoint/Miscompilation.cpp b/tools/bugpoint/Miscompilation.cpp index b88f87c9b1e..b3e616111e8 100644 --- a/tools/bugpoint/Miscompilation.cpp +++ b/tools/bugpoint/Miscompilation.cpp @@ -1,26 +1,51 @@ //===- Miscompilation.cpp - Debug program miscompilations -----------------===// // -// This file implements program miscompilation debugging support. +// The LLVM Compiler Infrastructure +// +// This file was developed by the LLVM research group and is distributed under +// the University of Illinois Open Source License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file implements optimizer and code generation miscompilation debugging +// support. // //===----------------------------------------------------------------------===// #include "BugDriver.h" #include "ListReducer.h" +#include "llvm/Constants.h" +#include "llvm/DerivedTypes.h" +#include "llvm/Instructions.h" +#include "llvm/Linker.h" #include "llvm/Module.h" #include "llvm/Pass.h" +#include "llvm/Analysis/Verifier.h" +#include "llvm/Support/Mangler.h" #include "llvm/Transforms/Utils/Cloning.h" -#include "llvm/Transforms/Utils/Linker.h" -#include "Support/FileUtilities.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/FileUtilities.h" +#include "llvm/Config/config.h" // for HAVE_LINK_R +using namespace llvm; -class ReduceMiscompilingPasses : public ListReducer { - BugDriver &BD; -public: - ReduceMiscompilingPasses(BugDriver &bd) : BD(bd) {} +namespace llvm { + extern cl::list InputArgv; +} - virtual TestResult doTest(std::vector &Prefix, - std::vector &Suffix); -}; +namespace { + class ReduceMiscompilingPasses : public ListReducer { + BugDriver &BD; + public: + ReduceMiscompilingPasses(BugDriver &bd) : BD(bd) {} + virtual TestResult doTest(std::vector &Prefix, + std::vector &Suffix); + }; +} + +/// TestResult - After passes have been split into a test group and a control +/// group, see if they still break the program. +/// ReduceMiscompilingPasses::TestResult ReduceMiscompilingPasses::doTest(std::vector &Prefix, std::vector &Suffix) { @@ -31,17 +56,24 @@ ReduceMiscompilingPasses::doTest(std::vector &Prefix, std::string BytecodeResult; if (BD.runPasses(Suffix, BytecodeResult, false/*delete*/, true/*quiet*/)) { - std::cerr << BD.getToolName() << ": Error running this sequence of passes" + std::cerr << " Error running this sequence of passes" << " on the input program!\n"; - exit(1); + BD.setPassesToRun(Suffix); + BD.EmitProgressBytecode("pass-error", false); + exit(BD.debugOptimizerCrash()); } // Check to see if the finished program matches the reference output... if (BD.diffProgram(BytecodeResult, "", true /*delete bytecode*/)) { - std::cout << "nope.\n"; - return KeepSuffix; // Miscompilation detected! + std::cout << " nope.\n"; + if (Suffix.empty()) { + std::cerr << BD.getToolName() << ": I'm confused: the test fails when " + << "no passes are run, nondeterministic program?\n"; + exit(1); + } + return KeepSuffix; // Miscompilation detected! } - std::cout << "yup.\n"; // No miscompilation! + std::cout << " yup.\n"; // No miscompilation! if (Prefix.empty()) return NoFailure; @@ -57,201 +89,485 @@ ReduceMiscompilingPasses::doTest(std::vector &Prefix, // prefix passes, then discard the prefix passes. // if (BD.runPasses(Prefix, BytecodeResult, false/*delete*/, true/*quiet*/)) { - std::cerr << BD.getToolName() << ": Error running this sequence of passes" + std::cerr << " Error running this sequence of passes" << " on the input program!\n"; - exit(1); + BD.setPassesToRun(Prefix); + BD.EmitProgressBytecode("pass-error", false); + exit(BD.debugOptimizerCrash()); } // If the prefix maintains the predicate by itself, only keep the prefix! if (BD.diffProgram(BytecodeResult)) { - std::cout << "nope.\n"; - removeFile(BytecodeResult); + std::cout << " nope.\n"; + sys::Path(BytecodeResult).eraseFromDisk(); return KeepPrefix; } - std::cout << "yup.\n"; // No miscompilation! + std::cout << " yup.\n"; // No miscompilation! // Ok, so now we know that the prefix passes work, try running the suffix // passes on the result of the prefix passes. // - Module *PrefixOutput = BD.ParseInputFile(BytecodeResult); + Module *PrefixOutput = ParseInputFile(BytecodeResult); if (PrefixOutput == 0) { std::cerr << BD.getToolName() << ": Error reading bytecode file '" << BytecodeResult << "'!\n"; exit(1); } - removeFile(BytecodeResult); // No longer need the file on disk - + sys::Path(BytecodeResult).eraseFromDisk(); // No longer need the file on disk + + // Don't check if there are no passes in the suffix. + if (Suffix.empty()) + return NoFailure; + std::cout << "Checking to see if '" << getPassesString(Suffix) << "' passes compile correctly after the '" << getPassesString(Prefix) << "' passes: "; - Module *OriginalInput = BD.Program; - BD.Program = PrefixOutput; + Module *OriginalInput = BD.swapProgramIn(PrefixOutput); if (BD.runPasses(Suffix, BytecodeResult, false/*delete*/, true/*quiet*/)) { - std::cerr << BD.getToolName() << ": Error running this sequence of passes" + std::cerr << " Error running this sequence of passes" << " on the input program!\n"; - exit(1); + BD.setPassesToRun(Suffix); + BD.EmitProgressBytecode("pass-error", false); + exit(BD.debugOptimizerCrash()); } // Run the result... if (BD.diffProgram(BytecodeResult, "", true/*delete bytecode*/)) { - std::cout << "nope.\n"; + std::cout << " nope.\n"; delete OriginalInput; // We pruned down the original input... return KeepSuffix; } // Otherwise, we must not be running the bad pass anymore. - std::cout << "yup.\n"; // No miscompilation! - BD.Program = OriginalInput; // Restore original program - delete PrefixOutput; // Free experiment + std::cout << " yup.\n"; // No miscompilation! + delete BD.swapProgramIn(OriginalInput); // Restore orig program & free test return NoFailure; } -class ReduceMiscompilingFunctions : public ListReducer { - BugDriver &BD; -public: - ReduceMiscompilingFunctions(BugDriver &bd) : BD(bd) {} - - virtual TestResult doTest(std::vector &Prefix, - std::vector &Suffix) { - if (!Suffix.empty() && TestFuncs(Suffix, false)) - return KeepSuffix; - if (!Prefix.empty() && TestFuncs(Prefix, false)) - return KeepPrefix; - return NoFailure; +namespace { + class ReduceMiscompilingFunctions : public ListReducer { + BugDriver &BD; + bool (*TestFn)(BugDriver &, Module *, Module *); + public: + ReduceMiscompilingFunctions(BugDriver &bd, + bool (*F)(BugDriver &, Module *, Module *)) + : BD(bd), TestFn(F) {} + + virtual TestResult doTest(std::vector &Prefix, + std::vector &Suffix) { + if (!Suffix.empty() && TestFuncs(Suffix)) + return KeepSuffix; + if (!Prefix.empty() && TestFuncs(Prefix)) + return KeepPrefix; + return NoFailure; + } + + bool TestFuncs(const std::vector &Prefix); + }; +} + +/// TestMergedProgram - Given two modules, link them together and run the +/// program, checking to see if the program matches the diff. If the diff +/// matches, return false, otherwise return true. If the DeleteInputs argument +/// is set to true then this function deletes both input modules before it +/// returns. +/// +static bool TestMergedProgram(BugDriver &BD, Module *M1, Module *M2, + bool DeleteInputs) { + // Link the two portions of the program back to together. + std::string ErrorMsg; + if (!DeleteInputs) { + M1 = CloneModule(M1); + M2 = CloneModule(M2); } - - bool TestFuncs(const std::vector &Prefix, bool EmitBytecode); -}; + if (Linker::LinkModules(M1, M2, &ErrorMsg)) { + std::cerr << BD.getToolName() << ": Error linking modules together:" + << ErrorMsg << '\n'; + exit(1); + } + delete M2; // We are done with this module. + + Module *OldProgram = BD.swapProgramIn(M1); + + // Execute the program. If it does not match the expected output, we must + // return true. + bool Broken = BD.diffProgram(); + + // Delete the linked module & restore the original + BD.swapProgramIn(OldProgram); + delete M1; + return Broken; +} -bool ReduceMiscompilingFunctions::TestFuncs(const std::vector &Funcs, - bool EmitBytecode) { +/// TestFuncs - split functions in a Module into two groups: those that are +/// under consideration for miscompilation vs. those that are not, and test +/// accordingly. Each group of functions becomes a separate Module. +/// +bool ReduceMiscompilingFunctions::TestFuncs(const std::vector&Funcs){ // Test to see if the function is misoptimized if we ONLY run it on the // functions listed in Funcs. - if (!EmitBytecode) { - std::cout << "Checking to see if the program is misoptimized when these " - << "functions are run\nthrough the passes: "; - BD.PrintFunctionList(Funcs); - std::cout << "\n"; - } else { - std::cout <<"Outputting reduced bytecode files which expose the problem:\n"; - } + std::cout << "Checking to see if the program is misoptimized when " + << (Funcs.size()==1 ? "this function is" : "these functions are") + << " run through the pass" + << (BD.getPassesToRun().size() == 1 ? "" : "es") << ":"; + PrintFunctionList(Funcs); + std::cout << '\n'; + + // Split the module into the two halves of the program we want. + Module *ToNotOptimize = CloneModule(BD.getProgram()); + Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize, Funcs); + + // Run the predicate, not that the predicate will delete both input modules. + return TestFn(BD, ToOptimize, ToNotOptimize); +} - // First step: clone the module for the two halves of the program we want. - Module *ToOptimize = CloneModule(BD.Program); +/// DisambiguateGlobalSymbols - Mangle symbols to guarantee uniqueness by +/// modifying predominantly internal symbols rather than external ones. +/// +static void DisambiguateGlobalSymbols(Module *M) { + // Try not to cause collisions by minimizing chances of renaming an + // already-external symbol, so take in external globals and functions as-is. + // The code should work correctly without disambiguation (assuming the same + // mangler is used by the two code generators), but having symbols with the + // same name causes warnings to be emitted by the code generator. + Mangler Mang(*M); + for (Module::global_iterator I = M->global_begin(), E = M->global_end(); I != E; ++I) + I->setName(Mang.getValueName(I)); + for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) + I->setName(Mang.getValueName(I)); +} - // Second step: Make sure functions & globals are all external so that linkage - // between the two modules will work. - for (Module::iterator I = ToOptimize->begin(), E = ToOptimize->end();I!=E;++I) - I->setLinkage(GlobalValue::ExternalLinkage); - for (Module::giterator I = ToOptimize->gbegin(), E = ToOptimize->gend(); - I != E; ++I) - I->setLinkage(GlobalValue::ExternalLinkage); - - // Third step: make a clone of the externalized program for the non-optimized - // part. - Module *ToNotOptimize = CloneModule(ToOptimize); - - // Fourth step: Remove the test functions from the ToNotOptimize module, and - // all of the global variables. - for (unsigned i = 0, e = Funcs.size(); i != e; ++i) { - Function *TNOF = ToNotOptimize->getFunction(Funcs[i]->getName(), - Funcs[i]->getFunctionType()); - assert(TNOF && "Function doesn't exist in module!"); - DeleteFunctionBody(TNOF); // Function is now external in this module! - } - for (Module::giterator I = ToNotOptimize->gbegin(), E = ToNotOptimize->gend(); - I != E; ++I) - I->setInitializer(0); // Delete the initializer to make it external +/// ExtractLoops - Given a reduced list of functions that still exposed the bug, +/// check to see if we can extract the loops in the region without obscuring the +/// bug. If so, it reduces the amount of code identified. +/// +static bool ExtractLoops(BugDriver &BD, + bool (*TestFn)(BugDriver &, Module *, Module *), + std::vector &MiscompiledFunctions) { + bool MadeChange = false; + while (1) { + if (BugpointIsInterrupted) return MadeChange; + + Module *ToNotOptimize = CloneModule(BD.getProgram()); + Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize, + MiscompiledFunctions); + Module *ToOptimizeLoopExtracted = BD.ExtractLoop(ToOptimize); + if (!ToOptimizeLoopExtracted) { + // If the loop extractor crashed or if there were no extractible loops, + // then this chapter of our odyssey is over with. + delete ToNotOptimize; + delete ToOptimize; + return MadeChange; + } - if (EmitBytecode) { - std::cout << " Non-optimized portion: "; - std::swap(BD.Program, ToNotOptimize); - BD.EmitProgressBytecode("tonotoptimize", true); - std::swap(BD.Program, ToNotOptimize); + std::cerr << "Extracted a loop from the breaking portion of the program.\n"; + + // Bugpoint is intentionally not very trusting of LLVM transformations. In + // particular, we're not going to assume that the loop extractor works, so + // we're going to test the newly loop extracted program to make sure nothing + // has broken. If something broke, then we'll inform the user and stop + // extraction. + AbstractInterpreter *AI = BD.switchToCBE(); + if (TestMergedProgram(BD, ToOptimizeLoopExtracted, ToNotOptimize, false)) { + BD.switchToInterpreter(AI); + + // Merged program doesn't work anymore! + std::cerr << " *** ERROR: Loop extraction broke the program. :(" + << " Please report a bug!\n"; + std::cerr << " Continuing on with un-loop-extracted version.\n"; + + BD.writeProgramToFile("bugpoint-loop-extract-fail-tno.bc", ToNotOptimize); + BD.writeProgramToFile("bugpoint-loop-extract-fail-to.bc", ToOptimize); + BD.writeProgramToFile("bugpoint-loop-extract-fail-to-le.bc", + ToOptimizeLoopExtracted); + + std::cerr << "Please submit the bugpoint-loop-extract-fail-*.bc files.\n"; + delete ToOptimize; + delete ToNotOptimize; + delete ToOptimizeLoopExtracted; + return MadeChange; + } + delete ToOptimize; + BD.switchToInterpreter(AI); + + std::cout << " Testing after loop extraction:\n"; + // Clone modules, the tester function will free them. + Module *TOLEBackup = CloneModule(ToOptimizeLoopExtracted); + Module *TNOBackup = CloneModule(ToNotOptimize); + if (!TestFn(BD, ToOptimizeLoopExtracted, ToNotOptimize)) { + std::cout << "*** Loop extraction masked the problem. Undoing.\n"; + // If the program is not still broken, then loop extraction did something + // that masked the error. Stop loop extraction now. + delete TOLEBackup; + delete TNOBackup; + return MadeChange; + } + ToOptimizeLoopExtracted = TOLEBackup; + ToNotOptimize = TNOBackup; + + std::cout << "*** Loop extraction successful!\n"; + + std::vector > MisCompFunctions; + for (Module::iterator I = ToOptimizeLoopExtracted->begin(), + E = ToOptimizeLoopExtracted->end(); I != E; ++I) + if (!I->isExternal()) + MisCompFunctions.push_back(std::make_pair(I->getName(), + I->getFunctionType())); + + // Okay, great! Now we know that we extracted a loop and that loop + // extraction both didn't break the program, and didn't mask the problem. + // Replace the current program with the loop extracted version, and try to + // extract another loop. + std::string ErrorMsg; + if (Linker::LinkModules(ToNotOptimize, ToOptimizeLoopExtracted, &ErrorMsg)){ + std::cerr << BD.getToolName() << ": Error linking modules together:" + << ErrorMsg << '\n'; + exit(1); + } + delete ToOptimizeLoopExtracted; + + // All of the Function*'s in the MiscompiledFunctions list are in the old + // module. Update this list to include all of the functions in the + // optimized and loop extracted module. + MiscompiledFunctions.clear(); + for (unsigned i = 0, e = MisCompFunctions.size(); i != e; ++i) { + Function *NewF = ToNotOptimize->getFunction(MisCompFunctions[i].first, + MisCompFunctions[i].second); + assert(NewF && "Function not found??"); + MiscompiledFunctions.push_back(NewF); + } + + BD.setNewProgram(ToNotOptimize); + MadeChange = true; } +} - // Fifth step: Remove all functions from the ToOptimize module EXCEPT for the - // ones specified in Funcs. We know which ones these are because they are - // non-external in ToOptimize, but external in ToNotOptimize. - // - for (Module::iterator I = ToOptimize->begin(), E = ToOptimize->end();I!=E;++I) - if (!I->isExternal()) { - Function *TNOF = ToNotOptimize->getFunction(I->getName(), - I->getFunctionType()); - assert(TNOF && "Function doesn't exist in ToNotOptimize module??"); - if (!TNOF->isExternal()) - DeleteFunctionBody(I); +namespace { + class ReduceMiscompiledBlocks : public ListReducer { + BugDriver &BD; + bool (*TestFn)(BugDriver &, Module *, Module *); + std::vector FunctionsBeingTested; + public: + ReduceMiscompiledBlocks(BugDriver &bd, + bool (*F)(BugDriver &, Module *, Module *), + const std::vector &Fns) + : BD(bd), TestFn(F), FunctionsBeingTested(Fns) {} + + virtual TestResult doTest(std::vector &Prefix, + std::vector &Suffix) { + if (!Suffix.empty() && TestFuncs(Suffix)) + return KeepSuffix; + if (TestFuncs(Prefix)) + return KeepPrefix; + return NoFailure; } - if (EmitBytecode) { - std::cout << " Portion that is input to optimizer: "; - std::swap(BD.Program, ToOptimize); - BD.EmitProgressBytecode("tooptimize"); - std::swap(BD.Program, ToOptimize); + bool TestFuncs(const std::vector &Prefix); + }; +} + +/// TestFuncs - Extract all blocks for the miscompiled functions except for the +/// specified blocks. If the problem still exists, return true. +/// +bool ReduceMiscompiledBlocks::TestFuncs(const std::vector &BBs) { + // Test to see if the function is misoptimized if we ONLY run it on the + // functions listed in Funcs. + std::cout << "Checking to see if the program is misoptimized when all "; + if (!BBs.empty()) { + std::cout << "but these " << BBs.size() << " blocks are extracted: "; + for (unsigned i = 0, e = BBs.size() < 10 ? BBs.size() : 10; i != e; ++i) + std::cout << BBs[i]->getName() << " "; + if (BBs.size() > 10) std::cout << "..."; + } else { + std::cout << "blocks are extracted."; + } + std::cout << '\n'; + + // Split the module into the two halves of the program we want. + Module *ToNotOptimize = CloneModule(BD.getProgram()); + Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize, + FunctionsBeingTested); + + // Try the extraction. If it doesn't work, then the block extractor crashed + // or something, in which case bugpoint can't chase down this possibility. + if (Module *New = BD.ExtractMappedBlocksFromModule(BBs, ToOptimize)) { + delete ToOptimize; + // Run the predicate, not that the predicate will delete both input modules. + return TestFn(BD, New, ToNotOptimize); } + delete ToOptimize; + delete ToNotOptimize; + return false; +} - // Sixth step: Run the optimization passes on ToOptimize, producing a - // transformed version of the functions being tested. - Module *OldProgram = BD.Program; - BD.Program = ToOptimize; - if (!EmitBytecode) - std::cout << " Optimizing functions being tested: "; - std::string BytecodeResult; - if (BD.runPasses(BD.PassesToRun, BytecodeResult, false/*delete*/, - true/*quiet*/)) { - std::cerr << BD.getToolName() << ": Error running this sequence of passes" - << " on the input program!\n"; - exit(1); +/// ExtractBlocks - Given a reduced list of functions that still expose the bug, +/// extract as many basic blocks from the region as possible without obscuring +/// the bug. +/// +static bool ExtractBlocks(BugDriver &BD, + bool (*TestFn)(BugDriver &, Module *, Module *), + std::vector &MiscompiledFunctions) { + if (BugpointIsInterrupted) return false; + + std::vector Blocks; + for (unsigned i = 0, e = MiscompiledFunctions.size(); i != e; ++i) + for (Function::iterator I = MiscompiledFunctions[i]->begin(), + E = MiscompiledFunctions[i]->end(); I != E; ++I) + Blocks.push_back(I); + + // Use the list reducer to identify blocks that can be extracted without + // obscuring the bug. The Blocks list will end up containing blocks that must + // be retained from the original program. + unsigned OldSize = Blocks.size(); + + // Check to see if all blocks are extractible first. + if (ReduceMiscompiledBlocks(BD, TestFn, + MiscompiledFunctions).TestFuncs(std::vector())) { + Blocks.clear(); + } else { + ReduceMiscompiledBlocks(BD, TestFn,MiscompiledFunctions).reduceList(Blocks); + if (Blocks.size() == OldSize) + return false; } - if (!EmitBytecode) - std::cout << "done.\n"; + Module *ProgClone = CloneModule(BD.getProgram()); + Module *ToExtract = SplitFunctionsOutOfModule(ProgClone, + MiscompiledFunctions); + Module *Extracted = BD.ExtractMappedBlocksFromModule(Blocks, ToExtract); + if (Extracted == 0) { + // Weird, extraction should have worked. + std::cerr << "Nondeterministic problem extracting blocks??\n"; + delete ProgClone; + delete ToExtract; + return false; + } - delete BD.Program; // Delete the old "ToOptimize" module - BD.Program = BD.ParseInputFile(BytecodeResult); + // Otherwise, block extraction succeeded. Link the two program fragments back + // together. + delete ToExtract; - if (EmitBytecode) { - std::cout << " 'tooptimize' after being optimized: "; - BD.EmitProgressBytecode("optimized", true); - } + std::vector > MisCompFunctions; + for (Module::iterator I = Extracted->begin(), E = Extracted->end(); + I != E; ++I) + if (!I->isExternal()) + MisCompFunctions.push_back(std::make_pair(I->getName(), + I->getFunctionType())); - if (BD.Program == 0) { - std::cerr << BD.getToolName() << ": Error reading bytecode file '" - << BytecodeResult << "'!\n"; + std::string ErrorMsg; + if (Linker::LinkModules(ProgClone, Extracted, &ErrorMsg)) { + std::cerr << BD.getToolName() << ": Error linking modules together:" + << ErrorMsg << '\n'; exit(1); } - removeFile(BytecodeResult); // No longer need the file on disk + delete Extracted; - // Seventh step: Link the optimized part of the program back to the - // unoptimized part of the program. - // - if (LinkModules(BD.Program, ToNotOptimize, &BytecodeResult)) { - std::cerr << BD.getToolName() << ": Error linking modules together:" - << BytecodeResult << "\n"; - exit(1); + // Set the new program and delete the old one. + BD.setNewProgram(ProgClone); + + // Update the list of miscompiled functions. + MiscompiledFunctions.clear(); + + for (unsigned i = 0, e = MisCompFunctions.size(); i != e; ++i) { + Function *NewF = ProgClone->getFunction(MisCompFunctions[i].first, + MisCompFunctions[i].second); + assert(NewF && "Function not found??"); + MiscompiledFunctions.push_back(NewF); } - delete ToNotOptimize; // We are done with this module... - - if (EmitBytecode) { - std::cout << " Program as tested: "; - BD.EmitProgressBytecode("linked", true); - delete BD.Program; - BD.Program = OldProgram; - return false; // We don't need to actually execute the program here. + + return true; +} + + +/// DebugAMiscompilation - This is a generic driver to narrow down +/// miscompilations, either in an optimization or a code generator. +/// +static std::vector +DebugAMiscompilation(BugDriver &BD, + bool (*TestFn)(BugDriver &, Module *, Module *)) { + // Okay, now that we have reduced the list of passes which are causing the + // failure, see if we can pin down which functions are being + // miscompiled... first build a list of all of the non-external functions in + // the program. + std::vector MiscompiledFunctions; + Module *Prog = BD.getProgram(); + for (Module::iterator I = Prog->begin(), E = Prog->end(); I != E; ++I) + if (!I->isExternal()) + MiscompiledFunctions.push_back(I); + + // Do the reduction... + if (!BugpointIsInterrupted) + ReduceMiscompilingFunctions(BD, TestFn).reduceList(MiscompiledFunctions); + + std::cout << "\n*** The following function" + << (MiscompiledFunctions.size() == 1 ? " is" : "s are") + << " being miscompiled: "; + PrintFunctionList(MiscompiledFunctions); + std::cout << '\n'; + + // See if we can rip any loops out of the miscompiled functions and still + // trigger the problem. + if (!BugpointIsInterrupted && + ExtractLoops(BD, TestFn, MiscompiledFunctions)) { + // Okay, we extracted some loops and the problem still appears. See if we + // can eliminate some of the created functions from being candidates. + + // Loop extraction can introduce functions with the same name (foo_code). + // Make sure to disambiguate the symbols so that when the program is split + // apart that we can link it back together again. + DisambiguateGlobalSymbols(BD.getProgram()); + + // Do the reduction... + if (!BugpointIsInterrupted) + ReduceMiscompilingFunctions(BD, TestFn).reduceList(MiscompiledFunctions); + + std::cout << "\n*** The following function" + << (MiscompiledFunctions.size() == 1 ? " is" : "s are") + << " being miscompiled: "; + PrintFunctionList(MiscompiledFunctions); + std::cout << '\n'; } - std::cout << " Checking to see if the merged program executes correctly: "; + if (!BugpointIsInterrupted && + ExtractBlocks(BD, TestFn, MiscompiledFunctions)) { + // Okay, we extracted some blocks and the problem still appears. See if we + // can eliminate some of the created functions from being candidates. - // Eighth step: Execute the program. If it does not match the expected - // output, then 'Funcs' are being misoptimized! - bool Broken = BD.diffProgram(); + // Block extraction can introduce functions with the same name (foo_code). + // Make sure to disambiguate the symbols so that when the program is split + // apart that we can link it back together again. + DisambiguateGlobalSymbols(BD.getProgram()); - delete BD.Program; // Delete the hacked up program - BD.Program = OldProgram; // Restore the original + // Do the reduction... + ReduceMiscompilingFunctions(BD, TestFn).reduceList(MiscompiledFunctions); - std::cout << (Broken ? "nope.\n" : "yup.\n"); + std::cout << "\n*** The following function" + << (MiscompiledFunctions.size() == 1 ? " is" : "s are") + << " being miscompiled: "; + PrintFunctionList(MiscompiledFunctions); + std::cout << '\n'; + } + + return MiscompiledFunctions; +} + +/// TestOptimizer - This is the predicate function used to check to see if the +/// "Test" portion of the program is misoptimized. If so, return true. In any +/// case, both module arguments are deleted. +/// +static bool TestOptimizer(BugDriver &BD, Module *Test, Module *Safe) { + // Run the optimization passes on ToOptimize, producing a transformed version + // of the functions being tested. + std::cout << " Optimizing functions being tested: "; + Module *Optimized = BD.runPassesOn(Test, BD.getPassesToRun(), + /*AutoDebugCrashes*/true); + std::cout << "done.\n"; + delete Test; + + std::cout << " Checking to see if the merged program executes correctly: "; + bool Broken = TestMergedProgram(BD, Optimized, Safe, true); + std::cout << (Broken ? " nope.\n" : " yup.\n"); return Broken; } @@ -261,43 +577,306 @@ bool ReduceMiscompilingFunctions::TestFuncs(const std::vector &Funcs, /// input. /// bool BugDriver::debugMiscompilation() { + // Make sure something was miscompiled... + if (!BugpointIsInterrupted) + if (!ReduceMiscompilingPasses(*this).reduceList(PassesToRun)) { + std::cerr << "*** Optimized program matches reference output! No problem" + << " detected...\nbugpoint can't help you with your problem!\n"; + return false; + } + + std::cout << "\n*** Found miscompiling pass" + << (getPassesToRun().size() == 1 ? "" : "es") << ": " + << getPassesString(getPassesToRun()) << '\n'; + EmitProgressBytecode("passinput"); + + std::vector MiscompiledFunctions = + DebugAMiscompilation(*this, TestOptimizer); + + // Output a bunch of bytecode files for the user... + std::cout << "Outputting reduced bytecode files which expose the problem:\n"; + Module *ToNotOptimize = CloneModule(getProgram()); + Module *ToOptimize = SplitFunctionsOutOfModule(ToNotOptimize, + MiscompiledFunctions); - if (diffProgram()) { - std::cout << "\n*** Input program does not match reference diff!\n" - << " Must be problem with input source!\n"; - return false; // Problem found + std::cout << " Non-optimized portion: "; + ToNotOptimize = swapProgramIn(ToNotOptimize); + EmitProgressBytecode("tonotoptimize", true); + setNewProgram(ToNotOptimize); // Delete hacked module. + + std::cout << " Portion that is input to optimizer: "; + ToOptimize = swapProgramIn(ToOptimize); + EmitProgressBytecode("tooptimize"); + setNewProgram(ToOptimize); // Delete hacked module. + + return false; +} + +/// CleanupAndPrepareModules - Get the specified modules ready for code +/// generator testing. +/// +static void CleanupAndPrepareModules(BugDriver &BD, Module *&Test, + Module *Safe) { + // Clean up the modules, removing extra cruft that we don't need anymore... + Test = BD.performFinalCleanups(Test); + + // If we are executing the JIT, we have several nasty issues to take care of. + if (!BD.isExecutingJIT()) return; + + // First, if the main function is in the Safe module, we must add a stub to + // the Test module to call into it. Thus, we create a new function `main' + // which just calls the old one. + if (Function *oldMain = Safe->getNamedFunction("main")) + if (!oldMain->isExternal()) { + // Rename it + oldMain->setName("llvm_bugpoint_old_main"); + // Create a NEW `main' function with same type in the test module. + Function *newMain = new Function(oldMain->getFunctionType(), + GlobalValue::ExternalLinkage, + "main", Test); + // Create an `oldmain' prototype in the test module, which will + // corresponds to the real main function in the same module. + Function *oldMainProto = new Function(oldMain->getFunctionType(), + GlobalValue::ExternalLinkage, + oldMain->getName(), Test); + // Set up and remember the argument list for the main function. + std::vector args; + for (Function::arg_iterator + I = newMain->arg_begin(), E = newMain->arg_end(), + OI = oldMain->arg_begin(); I != E; ++I, ++OI) { + I->setName(OI->getName()); // Copy argument names from oldMain + args.push_back(I); + } + + // Call the old main function and return its result + BasicBlock *BB = new BasicBlock("entry", newMain); + CallInst *call = new CallInst(oldMainProto, args, "", BB); + + // If the type of old function wasn't void, return value of call + new ReturnInst(call, BB); + } + + // The second nasty issue we must deal with in the JIT is that the Safe + // module cannot directly reference any functions defined in the test + // module. Instead, we use a JIT API call to dynamically resolve the + // symbol. + + // Add the resolver to the Safe module. + // Prototype: void *getPointerToNamedFunction(const char* Name) + Function *resolverFunc = + Safe->getOrInsertFunction("getPointerToNamedFunction", + PointerType::get(Type::SByteTy), + PointerType::get(Type::SByteTy), (Type *)0); + + // Use the function we just added to get addresses of functions we need. + for (Module::iterator F = Safe->begin(), E = Safe->end(); F != E; ++F) { + if (F->isExternal() && !F->use_empty() && &*F != resolverFunc && + F->getIntrinsicID() == 0 /* ignore intrinsics */) { + Function *TestFn = Test->getNamedFunction(F->getName()); + + // Don't forward functions which are external in the test module too. + if (TestFn && !TestFn->isExternal()) { + // 1. Add a string constant with its name to the global file + Constant *InitArray = ConstantArray::get(F->getName()); + GlobalVariable *funcName = + new GlobalVariable(InitArray->getType(), true /*isConstant*/, + GlobalValue::InternalLinkage, InitArray, + F->getName() + "_name", Safe); + + // 2. Use `GetElementPtr *funcName, 0, 0' to convert the string to an + // sbyte* so it matches the signature of the resolver function. + + // GetElementPtr *funcName, ulong 0, ulong 0 + std::vector GEPargs(2,Constant::getNullValue(Type::IntTy)); + Value *GEP = + ConstantExpr::getGetElementPtr(funcName, GEPargs); + std::vector ResolverArgs; + ResolverArgs.push_back(GEP); + + // Rewrite uses of F in global initializers, etc. to uses of a wrapper + // function that dynamically resolves the calls to F via our JIT API + if (!F->use_empty()) { + // Create a new global to hold the cached function pointer. + Constant *NullPtr = ConstantPointerNull::get(F->getType()); + GlobalVariable *Cache = + new GlobalVariable(F->getType(), false,GlobalValue::InternalLinkage, + NullPtr,F->getName()+".fpcache", F->getParent()); + + // Construct a new stub function that will re-route calls to F + const FunctionType *FuncTy = F->getFunctionType(); + Function *FuncWrapper = new Function(FuncTy, + GlobalValue::InternalLinkage, + F->getName() + "_wrapper", + F->getParent()); + BasicBlock *EntryBB = new BasicBlock("entry", FuncWrapper); + BasicBlock *DoCallBB = new BasicBlock("usecache", FuncWrapper); + BasicBlock *LookupBB = new BasicBlock("lookupfp", FuncWrapper); + + // Check to see if we already looked up the value. + Value *CachedVal = new LoadInst(Cache, "fpcache", EntryBB); + Value *IsNull = new SetCondInst(Instruction::SetEQ, CachedVal, + NullPtr, "isNull", EntryBB); + new BranchInst(LookupBB, DoCallBB, IsNull, EntryBB); + + // Resolve the call to function F via the JIT API: + // + // call resolver(GetElementPtr...) + CallInst *Resolver = new CallInst(resolverFunc, ResolverArgs, + "resolver", LookupBB); + // cast the result from the resolver to correctly-typed function + CastInst *CastedResolver = + new CastInst(Resolver, PointerType::get(F->getFunctionType()), + "resolverCast", LookupBB); + // Save the value in our cache. + new StoreInst(CastedResolver, Cache, LookupBB); + new BranchInst(DoCallBB, LookupBB); + + PHINode *FuncPtr = new PHINode(NullPtr->getType(), "fp", DoCallBB); + FuncPtr->addIncoming(CastedResolver, LookupBB); + FuncPtr->addIncoming(CachedVal, EntryBB); + + // Save the argument list. + std::vector Args; + for (Function::arg_iterator i = FuncWrapper->arg_begin(), + e = FuncWrapper->arg_end(); i != e; ++i) + Args.push_back(i); + + // Pass on the arguments to the real function, return its result + if (F->getReturnType() == Type::VoidTy) { + CallInst *Call = new CallInst(FuncPtr, Args, "", DoCallBB); + new ReturnInst(DoCallBB); + } else { + CallInst *Call = new CallInst(FuncPtr, Args, "retval", DoCallBB); + new ReturnInst(Call, DoCallBB); + } + + // Use the wrapper function instead of the old function + F->replaceAllUsesWith(FuncWrapper); + } + } + } } - // Make sure something was miscompiled... - if (!ReduceMiscompilingPasses(*this).reduceList(PassesToRun)) { - std::cerr << "*** Optimized program matches reference output! No problem " - << "detected...\nbugpoint can't help you with your problem!\n"; - return false; + if (verifyModule(*Test) || verifyModule(*Safe)) { + std::cerr << "Bugpoint has a bug, which corrupted a module!!\n"; + abort(); } +} - std::cout << "\n*** Found miscompiling pass" - << (PassesToRun.size() == 1 ? "" : "es") << ": " - << getPassesString(PassesToRun) << "\n"; - EmitProgressBytecode("passinput"); - // Okay, now that we have reduced the list of passes which are causing the - // failure, see if we can pin down which functions are being - // miscompiled... first build a list of all of the non-external functions in - // the program. - std::vector MiscompiledFunctions; - for (Module::iterator I = Program->begin(), E = Program->end(); I != E; ++I) - if (!I->isExternal()) - MiscompiledFunctions.push_back(I); - // Do the reduction... - ReduceMiscompilingFunctions(*this).reduceList(MiscompiledFunctions); +/// TestCodeGenerator - This is the predicate function used to check to see if +/// the "Test" portion of the program is miscompiled by the code generator under +/// test. If so, return true. In any case, both module arguments are deleted. +/// +static bool TestCodeGenerator(BugDriver &BD, Module *Test, Module *Safe) { + CleanupAndPrepareModules(BD, Test, Safe); - std::cout << "\n*** The following functions are being miscompiled: "; - PrintFunctionList(MiscompiledFunctions); - std::cout << "\n"; + sys::Path TestModuleBC("bugpoint.test.bc"); + TestModuleBC.makeUnique(); + if (BD.writeProgramToFile(TestModuleBC.toString(), Test)) { + std::cerr << "Error writing bytecode to `" << TestModuleBC << "'\nExiting."; + exit(1); + } + delete Test; - // Output a bunch of bytecode files for the user... - ReduceMiscompilingFunctions(*this).TestFuncs(MiscompiledFunctions, true); + // Make the shared library + sys::Path SafeModuleBC("bugpoint.safe.bc"); + SafeModuleBC.makeUnique(); + + if (BD.writeProgramToFile(SafeModuleBC.toString(), Safe)) { + std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting."; + exit(1); + } + std::string SharedObject = BD.compileSharedObject(SafeModuleBC.toString()); + delete Safe; + + // Run the code generator on the `Test' code, loading the shared library. + // The function returns whether or not the new output differs from reference. + int Result = BD.diffProgram(TestModuleBC.toString(), SharedObject, false); + + if (Result) + std::cerr << ": still failing!\n"; + else + std::cerr << ": didn't fail.\n"; + TestModuleBC.eraseFromDisk(); + SafeModuleBC.eraseFromDisk(); + sys::Path(SharedObject).eraseFromDisk(); + + return Result; +} + + +/// debugCodeGenerator - debug errors in LLC, LLI, or CBE. +/// +bool BugDriver::debugCodeGenerator() { + if ((void*)cbe == (void*)Interpreter) { + std::string Result = executeProgramWithCBE("bugpoint.cbe.out"); + std::cout << "\n*** The C backend cannot match the reference diff, but it " + << "is used as the 'known good'\n code generator, so I can't" + << " debug it. Perhaps you have a front-end problem?\n As a" + << " sanity check, I left the result of executing the program " + << "with the C backend\n in this file for you: '" + << Result << "'.\n"; + return true; + } + + DisambiguateGlobalSymbols(Program); + + std::vector Funcs = DebugAMiscompilation(*this, TestCodeGenerator); + + // Split the module into the two halves of the program we want. + Module *ToNotCodeGen = CloneModule(getProgram()); + Module *ToCodeGen = SplitFunctionsOutOfModule(ToNotCodeGen, Funcs); + + // Condition the modules + CleanupAndPrepareModules(*this, ToCodeGen, ToNotCodeGen); + + sys::Path TestModuleBC("bugpoint.test.bc"); + TestModuleBC.makeUnique(); + + if (writeProgramToFile(TestModuleBC.toString(), ToCodeGen)) { + std::cerr << "Error writing bytecode to `" << TestModuleBC << "'\nExiting."; + exit(1); + } + delete ToCodeGen; + + // Make the shared library + sys::Path SafeModuleBC("bugpoint.safe.bc"); + SafeModuleBC.makeUnique(); + + if (writeProgramToFile(SafeModuleBC.toString(), ToNotCodeGen)) { + std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting."; + exit(1); + } + std::string SharedObject = compileSharedObject(SafeModuleBC.toString()); + delete ToNotCodeGen; + + std::cout << "You can reproduce the problem with the command line: \n"; + if (isExecutingJIT()) { + std::cout << " lli -load " << SharedObject << " " << TestModuleBC; + } else { + std::cout << " llc -f " << TestModuleBC << " -o " << TestModuleBC<< ".s\n"; + std::cout << " gcc " << SharedObject << " " << TestModuleBC + << ".s -o " << TestModuleBC << ".exe"; +#if defined (HAVE_LINK_R) + std::cout << "-Wl,-R."; +#endif + std::cout << "\n"; + std::cout << " " << TestModuleBC << ".exe"; + } + for (unsigned i=0, e = InputArgv.size(); i != e; ++i) + std::cout << " " << InputArgv[i]; + std::cout << '\n'; + std::cout << "The shared object was created with:\n llc -march=c " + << SafeModuleBC << " -o temporary.c\n" + << " gcc -xc temporary.c -O2 -o " << SharedObject +#if defined(sparc) || defined(__sparc__) || defined(__sparcv9) + << " -G" // Compile a shared library, `-G' for Sparc +#else + << " -shared" // `-shared' for Linux/X86, maybe others +#endif + << " -fno-strict-aliasing\n"; return false; }