1 //===- CodeGeneratorBug.cpp - Debug code generation bugs ------------------===//
3 // This file implements program code generation debugging support.
5 //===----------------------------------------------------------------------===//
8 #include "ListReducer.h"
9 #include "llvm/Constants.h"
10 #include "llvm/DerivedTypes.h"
11 #include "llvm/GlobalValue.h"
12 #include "llvm/iMemory.h"
13 #include "llvm/iTerminators.h"
14 #include "llvm/iOther.h"
15 #include "llvm/Module.h"
16 #include "llvm/Pass.h"
17 #include "llvm/Analysis/Verifier.h"
18 #include "llvm/Support/Mangler.h"
19 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
20 #include "llvm/Transforms/Utils/Cloning.h"
21 #include "llvm/Transforms/Utils/Linker.h"
22 #include "Support/CommandLine.h"
23 #include "Support/Debug.h"
24 #include "Support/StringExtras.h"
25 #include "Support/FileUtilities.h"
29 extern cl::list<std::string> InputArgv;
31 class ReduceMisCodegenFunctions : public ListReducer<Function*> {
34 ReduceMisCodegenFunctions(BugDriver &bd) : BD(bd) {}
36 virtual TestResult doTest(std::vector<Function*> &Prefix,
37 std::vector<Function*> &Suffix) {
38 if (!Prefix.empty() && TestFuncs(Prefix))
40 if (!Suffix.empty() && TestFuncs(Suffix))
45 bool TestFuncs(const std::vector<Function*> &CodegenTest,
46 bool KeepFiles = false);
50 bool ReduceMisCodegenFunctions::TestFuncs(const std::vector<Function*> &Funcs,
52 std::cout << "Testing functions: ";
53 BD.PrintFunctionList(Funcs);
56 // Clone the module for the two halves of the program we want.
57 Module *SafeModule = CloneModule(BD.Program);
59 // Make sure functions & globals are all external so that linkage
60 // between the two modules will work.
61 for (Module::iterator I = SafeModule->begin(), E = SafeModule->end();I!=E;++I)
62 I->setLinkage(GlobalValue::ExternalLinkage);
63 for (Module::giterator I=SafeModule->gbegin(),E = SafeModule->gend();I!=E;++I)
64 I->setLinkage(GlobalValue::ExternalLinkage);
66 Module *TestModule = CloneModule(SafeModule);
68 // Make sure global initializers exist only in the safe module (CBE->.so)
69 for (Module::giterator I=TestModule->gbegin(),E = TestModule->gend();I!=E;++I)
70 I->setInitializer(0); // Delete the initializer to make it external
72 // Remove the Test functions from the Safe module
73 for (unsigned i = 0, e = Funcs.size(); i != e; ++i) {
74 Function *TNOF = SafeModule->getFunction(Funcs[i]->getName(),
75 Funcs[i]->getFunctionType());
76 DEBUG(std::cerr << "Removing function " << Funcs[i]->getName() << "\n");
77 assert(TNOF && "Function doesn't exist in module!");
78 DeleteFunctionBody(TNOF); // Function is now external in this module!
81 // Remove the Safe functions from the Test module
82 for (Module::iterator I=TestModule->begin(),E=TestModule->end(); I!=E; ++I) {
83 bool funcFound = false;
84 for (std::vector<Function*>::const_iterator F=Funcs.begin(),Fe=Funcs.end();
86 if (I->getName() == (*F)->getName()) funcFound = true;
88 if (!funcFound && !(BD.isExecutingJIT() && I->getName() == "main"))
89 DeleteFunctionBody(I);
92 // This is only applicable if we are debugging the JIT:
93 // Find all external functions in the Safe modules that are actually used
94 // (called or taken address of), and make them call the JIT wrapper instead
95 if (BD.isExecutingJIT()) {
96 // Must delete `main' from Safe module if it has it
97 Function *safeMain = SafeModule->getNamedFunction("main");
98 assert(safeMain && "`main' function not found in safe module!");
99 DeleteFunctionBody(safeMain);
101 // Add an external function "getPointerToNamedFunction" that JIT provides
102 // Prototype: void *getPointerToNamedFunction(const char* Name)
103 std::vector<const Type*> Params;
104 Params.push_back(PointerType::get(Type::SByteTy)); // std::string&
105 FunctionType *resolverTy = FunctionType::get(PointerType::get(Type::VoidTy),
106 Params, false /* isVarArg */);
107 Function *resolverFunc = new Function(resolverTy,
108 GlobalValue::ExternalLinkage,
109 "getPointerToNamedFunction",
112 // Use the function we just added to get addresses of functions we need
113 // Iterate over the global declarations in the Safe module
114 for (Module::iterator F=SafeModule->begin(),E=SafeModule->end(); F!=E; ++F){
115 if (F->isExternal() && !F->use_empty() && &(*F) != resolverFunc &&
116 F->getIntrinsicID() == 0 /* ignore intrinsics */) {
117 // If it has a non-zero use list,
118 // 1. Add a string constant with its name to the global file
119 // The correct type is `const [ NUM x sbyte ]' where NUM is length of
121 const std::string &Name = F->getName();
122 GlobalVariable *funcName =
123 new GlobalVariable(ArrayType::get(Type::SByteTy, Name.length()+1),
124 true /* isConstant */,
125 GlobalValue::InternalLinkage,
126 ConstantArray::get(Name),
130 // 2. Use `GetElementPtr *funcName, 0, 0' to convert the string to an
131 // sbyte* so it matches the signature of the resolver function.
132 std::vector<Constant*> GEPargs(2, Constant::getNullValue(Type::LongTy));
134 // 3. Replace all uses of `func' with calls to resolver by:
135 // (a) Iterating through the list of uses of this function
136 // (b) Insert a cast instruction in front of each use
137 // (c) Replace use of old call with new call
139 // GetElementPtr *funcName, ulong 0, ulong 0
141 ConstantExpr::getGetElementPtr(ConstantPointerRef::get(funcName),
143 std::vector<Value*> ResolverArgs;
144 ResolverArgs.push_back(GEP);
146 // Insert code at the beginning of the function
147 while (!F->use_empty())
148 if (Instruction *Inst = dyn_cast<Instruction>(F->use_back())) {
149 // call resolver(GetElementPtr...)
150 CallInst *resolve = new CallInst(resolverFunc, ResolverArgs,
152 // cast the result from the resolver to correctly-typed function
153 CastInst *castResolver =
154 new CastInst(resolve, PointerType::get(F->getFunctionType()),
155 "resolverCast", Inst);
156 // actually use the resolved function
157 Inst->replaceUsesOfWith(F, castResolver);
159 // FIXME: need to take care of cases where a function is used that
160 // is not an instruction, e.g. global variable initializer...
161 std::cerr << "Non-instruction is using an external function!\n";
168 if (verifyModule(*SafeModule) || verifyModule(*TestModule)) {
169 std::cerr << "Bugpoint has a bug, an corrupted a module!!\n";
173 DEBUG(std::cerr << "Safe module:\n";
174 typedef Module::iterator MI;
175 typedef Module::giterator MGI;
177 for (MI I = SafeModule->begin(), E = SafeModule->end(); I != E; ++I)
178 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
179 for (MGI I = SafeModule->gbegin(), E = SafeModule->gend(); I!=E; ++I)
180 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
182 std::cerr << "Test module:\n";
183 for (MI I = TestModule->begin(), E = TestModule->end(); I != E; ++I)
184 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
185 for (MGI I=TestModule->gbegin(),E = TestModule->gend(); I!= E; ++I)
186 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
189 // Write out the bytecode to be sent to CBE
190 std::string SafeModuleBC = getUniqueFilename("bugpoint.safe.bc");
192 if (BD.writeProgramToFile(SafeModuleBC, SafeModule)) {
193 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
197 // Remove all functions from the Test module EXCEPT for the ones specified in
198 // Funcs. We know which ones these are because they are non-external in
199 // ToOptimize, but external in ToNotOptimize.
201 for (Module::iterator I = TestModule->begin(), E = TestModule->end();I!=E;++I)
202 if (!I->isExternal()) {
203 Function *TNOF = SafeModule->getFunction(I->getName(),
204 I->getFunctionType());
205 assert(TNOF && "Function doesn't exist in ToNotOptimize module??");
206 if (!TNOF->isExternal())
207 DeleteFunctionBody(I);
210 std::string TestModuleBC = getUniqueFilename("bugpoint.test.bc");
211 if (verifyModule(*TestModule)) {
212 std::cerr << "Bytecode file corrupted!\n";
216 // Clean up the modules, removing extra cruft that we don't need anymore...
217 SafeModule = BD.performFinalCleanups(SafeModule);
218 TestModule = BD.performFinalCleanups(TestModule);
220 if (BD.writeProgramToFile(TestModuleBC, TestModule)) {
221 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
225 // Make a shared library
226 std::string SharedObject = BD.compileSharedObject(SafeModuleBC);
231 // Run the code generator on the `Test' code, loading the shared library.
232 // The function returns whether or not the new output differs from reference.
233 int Result = BD.diffProgram(TestModuleBC, SharedObject, false);
236 std::cerr << ": still failing!\n";
238 std::cerr << ": didn't fail.\n";
241 std::cout << "You can reproduce the problem with the command line: \n";
242 if (BD.isExecutingJIT()) {
243 std::cout << " lli -load " << SharedObject << " " << TestModuleBC;
245 std::cout << " llc " << TestModuleBC << " -o " << TestModuleBC << ".s\n";
246 std::cout << " gcc " << SharedObject << " " << TestModuleBC
247 << ".s -o " << TestModuleBC << ".exe -Wl,-R.\n";
248 std::cout << " " << TestModuleBC << ".exe";
250 for (unsigned i=0, e = InputArgv.size(); i != e; ++i)
251 std::cout << " " << InputArgv[i];
253 std::cout << "The shared object was created with:\n llvm-dis -c "
254 << SafeModuleBC << " -o temporary.c\n"
255 << " gcc -xc temporary.c -O2 -o " << SharedObject
256 #if defined(sparc) || defined(__sparc__) || defined(__sparcv9)
257 << " -G" // Compile a shared library, `-G' for Sparc
259 << " -shared" // `-shared' for Linux/X86, maybe others
261 << " -fno-strict-aliasing\n";
263 removeFile(TestModuleBC);
264 removeFile(SafeModuleBC);
265 removeFile(SharedObject);
271 struct Disambiguator {
272 std::set<std::string> SymbolNames;
273 std::set<GlobalValue*> Symbols;
274 uint64_t uniqueCounter;
277 Disambiguator() : uniqueCounter(0), externalOnly(true) {}
278 void setExternalOnly(bool value) { externalOnly = value; }
279 void add(GlobalValue &V) {
280 // If we're only processing externals and this isn't external, bail
281 if (externalOnly && !V.isExternal()) return;
282 // If we're already processed this symbol, don't add it again
283 if (Symbols.count(&V) != 0) return;
284 // Ignore intrinsic functions
285 if (Function *F = dyn_cast<Function>(&V))
286 if (F->getIntrinsicID() != 0)
289 std::string SymName = V.getName();
291 // Use the Mangler facility to make symbol names that will be valid in
293 SymName = Mangler::makeNameProper(SymName);
296 if (SymbolNames.count(SymName) == 0) {
297 DEBUG(std::cerr << "Disambiguator: adding " << SymName
298 << ", no conflicts.\n");
299 SymbolNames.insert(SymName);
301 // Mangle name before adding
304 newName = SymName + "_" + utostr(uniqueCounter);
305 if (SymbolNames.count(newName) == 0) break;
306 else ++uniqueCounter;
308 //while (SymbolNames.count(V->getName()+utostr(uniqueCounter++))==0);
309 DEBUG(std::cerr << "Disambiguator: conflict: " << SymName
310 << ", adding: " << newName << "\n");
312 SymbolNames.insert(newName);
319 void DisambiguateGlobalSymbols(Module *M) {
320 // First, try not to cause collisions by minimizing chances of renaming an
321 // already-external symbol, so take in external globals and functions as-is.
323 DEBUG(std::cerr << "Disambiguating globals (external-only)\n");
324 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) D.add(*I);
325 DEBUG(std::cerr << "Disambiguating functions (external-only)\n");
326 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) D.add(*I);
328 // Now just rename functions and globals as necessary, keeping what's already
329 // in the set unique.
330 D.setExternalOnly(false);
331 DEBUG(std::cerr << "Disambiguating globals\n");
332 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) D.add(*I);
333 DEBUG(std::cerr << "Disambiguating globals\n");
334 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) D.add(*I);
338 bool BugDriver::debugCodeGenerator() {
339 // See if we can pin down which functions are being miscompiled...
340 //First, build a list of all of the non-external functions in the program.
341 std::vector<Function*> MisCodegenFunctions;
342 for (Module::iterator I = Program->begin(), E = Program->end(); I != E; ++I)
343 if (!I->isExternal())
344 MisCodegenFunctions.push_back(I);
346 // If we are executing the JIT, we *must* keep the function `main' in the
347 // module that is passed in, and not the shared library. However, we still
348 // want to be able to debug the `main' function alone. Thus, we create a new
349 // function `main' which just calls the old one.
350 if (isExecutingJIT()) {
351 // Get the `main' function
352 Function *oldMain = Program->getNamedFunction("main");
353 assert(oldMain && "`main' function not found in program!");
355 oldMain->setName("llvm_old_main");
356 // Create a NEW `main' function with same type
357 Function *newMain = new Function(oldMain->getFunctionType(),
358 GlobalValue::ExternalLinkage,
360 // Call the old main function and return its result
361 BasicBlock *BB = new BasicBlock("entry", newMain);
362 std::vector<Value*> args;
363 for (Function::aiterator I = newMain->abegin(), E = newMain->aend(),
364 OI = oldMain->abegin(); I != E; ++I, ++OI) {
365 I->setName(OI->getName()); // Copy argument names from oldMain
368 CallInst *call = new CallInst(oldMain, args);
369 BB->getInstList().push_back(call);
371 // if the type of old function wasn't void, return value of call
373 if (oldMain->getReturnType() != Type::VoidTy) {
374 ret = new ReturnInst(call);
376 ret = new ReturnInst();
379 // Add the return instruction to the BasicBlock
380 BB->getInstList().push_back(ret);
383 DisambiguateGlobalSymbols(Program);
385 // Do the reduction...
386 if (!ReduceMisCodegenFunctions(*this).reduceList(MisCodegenFunctions)) {
387 std::cerr << "*** Execution matches reference output! "
388 << "bugpoint can't help you with your problem!\n";
392 std::cout << "\n*** The following functions are being miscompiled: ";
393 PrintFunctionList(MisCodegenFunctions);
396 // Output a bunch of bytecode files for the user...
397 ReduceMisCodegenFunctions(*this).TestFuncs(MisCodegenFunctions, true);