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,
53 std::cout << "Testing functions: ";
54 BD.PrintFunctionList(Funcs);
57 // Clone the module for the two halves of the program we want.
58 Module *SafeModule = CloneModule(BD.Program);
60 // Make sure functions & globals are all external so that linkage
61 // between the two modules will work.
62 for (Module::iterator I = SafeModule->begin(), E = SafeModule->end();I!=E;++I)
63 I->setLinkage(GlobalValue::ExternalLinkage);
64 for (Module::giterator I=SafeModule->gbegin(),E = SafeModule->gend();I!=E;++I)
65 I->setLinkage(GlobalValue::ExternalLinkage);
67 Module *TestModule = CloneModule(SafeModule);
69 // Make sure global initializers exist only in the safe module (CBE->.so)
70 for (Module::giterator I=TestModule->gbegin(),E = TestModule->gend();I!=E;++I)
71 I->setInitializer(0); // Delete the initializer to make it external
73 // Remove the Test functions from the Safe module
74 for (unsigned i = 0, e = Funcs.size(); i != e; ++i) {
75 Function *TNOF = SafeModule->getFunction(Funcs[i]->getName(),
76 Funcs[i]->getFunctionType());
77 DEBUG(std::cerr << "Removing function " << Funcs[i]->getName() << "\n");
78 assert(TNOF && "Function doesn't exist in module!");
79 DeleteFunctionBody(TNOF); // Function is now external in this module!
82 // Remove the Safe functions from the Test module
83 for (Module::iterator I=TestModule->begin(),E=TestModule->end(); I!=E; ++I) {
84 bool funcFound = false;
85 for (std::vector<Function*>::const_iterator F=Funcs.begin(),Fe=Funcs.end();
87 if (I->getName() == (*F)->getName()) funcFound = true;
89 if (!funcFound && !(BD.isExecutingJIT() && I->getName() == "main"))
90 DeleteFunctionBody(I);
93 // This is only applicable if we are debugging the JIT:
94 // Find all external functions in the Safe modules that are actually used
95 // (called or taken address of), and make them call the JIT wrapper instead
96 if (BD.isExecutingJIT()) {
97 // Must delete `main' from Safe module if it has it
98 Function *safeMain = SafeModule->getNamedFunction("main");
99 assert(safeMain && "`main' function not found in safe module!");
100 DeleteFunctionBody(safeMain);
102 // Add an external function "getPointerToNamedFunction" that JIT provides
103 // Prototype: void *getPointerToNamedFunction(const char* Name)
104 std::vector<const Type*> Params;
105 Params.push_back(PointerType::get(Type::SByteTy)); // std::string&
106 FunctionType *resolverTy = FunctionType::get(PointerType::get(Type::VoidTy),
107 Params, false /* isVarArg */);
108 Function *resolverFunc = new Function(resolverTy,
109 GlobalValue::ExternalLinkage,
110 "getPointerToNamedFunction",
113 // Use the function we just added to get addresses of functions we need
114 // Iterate over the global declarations in the Safe module
115 for (Module::iterator F=SafeModule->begin(),E=SafeModule->end(); F!=E; ++F){
116 if (F->isExternal() && !F->use_empty() && &(*F) != resolverFunc &&
117 F->getIntrinsicID() == 0 /* ignore intrinsics */) {
118 // If it has a non-zero use list,
119 // 1. Add a string constant with its name to the global file
120 // The correct type is `const [ NUM x sbyte ]' where NUM is length of
122 const std::string &Name = F->getName();
123 GlobalVariable *funcName =
124 new GlobalVariable(ArrayType::get(Type::SByteTy, Name.length()+1),
125 true /* isConstant */,
126 GlobalValue::InternalLinkage,
127 ConstantArray::get(Name),
131 // 2. Use `GetElementPtr *funcName, 0, 0' to convert the string to an
132 // sbyte* so it matches the signature of the resolver function.
133 std::vector<Constant*> GEPargs(2, Constant::getNullValue(Type::LongTy));
135 // 3. Replace all uses of `func' with calls to resolver by:
136 // (a) Iterating through the list of uses of this function
137 // (b) Insert a cast instruction in front of each use
138 // (c) Replace use of old call with new call
140 // GetElementPtr *funcName, ulong 0, ulong 0
142 ConstantExpr::getGetElementPtr(ConstantPointerRef::get(funcName),
144 std::vector<Value*> ResolverArgs;
145 ResolverArgs.push_back(GEP);
147 // Insert code at the beginning of the function
148 for (Value::use_iterator i=F->use_begin(), e=F->use_end(); i!=e; ++i) {
149 if (Instruction* Inst = dyn_cast<Instruction>(*i)) {
150 // call resolver(GetElementPtr...)
151 CallInst *resolve = new CallInst(resolverFunc, ResolverArgs,
153 // cast the result from the resolver to correctly-typed function
154 CastInst *castResolver =
155 new CastInst(resolve, PointerType::get(F->getFunctionType()),
156 "resolverCast", Inst);
157 // actually use the resolved function
158 Inst->replaceUsesOfWith(F, castResolver);
160 // FIXME: need to take care of cases where a function is used that
161 // is not an instruction, e.g. global variable initializer...
162 std::cerr << "Non-instruction is using an external function!\n";
170 if (verifyModule(*SafeModule) || verifyModule(*TestModule)) {
171 std::cerr << "Bugpoint has a bug, an corrupted a module!!\n";
175 DEBUG(std::cerr << "Safe module:\n";
176 typedef Module::iterator MI;
177 typedef Module::giterator MGI;
179 for (MI I = SafeModule->begin(), E = SafeModule->end(); I != E; ++I)
180 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
181 for (MGI I = SafeModule->gbegin(), E = SafeModule->gend(); I!=E; ++I)
182 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
184 std::cerr << "Test module:\n";
185 for (MI I = TestModule->begin(), E = TestModule->end(); I != E; ++I)
186 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
187 for (MGI I=TestModule->gbegin(),E = TestModule->gend(); I!= E; ++I)
188 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
191 // Write out the bytecode to be sent to CBE
192 std::string SafeModuleBC = getUniqueFilename("bugpoint.safe.bc");
194 if (BD.writeProgramToFile(SafeModuleBC, SafeModule)) {
195 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
199 // Remove all functions from the Test module EXCEPT for the ones specified in
200 // Funcs. We know which ones these are because they are non-external in
201 // ToOptimize, but external in ToNotOptimize.
203 for (Module::iterator I = TestModule->begin(), E = TestModule->end();I!=E;++I)
204 if (!I->isExternal()) {
205 Function *TNOF = SafeModule->getFunction(I->getName(),
206 I->getFunctionType());
207 assert(TNOF && "Function doesn't exist in ToNotOptimize module??");
208 if (!TNOF->isExternal())
209 DeleteFunctionBody(I);
212 std::string TestModuleBC = getUniqueFilename("bugpoint.test.bc");
213 if (verifyModule(*TestModule)) {
214 std::cerr << "Bytecode file corrupted!\n";
218 // Clean up the modules, removing extra cruft that we don't need anymore...
219 SafeModule = BD.performFinalCleanups(SafeModule);
220 TestModule = BD.performFinalCleanups(TestModule);
222 if (BD.writeProgramToFile(TestModuleBC, TestModule)) {
223 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
227 // Make a shared library
228 std::string SharedObject;
229 BD.compileSharedObject(SafeModuleBC, SharedObject);
234 // Run the code generator on the `Test' code, loading the shared library.
235 // The function returns whether or not the new output differs from reference.
236 int Result = BD.diffProgram(TestModuleBC, SharedObject, false);
239 std::cerr << ": still failing!\n";
241 std::cerr << ": didn't fail.\n";
244 std::cout << "You can reproduce the problem with the command line: \n";
245 if (BD.isExecutingJIT()) {
246 std::cout << " lli -load " << SharedObject << " " << TestModuleBC;
248 std::cout << " llc " << TestModuleBC << " -o " << TestModuleBC << ".s\n";
249 std::cout << " gcc " << SharedObject << " " << TestModuleBC
250 << ".s -o " << TestModuleBC << ".exe\n";
251 std::cout << " " << TestModuleBC << ".exe";
253 for (unsigned i=0, e = InputArgv.size(); i != e; ++i)
254 std::cout << " " << InputArgv[i];
256 std::cout << "The shared object was created with:\n dis -c "
257 << SafeModuleBC << " -o temporary.c\n"
258 << " gcc -shared temporary.c -o " << SharedObject << "\n";
260 removeFile(TestModuleBC);
261 removeFile(SafeModuleBC);
262 removeFile(SharedObject);
268 struct Disambiguator {
269 std::set<std::string> SymbolNames;
270 std::set<GlobalValue*> Symbols;
271 uint64_t uniqueCounter;
274 Disambiguator() : uniqueCounter(0), externalOnly(true) {}
275 void setExternalOnly(bool value) { externalOnly = value; }
276 void add(GlobalValue &V) {
277 // If we're only processing externals and this isn't external, bail
278 if (externalOnly && !V.isExternal()) return;
279 // If we're already processed this symbol, don't add it again
280 if (Symbols.count(&V) != 0) return;
281 // Ignore intrinsic functions
282 if (Function *F = dyn_cast<Function>(&V))
283 if (F->getIntrinsicID() != 0)
286 std::string SymName = V.getName();
288 // Use the Mangler facility to make symbol names that will be valid in
290 SymName = Mangler::makeNameProper(SymName);
293 if (SymbolNames.count(SymName) == 0) {
294 DEBUG(std::cerr << "Disambiguator: adding " << SymName
295 << ", no conflicts.\n");
296 SymbolNames.insert(SymName);
298 // Mangle name before adding
301 newName = SymName + "_" + utostr(uniqueCounter);
302 if (SymbolNames.count(newName) == 0) break;
303 else ++uniqueCounter;
305 //while (SymbolNames.count(V->getName()+utostr(uniqueCounter++))==0);
306 DEBUG(std::cerr << "Disambiguator: conflict: " << SymName
307 << ", adding: " << newName << "\n");
309 SymbolNames.insert(newName);
316 void DisambiguateGlobalSymbols(Module *M) {
317 // First, try not to cause collisions by minimizing chances of renaming an
318 // already-external symbol, so take in external globals and functions as-is.
320 DEBUG(std::cerr << "Disambiguating globals (external-only)\n");
321 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) D.add(*I);
322 DEBUG(std::cerr << "Disambiguating functions (external-only)\n");
323 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) D.add(*I);
325 // Now just rename functions and globals as necessary, keeping what's already
326 // in the set unique.
327 D.setExternalOnly(false);
328 DEBUG(std::cerr << "Disambiguating globals\n");
329 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) D.add(*I);
330 DEBUG(std::cerr << "Disambiguating globals\n");
331 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) D.add(*I);
335 bool BugDriver::debugCodeGenerator() {
336 // See if we can pin down which functions are being miscompiled...
337 //First, build a list of all of the non-external functions in the program.
338 std::vector<Function*> MisCodegenFunctions;
339 for (Module::iterator I = Program->begin(), E = Program->end(); I != E; ++I)
340 if (!I->isExternal())
341 MisCodegenFunctions.push_back(I);
343 // If we are executing the JIT, we *must* keep the function `main' in the
344 // module that is passed in, and not the shared library. However, we still
345 // want to be able to debug the `main' function alone. Thus, we create a new
346 // function `main' which just calls the old one.
347 if (isExecutingJIT()) {
348 // Get the `main' function
349 Function *oldMain = Program->getNamedFunction("main");
350 assert(oldMain && "`main' function not found in program!");
352 oldMain->setName("llvm_old_main");
353 // Create a NEW `main' function with same type
354 Function *newMain = new Function(oldMain->getFunctionType(),
355 GlobalValue::ExternalLinkage,
357 // Call the old main function and return its result
358 BasicBlock *BB = new BasicBlock("entry", newMain);
359 std::vector<Value*> args;
360 for (Function::aiterator I = newMain->abegin(), E = newMain->aend(),
361 OI = oldMain->abegin(); I != E; ++I, ++OI) {
362 I->setName(OI->getName()); // Copy argument names from oldMain
365 CallInst *call = new CallInst(oldMain, args);
366 BB->getInstList().push_back(call);
368 // if the type of old function wasn't void, return value of call
370 if (oldMain->getReturnType() != Type::VoidTy) {
371 ret = new ReturnInst(call);
373 ret = new ReturnInst();
376 // Add the return instruction to the BasicBlock
377 BB->getInstList().push_back(ret);
380 DisambiguateGlobalSymbols(Program);
382 // Do the reduction...
383 if (!ReduceMisCodegenFunctions(*this).reduceList(MisCodegenFunctions)) {
384 std::cerr << "*** Execution matches reference output! "
385 << "bugpoint can't help you with your problem!\n";
389 std::cout << "\n*** The following functions are being miscompiled: ";
390 PrintFunctionList(MisCodegenFunctions);
393 // Output a bunch of bytecode files for the user...
394 ReduceMisCodegenFunctions(*this).TestFuncs(MisCodegenFunctions, true);