1 //===- CodeGeneratorBug.cpp - Debug code generation bugs ------------------===//
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 program code generation debugging support.
12 //===----------------------------------------------------------------------===//
14 #include "BugDriver.h"
15 #include "ListReducer.h"
16 #include "llvm/Constants.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/GlobalValue.h"
19 #include "llvm/iMemory.h"
20 #include "llvm/iTerminators.h"
21 #include "llvm/iOther.h"
22 #include "llvm/Module.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Analysis/Verifier.h"
25 #include "llvm/Support/Mangler.h"
26 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
27 #include "llvm/Transforms/Utils/Cloning.h"
28 #include "llvm/Transforms/Utils/Linker.h"
29 #include "Support/CommandLine.h"
30 #include "Support/Debug.h"
31 #include "Support/StringExtras.h"
32 #include "Support/FileUtilities.h"
38 extern cl::list<std::string> InputArgv;
40 class ReduceMisCodegenFunctions : public ListReducer<Function*> {
43 ReduceMisCodegenFunctions(BugDriver &bd) : BD(bd) {}
45 virtual TestResult doTest(std::vector<Function*> &Prefix,
46 std::vector<Function*> &Suffix) {
47 if (!Prefix.empty() && TestFuncs(Prefix))
49 if (!Suffix.empty() && TestFuncs(Suffix))
54 bool TestFuncs(const std::vector<Function*> &CodegenTest,
55 bool KeepFiles = false);
59 bool ReduceMisCodegenFunctions::TestFuncs(const std::vector<Function*> &Funcs,
61 std::cout << "Testing functions: ";
62 PrintFunctionList(Funcs);
65 // Clone the module for the two halves of the program we want.
66 Module *SafeModule = CloneModule(BD.getProgram());
68 // The JIT must extract the 'main' function.
69 std::vector<Function*> RealFuncs(Funcs);
70 if (BD.isExecutingJIT()) {
71 if (Function *F = BD.Program->getMainFunction())
72 RealFuncs.push_back(F);
74 Module *TestModule = SplitFunctionsOutOfModule(SafeModule, RealFuncs);
76 // This is only applicable if we are debugging the JIT:
77 // Find all external functions in the Safe modules that are actually used
78 // (called or taken address of), and make them call the JIT wrapper instead
79 if (BD.isExecutingJIT()) {
80 // Must delete `main' from Safe module if it has it
81 Function *safeMain = SafeModule->getNamedFunction("main");
82 assert(safeMain && "`main' function not found in safe module!");
83 DeleteFunctionBody(safeMain);
85 // Add an external function "getPointerToNamedFunction" that JIT provides
86 // Prototype: void *getPointerToNamedFunction(const char* Name)
87 std::vector<const Type*> Params;
88 Params.push_back(PointerType::get(Type::SByteTy)); // std::string&
89 FunctionType *resolverTy = FunctionType::get(PointerType::get(Type::VoidTy),
90 Params, false /* isVarArg */);
91 Function *resolverFunc = new Function(resolverTy,
92 GlobalValue::ExternalLinkage,
93 "getPointerToNamedFunction",
96 // Use the function we just added to get addresses of functions we need
97 // Iterate over the global declarations in the Safe module
98 for (Module::iterator F=SafeModule->begin(),E=SafeModule->end(); F!=E; ++F){
99 if (F->isExternal() && !F->use_empty() && &*F != resolverFunc &&
100 F->getIntrinsicID() == 0 /* ignore intrinsics */ &&
101 // Don't forward functions which are external in the test module too.
102 !TestModule->getNamedFunction(F->getName())->isExternal()) {
103 // If it has a non-zero use list,
104 // 1. Add a string constant with its name to the global file
105 // The correct type is `const [ NUM x sbyte ]' where NUM is length of
107 const std::string &Name = F->getName();
108 GlobalVariable *funcName =
109 new GlobalVariable(ArrayType::get(Type::SByteTy, Name.length()+1),
110 true /* isConstant */,
111 GlobalValue::InternalLinkage,
112 ConstantArray::get(Name),
116 // 2. Use `GetElementPtr *funcName, 0, 0' to convert the string to an
117 // sbyte* so it matches the signature of the resolver function.
118 std::vector<Constant*> GEPargs(2, Constant::getNullValue(Type::LongTy));
120 // 3. Replace all uses of `func' with calls to resolver by:
121 // (a) Iterating through the list of uses of this function
122 // (b) Insert a cast instruction in front of each use
123 // (c) Replace use of old call with new call
125 // GetElementPtr *funcName, ulong 0, ulong 0
127 ConstantExpr::getGetElementPtr(ConstantPointerRef::get(funcName),
129 std::vector<Value*> ResolverArgs;
130 ResolverArgs.push_back(GEP);
132 // Insert code at the beginning of the function
133 while (!F->use_empty())
134 if (Instruction *Inst = dyn_cast<Instruction>(F->use_back())) {
135 // call resolver(GetElementPtr...)
136 CallInst *resolve = new CallInst(resolverFunc, ResolverArgs,
138 // cast the result from the resolver to correctly-typed function
139 CastInst *castResolver =
140 new CastInst(resolve, PointerType::get(F->getFunctionType()),
141 "resolverCast", Inst);
142 // actually use the resolved function
143 Inst->replaceUsesOfWith(F, castResolver);
145 // FIXME: need to take care of cases where a function is used by
146 // something other than an instruction; e.g., global variable
147 // initializers and constant expressions.
148 std::cerr << "UNSUPPORTED: Non-instruction is using an external "
149 << "function, " << F->getName() << "().\n";
156 if (verifyModule(*SafeModule) || verifyModule(*TestModule)) {
157 std::cerr << "Bugpoint has a bug, an corrupted a module!!\n";
161 DEBUG(std::cerr << "Safe module:\n";
162 typedef Module::iterator MI;
163 typedef Module::giterator MGI;
165 for (MI I = SafeModule->begin(), E = SafeModule->end(); I != E; ++I)
166 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
167 for (MGI I = SafeModule->gbegin(), E = SafeModule->gend(); I!=E; ++I)
168 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
170 std::cerr << "Test module:\n";
171 for (MI I = TestModule->begin(), E = TestModule->end(); I != E; ++I)
172 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
173 for (MGI I=TestModule->gbegin(),E = TestModule->gend(); I!= E; ++I)
174 if (!I->isExternal()) std::cerr << "\t" << I->getName() << "\n";
177 // Write out the bytecode to be sent to CBE
178 std::string SafeModuleBC = getUniqueFilename("bugpoint.safe.bc");
180 if (BD.writeProgramToFile(SafeModuleBC, SafeModule)) {
181 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
185 // Remove all functions from the Test module EXCEPT for the ones specified in
186 // Funcs. We know which ones these are because they are non-external in
187 // ToOptimize, but external in ToNotOptimize.
189 for (Module::iterator I = TestModule->begin(), E = TestModule->end();I!=E;++I)
190 if (!I->isExternal()) {
191 Function *TNOF = SafeModule->getFunction(I->getName(),
192 I->getFunctionType());
193 assert(TNOF && "Function doesn't exist in ToNotOptimize module??");
194 if (!TNOF->isExternal())
195 DeleteFunctionBody(I);
198 std::string TestModuleBC = getUniqueFilename("bugpoint.test.bc");
199 if (verifyModule(*TestModule)) {
200 std::cerr << "Bytecode file corrupted!\n";
204 // Clean up the modules, removing extra cruft that we don't need anymore...
205 SafeModule = BD.performFinalCleanups(SafeModule);
206 TestModule = BD.performFinalCleanups(TestModule);
208 if (BD.writeProgramToFile(TestModuleBC, TestModule)) {
209 std::cerr << "Error writing bytecode to `" << SafeModuleBC << "'\nExiting.";
213 // Make a shared library
214 std::string SharedObject = BD.compileSharedObject(SafeModuleBC);
219 // Run the code generator on the `Test' code, loading the shared library.
220 // The function returns whether or not the new output differs from reference.
221 int Result = BD.diffProgram(TestModuleBC, SharedObject, false);
224 std::cerr << ": still failing!\n";
226 std::cerr << ": didn't fail.\n";
229 std::cout << "You can reproduce the problem with the command line: \n";
230 if (BD.isExecutingJIT()) {
231 std::cout << " lli -load " << SharedObject << " " << TestModuleBC;
233 std::cout << " llc " << TestModuleBC << " -o " << TestModuleBC << ".s\n";
234 std::cout << " gcc " << SharedObject << " " << TestModuleBC
235 << ".s -o " << TestModuleBC << ".exe -Wl,-R.\n";
236 std::cout << " " << TestModuleBC << ".exe";
238 for (unsigned i=0, e = InputArgv.size(); i != e; ++i)
239 std::cout << " " << InputArgv[i];
241 std::cout << "The shared object was created with:\n llc -march=c "
242 << SafeModuleBC << " -o temporary.c\n"
243 << " gcc -xc temporary.c -O2 -o " << SharedObject
244 #if defined(sparc) || defined(__sparc__) || defined(__sparcv9)
245 << " -G" // Compile a shared library, `-G' for Sparc
247 << " -shared" // `-shared' for Linux/X86, maybe others
249 << " -fno-strict-aliasing\n";
251 removeFile(TestModuleBC);
252 removeFile(SafeModuleBC);
253 removeFile(SharedObject);
259 struct Disambiguator {
260 std::set<std::string> SymbolNames;
261 std::set<GlobalValue*> Symbols;
262 uint64_t uniqueCounter;
265 Disambiguator() : uniqueCounter(0), externalOnly(true) {}
266 void setExternalOnly(bool value) { externalOnly = value; }
267 void add(GlobalValue &V) {
268 // If we're only processing externals and this isn't external, bail
269 if (externalOnly && !V.isExternal()) return;
270 // If we're already processed this symbol, don't add it again
271 if (Symbols.count(&V) != 0) return;
272 // Ignore intrinsic functions
273 if (Function *F = dyn_cast<Function>(&V))
274 if (F->getIntrinsicID() != 0)
277 std::string SymName = V.getName();
279 // Use the Mangler facility to make symbol names that will be valid in
281 SymName = Mangler::makeNameProper(SymName);
284 if (SymbolNames.count(SymName) == 0) {
285 DEBUG(std::cerr << "Disambiguator: adding " << SymName
286 << ", no conflicts.\n");
287 SymbolNames.insert(SymName);
289 // Mangle name before adding
292 newName = SymName + "_" + utostr(uniqueCounter);
293 if (SymbolNames.count(newName) == 0) break;
294 else ++uniqueCounter;
296 //while (SymbolNames.count(V->getName()+utostr(uniqueCounter++))==0);
297 DEBUG(std::cerr << "Disambiguator: conflict: " << SymName
298 << ", adding: " << newName << "\n");
300 SymbolNames.insert(newName);
307 static void DisambiguateGlobalSymbols(Module *M) {
308 // First, try not to cause collisions by minimizing chances of renaming an
309 // already-external symbol, so take in external globals and functions as-is.
311 DEBUG(std::cerr << "Disambiguating globals (external-only)\n");
312 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) D.add(*I);
313 DEBUG(std::cerr << "Disambiguating functions (external-only)\n");
314 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) D.add(*I);
316 // Now just rename functions and globals as necessary, keeping what's already
317 // in the set unique.
318 D.setExternalOnly(false);
319 DEBUG(std::cerr << "Disambiguating globals\n");
320 for (Module::giterator I = M->gbegin(), E = M->gend(); I != E; ++I) D.add(*I);
321 DEBUG(std::cerr << "Disambiguating globals\n");
322 for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) D.add(*I);
326 bool BugDriver::debugCodeGenerator() {
327 if ((void*)cbe == (void*)Interpreter) {
328 std::string Result = executeProgramWithCBE("bugpoint.cbe.out");
329 std::cout << "\n*** The C backend cannot match the reference diff, but it "
330 << "is used as the 'known good'\n code generator, so I can't"
331 << " debug it. Perhaps you have a front-end problem?\n As a"
332 << " sanity check, I left the result of executing the program "
333 << "with the C backend\n in this file for you: '"
338 // See if we can pin down which functions are being miscompiled...
339 // First, build a list of all of the non-external functions in the program.
340 std::vector<Function*> MisCodegenFunctions;
341 for (Module::iterator I = Program->begin(), E = Program->end(); I != E; ++I)
342 if (!I->isExternal())
343 MisCodegenFunctions.push_back(I);
345 // If we are executing the JIT, we *must* keep the function `main' in the
346 // module that is passed in, and not the shared library. However, we still
347 // want to be able to debug the `main' function alone. Thus, we create a new
348 // function `main' which just calls the old one.
349 if (isExecutingJIT()) {
350 // Get the `main' function
351 Function *oldMain = Program->getNamedFunction("main");
352 assert(oldMain && "`main' function not found in program!");
354 oldMain->setName("llvm_old_main");
355 // Create a NEW `main' function with same type
356 Function *newMain = new Function(oldMain->getFunctionType(),
357 GlobalValue::ExternalLinkage,
359 // Call the old main function and return its result
360 BasicBlock *BB = new BasicBlock("entry", newMain);
361 std::vector<Value*> args;
362 for (Function::aiterator I = newMain->abegin(), E = newMain->aend(),
363 OI = oldMain->abegin(); I != E; ++I, ++OI) {
364 I->setName(OI->getName()); // Copy argument names from oldMain
367 CallInst *call = new CallInst(oldMain, args);
368 BB->getInstList().push_back(call);
370 // if the type of old function wasn't void, return value of call
371 if (oldMain->getReturnType() != Type::VoidTy) {
372 new ReturnInst(call, BB);
374 new ReturnInst(0, BB);
378 DisambiguateGlobalSymbols(Program);
380 // Do the reduction...
381 if (!ReduceMisCodegenFunctions(*this).reduceList(MisCodegenFunctions)) {
382 std::cerr << "*** Execution matches reference output! "
383 << "bugpoint can't help you with your problem!\n";
387 std::cout << "\n*** The following functions are being miscompiled: ";
388 PrintFunctionList(MisCodegenFunctions);
391 // Output a bunch of bytecode files for the user...
392 ReduceMisCodegenFunctions(*this).TestFuncs(MisCodegenFunctions, true);