1 //===- CleanupGCCOutput.cpp - Cleanup GCC Output ----------------------------=//
3 // This pass is used to cleanup the output of GCC. GCC's output is
4 // unneccessarily gross for a couple of reasons. This pass does the following
5 // things to try to clean it up:
7 // * Eliminate names for GCC types that we know can't be needed by the user.
8 // - Eliminate names for types that are unused in the entire translation unit
9 // but only if they do not name a structure type!
10 // - Replace calls to 'sbyte *%malloc(uint)' and 'void %free(sbyte *)' with
11 // malloc and free instructions.
13 // Note: This code produces dead declarations, it is a good idea to run DCE
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/CleanupGCCOutput.h"
19 #include "llvm/SymbolTable.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/iOther.h"
22 #include "llvm/iMemory.h"
23 #include "llvm/iTerminators.h"
27 static const Type *PtrArrSByte = 0; // '[sbyte]*' type
28 static const Type *PtrSByte = 0; // 'sbyte*' type
31 // ReplaceInstWithValue - Replace all uses of an instruction (specified by BI)
32 // with a value, then remove and delete the original instruction.
34 static void ReplaceInstWithValue(BasicBlock::InstListType &BIL,
35 BasicBlock::iterator &BI, Value *V) {
37 // Replaces all of the uses of the instruction with uses of the value
38 I->replaceAllUsesWith(V);
40 // Remove the unneccesary instruction now...
43 // Make sure to propogate a name if there is one already...
44 if (I->hasName() && !V->hasName())
45 V->setName(I->getName(), BIL.getParent()->getSymbolTable());
47 // Remove the dead instruction now...
52 // ReplaceInstWithInst - Replace the instruction specified by BI with the
53 // instruction specified by I. The original instruction is deleted and BI is
54 // updated to point to the new instruction.
56 static void ReplaceInstWithInst(BasicBlock::InstListType &BIL,
57 BasicBlock::iterator &BI, Instruction *I) {
58 assert(I->getParent() == 0 &&
59 "ReplaceInstWithInst: Instruction already inserted into basic block!");
61 // Insert the new instruction into the basic block...
62 BI = BIL.insert(BI, I)+1;
64 // Replace all uses of the old instruction, and delete it.
65 ReplaceInstWithValue(BIL, BI, I);
67 // Reexamine the instruction just inserted next time around the cleanup pass
74 // ConvertCallTo - Convert a call to a varargs function with no arg types
75 // specified to a concrete nonvarargs method.
77 static void ConvertCallTo(CallInst *CI, Method *Dest) {
78 const MethodType::ParamTypes &ParamTys =
79 Dest->getMethodType()->getParamTypes();
80 BasicBlock *BB = CI->getParent();
82 // Get an iterator to where we want to insert cast instructions if the
83 // argument types don't agree.
85 BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
86 assert(BBI != BB->end() && "CallInst not in parent block?");
88 assert(CI->getNumOperands()-1 == ParamTys.size()&&
89 "Method calls resolved funny somehow, incompatible number of args");
91 vector<Value*> Params;
93 // Convert all of the call arguments over... inserting cast instructions if
94 // the types are not compatible.
95 for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
96 Value *V = CI->getOperand(i);
98 if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
99 Instruction *Cast = new CastInst(V, ParamTys[i-1]);
100 BBI = BB->getInstList().insert(BBI, Cast)+1;
107 // Replace the old call instruction with a new call instruction that calls
110 ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
114 // PatchUpMethodReferences - Go over the methods that are in the module and
115 // look for methods that have the same name. More often than not, there will
118 // void "foo"(int, int)
119 // because of the way things are declared in C. If this is the case, patch
122 bool CleanupGCCOutput::PatchUpMethodReferences(Module *M) {
123 SymbolTable *ST = M->getSymbolTable();
124 if (!ST) return false;
126 map<string, vector<Method*> > Methods;
128 // Loop over the entries in the symbol table. If an entry is a method pointer,
129 // then add it to the Methods map. We do a two pass algorithm here to avoid
130 // problems with iterators getting invalidated if we did a one pass scheme.
132 for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
133 if (const PointerType *PT = dyn_cast<PointerType>(I->first))
134 if (const MethodType *MT = dyn_cast<MethodType>(PT->getValueType())) {
135 SymbolTable::VarMap &Plane = I->second;
136 for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
138 const string &Name = PI->first;
139 Method *M = cast<Method>(PI->second);
140 Methods[Name].push_back(M);
144 bool Changed = false;
146 // Now we have a list of all methods with a particular name. If there is more
147 // than one entry in a list, merge the methods together.
149 for (map<string, vector<Method*> >::iterator I = Methods.begin(),
150 E = Methods.end(); I != E; ++I) {
151 vector<Method*> &Methods = I->second;
152 if (Methods.size() > 1) { // Found a multiply defined method.
153 Method *Implementation = 0; // Find the implementation
154 Method *Concrete = 0;
155 for (unsigned i = 0; i < Methods.size(); ++i) {
156 // TODO: Ignore methods that are never USED! DCE them.
157 // Remove their name. this should fix a majority of problems here.
159 if (!Methods[i]->isExternal()) { // Found an implementation
160 assert(Implementation == 0 && "Multiple definitions of the same"
161 " method. Case not handled yet!");
162 Implementation = Methods[i];
165 if (!Methods[i]->getMethodType()->isVarArg() ||
166 Methods[i]->getMethodType()->getParamTypes().size()) {
167 if (Concrete) { // Found two different methods types. Can't choose
171 Concrete = Methods[i];
175 // We should find exactly one non-vararg method definition, which is
176 // probably the implementation. Change all of the method definitions
177 // and uses to use it instead.
180 cerr << "Warning: Found methods types that are not compatible:\n";
181 for (unsigned i = 0; i < Methods.size(); ++i) {
182 cerr << "\t" << Methods[i]->getType()->getDescription() << " %"
183 << Methods[i]->getName() << endl;
185 cerr << " No linkage of methods named '" << Methods[0]->getName()
188 for (unsigned i = 0; i < Methods.size(); ++i)
189 if (Methods[i] != Concrete) {
190 Method *Old = Methods[i];
191 assert(Old->getReturnType() == Concrete->getReturnType() &&
192 "Differing return types not handled yet!");
193 assert(Old->getMethodType()->getParamTypes().size() == 0 &&
194 "Cannot handle varargs fn's with specified element types!");
196 // Attempt to convert all of the uses of the old method to the
197 // concrete form of the method. If there is a use of the method
198 // that we don't understand here we punt to avoid making a bad
201 // At this point, we know that the return values are the same for
202 // our two functions and that the Old method has no varargs methods
203 // specified. In otherwords it's just <retty> (...)
205 for (unsigned i = 0; i < Old->use_size(); ) {
206 User *U = *(Old->use_begin()+i);
207 if (CastInst *CI = dyn_cast<CastInst>(U)) {
208 // Convert casts directly
209 assert(CI->getOperand(0) == Old);
210 CI->setOperand(0, Concrete);
212 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
213 // Can only fix up calls TO the argument, not args passed in.
214 if (CI->getCalledValue() == Old) {
215 ConvertCallTo(CI, Concrete);
218 cerr << "Couldn't cleanup this function call, must be an"
219 << " argument or something!" << CI;
223 cerr << "Cannot convert use of method: " << U << endl;
236 // ShouldNukSymtabEntry - Return true if this module level symbol table entry
237 // should be eliminated.
239 static inline bool ShouldNukeSymtabEntry(const pair<string, Value*> &E) {
240 // Nuke all names for primitive types!
241 if (cast<Type>(E.second)->isPrimitiveType()) return true;
243 // The only types that could contain .'s in the program are things generated
244 // by GCC itself, including "complex.float" and friends. Nuke them too.
245 if (E.first.find('.') != string::npos) return true;
250 // doPassInitialization - For this pass, it removes global symbol table
251 // entries for primitive types. These are never used for linking in GCC and
252 // they make the output uglier to look at, so we nuke them.
254 bool CleanupGCCOutput::doPassInitialization(Module *M) {
255 bool Changed = false;
257 if (PtrArrSByte == 0) {
258 PtrArrSByte = PointerType::get(ArrayType::get(Type::SByteTy));
259 PtrSByte = PointerType::get(Type::SByteTy);
262 if (M->hasSymbolTable()) {
263 SymbolTable *ST = M->getSymbolTable();
265 // Go over the methods that are in the module and look for methods that have
266 // the same name. More often than not, there will be things like:
267 // void "foo"(...) and void "foo"(int, int) because of the way things are
268 // declared in C. If this is the case, patch things up.
270 Changed |= PatchUpMethodReferences(M);
273 // If the module has a symbol table, they might be referring to the malloc
274 // and free functions. If this is the case, grab the method pointers that
275 // the module is using.
277 // Lookup %malloc and %free in the symbol table, for later use. If they
278 // don't exist, or are not external, we do not worry about converting calls
279 // to that function into the appropriate instruction.
281 const PointerType *MallocType = // Get the type for malloc
282 PointerType::get(MethodType::get(PointerType::get(Type::SByteTy),
283 vector<const Type*>(1, Type::UIntTy), false));
284 Malloc = cast_or_null<Method>(ST->lookup(MallocType, "malloc"));
285 if (Malloc && !Malloc->isExternal())
286 Malloc = 0; // Don't mess with locally defined versions of the fn
288 const PointerType *FreeType = // Get the type for free
289 PointerType::get(MethodType::get(Type::VoidTy,
290 vector<const Type*>(1, PointerType::get(Type::SByteTy)), false));
291 Free = cast_or_null<Method>(ST->lookup(FreeType, "free"));
292 if (Free && !Free->isExternal())
293 Free = 0; // Don't mess with locally defined versions of the fn
296 // Check the symbol table for superfluous type entries...
298 // Grab the 'type' plane of the module symbol...
299 SymbolTable::iterator STI = ST->find(Type::TypeTy);
300 if (STI != ST->end()) {
301 // Loop over all entries in the type plane...
302 SymbolTable::VarMap &Plane = STI->second;
303 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
304 if (ShouldNukeSymtabEntry(*PI)) { // Should we remove this entry?
305 #if MAP_IS_NOT_BRAINDEAD
306 PI = Plane.erase(PI); // STD C++ Map should support this!
308 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
322 // doOneCleanupPass - Do one pass over the input method, fixing stuff up.
324 bool CleanupGCCOutput::doOneCleanupPass(Method *M) {
325 bool Changed = false;
326 for (Method::iterator MI = M->begin(), ME = M->end(); MI != ME; ++MI) {
327 BasicBlock *BB = *MI;
328 BasicBlock::InstListType &BIL = BB->getInstList();
330 for (BasicBlock::iterator BI = BB->begin(); BI != BB->end();) {
331 Instruction *I = *BI;
333 if (CallInst *CI = dyn_cast<CallInst>(I)) {
334 if (CI->getCalledValue() == Malloc) { // Replace call to malloc?
335 MallocInst *MallocI = new MallocInst(PtrArrSByte, CI->getOperand(1),
338 BI = BIL.insert(BI, MallocI)+1;
339 ReplaceInstWithInst(BIL, BI, new CastInst(MallocI, PtrSByte));
341 continue; // Skip the ++BI
342 } else if (CI->getCalledValue() == Free) { // Replace call to free?
343 ReplaceInstWithInst(BIL, BI, new FreeInst(CI->getOperand(1)));
345 continue; // Skip the ++BI
358 // RefactorPredecessor - When we find out that a basic block is a repeated
359 // predecessor in a PHI node, we have to refactor the method until there is at
360 // most a single instance of a basic block in any predecessor list.
362 static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
363 Method *M = BB->getParent();
364 assert(find(BB->pred_begin(), BB->pred_end(), Pred) != BB->pred_end() &&
365 "Pred is not a predecessor of BB!");
367 // Create a new basic block, adding it to the end of the method.
368 BasicBlock *NewBB = new BasicBlock("", M);
370 // Add an unconditional branch to BB to the new block.
371 NewBB->getInstList().push_back(new BranchInst(BB));
373 // Get the terminator that causes a branch to BB from Pred.
374 TerminatorInst *TI = Pred->getTerminator();
376 // Find the first use of BB in the terminator...
377 User::op_iterator OI = find(TI->op_begin(), TI->op_end(), BB);
378 assert(OI != TI->op_end() && "Pred does not branch to BB!!!");
380 // Change the use of BB to point to the new stub basic block
383 // Now we need to loop through all of the PHI nodes in BB and convert their
384 // first incoming value for Pred to reference the new basic block instead.
386 for (BasicBlock::iterator I = BB->begin();
387 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
388 int BBIdx = PN->getBasicBlockIndex(Pred);
389 assert(BBIdx != -1 && "PHI node doesn't have an entry for Pred!");
391 // The value that used to look like it came from Pred now comes from NewBB
392 PN->setIncomingBlock((unsigned)BBIdx, NewBB);
397 // CheckIncomingValueFor - Make sure that the specified PHI node has an entry
398 // for the provided basic block. If it doesn't, add one and return true.
400 static inline void CheckIncomingValueFor(PHINode *PN, BasicBlock *BB) {
401 if (PN->getBasicBlockIndex(BB) != -1) return; // Already has value
404 const Type *Ty = PN->getType();
406 if (const PointerType *PT = dyn_cast<PointerType>(Ty))
407 NewVal = ConstPoolPointerNull::get(PT);
408 else if (Ty == Type::BoolTy)
409 NewVal = ConstPoolBool::True;
410 else if (Ty == Type::FloatTy || Ty == Type::DoubleTy)
411 NewVal = ConstPoolFP::get(Ty, 42);
412 else if (Ty->isIntegral())
413 NewVal = ConstPoolInt::get(Ty, 42);
415 assert(NewVal && "Unknown PHI node type!");
416 PN->addIncoming(NewVal, BB);
419 // fixLocalProblems - Loop through the method and fix problems with the PHI
420 // nodes in the current method. The two problems that are handled are:
422 // 1. PHI nodes with multiple entries for the same predecessor. GCC sometimes
423 // generates code that looks like this:
425 // bb7: br bool %cond1004, label %bb8, label %bb8
426 // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
428 // which is completely illegal LLVM code. To compensate for this, we insert
429 // an extra basic block, and convert the code to look like this:
431 // bb7: br bool %cond1004, label %bbX, label %bb8
433 // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
436 // 2. PHI nodes with fewer arguments than predecessors.
437 // These can be generated by GCC if a variable is uninitalized over a path
438 // in the CFG. We fix this by adding an entry for the missing predecessors
439 // that is initialized to either 42 for a numeric/FP value, or null if it's
440 // a pointer value. This problem can be generated by code that looks like
448 static bool fixLocalProblems(Method *M) {
449 bool Changed = false;
450 // Don't use iterators because invalidation gets messy...
451 for (unsigned MI = 0; MI < M->size(); ++MI) {
452 BasicBlock *BB = M->getBasicBlocks()[MI];
454 if (isa<PHINode>(BB->front())) {
455 const vector<BasicBlock*> Preds(BB->pred_begin(), BB->pred_end());
457 // Handle Problem #1. Sort the list of predecessors so that it is easy to
458 // decide whether or not duplicate predecessors exist.
459 vector<BasicBlock*> SortedPreds(Preds);
460 sort(SortedPreds.begin(), SortedPreds.end());
462 // Loop over the predecessors, looking for adjacent BB's that are equal.
463 BasicBlock *LastOne = 0;
464 for (unsigned i = 0; i < Preds.size(); ++i) {
465 if (SortedPreds[i] == LastOne) { // Found a duplicate.
466 RefactorPredecessor(BB, SortedPreds[i]);
469 LastOne = SortedPreds[i];
472 // Loop over all of the PHI nodes in the current BB. These PHI nodes are
473 // guaranteed to be at the beginning of the basic block.
475 for (BasicBlock::iterator I = BB->begin();
476 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
478 // Handle problem #2.
479 if (PN->getNumIncomingValues() != Preds.size()) {
480 assert(PN->getNumIncomingValues() <= Preds.size() &&
481 "Can't handle extra arguments to PHI nodes!");
482 for (unsigned i = 0; i < Preds.size(); ++i)
483 CheckIncomingValueFor(PN, Preds[i]);
495 // doPerMethodWork - This method simplifies the specified method hopefully.
497 bool CleanupGCCOutput::doPerMethodWork(Method *M) {
498 bool Changed = fixLocalProblems(M);
499 while (doOneCleanupPass(M)) Changed = true;