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
10 // Note: This code produces dead declarations, it is a good idea to run DCE
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/CleanupGCCOutput.h"
16 #include "llvm/Analysis/FindUsedTypes.h"
17 #include "TransformInternals.h"
18 #include "llvm/Module.h"
19 #include "llvm/SymbolTable.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/iPHINode.h"
22 #include "llvm/iMemory.h"
23 #include "llvm/iTerminators.h"
24 #include "llvm/iOther.h"
25 #include "llvm/Support/CFG.h"
26 #include "llvm/Pass.h"
33 static const Type *PtrSByte = 0; // 'sbyte*' type
36 struct CleanupGCCOutput : public MethodPass {
37 // doPassInitialization - For this pass, it removes global symbol table
38 // entries for primitive types. These are never used for linking in GCC and
39 // they make the output uglier to look at, so we nuke them.
41 // Also, initialize instance variables.
43 bool doInitialization(Module *M);
45 // doPerMethodWork - This method simplifies the specified method hopefully.
47 bool runOnMethod(Method *M);
49 // doPassFinalization - Strip out type names that are unused by the program
50 bool doFinalization(Module *M);
52 // getAnalysisUsageInfo - This function needs FindUsedTypes to do its job...
54 virtual void getAnalysisUsageInfo(Pass::AnalysisSet &Required,
55 Pass::AnalysisSet &Destroyed,
56 Pass::AnalysisSet &Provided) {
57 // FIXME: Invalidates the CFG
58 Required.push_back(FindUsedTypes::ID);
65 // ConvertCallTo - Convert a call to a varargs function with no arg types
66 // specified to a concrete nonvarargs method.
68 static void ConvertCallTo(CallInst *CI, Method *Dest) {
69 const MethodType::ParamTypes &ParamTys =
70 Dest->getMethodType()->getParamTypes();
71 BasicBlock *BB = CI->getParent();
73 // Get an iterator to where we want to insert cast instructions if the
74 // argument types don't agree.
76 BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
77 assert(BBI != BB->end() && "CallInst not in parent block?");
79 assert(CI->getNumOperands()-1 == ParamTys.size()&&
80 "Method calls resolved funny somehow, incompatible number of args");
82 vector<Value*> Params;
84 // Convert all of the call arguments over... inserting cast instructions if
85 // the types are not compatible.
86 for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
87 Value *V = CI->getOperand(i);
89 if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
90 Instruction *Cast = new CastInst(V, ParamTys[i-1]);
91 BBI = BB->getInstList().insert(BBI, Cast)+1;
98 // Replace the old call instruction with a new call instruction that calls
101 ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
105 // PatchUpMethodReferences - Go over the methods that are in the module and
106 // look for methods that have the same name. More often than not, there will
109 // void "foo"(int, int)
110 // because of the way things are declared in C. If this is the case, patch
113 static bool PatchUpMethodReferences(Module *M) {
114 SymbolTable *ST = M->getSymbolTable();
115 if (!ST) return false;
117 std::map<string, vector<Method*> > Methods;
119 // Loop over the entries in the symbol table. If an entry is a method pointer,
120 // then add it to the Methods map. We do a two pass algorithm here to avoid
121 // problems with iterators getting invalidated if we did a one pass scheme.
123 for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
124 if (const PointerType *PT = dyn_cast<PointerType>(I->first))
125 if (isa<MethodType>(PT->getElementType())) {
126 SymbolTable::VarMap &Plane = I->second;
127 for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
129 const string &Name = PI->first;
130 Method *M = cast<Method>(PI->second);
131 Methods[Name].push_back(M);
135 bool Changed = false;
137 // Now we have a list of all methods with a particular name. If there is more
138 // than one entry in a list, merge the methods together.
140 for (std::map<string, vector<Method*> >::iterator I = Methods.begin(),
141 E = Methods.end(); I != E; ++I) {
142 vector<Method*> &Methods = I->second;
143 Method *Implementation = 0; // Find the implementation
144 Method *Concrete = 0;
145 for (unsigned i = 0; i < Methods.size(); ) {
146 if (!Methods[i]->isExternal()) { // Found an implementation
147 assert(Implementation == 0 && "Multiple definitions of the same"
148 " method. Case not handled yet!");
149 Implementation = Methods[i];
151 // Ignore methods that are never used so they don't cause spurious
152 // warnings... here we will actually DCE the function so that it isn't
155 if (Methods[i]->use_size() == 0) {
156 M->getMethodList().remove(Methods[i]);
158 Methods.erase(Methods.begin()+i);
164 if (Methods[i] && (!Methods[i]->getMethodType()->isVarArg())) {
165 if (Concrete) { // Found two different methods types. Can't choose
169 Concrete = Methods[i];
174 if (Methods.size() > 1) { // Found a multiply defined method.
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() << "\n";
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 const MethodType *OldMT = Old->getMethodType();
192 const MethodType *ConcreteMT = Concrete->getMethodType();
195 assert(Old->getReturnType() == Concrete->getReturnType() &&
196 "Differing return types not handled yet!");
197 assert(OldMT->getParamTypes().size() <=
198 ConcreteMT->getParamTypes().size() &&
199 "Concrete type must have more specified parameters!");
201 // Check to make sure that if there are specified types, that they
204 for (unsigned i = 0; i < OldMT->getParamTypes().size(); ++i)
205 if (OldMT->getParamTypes()[i] != ConcreteMT->getParamTypes()[i]) {
206 cerr << "Parameter types conflict for" << OldMT
207 << " and " << ConcreteMT;
210 if (Broken) break; // Can't process this one!
213 // Attempt to convert all of the uses of the old method to the
214 // concrete form of the method. If there is a use of the method
215 // that we don't understand here we punt to avoid making a bad
218 // At this point, we know that the return values are the same for
219 // our two functions and that the Old method has no varargs methods
220 // specified. In otherwords it's just <retty> (...)
222 for (unsigned i = 0; i < Old->use_size(); ) {
223 User *U = *(Old->use_begin()+i);
224 if (CastInst *CI = dyn_cast<CastInst>(U)) {
225 // Convert casts directly
226 assert(CI->getOperand(0) == Old);
227 CI->setOperand(0, Concrete);
229 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
230 // Can only fix up calls TO the argument, not args passed in.
231 if (CI->getCalledValue() == Old) {
232 ConvertCallTo(CI, Concrete);
235 cerr << "Couldn't cleanup this function call, must be an"
236 << " argument or something!" << CI;
240 cerr << "Cannot convert use of method: " << U << "\n";
253 // ShouldNukSymtabEntry - Return true if this module level symbol table entry
254 // should be eliminated.
256 static inline bool ShouldNukeSymtabEntry(const std::pair<string, Value*> &E) {
257 // Nuke all names for primitive types!
258 if (cast<Type>(E.second)->isPrimitiveType()) return true;
260 // Nuke all pointers to primitive types as well...
261 if (const PointerType *PT = dyn_cast<PointerType>(E.second))
262 if (PT->getElementType()->isPrimitiveType()) return true;
264 // The only types that could contain .'s in the program are things generated
265 // by GCC itself, including "complex.float" and friends. Nuke them too.
266 if (E.first.find('.') != string::npos) return true;
271 // doInitialization - For this pass, it removes global symbol table
272 // entries for primitive types. These are never used for linking in GCC and
273 // they make the output uglier to look at, so we nuke them.
275 bool CleanupGCCOutput::doInitialization(Module *M) {
276 bool Changed = false;
279 PtrSByte = PointerType::get(Type::SByteTy);
281 if (M->hasSymbolTable()) {
282 SymbolTable *ST = M->getSymbolTable();
284 // Go over the methods that are in the module and look for methods that have
285 // the same name. More often than not, there will be things like:
286 // void "foo"(...) and void "foo"(int, int) because of the way things are
287 // declared in C. If this is the case, patch things up.
289 Changed |= PatchUpMethodReferences(M);
291 // Check the symbol table for superfluous type entries...
293 // Grab the 'type' plane of the module symbol...
294 SymbolTable::iterator STI = ST->find(Type::TypeTy);
295 if (STI != ST->end()) {
296 // Loop over all entries in the type plane...
297 SymbolTable::VarMap &Plane = STI->second;
298 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
299 if (ShouldNukeSymtabEntry(*PI)) { // Should we remove this entry?
300 #if MAP_IS_NOT_BRAINDEAD
301 PI = Plane.erase(PI); // STD C++ Map should support this!
303 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
317 // FixCastsAndPHIs - The LLVM GCC has a tendancy to intermix Cast instructions
318 // in with the PHI nodes. These cast instructions are potentially there for two
319 // different reasons:
321 // 1. The cast could be for an early PHI, and be accidentally inserted before
322 // another PHI node. In this case, the PHI node should be moved to the end
323 // of the PHI nodes in the basic block. We know that it is this case if
324 // the source for the cast is a PHI node in this basic block.
326 // 2. If not #1, the cast must be a source argument for one of the PHI nodes
327 // in the current basic block. If this is the case, the cast should be
328 // lifted into the basic block for the appropriate predecessor.
330 static inline bool FixCastsAndPHIs(BasicBlock *BB) {
331 bool Changed = false;
333 BasicBlock::iterator InsertPos = BB->begin();
335 // Find the end of the interesting instructions...
336 while (isa<PHINode>(*InsertPos) || isa<CastInst>(*InsertPos)) ++InsertPos;
338 // Back the InsertPos up to right after the last PHI node.
339 while (InsertPos != BB->begin() && isa<CastInst>(*(InsertPos-1))) --InsertPos;
341 // No PHI nodes, quick exit.
342 if (InsertPos == BB->begin()) return false;
344 // Loop over all casts trapped between the PHI's...
345 BasicBlock::iterator I = BB->begin();
346 while (I != InsertPos) {
347 if (CastInst *CI = dyn_cast<CastInst>(*I)) { // Fix all cast instructions
348 Value *Src = CI->getOperand(0);
350 // Move the cast instruction to the current insert position...
351 --InsertPos; // New position for cast to go...
352 std::swap(*InsertPos, *I); // Cast goes down, PHI goes up
354 if (isa<PHINode>(Src) && // Handle case #1
355 cast<PHINode>(Src)->getParent() == BB) {
356 // We're done for case #1
357 } else { // Handle case #2
358 // In case #2, we have to do a few things:
359 // 1. Remove the cast from the current basic block.
360 // 2. Identify the PHI node that the cast is for.
361 // 3. Find out which predecessor the value is for.
362 // 4. Move the cast to the end of the basic block that it SHOULD be
365 // Remove the cast instruction from the basic block. The remove only
366 // invalidates iterators in the basic block that are AFTER the removed
367 // element. Because we just moved the CastInst to the InsertPos, no
368 // iterators get invalidated.
370 BB->getInstList().remove(InsertPos);
372 // Find the PHI node. Since this cast was generated specifically for a
373 // PHI node, there can only be a single PHI node using it.
375 assert(CI->use_size() == 1 && "Exactly one PHI node should use cast!");
376 PHINode *PN = cast<PHINode>(*CI->use_begin());
378 // Find out which operand of the PHI it is...
380 for (i = 0; i < PN->getNumIncomingValues(); ++i)
381 if (PN->getIncomingValue(i) == CI)
383 assert(i != PN->getNumIncomingValues() && "PHI doesn't use cast!");
385 // Get the predecessor the value is for...
386 BasicBlock *Pred = PN->getIncomingBlock(i);
388 // Reinsert the cast right before the terminator in Pred.
389 Pred->getInstList().insert(Pred->end()-1, CI);
399 // RefactorPredecessor - When we find out that a basic block is a repeated
400 // predecessor in a PHI node, we have to refactor the method until there is at
401 // most a single instance of a basic block in any predecessor list.
403 static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
404 Method *M = BB->getParent();
405 assert(find(pred_begin(BB), pred_end(BB), Pred) != pred_end(BB) &&
406 "Pred is not a predecessor of BB!");
408 // Create a new basic block, adding it to the end of the method.
409 BasicBlock *NewBB = new BasicBlock("", M);
411 // Add an unconditional branch to BB to the new block.
412 NewBB->getInstList().push_back(new BranchInst(BB));
414 // Get the terminator that causes a branch to BB from Pred.
415 TerminatorInst *TI = Pred->getTerminator();
417 // Find the first use of BB in the terminator...
418 User::op_iterator OI = find(TI->op_begin(), TI->op_end(), BB);
419 assert(OI != TI->op_end() && "Pred does not branch to BB!!!");
421 // Change the use of BB to point to the new stub basic block
424 // Now we need to loop through all of the PHI nodes in BB and convert their
425 // first incoming value for Pred to reference the new basic block instead.
427 for (BasicBlock::iterator I = BB->begin();
428 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
429 int BBIdx = PN->getBasicBlockIndex(Pred);
430 assert(BBIdx != -1 && "PHI node doesn't have an entry for Pred!");
432 // The value that used to look like it came from Pred now comes from NewBB
433 PN->setIncomingBlock((unsigned)BBIdx, NewBB);
438 // fixLocalProblems - Loop through the method and fix problems with the PHI
439 // nodes in the current method. The problem is that PHI nodes might exist with
440 // multiple entries for the same predecessor. GCC sometimes generates code
441 // that looks like this:
443 // bb7: br bool %cond1004, label %bb8, label %bb8
444 // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
446 // which is completely illegal LLVM code. To compensate for this, we insert
447 // an extra basic block, and convert the code to look like this:
449 // bb7: br bool %cond1004, label %bbX, label %bb8
451 // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
454 static bool fixLocalProblems(Method *M) {
455 bool Changed = false;
456 // Don't use iterators because invalidation gets messy...
457 for (unsigned MI = 0; MI < M->size(); ++MI) {
458 BasicBlock *BB = M->getBasicBlocks()[MI];
460 Changed |= FixCastsAndPHIs(BB);
462 if (isa<PHINode>(BB->front())) {
463 const vector<BasicBlock*> Preds(pred_begin(BB), pred_end(BB));
465 // Handle the problem. Sort the list of predecessors so that it is easy
466 // to decide whether or not duplicate predecessors exist.
467 vector<BasicBlock*> SortedPreds(Preds);
468 sort(SortedPreds.begin(), SortedPreds.end());
470 // Loop over the predecessors, looking for adjacent BB's that are equal.
471 BasicBlock *LastOne = 0;
472 for (unsigned i = 0; i < Preds.size(); ++i) {
473 if (SortedPreds[i] == LastOne) { // Found a duplicate.
474 RefactorPredecessor(BB, SortedPreds[i]);
477 LastOne = SortedPreds[i];
487 // doPerMethodWork - This method simplifies the specified method hopefully.
489 bool CleanupGCCOutput::runOnMethod(Method *M) {
490 return fixLocalProblems(M);
493 bool CleanupGCCOutput::doFinalization(Module *M) {
494 bool Changed = false;
497 if (M->hasSymbolTable()) {
498 SymbolTable *ST = M->getSymbolTable();
499 const std::set<const Type *> &UsedTypes =
500 getAnalysis<FindUsedTypes>().getTypes();
502 // Check the symbol table for superfluous type entries that aren't used in
505 // Grab the 'type' plane of the module symbol...
506 SymbolTable::iterator STI = ST->find(Type::TypeTy);
507 if (STI != ST->end()) {
508 // Loop over all entries in the type plane...
509 SymbolTable::VarMap &Plane = STI->second;
510 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
511 if (!UsedTypes.count(cast<Type>(PI->second))) {
512 #if MAP_IS_NOT_BRAINDEAD
513 PI = Plane.erase(PI); // STD C++ Map should support this!
515 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
516 PI = Plane.begin(); // N^2 algorithms are fun. :(
527 Pass *createCleanupGCCOutputPass() {
528 return new CleanupGCCOutput();