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"
31 static const Type *PtrSByte = 0; // 'sbyte*' type
33 // ConvertCallTo - Convert a call to a varargs function with no arg types
34 // specified to a concrete nonvarargs method.
36 static void ConvertCallTo(CallInst *CI, Method *Dest) {
37 const MethodType::ParamTypes &ParamTys =
38 Dest->getMethodType()->getParamTypes();
39 BasicBlock *BB = CI->getParent();
41 // Get an iterator to where we want to insert cast instructions if the
42 // argument types don't agree.
44 BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
45 assert(BBI != BB->end() && "CallInst not in parent block?");
47 assert(CI->getNumOperands()-1 == ParamTys.size()&&
48 "Method calls resolved funny somehow, incompatible number of args");
50 vector<Value*> Params;
52 // Convert all of the call arguments over... inserting cast instructions if
53 // the types are not compatible.
54 for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
55 Value *V = CI->getOperand(i);
57 if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
58 Instruction *Cast = new CastInst(V, ParamTys[i-1]);
59 BBI = BB->getInstList().insert(BBI, Cast)+1;
66 // Replace the old call instruction with a new call instruction that calls
69 ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
73 // PatchUpMethodReferences - Go over the methods that are in the module and
74 // look for methods that have the same name. More often than not, there will
77 // void "foo"(int, int)
78 // because of the way things are declared in C. If this is the case, patch
81 bool CleanupGCCOutput::PatchUpMethodReferences(Module *M) {
82 SymbolTable *ST = M->getSymbolTable();
83 if (!ST) return false;
85 std::map<string, vector<Method*> > Methods;
87 // Loop over the entries in the symbol table. If an entry is a method pointer,
88 // then add it to the Methods map. We do a two pass algorithm here to avoid
89 // problems with iterators getting invalidated if we did a one pass scheme.
91 for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
92 if (const PointerType *PT = dyn_cast<PointerType>(I->first))
93 if (isa<MethodType>(PT->getElementType())) {
94 SymbolTable::VarMap &Plane = I->second;
95 for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
97 const string &Name = PI->first;
98 Method *M = cast<Method>(PI->second);
99 Methods[Name].push_back(M);
103 bool Changed = false;
105 // Now we have a list of all methods with a particular name. If there is more
106 // than one entry in a list, merge the methods together.
108 for (std::map<string, vector<Method*> >::iterator I = Methods.begin(),
109 E = Methods.end(); I != E; ++I) {
110 vector<Method*> &Methods = I->second;
111 Method *Implementation = 0; // Find the implementation
112 Method *Concrete = 0;
113 for (unsigned i = 0; i < Methods.size(); ) {
114 if (!Methods[i]->isExternal()) { // Found an implementation
115 assert(Implementation == 0 && "Multiple definitions of the same"
116 " method. Case not handled yet!");
117 Implementation = Methods[i];
119 // Ignore methods that are never used so they don't cause spurious
120 // warnings... here we will actually DCE the function so that it isn't
123 if (Methods[i]->use_size() == 0) {
124 M->getMethodList().remove(Methods[i]);
126 Methods.erase(Methods.begin()+i);
132 if (Methods[i] && (!Methods[i]->getMethodType()->isVarArg() ||
133 Methods[i]->getMethodType()->getParamTypes().size())) {
134 if (Concrete) { // Found two different methods types. Can't choose
138 Concrete = Methods[i];
143 if (Methods.size() > 1) { // Found a multiply defined method.
144 // We should find exactly one non-vararg method definition, which is
145 // probably the implementation. Change all of the method definitions
146 // and uses to use it instead.
149 cerr << "Warning: Found methods types that are not compatible:\n";
150 for (unsigned i = 0; i < Methods.size(); ++i) {
151 cerr << "\t" << Methods[i]->getType()->getDescription() << " %"
152 << Methods[i]->getName() << "\n";
154 cerr << " No linkage of methods named '" << Methods[0]->getName()
157 for (unsigned i = 0; i < Methods.size(); ++i)
158 if (Methods[i] != Concrete) {
159 Method *Old = Methods[i];
160 assert(Old->getReturnType() == Concrete->getReturnType() &&
161 "Differing return types not handled yet!");
162 assert(Old->getMethodType()->getParamTypes().size() == 0 &&
163 "Cannot handle varargs fn's with specified element types!");
165 // Attempt to convert all of the uses of the old method to the
166 // concrete form of the method. If there is a use of the method
167 // that we don't understand here we punt to avoid making a bad
170 // At this point, we know that the return values are the same for
171 // our two functions and that the Old method has no varargs methods
172 // specified. In otherwords it's just <retty> (...)
174 for (unsigned i = 0; i < Old->use_size(); ) {
175 User *U = *(Old->use_begin()+i);
176 if (CastInst *CI = dyn_cast<CastInst>(U)) {
177 // Convert casts directly
178 assert(CI->getOperand(0) == Old);
179 CI->setOperand(0, Concrete);
181 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
182 // Can only fix up calls TO the argument, not args passed in.
183 if (CI->getCalledValue() == Old) {
184 ConvertCallTo(CI, Concrete);
187 cerr << "Couldn't cleanup this function call, must be an"
188 << " argument or something!" << CI;
192 cerr << "Cannot convert use of method: " << U << "\n";
205 // ShouldNukSymtabEntry - Return true if this module level symbol table entry
206 // should be eliminated.
208 static inline bool ShouldNukeSymtabEntry(const std::pair<string, Value*> &E) {
209 // Nuke all names for primitive types!
210 if (cast<Type>(E.second)->isPrimitiveType()) return true;
212 // Nuke all pointers to primitive types as well...
213 if (const PointerType *PT = dyn_cast<PointerType>(E.second))
214 if (PT->getElementType()->isPrimitiveType()) return true;
216 // The only types that could contain .'s in the program are things generated
217 // by GCC itself, including "complex.float" and friends. Nuke them too.
218 if (E.first.find('.') != string::npos) return true;
223 // doInitialization - For this pass, it removes global symbol table
224 // entries for primitive types. These are never used for linking in GCC and
225 // they make the output uglier to look at, so we nuke them.
227 bool CleanupGCCOutput::doInitialization(Module *M) {
228 bool Changed = false;
231 PtrSByte = PointerType::get(Type::SByteTy);
233 if (M->hasSymbolTable()) {
234 SymbolTable *ST = M->getSymbolTable();
236 // Go over the methods that are in the module and look for methods that have
237 // the same name. More often than not, there will be things like:
238 // void "foo"(...) and void "foo"(int, int) because of the way things are
239 // declared in C. If this is the case, patch things up.
241 Changed |= PatchUpMethodReferences(M);
243 // Check the symbol table for superfluous type entries...
245 // Grab the 'type' plane of the module symbol...
246 SymbolTable::iterator STI = ST->find(Type::TypeTy);
247 if (STI != ST->end()) {
248 // Loop over all entries in the type plane...
249 SymbolTable::VarMap &Plane = STI->second;
250 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
251 if (ShouldNukeSymtabEntry(*PI)) { // Should we remove this entry?
252 #if MAP_IS_NOT_BRAINDEAD
253 PI = Plane.erase(PI); // STD C++ Map should support this!
255 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
269 // FixCastsAndPHIs - The LLVM GCC has a tendancy to intermix Cast instructions
270 // in with the PHI nodes. These cast instructions are potentially there for two
271 // different reasons:
273 // 1. The cast could be for an early PHI, and be accidentally inserted before
274 // another PHI node. In this case, the PHI node should be moved to the end
275 // of the PHI nodes in the basic block. We know that it is this case if
276 // the source for the cast is a PHI node in this basic block.
278 // 2. If not #1, the cast must be a source argument for one of the PHI nodes
279 // in the current basic block. If this is the case, the cast should be
280 // lifted into the basic block for the appropriate predecessor.
282 static inline bool FixCastsAndPHIs(BasicBlock *BB) {
283 bool Changed = false;
285 BasicBlock::iterator InsertPos = BB->begin();
287 // Find the end of the interesting instructions...
288 while (isa<PHINode>(*InsertPos) || isa<CastInst>(*InsertPos)) ++InsertPos;
290 // Back the InsertPos up to right after the last PHI node.
291 while (InsertPos != BB->begin() && isa<CastInst>(*(InsertPos-1))) --InsertPos;
293 // No PHI nodes, quick exit.
294 if (InsertPos == BB->begin()) return false;
296 // Loop over all casts trapped between the PHI's...
297 BasicBlock::iterator I = BB->begin();
298 while (I != InsertPos) {
299 if (CastInst *CI = dyn_cast<CastInst>(*I)) { // Fix all cast instructions
300 Value *Src = CI->getOperand(0);
302 // Move the cast instruction to the current insert position...
303 --InsertPos; // New position for cast to go...
304 std::swap(*InsertPos, *I); // Cast goes down, PHI goes up
306 if (isa<PHINode>(Src) && // Handle case #1
307 cast<PHINode>(Src)->getParent() == BB) {
308 // We're done for case #1
309 } else { // Handle case #2
310 // In case #2, we have to do a few things:
311 // 1. Remove the cast from the current basic block.
312 // 2. Identify the PHI node that the cast is for.
313 // 3. Find out which predecessor the value is for.
314 // 4. Move the cast to the end of the basic block that it SHOULD be
317 // Remove the cast instruction from the basic block. The remove only
318 // invalidates iterators in the basic block that are AFTER the removed
319 // element. Because we just moved the CastInst to the InsertPos, no
320 // iterators get invalidated.
322 BB->getInstList().remove(InsertPos);
324 // Find the PHI node. Since this cast was generated specifically for a
325 // PHI node, there can only be a single PHI node using it.
327 assert(CI->use_size() == 1 && "Exactly one PHI node should use cast!");
328 PHINode *PN = cast<PHINode>(*CI->use_begin());
330 // Find out which operand of the PHI it is...
332 for (i = 0; i < PN->getNumIncomingValues(); ++i)
333 if (PN->getIncomingValue(i) == CI)
335 assert(i != PN->getNumIncomingValues() && "PHI doesn't use cast!");
337 // Get the predecessor the value is for...
338 BasicBlock *Pred = PN->getIncomingBlock(i);
340 // Reinsert the cast right before the terminator in Pred.
341 Pred->getInstList().insert(Pred->end()-1, CI);
352 // RefactorPredecessor - When we find out that a basic block is a repeated
353 // predecessor in a PHI node, we have to refactor the method until there is at
354 // most a single instance of a basic block in any predecessor list.
356 static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
357 Method *M = BB->getParent();
358 assert(find(BB->pred_begin(), BB->pred_end(), Pred) != BB->pred_end() &&
359 "Pred is not a predecessor of BB!");
361 // Create a new basic block, adding it to the end of the method.
362 BasicBlock *NewBB = new BasicBlock("", M);
364 // Add an unconditional branch to BB to the new block.
365 NewBB->getInstList().push_back(new BranchInst(BB));
367 // Get the terminator that causes a branch to BB from Pred.
368 TerminatorInst *TI = Pred->getTerminator();
370 // Find the first use of BB in the terminator...
371 User::op_iterator OI = find(TI->op_begin(), TI->op_end(), BB);
372 assert(OI != TI->op_end() && "Pred does not branch to BB!!!");
374 // Change the use of BB to point to the new stub basic block
377 // Now we need to loop through all of the PHI nodes in BB and convert their
378 // first incoming value for Pred to reference the new basic block instead.
380 for (BasicBlock::iterator I = BB->begin();
381 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
382 int BBIdx = PN->getBasicBlockIndex(Pred);
383 assert(BBIdx != -1 && "PHI node doesn't have an entry for Pred!");
385 // The value that used to look like it came from Pred now comes from NewBB
386 PN->setIncomingBlock((unsigned)BBIdx, NewBB);
391 // CheckIncomingValueFor - Make sure that the specified PHI node has an entry
392 // for the provided basic block. If it doesn't, add one and return true.
394 static inline void CheckIncomingValueFor(PHINode *PN, BasicBlock *BB) {
395 if (PN->getBasicBlockIndex(BB) != -1) return; // Already has value
398 const Type *Ty = PN->getType();
400 if (const PointerType *PT = dyn_cast<PointerType>(Ty))
401 NewVal = ConstantPointerNull::get(PT);
402 else if (Ty == Type::BoolTy)
403 NewVal = ConstantBool::True;
404 else if (Ty == Type::FloatTy || Ty == Type::DoubleTy)
405 NewVal = ConstantFP::get(Ty, 42);
406 else if (Ty->isIntegral())
407 NewVal = ConstantInt::get(Ty, 42);
409 assert(NewVal && "Unknown PHI node type!");
410 PN->addIncoming(NewVal, BB);
413 // fixLocalProblems - Loop through the method and fix problems with the PHI
414 // nodes in the current method. The two problems that are handled are:
416 // 1. PHI nodes with multiple entries for the same predecessor. GCC sometimes
417 // generates code that looks like this:
419 // bb7: br bool %cond1004, label %bb8, label %bb8
420 // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
422 // which is completely illegal LLVM code. To compensate for this, we insert
423 // an extra basic block, and convert the code to look like this:
425 // bb7: br bool %cond1004, label %bbX, label %bb8
427 // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
430 // 2. PHI nodes with fewer arguments than predecessors.
431 // These can be generated by GCC if a variable is uninitalized over a path
432 // in the CFG. We fix this by adding an entry for the missing predecessors
433 // that is initialized to either 42 for a numeric/FP value, or null if it's
434 // a pointer value. This problem can be generated by code that looks like
442 static bool fixLocalProblems(Method *M) {
443 bool Changed = false;
444 // Don't use iterators because invalidation gets messy...
445 for (unsigned MI = 0; MI < M->size(); ++MI) {
446 BasicBlock *BB = M->getBasicBlocks()[MI];
448 Changed |= FixCastsAndPHIs(BB);
450 if (isa<PHINode>(BB->front())) {
451 const vector<BasicBlock*> Preds(BB->pred_begin(), BB->pred_end());
453 // Handle Problem #1. Sort the list of predecessors so that it is easy to
454 // decide whether or not duplicate predecessors exist.
455 vector<BasicBlock*> SortedPreds(Preds);
456 sort(SortedPreds.begin(), SortedPreds.end());
458 // Loop over the predecessors, looking for adjacent BB's that are equal.
459 BasicBlock *LastOne = 0;
460 for (unsigned i = 0; i < Preds.size(); ++i) {
461 if (SortedPreds[i] == LastOne) { // Found a duplicate.
462 RefactorPredecessor(BB, SortedPreds[i]);
465 LastOne = SortedPreds[i];
468 // Loop over all of the PHI nodes in the current BB. These PHI nodes are
469 // guaranteed to be at the beginning of the basic block.
471 for (BasicBlock::iterator I = BB->begin();
472 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
474 // Handle problem #2.
475 if (PN->getNumIncomingValues() != Preds.size()) {
476 assert(PN->getNumIncomingValues() <= Preds.size() &&
477 "Can't handle extra arguments to PHI nodes!");
478 for (unsigned i = 0; i < Preds.size(); ++i)
479 CheckIncomingValueFor(PN, Preds[i]);
491 // doPerMethodWork - This method simplifies the specified method hopefully.
493 bool CleanupGCCOutput::runOnMethod(Method *M) {
494 return fixLocalProblems(M);
497 bool CleanupGCCOutput::doFinalization(Module *M) {
498 bool Changed = false;
501 if (M->hasSymbolTable()) {
502 SymbolTable *ST = M->getSymbolTable();
503 const std::set<const Type *> &UsedTypes =
504 getAnalysis<FindUsedTypes>().getTypes();
506 // Check the symbol table for superfluous type entries that aren't used in
509 // Grab the 'type' plane of the module symbol...
510 SymbolTable::iterator STI = ST->find(Type::TypeTy);
511 if (STI != ST->end()) {
512 // Loop over all entries in the type plane...
513 SymbolTable::VarMap &Plane = STI->second;
514 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
515 if (!UsedTypes.count(cast<Type>(PI->second))) {
516 #if MAP_IS_NOT_BRAINDEAD
517 PI = Plane.erase(PI); // STD C++ Map should support this!
519 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
520 PI = Plane.begin(); // N^2 algorithms are fun. :(
531 // getAnalysisUsageInfo - This function needs the results of the
532 // FindUsedTypes and FindUnsafePointerTypes analysis passes...
534 void CleanupGCCOutput::getAnalysisUsageInfo(Pass::AnalysisSet &Required,
535 Pass::AnalysisSet &Destroyed,
536 Pass::AnalysisSet &Provided) {
537 // FIXME: Invalidates the CFG
538 Required.push_back(FindUsedTypes::ID);