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"
32 static const Type *PtrSByte = 0; // 'sbyte*' type
34 // ConvertCallTo - Convert a call to a varargs function with no arg types
35 // specified to a concrete nonvarargs method.
37 static void ConvertCallTo(CallInst *CI, Method *Dest) {
38 const MethodType::ParamTypes &ParamTys =
39 Dest->getMethodType()->getParamTypes();
40 BasicBlock *BB = CI->getParent();
42 // Get an iterator to where we want to insert cast instructions if the
43 // argument types don't agree.
45 BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
46 assert(BBI != BB->end() && "CallInst not in parent block?");
48 assert(CI->getNumOperands()-1 == ParamTys.size()&&
49 "Method calls resolved funny somehow, incompatible number of args");
51 vector<Value*> Params;
53 // Convert all of the call arguments over... inserting cast instructions if
54 // the types are not compatible.
55 for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
56 Value *V = CI->getOperand(i);
58 if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
59 Instruction *Cast = new CastInst(V, ParamTys[i-1]);
60 BBI = BB->getInstList().insert(BBI, Cast)+1;
67 // Replace the old call instruction with a new call instruction that calls
70 ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
74 // PatchUpMethodReferences - Go over the methods that are in the module and
75 // look for methods that have the same name. More often than not, there will
78 // void "foo"(int, int)
79 // because of the way things are declared in C. If this is the case, patch
82 bool CleanupGCCOutput::PatchUpMethodReferences(Module *M) {
83 SymbolTable *ST = M->getSymbolTable();
84 if (!ST) return false;
86 std::map<string, vector<Method*> > Methods;
88 // Loop over the entries in the symbol table. If an entry is a method pointer,
89 // then add it to the Methods map. We do a two pass algorithm here to avoid
90 // problems with iterators getting invalidated if we did a one pass scheme.
92 for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
93 if (const PointerType *PT = dyn_cast<PointerType>(I->first))
94 if (isa<MethodType>(PT->getElementType())) {
95 SymbolTable::VarMap &Plane = I->second;
96 for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
98 const string &Name = PI->first;
99 Method *M = cast<Method>(PI->second);
100 Methods[Name].push_back(M);
104 bool Changed = false;
106 // Now we have a list of all methods with a particular name. If there is more
107 // than one entry in a list, merge the methods together.
109 for (std::map<string, vector<Method*> >::iterator I = Methods.begin(),
110 E = Methods.end(); I != E; ++I) {
111 vector<Method*> &Methods = I->second;
112 Method *Implementation = 0; // Find the implementation
113 Method *Concrete = 0;
114 for (unsigned i = 0; i < Methods.size(); ) {
115 if (!Methods[i]->isExternal()) { // Found an implementation
116 assert(Implementation == 0 && "Multiple definitions of the same"
117 " method. Case not handled yet!");
118 Implementation = Methods[i];
120 // Ignore methods that are never used so they don't cause spurious
121 // warnings... here we will actually DCE the function so that it isn't
124 if (Methods[i]->use_size() == 0) {
125 M->getMethodList().remove(Methods[i]);
127 Methods.erase(Methods.begin()+i);
133 if (Methods[i] && (!Methods[i]->getMethodType()->isVarArg() ||
134 Methods[i]->getMethodType()->getParamTypes().size())) {
135 if (Concrete) { // Found two different methods types. Can't choose
139 Concrete = Methods[i];
144 if (Methods.size() > 1) { // Found a multiply defined method.
145 // We should find exactly one non-vararg method definition, which is
146 // probably the implementation. Change all of the method definitions
147 // and uses to use it instead.
150 cerr << "Warning: Found methods types that are not compatible:\n";
151 for (unsigned i = 0; i < Methods.size(); ++i) {
152 cerr << "\t" << Methods[i]->getType()->getDescription() << " %"
153 << Methods[i]->getName() << "\n";
155 cerr << " No linkage of methods named '" << Methods[0]->getName()
158 for (unsigned i = 0; i < Methods.size(); ++i)
159 if (Methods[i] != Concrete) {
160 Method *Old = Methods[i];
161 assert(Old->getReturnType() == Concrete->getReturnType() &&
162 "Differing return types not handled yet!");
163 assert(Old->getMethodType()->getParamTypes().size() == 0 &&
164 "Cannot handle varargs fn's with specified element types!");
166 // Attempt to convert all of the uses of the old method to the
167 // concrete form of the method. If there is a use of the method
168 // that we don't understand here we punt to avoid making a bad
171 // At this point, we know that the return values are the same for
172 // our two functions and that the Old method has no varargs methods
173 // specified. In otherwords it's just <retty> (...)
175 for (unsigned i = 0; i < Old->use_size(); ) {
176 User *U = *(Old->use_begin()+i);
177 if (CastInst *CI = dyn_cast<CastInst>(U)) {
178 // Convert casts directly
179 assert(CI->getOperand(0) == Old);
180 CI->setOperand(0, Concrete);
182 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
183 // Can only fix up calls TO the argument, not args passed in.
184 if (CI->getCalledValue() == Old) {
185 ConvertCallTo(CI, Concrete);
188 cerr << "Couldn't cleanup this function call, must be an"
189 << " argument or something!" << CI;
193 cerr << "Cannot convert use of method: " << U << "\n";
206 // ShouldNukSymtabEntry - Return true if this module level symbol table entry
207 // should be eliminated.
209 static inline bool ShouldNukeSymtabEntry(const std::pair<string, Value*> &E) {
210 // Nuke all names for primitive types!
211 if (cast<Type>(E.second)->isPrimitiveType()) return true;
213 // Nuke all pointers to primitive types as well...
214 if (const PointerType *PT = dyn_cast<PointerType>(E.second))
215 if (PT->getElementType()->isPrimitiveType()) return true;
217 // The only types that could contain .'s in the program are things generated
218 // by GCC itself, including "complex.float" and friends. Nuke them too.
219 if (E.first.find('.') != string::npos) return true;
224 // doInitialization - For this pass, it removes global symbol table
225 // entries for primitive types. These are never used for linking in GCC and
226 // they make the output uglier to look at, so we nuke them.
228 bool CleanupGCCOutput::doInitialization(Module *M) {
229 bool Changed = false;
232 PtrSByte = PointerType::get(Type::SByteTy);
234 if (M->hasSymbolTable()) {
235 SymbolTable *ST = M->getSymbolTable();
237 // Go over the methods that are in the module and look for methods that have
238 // the same name. More often than not, there will be things like:
239 // void "foo"(...) and void "foo"(int, int) because of the way things are
240 // declared in C. If this is the case, patch things up.
242 Changed |= PatchUpMethodReferences(M);
244 // Check the symbol table for superfluous type entries...
246 // Grab the 'type' plane of the module symbol...
247 SymbolTable::iterator STI = ST->find(Type::TypeTy);
248 if (STI != ST->end()) {
249 // Loop over all entries in the type plane...
250 SymbolTable::VarMap &Plane = STI->second;
251 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
252 if (ShouldNukeSymtabEntry(*PI)) { // Should we remove this entry?
253 #if MAP_IS_NOT_BRAINDEAD
254 PI = Plane.erase(PI); // STD C++ Map should support this!
256 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
270 // FixCastsAndPHIs - The LLVM GCC has a tendancy to intermix Cast instructions
271 // in with the PHI nodes. These cast instructions are potentially there for two
272 // different reasons:
274 // 1. The cast could be for an early PHI, and be accidentally inserted before
275 // another PHI node. In this case, the PHI node should be moved to the end
276 // of the PHI nodes in the basic block. We know that it is this case if
277 // the source for the cast is a PHI node in this basic block.
279 // 2. If not #1, the cast must be a source argument for one of the PHI nodes
280 // in the current basic block. If this is the case, the cast should be
281 // lifted into the basic block for the appropriate predecessor.
283 static inline bool FixCastsAndPHIs(BasicBlock *BB) {
284 bool Changed = false;
286 BasicBlock::iterator InsertPos = BB->begin();
288 // Find the end of the interesting instructions...
289 while (isa<PHINode>(*InsertPos) || isa<CastInst>(*InsertPos)) ++InsertPos;
291 // Back the InsertPos up to right after the last PHI node.
292 while (InsertPos != BB->begin() && isa<CastInst>(*(InsertPos-1))) --InsertPos;
294 // No PHI nodes, quick exit.
295 if (InsertPos == BB->begin()) return false;
297 // Loop over all casts trapped between the PHI's...
298 BasicBlock::iterator I = BB->begin();
299 while (I != InsertPos) {
300 if (CastInst *CI = dyn_cast<CastInst>(*I)) { // Fix all cast instructions
301 Value *Src = CI->getOperand(0);
303 // Move the cast instruction to the current insert position...
304 --InsertPos; // New position for cast to go...
305 std::swap(*InsertPos, *I); // Cast goes down, PHI goes up
307 if (isa<PHINode>(Src) && // Handle case #1
308 cast<PHINode>(Src)->getParent() == BB) {
309 // We're done for case #1
310 } else { // Handle case #2
311 // In case #2, we have to do a few things:
312 // 1. Remove the cast from the current basic block.
313 // 2. Identify the PHI node that the cast is for.
314 // 3. Find out which predecessor the value is for.
315 // 4. Move the cast to the end of the basic block that it SHOULD be
318 // Remove the cast instruction from the basic block. The remove only
319 // invalidates iterators in the basic block that are AFTER the removed
320 // element. Because we just moved the CastInst to the InsertPos, no
321 // iterators get invalidated.
323 BB->getInstList().remove(InsertPos);
325 // Find the PHI node. Since this cast was generated specifically for a
326 // PHI node, there can only be a single PHI node using it.
328 assert(CI->use_size() == 1 && "Exactly one PHI node should use cast!");
329 PHINode *PN = cast<PHINode>(*CI->use_begin());
331 // Find out which operand of the PHI it is...
333 for (i = 0; i < PN->getNumIncomingValues(); ++i)
334 if (PN->getIncomingValue(i) == CI)
336 assert(i != PN->getNumIncomingValues() && "PHI doesn't use cast!");
338 // Get the predecessor the value is for...
339 BasicBlock *Pred = PN->getIncomingBlock(i);
341 // Reinsert the cast right before the terminator in Pred.
342 Pred->getInstList().insert(Pred->end()-1, CI);
353 // RefactorPredecessor - When we find out that a basic block is a repeated
354 // predecessor in a PHI node, we have to refactor the method until there is at
355 // most a single instance of a basic block in any predecessor list.
357 static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
358 Method *M = BB->getParent();
359 assert(find(pred_begin(BB), pred_end(BB), Pred) != pred_end(BB) &&
360 "Pred is not a predecessor of BB!");
362 // Create a new basic block, adding it to the end of the method.
363 BasicBlock *NewBB = new BasicBlock("", M);
365 // Add an unconditional branch to BB to the new block.
366 NewBB->getInstList().push_back(new BranchInst(BB));
368 // Get the terminator that causes a branch to BB from Pred.
369 TerminatorInst *TI = Pred->getTerminator();
371 // Find the first use of BB in the terminator...
372 User::op_iterator OI = find(TI->op_begin(), TI->op_end(), BB);
373 assert(OI != TI->op_end() && "Pred does not branch to BB!!!");
375 // Change the use of BB to point to the new stub basic block
378 // Now we need to loop through all of the PHI nodes in BB and convert their
379 // first incoming value for Pred to reference the new basic block instead.
381 for (BasicBlock::iterator I = BB->begin();
382 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
383 int BBIdx = PN->getBasicBlockIndex(Pred);
384 assert(BBIdx != -1 && "PHI node doesn't have an entry for Pred!");
386 // The value that used to look like it came from Pred now comes from NewBB
387 PN->setIncomingBlock((unsigned)BBIdx, NewBB);
392 // CheckIncomingValueFor - Make sure that the specified PHI node has an entry
393 // for the provided basic block. If it doesn't, add one and return true.
395 static inline void CheckIncomingValueFor(PHINode *PN, BasicBlock *BB) {
396 if (PN->getBasicBlockIndex(BB) != -1) return; // Already has value
399 const Type *Ty = PN->getType();
401 if (const PointerType *PT = dyn_cast<PointerType>(Ty))
402 NewVal = ConstantPointerNull::get(PT);
403 else if (Ty == Type::BoolTy)
404 NewVal = ConstantBool::True;
405 else if (Ty == Type::FloatTy || Ty == Type::DoubleTy)
406 NewVal = ConstantFP::get(Ty, 42);
407 else if (Ty->isIntegral())
408 NewVal = ConstantInt::get(Ty, 42);
410 assert(NewVal && "Unknown PHI node type!");
411 PN->addIncoming(NewVal, BB);
414 // fixLocalProblems - Loop through the method and fix problems with the PHI
415 // nodes in the current method. The two problems that are handled are:
417 // 1. PHI nodes with multiple entries for the same predecessor. GCC sometimes
418 // generates code that looks like this:
420 // bb7: br bool %cond1004, label %bb8, label %bb8
421 // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
423 // which is completely illegal LLVM code. To compensate for this, we insert
424 // an extra basic block, and convert the code to look like this:
426 // bb7: br bool %cond1004, label %bbX, label %bb8
428 // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
431 // 2. PHI nodes with fewer arguments than predecessors.
432 // These can be generated by GCC if a variable is uninitalized over a path
433 // in the CFG. We fix this by adding an entry for the missing predecessors
434 // that is initialized to either 42 for a numeric/FP value, or null if it's
435 // a pointer value. This problem can be generated by code that looks like
443 static bool fixLocalProblems(Method *M) {
444 bool Changed = false;
445 // Don't use iterators because invalidation gets messy...
446 for (unsigned MI = 0; MI < M->size(); ++MI) {
447 BasicBlock *BB = M->getBasicBlocks()[MI];
449 Changed |= FixCastsAndPHIs(BB);
451 if (isa<PHINode>(BB->front())) {
452 const vector<BasicBlock*> Preds(pred_begin(BB), pred_end(BB));
454 // Handle Problem #1. Sort the list of predecessors so that it is easy to
455 // decide whether or not duplicate predecessors exist.
456 vector<BasicBlock*> SortedPreds(Preds);
457 sort(SortedPreds.begin(), SortedPreds.end());
459 // Loop over the predecessors, looking for adjacent BB's that are equal.
460 BasicBlock *LastOne = 0;
461 for (unsigned i = 0; i < Preds.size(); ++i) {
462 if (SortedPreds[i] == LastOne) { // Found a duplicate.
463 RefactorPredecessor(BB, SortedPreds[i]);
466 LastOne = SortedPreds[i];
469 // Loop over all of the PHI nodes in the current BB. These PHI nodes are
470 // guaranteed to be at the beginning of the basic block.
472 for (BasicBlock::iterator I = BB->begin();
473 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
475 // Handle problem #2.
476 if (PN->getNumIncomingValues() != Preds.size()) {
477 assert(PN->getNumIncomingValues() <= Preds.size() &&
478 "Can't handle extra arguments to PHI nodes!");
479 for (unsigned i = 0; i < Preds.size(); ++i)
480 CheckIncomingValueFor(PN, Preds[i]);
492 // doPerMethodWork - This method simplifies the specified method hopefully.
494 bool CleanupGCCOutput::runOnMethod(Method *M) {
495 return fixLocalProblems(M);
498 bool CleanupGCCOutput::doFinalization(Module *M) {
499 bool Changed = false;
502 if (M->hasSymbolTable()) {
503 SymbolTable *ST = M->getSymbolTable();
504 const std::set<const Type *> &UsedTypes =
505 getAnalysis<FindUsedTypes>().getTypes();
507 // Check the symbol table for superfluous type entries that aren't used in
510 // Grab the 'type' plane of the module symbol...
511 SymbolTable::iterator STI = ST->find(Type::TypeTy);
512 if (STI != ST->end()) {
513 // Loop over all entries in the type plane...
514 SymbolTable::VarMap &Plane = STI->second;
515 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
516 if (!UsedTypes.count(cast<Type>(PI->second))) {
517 #if MAP_IS_NOT_BRAINDEAD
518 PI = Plane.erase(PI); // STD C++ Map should support this!
520 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
521 PI = Plane.begin(); // N^2 algorithms are fun. :(
532 // getAnalysisUsageInfo - This function needs the results of the
533 // FindUsedTypes and FindUnsafePointerTypes analysis passes...
535 void CleanupGCCOutput::getAnalysisUsageInfo(Pass::AnalysisSet &Required,
536 Pass::AnalysisSet &Destroyed,
537 Pass::AnalysisSet &Provided) {
538 // FIXME: Invalidates the CFG
539 Required.push_back(FindUsedTypes::ID);