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 // * Replace calls to 'sbyte *%malloc(uint)' and 'void %free(sbyte *)' with
10 // malloc and free instructions.
12 // Note: This code produces dead declarations, it is a good idea to run DCE
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Transforms/CleanupGCCOutput.h"
18 #include "TransformInternals.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"
27 static const Type *PtrSByte = 0; // 'sbyte*' type
29 // ConvertCallTo - Convert a call to a varargs function with no arg types
30 // specified to a concrete nonvarargs method.
32 static void ConvertCallTo(CallInst *CI, Method *Dest) {
33 const MethodType::ParamTypes &ParamTys =
34 Dest->getMethodType()->getParamTypes();
35 BasicBlock *BB = CI->getParent();
37 // Get an iterator to where we want to insert cast instructions if the
38 // argument types don't agree.
40 BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
41 assert(BBI != BB->end() && "CallInst not in parent block?");
43 assert(CI->getNumOperands()-1 == ParamTys.size()&&
44 "Method calls resolved funny somehow, incompatible number of args");
46 vector<Value*> Params;
48 // Convert all of the call arguments over... inserting cast instructions if
49 // the types are not compatible.
50 for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
51 Value *V = CI->getOperand(i);
53 if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
54 Instruction *Cast = new CastInst(V, ParamTys[i-1]);
55 BBI = BB->getInstList().insert(BBI, Cast)+1;
62 // Replace the old call instruction with a new call instruction that calls
65 ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
69 // PatchUpMethodReferences - Go over the methods that are in the module and
70 // look for methods that have the same name. More often than not, there will
73 // void "foo"(int, int)
74 // because of the way things are declared in C. If this is the case, patch
77 bool CleanupGCCOutput::PatchUpMethodReferences(Module *M) {
78 SymbolTable *ST = M->getSymbolTable();
79 if (!ST) return false;
81 map<string, vector<Method*> > Methods;
83 // Loop over the entries in the symbol table. If an entry is a method pointer,
84 // then add it to the Methods map. We do a two pass algorithm here to avoid
85 // problems with iterators getting invalidated if we did a one pass scheme.
87 for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
88 if (const PointerType *PT = dyn_cast<PointerType>(I->first))
89 if (const MethodType *MT = dyn_cast<MethodType>(PT->getElementType())) {
90 SymbolTable::VarMap &Plane = I->second;
91 for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
93 const string &Name = PI->first;
94 Method *M = cast<Method>(PI->second);
95 Methods[Name].push_back(M);
101 // Now we have a list of all methods with a particular name. If there is more
102 // than one entry in a list, merge the methods together.
104 for (map<string, vector<Method*> >::iterator I = Methods.begin(),
105 E = Methods.end(); I != E; ++I) {
106 vector<Method*> &Methods = I->second;
107 Method *Implementation = 0; // Find the implementation
108 Method *Concrete = 0;
109 for (unsigned i = 0; i < Methods.size(); ) {
110 if (!Methods[i]->isExternal()) { // Found an implementation
111 assert(Implementation == 0 && "Multiple definitions of the same"
112 " method. Case not handled yet!");
113 Implementation = Methods[i];
115 // Ignore methods that are never used so they don't cause spurious
116 // warnings... here we will actually DCE the function so that it isn't
119 if (Methods[i]->use_size() == 0) {
120 M->getMethodList().remove(Methods[i]);
122 Methods.erase(Methods.begin()+i);
128 if (Methods[i] && (!Methods[i]->getMethodType()->isVarArg() ||
129 Methods[i]->getMethodType()->getParamTypes().size())) {
130 if (Concrete) { // Found two different methods types. Can't choose
134 Concrete = Methods[i];
139 if (Methods.size() > 1) { // Found a multiply defined method.
140 // We should find exactly one non-vararg method definition, which is
141 // probably the implementation. Change all of the method definitions
142 // and uses to use it instead.
145 cerr << "Warning: Found methods types that are not compatible:\n";
146 for (unsigned i = 0; i < Methods.size(); ++i) {
147 cerr << "\t" << Methods[i]->getType()->getDescription() << " %"
148 << Methods[i]->getName() << endl;
150 cerr << " No linkage of methods named '" << Methods[0]->getName()
153 for (unsigned i = 0; i < Methods.size(); ++i)
154 if (Methods[i] != Concrete) {
155 Method *Old = Methods[i];
156 assert(Old->getReturnType() == Concrete->getReturnType() &&
157 "Differing return types not handled yet!");
158 assert(Old->getMethodType()->getParamTypes().size() == 0 &&
159 "Cannot handle varargs fn's with specified element types!");
161 // Attempt to convert all of the uses of the old method to the
162 // concrete form of the method. If there is a use of the method
163 // that we don't understand here we punt to avoid making a bad
166 // At this point, we know that the return values are the same for
167 // our two functions and that the Old method has no varargs methods
168 // specified. In otherwords it's just <retty> (...)
170 for (unsigned i = 0; i < Old->use_size(); ) {
171 User *U = *(Old->use_begin()+i);
172 if (CastInst *CI = dyn_cast<CastInst>(U)) {
173 // Convert casts directly
174 assert(CI->getOperand(0) == Old);
175 CI->setOperand(0, Concrete);
177 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
178 // Can only fix up calls TO the argument, not args passed in.
179 if (CI->getCalledValue() == Old) {
180 ConvertCallTo(CI, Concrete);
183 cerr << "Couldn't cleanup this function call, must be an"
184 << " argument or something!" << CI;
188 cerr << "Cannot convert use of method: " << U << endl;
201 // ShouldNukSymtabEntry - Return true if this module level symbol table entry
202 // should be eliminated.
204 static inline bool ShouldNukeSymtabEntry(const pair<string, Value*> &E) {
205 // Nuke all names for primitive types!
206 if (cast<Type>(E.second)->isPrimitiveType()) return true;
208 // Nuke all pointers to primitive types as well...
209 if (const PointerType *PT = dyn_cast<PointerType>(E.second))
210 if (PT->getElementType()->isPrimitiveType()) return true;
212 // The only types that could contain .'s in the program are things generated
213 // by GCC itself, including "complex.float" and friends. Nuke them too.
214 if (E.first.find('.') != string::npos) return true;
219 // doPassInitialization - For this pass, it removes global symbol table
220 // entries for primitive types. These are never used for linking in GCC and
221 // they make the output uglier to look at, so we nuke them.
223 bool CleanupGCCOutput::doPassInitialization(Module *M) {
224 bool Changed = false;
226 FUT.doPassInitialization(M);
229 PtrSByte = PointerType::get(Type::SByteTy);
231 if (M->hasSymbolTable()) {
232 SymbolTable *ST = M->getSymbolTable();
234 // Go over the methods that are in the module and look for methods that have
235 // the same name. More often than not, there will be things like:
236 // void "foo"(...) and void "foo"(int, int) because of the way things are
237 // declared in C. If this is the case, patch things up.
239 Changed |= PatchUpMethodReferences(M);
242 // If the module has a symbol table, they might be referring to the malloc
243 // and free functions. If this is the case, grab the method pointers that
244 // the module is using.
246 // Lookup %malloc and %free in the symbol table, for later use. If they
247 // don't exist, or are not external, we do not worry about converting calls
248 // to that function into the appropriate instruction.
250 const PointerType *MallocType = // Get the type for malloc
251 PointerType::get(MethodType::get(PointerType::get(Type::SByteTy),
252 vector<const Type*>(1, Type::UIntTy), false));
253 Malloc = cast_or_null<Method>(ST->lookup(MallocType, "malloc"));
254 if (Malloc && !Malloc->isExternal())
255 Malloc = 0; // Don't mess with locally defined versions of the fn
257 const PointerType *FreeType = // Get the type for free
258 PointerType::get(MethodType::get(Type::VoidTy,
259 vector<const Type*>(1, PointerType::get(Type::SByteTy)), false));
260 Free = cast_or_null<Method>(ST->lookup(FreeType, "free"));
261 if (Free && !Free->isExternal())
262 Free = 0; // Don't mess with locally defined versions of the fn
265 // Check the symbol table for superfluous type entries...
267 // Grab the 'type' plane of the module symbol...
268 SymbolTable::iterator STI = ST->find(Type::TypeTy);
269 if (STI != ST->end()) {
270 // Loop over all entries in the type plane...
271 SymbolTable::VarMap &Plane = STI->second;
272 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
273 if (ShouldNukeSymtabEntry(*PI)) { // Should we remove this entry?
274 #if MAP_IS_NOT_BRAINDEAD
275 PI = Plane.erase(PI); // STD C++ Map should support this!
277 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
291 // doOneCleanupPass - Do one pass over the input method, fixing stuff up.
293 bool CleanupGCCOutput::doOneCleanupPass(Method *M) {
294 bool Changed = false;
295 for (Method::iterator MI = M->begin(), ME = M->end(); MI != ME; ++MI) {
296 BasicBlock *BB = *MI;
297 BasicBlock::InstListType &BIL = BB->getInstList();
299 for (BasicBlock::iterator BI = BB->begin(); BI != BB->end();) {
300 Instruction *I = *BI;
302 if (CallInst *CI = dyn_cast<CallInst>(I)) {
303 if (CI->getCalledValue() == Malloc) { // Replace call to malloc?
304 MallocInst *MallocI = new MallocInst(PtrSByte, CI->getOperand(1),
307 BI = BIL.insert(BI, MallocI)+1;
308 ReplaceInstWithInst(BIL, BI, new CastInst(MallocI, PtrSByte));
310 continue; // Skip the ++BI
311 } else if (CI->getCalledValue() == Free) { // Replace call to free?
312 ReplaceInstWithInst(BIL, BI, new FreeInst(CI->getOperand(1)));
314 continue; // Skip the ++BI
326 // FixCastsAndPHIs - The LLVM GCC has a tendancy to intermix Cast instructions
327 // in with the PHI nodes. These cast instructions are potentially there for two
328 // different reasons:
330 // 1. The cast could be for an early PHI, and be accidentally inserted before
331 // another PHI node. In this case, the PHI node should be moved to the end
332 // of the PHI nodes in the basic block. We know that it is this case if
333 // the source for the cast is a PHI node in this basic block.
335 // 2. If not #1, the cast must be a source argument for one of the PHI nodes
336 // in the current basic block. If this is the case, the cast should be
337 // lifted into the basic block for the appropriate predecessor.
339 static inline bool FixCastsAndPHIs(BasicBlock *BB) {
340 bool Changed = false;
342 BasicBlock::iterator InsertPos = BB->begin();
344 // Find the end of the interesting instructions...
345 while (isa<PHINode>(*InsertPos) || isa<CastInst>(*InsertPos)) ++InsertPos;
347 // Back the InsertPos up to right after the last PHI node.
348 while (InsertPos != BB->begin() && isa<CastInst>(*(InsertPos-1))) --InsertPos;
350 // No PHI nodes, quick exit.
351 if (InsertPos == BB->begin()) return false;
353 // Loop over all casts trapped between the PHI's...
354 BasicBlock::iterator I = BB->begin();
355 while (I != InsertPos) {
356 if (CastInst *CI = dyn_cast<CastInst>(*I)) { // Fix all cast instructions
357 Value *Src = CI->getOperand(0);
359 // Move the cast instruction to the current insert position...
360 --InsertPos; // New position for cast to go...
361 swap(*InsertPos, *I); // Cast goes down, PHI goes up
363 if (isa<PHINode>(Src) && // Handle case #1
364 cast<PHINode>(Src)->getParent() == BB) {
365 // We're done for case #1
366 } else { // Handle case #2
367 // In case #2, we have to do a few things:
368 // 1. Remove the cast from the current basic block.
369 // 2. Identify the PHI node that the cast is for.
370 // 3. Find out which predecessor the value is for.
371 // 4. Move the cast to the end of the basic block that it SHOULD be
374 // Remove the cast instruction from the basic block. The remove only
375 // invalidates iterators in the basic block that are AFTER the removed
376 // element. Because we just moved the CastInst to the InsertPos, no
377 // iterators get invalidated.
379 BB->getInstList().remove(InsertPos);
381 // Find the PHI node. Since this cast was generated specifically for a
382 // PHI node, there can only be a single PHI node using it.
384 assert(CI->use_size() == 1 && "Exactly one PHI node should use cast!");
385 PHINode *PN = cast<PHINode>(*CI->use_begin());
387 // Find out which operand of the PHI it is...
389 for (i = 0; i < PN->getNumIncomingValues(); ++i)
390 if (PN->getIncomingValue(i) == CI)
392 assert(i != PN->getNumIncomingValues() && "PHI doesn't use cast!");
394 // Get the predecessor the value is for...
395 BasicBlock *Pred = PN->getIncomingBlock(i);
397 // Reinsert the cast right before the terminator in Pred.
398 Pred->getInstList().insert(Pred->end()-1, CI);
409 // RefactorPredecessor - When we find out that a basic block is a repeated
410 // predecessor in a PHI node, we have to refactor the method until there is at
411 // most a single instance of a basic block in any predecessor list.
413 static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
414 Method *M = BB->getParent();
415 assert(find(BB->pred_begin(), BB->pred_end(), Pred) != BB->pred_end() &&
416 "Pred is not a predecessor of BB!");
418 // Create a new basic block, adding it to the end of the method.
419 BasicBlock *NewBB = new BasicBlock("", M);
421 // Add an unconditional branch to BB to the new block.
422 NewBB->getInstList().push_back(new BranchInst(BB));
424 // Get the terminator that causes a branch to BB from Pred.
425 TerminatorInst *TI = Pred->getTerminator();
427 // Find the first use of BB in the terminator...
428 User::op_iterator OI = find(TI->op_begin(), TI->op_end(), BB);
429 assert(OI != TI->op_end() && "Pred does not branch to BB!!!");
431 // Change the use of BB to point to the new stub basic block
434 // Now we need to loop through all of the PHI nodes in BB and convert their
435 // first incoming value for Pred to reference the new basic block instead.
437 for (BasicBlock::iterator I = BB->begin();
438 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
439 int BBIdx = PN->getBasicBlockIndex(Pred);
440 assert(BBIdx != -1 && "PHI node doesn't have an entry for Pred!");
442 // The value that used to look like it came from Pred now comes from NewBB
443 PN->setIncomingBlock((unsigned)BBIdx, NewBB);
448 // CheckIncomingValueFor - Make sure that the specified PHI node has an entry
449 // for the provided basic block. If it doesn't, add one and return true.
451 static inline void CheckIncomingValueFor(PHINode *PN, BasicBlock *BB) {
452 if (PN->getBasicBlockIndex(BB) != -1) return; // Already has value
455 const Type *Ty = PN->getType();
457 if (const PointerType *PT = dyn_cast<PointerType>(Ty))
458 NewVal = ConstantPointerNull::get(PT);
459 else if (Ty == Type::BoolTy)
460 NewVal = ConstantBool::True;
461 else if (Ty == Type::FloatTy || Ty == Type::DoubleTy)
462 NewVal = ConstantFP::get(Ty, 42);
463 else if (Ty->isIntegral())
464 NewVal = ConstantInt::get(Ty, 42);
466 assert(NewVal && "Unknown PHI node type!");
467 PN->addIncoming(NewVal, BB);
470 // fixLocalProblems - Loop through the method and fix problems with the PHI
471 // nodes in the current method. The two problems that are handled are:
473 // 1. PHI nodes with multiple entries for the same predecessor. GCC sometimes
474 // generates code that looks like this:
476 // bb7: br bool %cond1004, label %bb8, label %bb8
477 // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
479 // which is completely illegal LLVM code. To compensate for this, we insert
480 // an extra basic block, and convert the code to look like this:
482 // bb7: br bool %cond1004, label %bbX, label %bb8
484 // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
487 // 2. PHI nodes with fewer arguments than predecessors.
488 // These can be generated by GCC if a variable is uninitalized over a path
489 // in the CFG. We fix this by adding an entry for the missing predecessors
490 // that is initialized to either 42 for a numeric/FP value, or null if it's
491 // a pointer value. This problem can be generated by code that looks like
499 static bool fixLocalProblems(Method *M) {
500 bool Changed = false;
501 // Don't use iterators because invalidation gets messy...
502 for (unsigned MI = 0; MI < M->size(); ++MI) {
503 BasicBlock *BB = M->getBasicBlocks()[MI];
505 Changed |= FixCastsAndPHIs(BB);
507 if (isa<PHINode>(BB->front())) {
508 const vector<BasicBlock*> Preds(BB->pred_begin(), BB->pred_end());
510 // Handle Problem #1. Sort the list of predecessors so that it is easy to
511 // decide whether or not duplicate predecessors exist.
512 vector<BasicBlock*> SortedPreds(Preds);
513 sort(SortedPreds.begin(), SortedPreds.end());
515 // Loop over the predecessors, looking for adjacent BB's that are equal.
516 BasicBlock *LastOne = 0;
517 for (unsigned i = 0; i < Preds.size(); ++i) {
518 if (SortedPreds[i] == LastOne) { // Found a duplicate.
519 RefactorPredecessor(BB, SortedPreds[i]);
522 LastOne = SortedPreds[i];
525 // Loop over all of the PHI nodes in the current BB. These PHI nodes are
526 // guaranteed to be at the beginning of the basic block.
528 for (BasicBlock::iterator I = BB->begin();
529 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
531 // Handle problem #2.
532 if (PN->getNumIncomingValues() != Preds.size()) {
533 assert(PN->getNumIncomingValues() <= Preds.size() &&
534 "Can't handle extra arguments to PHI nodes!");
535 for (unsigned i = 0; i < Preds.size(); ++i)
536 CheckIncomingValueFor(PN, Preds[i]);
548 // doPerMethodWork - This method simplifies the specified method hopefully.
550 bool CleanupGCCOutput::doPerMethodWork(Method *M) {
551 bool Changed = fixLocalProblems(M);
552 while (doOneCleanupPass(M)) Changed = true;
554 FUT.doPerMethodWork(M);
558 bool CleanupGCCOutput::doPassFinalization(Module *M) {
559 bool Changed = false;
560 FUT.doPassFinalization(M);
562 if (M->hasSymbolTable()) {
563 SymbolTable *ST = M->getSymbolTable();
564 const set<const Type *> &UsedTypes = FUT.getTypes();
566 // Check the symbol table for superfluous type entries that aren't used in
569 // Grab the 'type' plane of the module symbol...
570 SymbolTable::iterator STI = ST->find(Type::TypeTy);
571 if (STI != ST->end()) {
572 // Loop over all entries in the type plane...
573 SymbolTable::VarMap &Plane = STI->second;
574 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
575 if (!UsedTypes.count(cast<Type>(PI->second))) {
576 #if MAP_IS_NOT_BRAINDEAD
577 PI = Plane.erase(PI); // STD C++ Map should support this!
579 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
580 PI = Plane.begin(); // N^2 algorithms are fun. :(