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"
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 // doPassInitialization - 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::doPassInitialization(Module *M) {
228 bool Changed = false;
230 FUT.doPassInitialization(M);
233 PtrSByte = PointerType::get(Type::SByteTy);
235 if (M->hasSymbolTable()) {
236 SymbolTable *ST = M->getSymbolTable();
238 // Go over the methods that are in the module and look for methods that have
239 // the same name. More often than not, there will be things like:
240 // void "foo"(...) and void "foo"(int, int) because of the way things are
241 // declared in C. If this is the case, patch things up.
243 Changed |= PatchUpMethodReferences(M);
246 // If the module has a symbol table, they might be referring to the malloc
247 // and free functions. If this is the case, grab the method pointers that
248 // the module is using.
250 // Lookup %malloc and %free in the symbol table, for later use. If they
251 // don't exist, or are not external, we do not worry about converting calls
252 // to that function into the appropriate instruction.
254 const PointerType *MallocType = // Get the type for malloc
255 PointerType::get(MethodType::get(PointerType::get(Type::SByteTy),
256 vector<const Type*>(1, Type::UIntTy), false));
257 Malloc = cast_or_null<Method>(ST->lookup(MallocType, "malloc"));
258 if (Malloc && !Malloc->isExternal())
259 Malloc = 0; // Don't mess with locally defined versions of the fn
261 const PointerType *FreeType = // Get the type for free
262 PointerType::get(MethodType::get(Type::VoidTy,
263 vector<const Type*>(1, PointerType::get(Type::SByteTy)), false));
264 Free = cast_or_null<Method>(ST->lookup(FreeType, "free"));
265 if (Free && !Free->isExternal())
266 Free = 0; // Don't mess with locally defined versions of the fn
269 // Check the symbol table for superfluous type entries...
271 // Grab the 'type' plane of the module symbol...
272 SymbolTable::iterator STI = ST->find(Type::TypeTy);
273 if (STI != ST->end()) {
274 // Loop over all entries in the type plane...
275 SymbolTable::VarMap &Plane = STI->second;
276 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
277 if (ShouldNukeSymtabEntry(*PI)) { // Should we remove this entry?
278 #if MAP_IS_NOT_BRAINDEAD
279 PI = Plane.erase(PI); // STD C++ Map should support this!
281 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
295 // doOneCleanupPass - Do one pass over the input method, fixing stuff up.
297 bool CleanupGCCOutput::doOneCleanupPass(Method *M) {
298 bool Changed = false;
299 for (Method::iterator MI = M->begin(), ME = M->end(); MI != ME; ++MI) {
300 BasicBlock *BB = *MI;
301 BasicBlock::InstListType &BIL = BB->getInstList();
303 for (BasicBlock::iterator BI = BB->begin(); BI != BB->end();) {
304 Instruction *I = *BI;
306 if (CallInst *CI = dyn_cast<CallInst>(I)) {
307 if (CI->getCalledValue() == Malloc) { // Replace call to malloc?
308 MallocInst *MallocI = new MallocInst(PtrSByte, CI->getOperand(1),
311 BI = BIL.insert(BI, MallocI)+1;
312 ReplaceInstWithInst(BIL, BI, new CastInst(MallocI, PtrSByte));
314 continue; // Skip the ++BI
315 } else if (CI->getCalledValue() == Free) { // Replace call to free?
316 ReplaceInstWithInst(BIL, BI, new FreeInst(CI->getOperand(1)));
318 continue; // Skip the ++BI
330 // FixCastsAndPHIs - The LLVM GCC has a tendancy to intermix Cast instructions
331 // in with the PHI nodes. These cast instructions are potentially there for two
332 // different reasons:
334 // 1. The cast could be for an early PHI, and be accidentally inserted before
335 // another PHI node. In this case, the PHI node should be moved to the end
336 // of the PHI nodes in the basic block. We know that it is this case if
337 // the source for the cast is a PHI node in this basic block.
339 // 2. If not #1, the cast must be a source argument for one of the PHI nodes
340 // in the current basic block. If this is the case, the cast should be
341 // lifted into the basic block for the appropriate predecessor.
343 static inline bool FixCastsAndPHIs(BasicBlock *BB) {
344 bool Changed = false;
346 BasicBlock::iterator InsertPos = BB->begin();
348 // Find the end of the interesting instructions...
349 while (isa<PHINode>(*InsertPos) || isa<CastInst>(*InsertPos)) ++InsertPos;
351 // Back the InsertPos up to right after the last PHI node.
352 while (InsertPos != BB->begin() && isa<CastInst>(*(InsertPos-1))) --InsertPos;
354 // No PHI nodes, quick exit.
355 if (InsertPos == BB->begin()) return false;
357 // Loop over all casts trapped between the PHI's...
358 BasicBlock::iterator I = BB->begin();
359 while (I != InsertPos) {
360 if (CastInst *CI = dyn_cast<CastInst>(*I)) { // Fix all cast instructions
361 Value *Src = CI->getOperand(0);
363 // Move the cast instruction to the current insert position...
364 --InsertPos; // New position for cast to go...
365 std::swap(*InsertPos, *I); // Cast goes down, PHI goes up
367 if (isa<PHINode>(Src) && // Handle case #1
368 cast<PHINode>(Src)->getParent() == BB) {
369 // We're done for case #1
370 } else { // Handle case #2
371 // In case #2, we have to do a few things:
372 // 1. Remove the cast from the current basic block.
373 // 2. Identify the PHI node that the cast is for.
374 // 3. Find out which predecessor the value is for.
375 // 4. Move the cast to the end of the basic block that it SHOULD be
378 // Remove the cast instruction from the basic block. The remove only
379 // invalidates iterators in the basic block that are AFTER the removed
380 // element. Because we just moved the CastInst to the InsertPos, no
381 // iterators get invalidated.
383 BB->getInstList().remove(InsertPos);
385 // Find the PHI node. Since this cast was generated specifically for a
386 // PHI node, there can only be a single PHI node using it.
388 assert(CI->use_size() == 1 && "Exactly one PHI node should use cast!");
389 PHINode *PN = cast<PHINode>(*CI->use_begin());
391 // Find out which operand of the PHI it is...
393 for (i = 0; i < PN->getNumIncomingValues(); ++i)
394 if (PN->getIncomingValue(i) == CI)
396 assert(i != PN->getNumIncomingValues() && "PHI doesn't use cast!");
398 // Get the predecessor the value is for...
399 BasicBlock *Pred = PN->getIncomingBlock(i);
401 // Reinsert the cast right before the terminator in Pred.
402 Pred->getInstList().insert(Pred->end()-1, CI);
413 // RefactorPredecessor - When we find out that a basic block is a repeated
414 // predecessor in a PHI node, we have to refactor the method until there is at
415 // most a single instance of a basic block in any predecessor list.
417 static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
418 Method *M = BB->getParent();
419 assert(find(BB->pred_begin(), BB->pred_end(), Pred) != BB->pred_end() &&
420 "Pred is not a predecessor of BB!");
422 // Create a new basic block, adding it to the end of the method.
423 BasicBlock *NewBB = new BasicBlock("", M);
425 // Add an unconditional branch to BB to the new block.
426 NewBB->getInstList().push_back(new BranchInst(BB));
428 // Get the terminator that causes a branch to BB from Pred.
429 TerminatorInst *TI = Pred->getTerminator();
431 // Find the first use of BB in the terminator...
432 User::op_iterator OI = find(TI->op_begin(), TI->op_end(), BB);
433 assert(OI != TI->op_end() && "Pred does not branch to BB!!!");
435 // Change the use of BB to point to the new stub basic block
438 // Now we need to loop through all of the PHI nodes in BB and convert their
439 // first incoming value for Pred to reference the new basic block instead.
441 for (BasicBlock::iterator I = BB->begin();
442 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
443 int BBIdx = PN->getBasicBlockIndex(Pred);
444 assert(BBIdx != -1 && "PHI node doesn't have an entry for Pred!");
446 // The value that used to look like it came from Pred now comes from NewBB
447 PN->setIncomingBlock((unsigned)BBIdx, NewBB);
452 // CheckIncomingValueFor - Make sure that the specified PHI node has an entry
453 // for the provided basic block. If it doesn't, add one and return true.
455 static inline void CheckIncomingValueFor(PHINode *PN, BasicBlock *BB) {
456 if (PN->getBasicBlockIndex(BB) != -1) return; // Already has value
459 const Type *Ty = PN->getType();
461 if (const PointerType *PT = dyn_cast<PointerType>(Ty))
462 NewVal = ConstantPointerNull::get(PT);
463 else if (Ty == Type::BoolTy)
464 NewVal = ConstantBool::True;
465 else if (Ty == Type::FloatTy || Ty == Type::DoubleTy)
466 NewVal = ConstantFP::get(Ty, 42);
467 else if (Ty->isIntegral())
468 NewVal = ConstantInt::get(Ty, 42);
470 assert(NewVal && "Unknown PHI node type!");
471 PN->addIncoming(NewVal, BB);
474 // fixLocalProblems - Loop through the method and fix problems with the PHI
475 // nodes in the current method. The two problems that are handled are:
477 // 1. PHI nodes with multiple entries for the same predecessor. GCC sometimes
478 // generates code that looks like this:
480 // bb7: br bool %cond1004, label %bb8, label %bb8
481 // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
483 // which is completely illegal LLVM code. To compensate for this, we insert
484 // an extra basic block, and convert the code to look like this:
486 // bb7: br bool %cond1004, label %bbX, label %bb8
488 // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
491 // 2. PHI nodes with fewer arguments than predecessors.
492 // These can be generated by GCC if a variable is uninitalized over a path
493 // in the CFG. We fix this by adding an entry for the missing predecessors
494 // that is initialized to either 42 for a numeric/FP value, or null if it's
495 // a pointer value. This problem can be generated by code that looks like
503 static bool fixLocalProblems(Method *M) {
504 bool Changed = false;
505 // Don't use iterators because invalidation gets messy...
506 for (unsigned MI = 0; MI < M->size(); ++MI) {
507 BasicBlock *BB = M->getBasicBlocks()[MI];
509 Changed |= FixCastsAndPHIs(BB);
511 if (isa<PHINode>(BB->front())) {
512 const vector<BasicBlock*> Preds(BB->pred_begin(), BB->pred_end());
514 // Handle Problem #1. Sort the list of predecessors so that it is easy to
515 // decide whether or not duplicate predecessors exist.
516 vector<BasicBlock*> SortedPreds(Preds);
517 sort(SortedPreds.begin(), SortedPreds.end());
519 // Loop over the predecessors, looking for adjacent BB's that are equal.
520 BasicBlock *LastOne = 0;
521 for (unsigned i = 0; i < Preds.size(); ++i) {
522 if (SortedPreds[i] == LastOne) { // Found a duplicate.
523 RefactorPredecessor(BB, SortedPreds[i]);
526 LastOne = SortedPreds[i];
529 // Loop over all of the PHI nodes in the current BB. These PHI nodes are
530 // guaranteed to be at the beginning of the basic block.
532 for (BasicBlock::iterator I = BB->begin();
533 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
535 // Handle problem #2.
536 if (PN->getNumIncomingValues() != Preds.size()) {
537 assert(PN->getNumIncomingValues() <= Preds.size() &&
538 "Can't handle extra arguments to PHI nodes!");
539 for (unsigned i = 0; i < Preds.size(); ++i)
540 CheckIncomingValueFor(PN, Preds[i]);
552 // doPerMethodWork - This method simplifies the specified method hopefully.
554 bool CleanupGCCOutput::doPerMethodWork(Method *M) {
555 bool Changed = fixLocalProblems(M);
556 while (doOneCleanupPass(M)) Changed = true;
558 FUT.doPerMethodWork(M);
562 bool CleanupGCCOutput::doPassFinalization(Module *M) {
563 bool Changed = false;
564 FUT.doPassFinalization(M);
566 if (M->hasSymbolTable()) {
567 SymbolTable *ST = M->getSymbolTable();
568 const std::set<const Type *> &UsedTypes = FUT.getTypes();
570 // Check the symbol table for superfluous type entries that aren't used in
573 // Grab the 'type' plane of the module symbol...
574 SymbolTable::iterator STI = ST->find(Type::TypeTy);
575 if (STI != ST->end()) {
576 // Loop over all entries in the type plane...
577 SymbolTable::VarMap &Plane = STI->second;
578 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
579 if (!UsedTypes.count(cast<Type>(PI->second))) {
580 #if MAP_IS_NOT_BRAINDEAD
581 PI = Plane.erase(PI); // STD C++ Map should support this!
583 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
584 PI = Plane.begin(); // N^2 algorithms are fun. :(