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 // * Fix various problems that we might have in PHI nodes and casts
10 // * Link uses of 'void %foo(...)' to 'void %foo(sometypes)'
12 // Note: This code produces dead declarations, it is a good idea to run DCE
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Transforms/CleanupGCCOutput.h"
18 #include "llvm/Analysis/FindUsedTypes.h"
19 #include "TransformInternals.h"
20 #include "llvm/Module.h"
21 #include "llvm/SymbolTable.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/iPHINode.h"
24 #include "llvm/iMemory.h"
25 #include "llvm/iTerminators.h"
26 #include "llvm/iOther.h"
27 #include "llvm/Support/CFG.h"
28 #include "llvm/Pass.h"
35 static const Type *PtrSByte = 0; // 'sbyte*' type
38 struct CleanupGCCOutput : public MethodPass {
39 // doPassInitialization - For this pass, it removes global symbol table
40 // entries for primitive types. These are never used for linking in GCC and
41 // they make the output uglier to look at, so we nuke them.
43 // Also, initialize instance variables.
45 bool doInitialization(Module *M);
47 // runOnFunction - This method simplifies the specified function hopefully.
49 bool runOnMethod(Function *F);
51 // doPassFinalization - Strip out type names that are unused by the program
52 bool doFinalization(Module *M);
54 // getAnalysisUsageInfo - This function needs FindUsedTypes to do its job...
56 virtual void getAnalysisUsageInfo(Pass::AnalysisSet &Required,
57 Pass::AnalysisSet &Destroyed,
58 Pass::AnalysisSet &Provided) {
59 // FIXME: Invalidates the CFG
60 Required.push_back(FindUsedTypes::ID);
65 Pass *createCleanupGCCOutputPass() {
66 return new CleanupGCCOutput();
71 // ShouldNukSymtabEntry - Return true if this module level symbol table entry
72 // should be eliminated.
74 static inline bool ShouldNukeSymtabEntry(const std::pair<string, Value*> &E) {
75 // Nuke all names for primitive types!
76 if (cast<Type>(E.second)->isPrimitiveType()) return true;
78 // Nuke all pointers to primitive types as well...
79 if (const PointerType *PT = dyn_cast<PointerType>(E.second))
80 if (PT->getElementType()->isPrimitiveType()) return true;
82 // The only types that could contain .'s in the program are things generated
83 // by GCC itself, including "complex.float" and friends. Nuke them too.
84 if (E.first.find('.') != string::npos) return true;
89 // doInitialization - For this pass, it removes global symbol table
90 // entries for primitive types. These are never used for linking in GCC and
91 // they make the output uglier to look at, so we nuke them.
93 bool CleanupGCCOutput::doInitialization(Module *M) {
97 PtrSByte = PointerType::get(Type::SByteTy);
99 if (M->hasSymbolTable()) {
100 SymbolTable *ST = M->getSymbolTable();
102 // Check the symbol table for superfluous type entries...
104 // Grab the 'type' plane of the module symbol...
105 SymbolTable::iterator STI = ST->find(Type::TypeTy);
106 if (STI != ST->end()) {
107 // Loop over all entries in the type plane...
108 SymbolTable::VarMap &Plane = STI->second;
109 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
110 if (ShouldNukeSymtabEntry(*PI)) { // Should we remove this entry?
111 #if MAP_IS_NOT_BRAINDEAD
112 PI = Plane.erase(PI); // STD C++ Map should support this!
114 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
128 // FixCastsAndPHIs - The LLVM GCC has a tendancy to intermix Cast instructions
129 // in with the PHI nodes. These cast instructions are potentially there for two
130 // different reasons:
132 // 1. The cast could be for an early PHI, and be accidentally inserted before
133 // another PHI node. In this case, the PHI node should be moved to the end
134 // of the PHI nodes in the basic block. We know that it is this case if
135 // the source for the cast is a PHI node in this basic block.
137 // 2. If not #1, the cast must be a source argument for one of the PHI nodes
138 // in the current basic block. If this is the case, the cast should be
139 // lifted into the basic block for the appropriate predecessor.
141 static inline bool FixCastsAndPHIs(BasicBlock *BB) {
142 bool Changed = false;
144 BasicBlock::iterator InsertPos = BB->begin();
146 // Find the end of the interesting instructions...
147 while (isa<PHINode>(*InsertPos) || isa<CastInst>(*InsertPos)) ++InsertPos;
149 // Back the InsertPos up to right after the last PHI node.
150 while (InsertPos != BB->begin() && isa<CastInst>(*(InsertPos-1))) --InsertPos;
152 // No PHI nodes, quick exit.
153 if (InsertPos == BB->begin()) return false;
155 // Loop over all casts trapped between the PHI's...
156 BasicBlock::iterator I = BB->begin();
157 while (I != InsertPos) {
158 if (CastInst *CI = dyn_cast<CastInst>(*I)) { // Fix all cast instructions
159 Value *Src = CI->getOperand(0);
161 // Move the cast instruction to the current insert position...
162 --InsertPos; // New position for cast to go...
163 std::swap(*InsertPos, *I); // Cast goes down, PHI goes up
165 if (isa<PHINode>(Src) && // Handle case #1
166 cast<PHINode>(Src)->getParent() == BB) {
167 // We're done for case #1
168 } else { // Handle case #2
169 // In case #2, we have to do a few things:
170 // 1. Remove the cast from the current basic block.
171 // 2. Identify the PHI node that the cast is for.
172 // 3. Find out which predecessor the value is for.
173 // 4. Move the cast to the end of the basic block that it SHOULD be
176 // Remove the cast instruction from the basic block. The remove only
177 // invalidates iterators in the basic block that are AFTER the removed
178 // element. Because we just moved the CastInst to the InsertPos, no
179 // iterators get invalidated.
181 BB->getInstList().remove(InsertPos);
183 // Find the PHI node. Since this cast was generated specifically for a
184 // PHI node, there can only be a single PHI node using it.
186 assert(CI->use_size() == 1 && "Exactly one PHI node should use cast!");
187 PHINode *PN = cast<PHINode>(*CI->use_begin());
189 // Find out which operand of the PHI it is...
191 for (i = 0; i < PN->getNumIncomingValues(); ++i)
192 if (PN->getIncomingValue(i) == CI)
194 assert(i != PN->getNumIncomingValues() && "PHI doesn't use cast!");
196 // Get the predecessor the value is for...
197 BasicBlock *Pred = PN->getIncomingBlock(i);
199 // Reinsert the cast right before the terminator in Pred.
200 Pred->getInstList().insert(Pred->end()-1, CI);
210 // RefactorPredecessor - When we find out that a basic block is a repeated
211 // predecessor in a PHI node, we have to refactor the function until there is at
212 // most a single instance of a basic block in any predecessor list.
214 static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
215 Function *M = BB->getParent();
216 assert(find(pred_begin(BB), pred_end(BB), Pred) != pred_end(BB) &&
217 "Pred is not a predecessor of BB!");
219 // Create a new basic block, adding it to the end of the function.
220 BasicBlock *NewBB = new BasicBlock("", M);
222 // Add an unconditional branch to BB to the new block.
223 NewBB->getInstList().push_back(new BranchInst(BB));
225 // Get the terminator that causes a branch to BB from Pred.
226 TerminatorInst *TI = Pred->getTerminator();
228 // Find the first use of BB in the terminator...
229 User::op_iterator OI = find(TI->op_begin(), TI->op_end(), BB);
230 assert(OI != TI->op_end() && "Pred does not branch to BB!!!");
232 // Change the use of BB to point to the new stub basic block
235 // Now we need to loop through all of the PHI nodes in BB and convert their
236 // first incoming value for Pred to reference the new basic block instead.
238 for (BasicBlock::iterator I = BB->begin();
239 PHINode *PN = dyn_cast<PHINode>(*I); ++I) {
240 int BBIdx = PN->getBasicBlockIndex(Pred);
241 assert(BBIdx != -1 && "PHI node doesn't have an entry for Pred!");
243 // The value that used to look like it came from Pred now comes from NewBB
244 PN->setIncomingBlock((unsigned)BBIdx, NewBB);
249 // runOnMethod - Loop through the function and fix problems with the PHI nodes
250 // in the current function. The problem is that PHI nodes might exist with
251 // multiple entries for the same predecessor. GCC sometimes generates code that
254 // bb7: br bool %cond1004, label %bb8, label %bb8
255 // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
257 // which is completely illegal LLVM code. To compensate for this, we insert
258 // an extra basic block, and convert the code to look like this:
260 // bb7: br bool %cond1004, label %bbX, label %bb8
262 // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
265 bool CleanupGCCOutput::runOnMethod(Function *M) {
266 bool Changed = false;
267 // Don't use iterators because invalidation gets messy...
268 for (unsigned MI = 0; MI < M->size(); ++MI) {
269 BasicBlock *BB = M->getBasicBlocks()[MI];
271 Changed |= FixCastsAndPHIs(BB);
273 if (isa<PHINode>(BB->front())) {
274 const vector<BasicBlock*> Preds(pred_begin(BB), pred_end(BB));
276 // Handle the problem. Sort the list of predecessors so that it is easy
277 // to decide whether or not duplicate predecessors exist.
278 vector<BasicBlock*> SortedPreds(Preds);
279 sort(SortedPreds.begin(), SortedPreds.end());
281 // Loop over the predecessors, looking for adjacent BB's that are equal.
282 BasicBlock *LastOne = 0;
283 for (unsigned i = 0; i < Preds.size(); ++i) {
284 if (SortedPreds[i] == LastOne) { // Found a duplicate.
285 RefactorPredecessor(BB, SortedPreds[i]);
288 LastOne = SortedPreds[i];
295 bool CleanupGCCOutput::doFinalization(Module *M) {
296 bool Changed = false;
298 if (M->hasSymbolTable()) {
299 SymbolTable *ST = M->getSymbolTable();
300 const std::set<const Type *> &UsedTypes =
301 getAnalysis<FindUsedTypes>().getTypes();
303 // Check the symbol table for superfluous type entries that aren't used in
306 // Grab the 'type' plane of the module symbol...
307 SymbolTable::iterator STI = ST->find(Type::TypeTy);
308 if (STI != ST->end()) {
309 // Loop over all entries in the type plane...
310 SymbolTable::VarMap &Plane = STI->second;
311 for (SymbolTable::VarMap::iterator PI = Plane.begin(); PI != Plane.end();)
312 if (!UsedTypes.count(cast<Type>(PI->second))) {
313 #if MAP_IS_NOT_BRAINDEAD
314 PI = Plane.erase(PI); // STD C++ Map should support this!
316 Plane.erase(PI); // Alas, GCC 2.95.3 doesn't *SIGH*
317 PI = Plane.begin(); // N^2 algorithms are fun. :(
329 //===----------------------------------------------------------------------===//
331 // FunctionResolvingPass - Go over the functions that are in the module and
332 // look for functions that have the same name. More often than not, there will
335 // void "foo"(int, int)
336 // because of the way things are declared in C. If this is the case, patch
339 //===----------------------------------------------------------------------===//
342 struct FunctionResolvingPass : public Pass {
347 // ConvertCallTo - Convert a call to a varargs function with no arg types
348 // specified to a concrete nonvarargs function.
350 static void ConvertCallTo(CallInst *CI, Function *Dest) {
351 const FunctionType::ParamTypes &ParamTys =
352 Dest->getFunctionType()->getParamTypes();
353 BasicBlock *BB = CI->getParent();
355 // Get an iterator to where we want to insert cast instructions if the
356 // argument types don't agree.
358 BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
359 assert(BBI != BB->end() && "CallInst not in parent block?");
361 assert(CI->getNumOperands()-1 == ParamTys.size()&&
362 "Function calls resolved funny somehow, incompatible number of args");
364 vector<Value*> Params;
366 // Convert all of the call arguments over... inserting cast instructions if
367 // the types are not compatible.
368 for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
369 Value *V = CI->getOperand(i);
371 if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
372 Instruction *Cast = new CastInst(V, ParamTys[i-1]);
373 BBI = BB->getInstList().insert(BBI, Cast)+1;
380 // Replace the old call instruction with a new call instruction that calls
381 // the real function.
383 ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
387 bool FunctionResolvingPass::run(Module *M) {
388 SymbolTable *ST = M->getSymbolTable();
389 if (!ST) return false;
391 std::map<string, vector<Function*> > Functions;
393 // Loop over the entries in the symbol table. If an entry is a func pointer,
394 // then add it to the Functions map. We do a two pass algorithm here to avoid
395 // problems with iterators getting invalidated if we did a one pass scheme.
397 for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
398 if (const PointerType *PT = dyn_cast<PointerType>(I->first))
399 if (isa<FunctionType>(PT->getElementType())) {
400 SymbolTable::VarMap &Plane = I->second;
401 for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
403 const string &Name = PI->first;
404 Functions[Name].push_back(cast<Function>(PI->second));
408 bool Changed = false;
410 // Now we have a list of all functions with a particular name. If there is
411 // more than one entry in a list, merge the functions together.
413 for (std::map<string, vector<Function*> >::iterator I = Functions.begin(),
414 E = Functions.end(); I != E; ++I) {
415 vector<Function*> &Functions = I->second;
416 Function *Implementation = 0; // Find the implementation
417 Function *Concrete = 0;
418 for (unsigned i = 0; i < Functions.size(); ) {
419 if (!Functions[i]->isExternal()) { // Found an implementation
420 assert(Implementation == 0 && "Multiple definitions of the same"
421 " function. Case not handled yet!");
422 Implementation = Functions[i];
424 // Ignore functions that are never used so they don't cause spurious
425 // warnings... here we will actually DCE the function so that it isn't
428 if (Functions[i]->use_size() == 0) {
429 M->getFunctionList().remove(Functions[i]);
431 Functions.erase(Functions.begin()+i);
437 if (Functions[i] && (!Functions[i]->getFunctionType()->isVarArg())) {
438 if (Concrete) { // Found two different functions types. Can't choose
442 Concrete = Functions[i];
447 if (Functions.size() > 1) { // Found a multiply defined function...
448 // We should find exactly one non-vararg function definition, which is
449 // probably the implementation. Change all of the function definitions
450 // and uses to use it instead.
453 cerr << "Warning: Found functions types that are not compatible:\n";
454 for (unsigned i = 0; i < Functions.size(); ++i) {
455 cerr << "\t" << Functions[i]->getType()->getDescription() << " %"
456 << Functions[i]->getName() << "\n";
458 cerr << " No linkage of functions named '" << Functions[0]->getName()
461 for (unsigned i = 0; i < Functions.size(); ++i)
462 if (Functions[i] != Concrete) {
463 Function *Old = Functions[i];
464 const FunctionType *OldMT = Old->getFunctionType();
465 const FunctionType *ConcreteMT = Concrete->getFunctionType();
468 assert(Old->getReturnType() == Concrete->getReturnType() &&
469 "Differing return types not handled yet!");
470 assert(OldMT->getParamTypes().size() <=
471 ConcreteMT->getParamTypes().size() &&
472 "Concrete type must have more specified parameters!");
474 // Check to make sure that if there are specified types, that they
477 for (unsigned i = 0; i < OldMT->getParamTypes().size(); ++i)
478 if (OldMT->getParamTypes()[i] != ConcreteMT->getParamTypes()[i]) {
479 cerr << "Parameter types conflict for" << OldMT
480 << " and " << ConcreteMT;
483 if (Broken) break; // Can't process this one!
486 // Attempt to convert all of the uses of the old function to the
487 // concrete form of the function. If there is a use of the fn
488 // that we don't understand here we punt to avoid making a bad
491 // At this point, we know that the return values are the same for
492 // our two functions and that the Old function has no varargs fns
493 // specified. In otherwords it's just <retty> (...)
495 for (unsigned i = 0; i < Old->use_size(); ) {
496 User *U = *(Old->use_begin()+i);
497 if (CastInst *CI = dyn_cast<CastInst>(U)) {
498 // Convert casts directly
499 assert(CI->getOperand(0) == Old);
500 CI->setOperand(0, Concrete);
502 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
503 // Can only fix up calls TO the argument, not args passed in.
504 if (CI->getCalledValue() == Old) {
505 ConvertCallTo(CI, Concrete);
508 cerr << "Couldn't cleanup this function call, must be an"
509 << " argument or something!" << CI;
513 cerr << "Cannot convert use of function: " << U << "\n";
525 Pass *createFunctionResolvingPass() {
526 return new FunctionResolvingPass();