6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-08-10 02:25:47 +00:00
Code Issues Projects Releases Wiki Activity

* Add documentation

Browse Source 
* Split the CleanGCC pass into two passes, a global pass and an IP pass.
  Before it was just a global pass, but it did illegal things to the
  module, which broke other passes that were being scheduled with it by
  gccld.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@2224 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Chris Lattner
2002-04-10 20:33:11 +00:00
parent 5648b58a19
commit 265b083e84
1 changed files with 208 additions and 210 deletions
Download Patch File Download Diff File Expand all files Collapse all files

418
lib/Transforms/IPO/DeadTypeElimination.cpp
View File

@@ -1,4 +1,4 @@
//===- CleanupGCCOutput.cpp - Cleanup GCC Output ----------------------------=// //===- CleanupGCCOutput.cpp - Cleanup GCC Output --------------------------===//
// //
// This pass is used to cleanup the output of GCC. GCC's output is // This pass is used to cleanup the output of GCC. GCC's output is
// unneccessarily gross for a couple of reasons. This pass does the following // unneccessarily gross for a couple of reasons. This pass does the following
@@ -6,6 +6,8 @@
// //
// * Eliminate names for GCC types that we know can't be needed by the user. // * Eliminate names for GCC types that we know can't be needed by the user.
// * Eliminate names for types that are unused in the entire translation unit // * Eliminate names for types that are unused in the entire translation unit
// * Fix various problems that we might have in PHI nodes and casts
// * Link uses of 'void %foo(...)' to 'void %foo(sometypes)'
// //
// Note: This code produces dead declarations, it is a good idea to run DCE // Note: This code produces dead declarations, it is a good idea to run DCE
// after this pass. // after this pass.
@@ -60,194 +62,11 @@ namespace {
}; };
} }
Pass *createCleanupGCCOutputPass() {
return new CleanupGCCOutput();
// ConvertCallTo - Convert a call to a varargs function with no arg types
// specified to a concrete nonvarargs function.
//
static void ConvertCallTo(CallInst *CI, Function *Dest) {
const FunctionType::ParamTypes &ParamTys =
Dest->getFunctionType()->getParamTypes();
BasicBlock *BB = CI->getParent();
// Get an iterator to where we want to insert cast instructions if the
// argument types don't agree.
//
BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
assert(BBI != BB->end() && "CallInst not in parent block?");
assert(CI->getNumOperands()-1 == ParamTys.size()&&
"Function calls resolved funny somehow, incompatible number of args");
vector<Value*> Params;
// Convert all of the call arguments over... inserting cast instructions if
// the types are not compatible.
for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
Value *V = CI->getOperand(i);
if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
Instruction *Cast = new CastInst(V, ParamTys[i-1]);
BBI = BB->getInstList().insert(BBI, Cast)+1;
V = Cast;
}
Params.push_back(V);
}
// Replace the old call instruction with a new call instruction that calls
// the real function.
//
ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
} }
// PatchUpFunctionReferences - Go over the functions that are in the module and
// look for functions that have the same name. More often than not, there will
// be things like:
// void "foo"(...)
// void "foo"(int, int)
// because of the way things are declared in C. If this is the case, patch
// things up.
//
static bool PatchUpFunctionReferences(Module *M) {
SymbolTable *ST = M->getSymbolTable();
if (!ST) return false;
std::map<string, vector<Function*> > Functions;
// Loop over the entries in the symbol table. If an entry is a func pointer,
// then add it to the Functions map. We do a two pass algorithm here to avoid
// problems with iterators getting invalidated if we did a one pass scheme.
//
for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
if (const PointerType *PT = dyn_cast<PointerType>(I->first))
if (isa<FunctionType>(PT->getElementType())) {
SymbolTable::VarMap &Plane = I->second;
for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
PI != PE; ++PI) {
const string &Name = PI->first;
Functions[Name].push_back(cast<Function>(PI->second));
}
}
bool Changed = false;
// Now we have a list of all functions with a particular name. If there is
// more than one entry in a list, merge the functions together.
//
for (std::map<string, vector<Function*> >::iterator I = Functions.begin(),
E = Functions.end(); I != E; ++I) {
vector<Function*> &Functions = I->second;
Function *Implementation = 0; // Find the implementation
Function *Concrete = 0;
for (unsigned i = 0; i < Functions.size(); ) {
if (!Functions[i]->isExternal()) { // Found an implementation
assert(Implementation == 0 && "Multiple definitions of the same"
" function. Case not handled yet!");
Implementation = Functions[i];
} else {
// Ignore functions that are never used so they don't cause spurious
// warnings... here we will actually DCE the function so that it isn't
// used later.
//
if (Functions[i]->use_size() == 0) {
M->getFunctionList().remove(Functions[i]);
delete Functions[i];
Functions.erase(Functions.begin()+i);
Changed = true;
continue;
}
}
if (Functions[i] && (!Functions[i]->getFunctionType()->isVarArg())) {
if (Concrete) { // Found two different functions types. Can't choose
Concrete = 0;
break;
}
Concrete = Functions[i];
}
++i;
}
if (Functions.size() > 1) { // Found a multiply defined function...
// We should find exactly one non-vararg function definition, which is
// probably the implementation. Change all of the function definitions
// and uses to use it instead.
//
if (!Concrete) {
cerr << "Warning: Found functions types that are not compatible:\n";
for (unsigned i = 0; i < Functions.size(); ++i) {
cerr << "\t" << Functions[i]->getType()->getDescription() << " %"
<< Functions[i]->getName() << "\n";
}
cerr << " No linkage of functions named '" << Functions[0]->getName()
<< "' performed!\n";
} else {
for (unsigned i = 0; i < Functions.size(); ++i)
if (Functions[i] != Concrete) {
Function *Old = Functions[i];
const FunctionType *OldMT = Old->getFunctionType();
const FunctionType *ConcreteMT = Concrete->getFunctionType();
bool Broken = false;
assert(Old->getReturnType() == Concrete->getReturnType() &&
"Differing return types not handled yet!");
assert(OldMT->getParamTypes().size() <=
ConcreteMT->getParamTypes().size() &&
"Concrete type must have more specified parameters!");
// Check to make sure that if there are specified types, that they
// match...
//
for (unsigned i = 0; i < OldMT->getParamTypes().size(); ++i)
if (OldMT->getParamTypes()[i] != ConcreteMT->getParamTypes()[i]) {
cerr << "Parameter types conflict for" << OldMT
<< " and " << ConcreteMT;
Broken = true;
}
if (Broken) break; // Can't process this one!
// Attempt to convert all of the uses of the old function to the
// concrete form of the function. If there is a use of the fn
// that we don't understand here we punt to avoid making a bad
// transformation.
//
// At this point, we know that the return values are the same for
// our two functions and that the Old function has no varargs fns
// specified. In otherwords it's just <retty> (...)
//
for (unsigned i = 0; i < Old->use_size(); ) {
User *U = *(Old->use_begin()+i);
if (CastInst *CI = dyn_cast<CastInst>(U)) {
// Convert casts directly
assert(CI->getOperand(0) == Old);
CI->setOperand(0, Concrete);
Changed = true;
} else if (CallInst *CI = dyn_cast<CallInst>(U)) {
// Can only fix up calls TO the argument, not args passed in.
if (CI->getCalledValue() == Old) {
ConvertCallTo(CI, Concrete);
Changed = true;
} else {
cerr << "Couldn't cleanup this function call, must be an"
<< " argument or something!" << CI;
++i;
}
} else {
cerr << "Cannot convert use of function: " << U << "\n";
++i;
}
}
}
}
}
}
return Changed;
}
// ShouldNukSymtabEntry - Return true if this module level symbol table entry // ShouldNukSymtabEntry - Return true if this module level symbol table entry
// should be eliminated. // should be eliminated.
@@ -280,13 +99,6 @@ bool CleanupGCCOutput::doInitialization(Module *M) {
if (M->hasSymbolTable()) { if (M->hasSymbolTable()) {
SymbolTable *ST = M->getSymbolTable(); SymbolTable *ST = M->getSymbolTable();
// Go over the functions that are in the module and look for methods that
// have the same name. More often than not, there will be things like:
// void "foo"(...) and void "foo"(int, int) because of the way things are
// declared in C. If this is the case, patch things up.
//
Changed |= PatchUpFunctionReferences(M);
// Check the symbol table for superfluous type entries... // Check the symbol table for superfluous type entries...
// //
// Grab the 'type' plane of the module symbol... // Grab the 'type' plane of the module symbol...
@@ -434,10 +246,10 @@ static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
} }
// fixLocalProblems - Loop through the function and fix problems with the PHI // runOnMethod - Loop through the function and fix problems with the PHI nodes
// nodes in the current function. The problem is that PHI nodes might exist // in the current function. The problem is that PHI nodes might exist with
// with multiple entries for the same predecessor. GCC sometimes generates code // multiple entries for the same predecessor. GCC sometimes generates code that
// that looks like this: // looks like this:
// //
// bb7: br bool %cond1004, label %bb8, label %bb8 // bb7: br bool %cond1004, label %bb8, label %bb8
// bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ] // bb8: %reg119 = phi uint [ 0, %bb7 ], [ 1, %bb7 ]
@@ -450,7 +262,7 @@ static inline void RefactorPredecessor(BasicBlock *BB, BasicBlock *Pred) {
// bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ] // bb8: %reg119 = phi uint [ 0, %bbX ], [ 1, %bb7 ]
// //
// //
static bool fixLocalProblems(Function *M) { bool CleanupGCCOutput::runOnMethod(Function *M) {
bool Changed = false; bool Changed = false;
// Don't use iterators because invalidation gets messy... // Don't use iterators because invalidation gets messy...
for (unsigned MI = 0; MI < M->size(); ++MI) { for (unsigned MI = 0; MI < M->size(); ++MI) {
@@ -480,19 +292,9 @@ static bool fixLocalProblems(Function *M) {
return Changed; return Changed;
} }
// runOnFunction - This method simplifies the specified function hopefully.
//
bool CleanupGCCOutput::runOnMethod(Function *F) {
return fixLocalProblems(F);
}
bool CleanupGCCOutput::doFinalization(Module *M) { bool CleanupGCCOutput::doFinalization(Module *M) {
bool Changed = false; bool Changed = false;
if (M->hasSymbolTable()) { if (M->hasSymbolTable()) {
SymbolTable *ST = M->getSymbolTable(); SymbolTable *ST = M->getSymbolTable();
const std::set<const Type *> &UsedTypes = const std::set<const Type *> &UsedTypes =
@@ -523,7 +325,203 @@ bool CleanupGCCOutput::doFinalization(Module *M) {
return Changed; return Changed;
} }
Pass *createCleanupGCCOutputPass() {
return new CleanupGCCOutput(); //===----------------------------------------------------------------------===//
//
// FunctionResolvingPass - Go over the functions that are in the module and
// look for functions that have the same name. More often than not, there will
// be things like:
// void "foo"(...)
// void "foo"(int, int)
// because of the way things are declared in C. If this is the case, patch
// things up.
//
//===----------------------------------------------------------------------===//
namespace {
struct FunctionResolvingPass : public Pass {
bool run(Module *M);
};
} }
// ConvertCallTo - Convert a call to a varargs function with no arg types
// specified to a concrete nonvarargs function.
//
static void ConvertCallTo(CallInst *CI, Function *Dest) {
const FunctionType::ParamTypes &ParamTys =
Dest->getFunctionType()->getParamTypes();
BasicBlock *BB = CI->getParent();
// Get an iterator to where we want to insert cast instructions if the
// argument types don't agree.
//
BasicBlock::iterator BBI = find(BB->begin(), BB->end(), CI);
assert(BBI != BB->end() && "CallInst not in parent block?");
assert(CI->getNumOperands()-1 == ParamTys.size()&&
"Function calls resolved funny somehow, incompatible number of args");
vector<Value*> Params;
// Convert all of the call arguments over... inserting cast instructions if
// the types are not compatible.
for (unsigned i = 1; i < CI->getNumOperands(); ++i) {
Value *V = CI->getOperand(i);
if (V->getType() != ParamTys[i-1]) { // Must insert a cast...
Instruction *Cast = new CastInst(V, ParamTys[i-1]);
BBI = BB->getInstList().insert(BBI, Cast)+1;
V = Cast;
}
Params.push_back(V);
}
// Replace the old call instruction with a new call instruction that calls
// the real function.
//
ReplaceInstWithInst(BB->getInstList(), BBI, new CallInst(Dest, Params));
}
bool FunctionResolvingPass::run(Module *M) {
SymbolTable *ST = M->getSymbolTable();
if (!ST) return false;
std::map<string, vector<Function*> > Functions;
// Loop over the entries in the symbol table. If an entry is a func pointer,
// then add it to the Functions map. We do a two pass algorithm here to avoid
// problems with iterators getting invalidated if we did a one pass scheme.
//
for (SymbolTable::iterator I = ST->begin(), E = ST->end(); I != E; ++I)
if (const PointerType *PT = dyn_cast<PointerType>(I->first))
if (isa<FunctionType>(PT->getElementType())) {
SymbolTable::VarMap &Plane = I->second;
for (SymbolTable::type_iterator PI = Plane.begin(), PE = Plane.end();
PI != PE; ++PI) {
const string &Name = PI->first;
Functions[Name].push_back(cast<Function>(PI->second));
}
}
bool Changed = false;
// Now we have a list of all functions with a particular name. If there is
// more than one entry in a list, merge the functions together.
//
for (std::map<string, vector<Function*> >::iterator I = Functions.begin(),
E = Functions.end(); I != E; ++I) {
vector<Function*> &Functions = I->second;
Function *Implementation = 0; // Find the implementation
Function *Concrete = 0;
for (unsigned i = 0; i < Functions.size(); ) {
if (!Functions[i]->isExternal()) { // Found an implementation
assert(Implementation == 0 && "Multiple definitions of the same"
" function. Case not handled yet!");
Implementation = Functions[i];
} else {
// Ignore functions that are never used so they don't cause spurious
// warnings... here we will actually DCE the function so that it isn't
// used later.
//
if (Functions[i]->use_size() == 0) {
M->getFunctionList().remove(Functions[i]);
delete Functions[i];
Functions.erase(Functions.begin()+i);
Changed = true;
continue;
}
}
if (Functions[i] && (!Functions[i]->getFunctionType()->isVarArg())) {
if (Concrete) { // Found two different functions types. Can't choose
Concrete = 0;
break;
}
Concrete = Functions[i];
}
++i;
}
if (Functions.size() > 1) { // Found a multiply defined function...
// We should find exactly one non-vararg function definition, which is
// probably the implementation. Change all of the function definitions
// and uses to use it instead.
//
if (!Concrete) {
cerr << "Warning: Found functions types that are not compatible:\n";
for (unsigned i = 0; i < Functions.size(); ++i) {
cerr << "\t" << Functions[i]->getType()->getDescription() << " %"
<< Functions[i]->getName() << "\n";
}
cerr << " No linkage of functions named '" << Functions[0]->getName()
<< "' performed!\n";
} else {
for (unsigned i = 0; i < Functions.size(); ++i)
if (Functions[i] != Concrete) {
Function *Old = Functions[i];
const FunctionType *OldMT = Old->getFunctionType();
const FunctionType *ConcreteMT = Concrete->getFunctionType();
bool Broken = false;
assert(Old->getReturnType() == Concrete->getReturnType() &&
"Differing return types not handled yet!");
assert(OldMT->getParamTypes().size() <=
ConcreteMT->getParamTypes().size() &&
"Concrete type must have more specified parameters!");
// Check to make sure that if there are specified types, that they
// match...
//
for (unsigned i = 0; i < OldMT->getParamTypes().size(); ++i)
if (OldMT->getParamTypes()[i] != ConcreteMT->getParamTypes()[i]) {
cerr << "Parameter types conflict for" << OldMT
<< " and " << ConcreteMT;
Broken = true;
}
if (Broken) break; // Can't process this one!
// Attempt to convert all of the uses of the old function to the
// concrete form of the function. If there is a use of the fn
// that we don't understand here we punt to avoid making a bad
// transformation.
//
// At this point, we know that the return values are the same for
// our two functions and that the Old function has no varargs fns
// specified. In otherwords it's just <retty> (...)
//
for (unsigned i = 0; i < Old->use_size(); ) {
User *U = *(Old->use_begin()+i);
if (CastInst *CI = dyn_cast<CastInst>(U)) {
// Convert casts directly
assert(CI->getOperand(0) == Old);
CI->setOperand(0, Concrete);
Changed = true;
} else if (CallInst *CI = dyn_cast<CallInst>(U)) {
// Can only fix up calls TO the argument, not args passed in.
if (CI->getCalledValue() == Old) {
ConvertCallTo(CI, Concrete);
Changed = true;
} else {
cerr << "Couldn't cleanup this function call, must be an"
<< " argument or something!" << CI;
++i;
}
} else {
cerr << "Cannot convert use of function: " << U << "\n";
++i;
}
}
}
}
}
}
return Changed;
}
Pass *createFunctionResolvingPass() {
return new FunctionResolvingPass();
}