using namespace llvm;
STATISTIC(NumMarked , "Number of globals marked constant");
+STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
STATISTIC(NumNestRemoved , "Number of nest attributes removed");
STATISTIC(NumAliasesResolved, "Number of global aliases resolved");
STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated");
+STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed");
namespace {
+ struct GlobalStatus;
struct GlobalOpt : public ModulePass {
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
}
static char ID; // Pass identification, replacement for typeid
- GlobalOpt() : ModulePass(ID) {}
+ GlobalOpt() : ModulePass(ID) {
+ initializeGlobalOptPass(*PassRegistry::getPassRegistry());
+ }
bool runOnModule(Module &M);
bool OptimizeGlobalVars(Module &M);
bool OptimizeGlobalAliases(Module &M);
bool OptimizeGlobalCtorsList(GlobalVariable *&GCL);
- bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
+ bool ProcessGlobal(GlobalVariable *GV,Module::global_iterator &GVI);
+ bool ProcessInternalGlobal(GlobalVariable *GV,Module::global_iterator &GVI,
+ const SmallPtrSet<const PHINode*, 16> &PHIUsers,
+ const GlobalStatus &GS);
+ bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn);
};
}
char GlobalOpt::ID = 0;
INITIALIZE_PASS(GlobalOpt, "globalopt",
- "Global Variable Optimizer", false, false);
+ "Global Variable Optimizer", false, false)
ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); }
/// about it. If we find out that the address of the global is taken, none of
/// this info will be accurate.
struct GlobalStatus {
+ /// isCompared - True if the global's address is used in a comparison.
+ bool isCompared;
+
/// isLoaded - True if the global is ever loaded. If the global isn't ever
/// loaded it can be deleted.
bool isLoaded;
/// HasPHIUser - Set to true if this global has a user that is a PHI node.
bool HasPHIUser;
-
- GlobalStatus() : isLoaded(false), StoredType(NotStored), StoredOnceValue(0),
- AccessingFunction(0), HasMultipleAccessingFunctions(false),
- HasNonInstructionUser(false), HasPHIUser(false) {}
+
+ GlobalStatus() : isCompared(false), isLoaded(false), StoredType(NotStored),
+ StoredOnceValue(0), AccessingFunction(0),
+ HasMultipleAccessingFunctions(false), HasNonInstructionUser(false),
+ HasPHIUser(false) {}
};
}
const User *U = *UI;
if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
GS.HasNonInstructionUser = true;
+
+ // If the result of the constantexpr isn't pointer type, then we won't
+ // know to expect it in various places. Just reject early.
+ if (!isa<PointerType>(CE->getType())) return true;
+
if (AnalyzeGlobal(CE, GS, PHIUsers)) return true;
} else if (const Instruction *I = dyn_cast<Instruction>(U)) {
if (!GS.HasMultipleAccessingFunctions) {
if (AnalyzeGlobal(I, GS, PHIUsers)) return true;
GS.HasPHIUser = true;
} else if (isa<CmpInst>(I)) {
- // Nothing to analyse.
+ GS.isCompared = true;
} else if (isa<MemTransferInst>(I)) {
const MemTransferInst *MTI = cast<MemTransferInst>(I);
if (MTI->getArgOperand(0) == V)
if (Init)
SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
Changed |= CleanupConstantGlobalUsers(CE, SubInit);
- } else if (CE->getOpcode() == Instruction::BitCast &&
+ } else if (CE->getOpcode() == Instruction::BitCast &&
CE->getType()->isPointerTy()) {
// Pointer cast, delete any stores and memsets to the global.
Changed |= CleanupConstantGlobalUsers(CE, 0);
// and will invalidate our notion of what Init is.
Constant *SubInit = 0;
if (!isa<ConstantExpr>(GEP->getOperand(0))) {
- ConstantExpr *CE =
+ ConstantExpr *CE =
dyn_cast_or_null<ConstantExpr>(ConstantFoldInstruction(GEP));
if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE);
// We might have a dead and dangling constant hanging off of here.
if (Constant *C = dyn_cast<Constant>(V))
return SafeToDestroyConstant(C);
-
+
Instruction *I = dyn_cast<Instruction>(V);
if (!I) return false;
// Stores *to* the pointer are ok.
if (StoreInst *SI = dyn_cast<StoreInst>(I))
return SI->getOperand(0) != V;
-
+
// Otherwise, it must be a GEP.
GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I);
if (GEPI == 0) return false;
-
+
if (GEPI->getNumOperands() < 3 || !isa<Constant>(GEPI->getOperand(1)) ||
!cast<Constant>(GEPI->getOperand(1))->isNullValue())
return false;
-
+
for (Value::use_iterator I = GEPI->use_begin(), E = GEPI->use_end();
I != E; ++I)
if (!isSafeSROAElementUse(*I))
///
static bool IsUserOfGlobalSafeForSRA(User *U, GlobalValue *GV) {
// The user of the global must be a GEP Inst or a ConstantExpr GEP.
- if (!isa<GetElementPtrInst>(U) &&
- (!isa<ConstantExpr>(U) ||
+ if (!isa<GetElementPtrInst>(U) &&
+ (!isa<ConstantExpr>(U) ||
cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
return false;
-
+
// Check to see if this ConstantExpr GEP is SRA'able. In particular, we
// don't like < 3 operand CE's, and we don't like non-constant integer
// indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
++GEPI; // Skip over the pointer index.
-
+
// If this is a use of an array allocation, do a bit more checking for sanity.
if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPI)) {
uint64_t NumElements = AT->getNumElements();
ConstantInt *Idx = cast<ConstantInt>(U->getOperand(2));
-
+
// Check to make sure that index falls within the array. If not,
// something funny is going on, so we won't do the optimization.
//
if (Idx->getZExtValue() >= NumElements)
return false;
-
+
// We cannot scalar repl this level of the array unless any array
// sub-indices are in-range constants. In particular, consider:
// A[0][i]. We cannot know that the user isn't doing invalid things like
"Indexed GEP type is not array, vector, or struct!");
continue;
}
-
+
ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
if (!IdxVal || IdxVal->getZExtValue() >= NumElements)
return false;
}
return true;
}
-
+
/// SRAGlobal - Perform scalar replacement of aggregates on the specified global
/// variable. This opens the door for other optimizations by exposing the
// Make sure this global only has simple uses that we can SRA.
if (!GlobalUsersSafeToSRA(GV))
return 0;
-
+
assert(GV->hasLocalLinkage() && !GV->isConstant());
Constant *Init = GV->getInitializer();
const Type *Ty = Init->getType();
unsigned StartAlignment = GV->getAlignment();
if (StartAlignment == 0)
StartAlignment = TD.getABITypeAlignment(GV->getType());
-
+
if (const StructType *STy = dyn_cast<StructType>(Ty)) {
NewGlobals.reserve(STy->getNumElements());
const StructLayout &Layout = *TD.getStructLayout(STy);
GV->getType()->getAddressSpace());
Globals.insert(GV, NGV);
NewGlobals.push_back(NGV);
-
+
// Calculate the known alignment of the field. If the original aggregate
// had 256 byte alignment for example, something might depend on that:
// propagate info to each field.
if (NumElements > 16 && GV->hasNUsesOrMore(16))
return 0; // It's not worth it.
NewGlobals.reserve(NumElements);
-
+
uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType());
unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType());
for (unsigned i = 0, e = NumElements; i != e; ++i) {
GV->getType()->getAddressSpace());
Globals.insert(GV, NGV);
NewGlobals.push_back(NGV);
-
+
// Calculate the known alignment of the field. If the original aggregate
// had 256 byte alignment for example, something might depend on that:
// propagate info to each field.
if (NewGlobals.empty())
return 0;
-
+
DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV);
Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
}
/// AllUsesOfValueWillTrapIfNull - Return true if all users of the specified
-/// value will trap if the value is dynamically null. PHIs keeps track of any
+/// value will trap if the value is dynamically null. PHIs keeps track of any
/// phi nodes we've seen to avoid reprocessing them.
static bool AllUsesOfValueWillTrapIfNull(const Value *V,
SmallPtrSet<const PHINode*, 8> &PHIs) {
// Keep track of whether we are able to remove all the uses of the global
// other than the store that defines it.
bool AllNonStoreUsesGone = true;
-
+
// Replace all uses of loads with uses of uses of the stored value.
for (Value::use_iterator GUI = GV->use_begin(), E = GV->use_end(); GUI != E;){
User *GlobalUser = *GUI++;
ConstantInt *NElements,
TargetData* TD) {
DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n');
-
+
const Type *GlobalType;
if (NElements->getZExtValue() == 1)
GlobalType = AllocTy;
// Create the new global variable. The contents of the malloc'd memory is
// undefined, so initialize with an undef value.
- GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
+ GlobalVariable *NewGV = new GlobalVariable(*GV->getParent(),
GlobalType, false,
GlobalValue::InternalLinkage,
UndefValue::get(GlobalType),
GV->getName()+".body",
GV,
GV->isThreadLocal());
-
+
// If there are bitcast users of the malloc (which is typical, usually we have
// a malloc + bitcast) then replace them with uses of the new global. Update
// other users to use the global as well.
User->replaceUsesOfWith(CI, TheBC);
}
}
-
+
Constant *RepValue = NewGV;
if (NewGV->getType() != GV->getType()->getElementType())
- RepValue = ConstantExpr::getBitCast(RepValue,
+ RepValue = ConstantExpr::getBitCast(RepValue,
GV->getType()->getElementType());
// If there is a comparison against null, we will insert a global bool to
SI->eraseFromParent();
continue;
}
-
+
LoadInst *LI = cast<LoadInst>(GV->use_back());
while (!LI->use_empty()) {
Use &LoadUse = LI->use_begin().getUse();
LoadUse = RepValue;
continue;
}
-
+
ICmpInst *ICI = cast<ICmpInst>(LoadUse.getUser());
// Replace the cmp X, 0 with a use of the bool value.
Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI);
if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
continue; // Fine, ignore.
}
-
+
if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
return false; // Storing the pointer itself... bad.
continue; // Otherwise, storing through it, or storing into GV... fine.
}
-
+
// Must index into the array and into the struct.
if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
return false;
continue;
}
-
+
if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
// PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
// cycles.
return false;
continue;
}
-
+
if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs))
return false;
continue;
}
-
+
return false;
}
return true;
/// ReplaceUsesOfMallocWithGlobal - The Alloc pointer is stored into GV
/// somewhere. Transform all uses of the allocation into loads from the
/// global and uses of the resultant pointer. Further, delete the store into
-/// GV. This assumes that these value pass the
+/// GV. This assumes that these value pass the
/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
-static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
+static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
GlobalVariable *GV) {
while (!Alloc->use_empty()) {
Instruction *U = cast<Instruction>(*Alloc->use_begin());
continue;
}
}
-
+
// Insert a load from the global, and use it instead of the malloc.
Value *NL = new LoadInst(GV, GV->getName()+".val", InsertPt);
U->replaceUsesOfWith(Alloc, NL);
for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;
++UI) {
const Instruction *User = cast<Instruction>(*UI);
-
+
// Comparison against null is ok.
if (const ICmpInst *ICI = dyn_cast<ICmpInst>(User)) {
if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
return false;
continue;
}
-
+
// getelementptr is also ok, but only a simple form.
if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
// Must index into the array and into the struct.
if (GEPI->getNumOperands() < 3)
return false;
-
+
// Otherwise the GEP is ok.
continue;
}
-
+
if (const PHINode *PN = dyn_cast<PHINode>(User)) {
if (!LoadUsingPHIsPerLoad.insert(PN))
// This means some phi nodes are dependent on each other.
if (!LoadUsingPHIs.insert(PN))
// If we have already analyzed this PHI, then it is safe.
continue;
-
+
// Make sure all uses of the PHI are simple enough to transform.
if (!LoadUsesSimpleEnoughForHeapSRA(PN,
LoadUsingPHIs, LoadUsingPHIsPerLoad))
return false;
-
+
continue;
}
-
+
// Otherwise we don't know what this is, not ok.
return false;
}
-
+
return true;
}
return false;
LoadUsingPHIsPerLoad.clear();
}
-
+
// If we reach here, we know that all uses of the loads and transitive uses
// (through PHI nodes) are simple enough to transform. However, we don't know
- // that all inputs the to the PHI nodes are in the same equivalence sets.
+ // that all inputs the to the PHI nodes are in the same equivalence sets.
// Check to verify that all operands of the PHIs are either PHIS that can be
// transformed, loads from GV, or MI itself.
for (SmallPtrSet<const PHINode*, 32>::const_iterator I = LoadUsingPHIs.begin()
const PHINode *PN = *I;
for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
Value *InVal = PN->getIncomingValue(op);
-
+
// PHI of the stored value itself is ok.
if (InVal == StoredVal) continue;
-
+
if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
// One of the PHIs in our set is (optimistically) ok.
if (LoadUsingPHIs.count(InPN))
continue;
return false;
}
-
+
// Load from GV is ok.
if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
if (LI->getOperand(0) == GV)
continue;
-
+
// UNDEF? NULL?
-
+
// Anything else is rejected.
return false;
}
}
-
+
return true;
}
DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
std::vector<Value*> &FieldVals = InsertedScalarizedValues[V];
-
+
if (FieldNo >= FieldVals.size())
FieldVals.resize(FieldNo+1);
-
+
// If we already have this value, just reuse the previously scalarized
// version.
if (Value *FieldVal = FieldVals[FieldNo])
return FieldVal;
-
+
// Depending on what instruction this is, we have several cases.
Value *Result;
if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
} else if (PHINode *PN = dyn_cast<PHINode>(V)) {
// PN's type is pointer to struct. Make a new PHI of pointer to struct
// field.
- const StructType *ST =
+ const StructType *ST =
cast<StructType>(cast<PointerType>(PN->getType())->getElementType());
-
- Result =
+
+ PHINode *NewPN =
PHINode::Create(PointerType::getUnqual(ST->getElementType(FieldNo)),
+ PN->getNumIncomingValues(),
PN->getName()+".f"+Twine(FieldNo), PN);
+ Result = NewPN;
PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
} else {
llvm_unreachable("Unknown usable value");
Result = 0;
}
-
+
return FieldVals[FieldNo] = Result;
}
/// RewriteHeapSROALoadUser - Given a load instruction and a value derived from
/// the load, rewrite the derived value to use the HeapSRoA'd load.
-static void RewriteHeapSROALoadUser(Instruction *LoadUser,
+static void RewriteHeapSROALoadUser(Instruction *LoadUser,
DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
// If this is a comparison against null, handle it.
// field.
Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
InsertedScalarizedValues, PHIsToRewrite);
-
+
Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
- Constant::getNullValue(NPtr->getType()),
+ Constant::getNullValue(NPtr->getType()),
SCI->getName());
SCI->replaceAllUsesWith(New);
SCI->eraseFromParent();
return;
}
-
+
// Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
&& "Unexpected GEPI!");
-
+
// Load the pointer for this field.
unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
InsertedScalarizedValues, PHIsToRewrite);
-
+
// Create the new GEP idx vector.
SmallVector<Value*, 8> GEPIdx;
GEPIdx.push_back(GEPI->getOperand(1));
GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
-
+
Value *NGEPI = GetElementPtrInst::Create(NewPtr,
GEPIdx.begin(), GEPIdx.end(),
GEPI->getName(), GEPI);
tie(InsertPos, Inserted) =
InsertedScalarizedValues.insert(std::make_pair(PN, std::vector<Value*>()));
if (!Inserted) return;
-
+
// If this is the first time we've seen this PHI, recursively process all
// users.
for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; ) {
/// is a value loaded from the global. Eliminate all uses of Ptr, making them
/// use FieldGlobals instead. All uses of loaded values satisfy
/// AllGlobalLoadUsesSimpleEnoughForHeapSRA.
-static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
+static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
DenseMap<Value*, std::vector<Value*> > &InsertedScalarizedValues,
std::vector<std::pair<PHINode*, unsigned> > &PHIsToRewrite) {
for (Value::use_iterator UI = Load->use_begin(), E = Load->use_end();
Instruction *User = cast<Instruction>(*UI++);
RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
}
-
+
if (Load->use_empty()) {
Load->eraseFromParent();
InsertedScalarizedValues.erase(Load);
// new mallocs at the same place as CI, and N globals.
std::vector<Value*> FieldGlobals;
std::vector<Value*> FieldMallocs;
-
+
for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
const Type *FieldTy = STy->getElementType(FieldNo);
const PointerType *PFieldTy = PointerType::getUnqual(FieldTy);
-
+
GlobalVariable *NGV =
new GlobalVariable(*GV->getParent(),
PFieldTy, false, GlobalValue::InternalLinkage,
GV->getName() + ".f" + Twine(FieldNo), GV,
GV->isThreadLocal());
FieldGlobals.push_back(NGV);
-
+
unsigned TypeSize = TD->getTypeAllocSize(FieldTy);
if (const StructType *ST = dyn_cast<StructType>(FieldTy))
TypeSize = TD->getStructLayout(ST)->getSizeInBytes();
FieldMallocs.push_back(NMI);
new StoreInst(NMI, NGV, CI);
}
-
+
// The tricky aspect of this transformation is handling the case when malloc
// fails. In the original code, malloc failing would set the result pointer
// of malloc to null. In this case, some mallocs could succeed and others
// Split the basic block at the old malloc.
BasicBlock *OrigBB = CI->getParent();
BasicBlock *ContBB = OrigBB->splitBasicBlock(CI, "malloc_cont");
-
+
// Create the block to check the first condition. Put all these blocks at the
// end of the function as they are unlikely to be executed.
BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
"malloc_ret_null",
OrigBB->getParent());
-
+
// Remove the uncond branch from OrigBB to ContBB, turning it into a cond
// branch on RunningOr.
OrigBB->getTerminator()->eraseFromParent();
BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
-
+
// Within the NullPtrBlock, we need to emit a comparison and branch for each
// pointer, because some may be null while others are not.
for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
Value *GVVal = new LoadInst(FieldGlobals[i], "tmp", NullPtrBlock);
- Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
+ Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
Constant::getNullValue(GVVal->getType()),
"tmp");
BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
FreeBlock);
BranchInst::Create(NextBlock, FreeBlock);
-
+
NullPtrBlock = NextBlock;
}
-
+
BranchInst::Create(ContBB, NullPtrBlock);
// CI is no longer needed, remove it.
/// inserted for a given load.
DenseMap<Value*, std::vector<Value*> > InsertedScalarizedValues;
InsertedScalarizedValues[GV] = FieldGlobals;
-
+
std::vector<std::pair<PHINode*, unsigned> > PHIsToRewrite;
-
+
// Okay, the malloc site is completely handled. All of the uses of GV are now
// loads, and all uses of those loads are simple. Rewrite them to use loads
// of the per-field globals instead.
for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); UI != E;) {
Instruction *User = cast<Instruction>(*UI++);
-
+
if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
continue;
}
-
+
// Must be a store of null.
StoreInst *SI = cast<StoreInst>(User);
assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
"Unexpected heap-sra user!");
-
+
// Insert a store of null into each global.
for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
const PointerType *PT = cast<PointerType>(FieldGlobals[i]->getType());
FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
}
}
-
+
// Drop all inter-phi links and any loads that made it this far.
for (DenseMap<Value*, std::vector<Value*> >::iterator
I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
LI->dropAllReferences();
}
-
+
// Delete all the phis and loads now that inter-references are dead.
for (DenseMap<Value*, std::vector<Value*> >::iterator
I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
LI->eraseFromParent();
}
-
+
// The old global is now dead, remove it.
GV->eraseFromParent();
TargetData *TD) {
if (!TD)
return false;
-
+
// If this is a malloc of an abstract type, don't touch it.
if (!AllocTy->isSized())
return false;
GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD);
return true;
}
-
+
// If the allocation is an array of structures, consider transforming this
// into multiple malloc'd arrays, one for each field. This is basically
// SRoA for malloc'd memory.
CI = dyn_cast<BitCastInst>(Malloc) ?
extractMallocCallFromBitCast(Malloc) : cast<CallInst>(Malloc);
}
-
+
GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD);
return true;
}
-
+
return false;
-}
+}
// OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge
// that only one value (besides its initializer) is ever stored to the global.
GV->getInitializer()->isNullValue()) {
if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
if (GV->getInitializer()->getType() != SOVC->getType())
- SOVC =
+ SOVC =
ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
// Optimize away any trapping uses of the loaded value.
return true;
} else if (CallInst *CI = extractMallocCall(StoredOnceVal)) {
const Type* MallocType = getMallocAllocatedType(CI);
- if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
+ if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
GVI, TD))
return true;
}
/// whenever it is used. This exposes the values to other scalar optimizations.
static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
const Type *GVElType = GV->getType()->getElementType();
-
+
// If GVElType is already i1, it is already shrunk. If the type of the GV is
// an FP value, pointer or vector, don't do this optimization because a select
// between them is very expensive and unlikely to lead to later
}
DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV);
-
+
// Create the new global, initializing it to false.
GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
false,
- GlobalValue::InternalLinkage,
+ GlobalValue::InternalLinkage,
ConstantInt::getFalse(GV->getContext()),
GV->getName()+".b",
GV->isThreadLocal());
/// ProcessInternalGlobal - Analyze the specified global variable and optimize
/// it if possible. If we make a change, return true.
-bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
- Module::global_iterator &GVI) {
- SmallPtrSet<const PHINode*, 16> PHIUsers;
- GlobalStatus GS;
+bool GlobalOpt::ProcessGlobal(GlobalVariable *GV,
+ Module::global_iterator &GVI) {
+ if (!GV->hasLocalLinkage())
+ return false;
+
+ // Do more involved optimizations if the global is internal.
GV->removeDeadConstantUsers();
if (GV->use_empty()) {
return true;
}
- if (!AnalyzeGlobal(GV, GS, PHIUsers)) {
-#if 0
- DEBUG(dbgs() << "Global: " << *GV);
- DEBUG(dbgs() << " isLoaded = " << GS.isLoaded << "\n");
- DEBUG(dbgs() << " StoredType = ");
- switch (GS.StoredType) {
- case GlobalStatus::NotStored: DEBUG(dbgs() << "NEVER STORED\n"); break;
- case GlobalStatus::isInitializerStored: DEBUG(dbgs() << "INIT STORED\n");
- break;
- case GlobalStatus::isStoredOnce: DEBUG(dbgs() << "STORED ONCE\n"); break;
- case GlobalStatus::isStored: DEBUG(dbgs() << "stored\n"); break;
- }
- if (GS.StoredType == GlobalStatus::isStoredOnce && GS.StoredOnceValue)
- DEBUG(dbgs() << " StoredOnceValue = " << *GS.StoredOnceValue << "\n");
- if (GS.AccessingFunction && !GS.HasMultipleAccessingFunctions)
- DEBUG(dbgs() << " AccessingFunction = "
- << GS.AccessingFunction->getName() << "\n");
- DEBUG(dbgs() << " HasMultipleAccessingFunctions = "
- << GS.HasMultipleAccessingFunctions << "\n");
- DEBUG(dbgs() << " HasNonInstructionUser = "
- << GS.HasNonInstructionUser<<"\n");
- DEBUG(dbgs() << "\n");
-#endif
-
- // If this is a first class global and has only one accessing function
- // and this function is main (which we know is not recursive we can make
- // this global a local variable) we replace the global with a local alloca
- // in this function.
- //
- // NOTE: It doesn't make sense to promote non single-value types since we
- // are just replacing static memory to stack memory.
- //
- // If the global is in different address space, don't bring it to stack.
- if (!GS.HasMultipleAccessingFunctions &&
- GS.AccessingFunction && !GS.HasNonInstructionUser &&
- GV->getType()->getElementType()->isSingleValueType() &&
- GS.AccessingFunction->getName() == "main" &&
- GS.AccessingFunction->hasExternalLinkage() &&
- GV->getType()->getAddressSpace() == 0) {
- DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
- Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
- ->getEntryBlock().begin());
- const Type* ElemTy = GV->getType()->getElementType();
- // FIXME: Pass Global's alignment when globals have alignment
- AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
- if (!isa<UndefValue>(GV->getInitializer()))
- new StoreInst(GV->getInitializer(), Alloca, &FirstI);
-
- GV->replaceAllUsesWith(Alloca);
+ SmallPtrSet<const PHINode*, 16> PHIUsers;
+ GlobalStatus GS;
+
+ if (AnalyzeGlobal(GV, GS, PHIUsers))
+ return false;
+
+ if (!GS.isCompared && !GV->hasUnnamedAddr()) {
+ GV->setUnnamedAddr(true);
+ NumUnnamed++;
+ }
+
+ if (GV->isConstant() || !GV->hasInitializer())
+ return false;
+
+ return ProcessInternalGlobal(GV, GVI, PHIUsers, GS);
+}
+
+/// ProcessInternalGlobal - Analyze the specified global variable and optimize
+/// it if possible. If we make a change, return true.
+bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV,
+ Module::global_iterator &GVI,
+ const SmallPtrSet<const PHINode*, 16> &PHIUsers,
+ const GlobalStatus &GS) {
+ // If this is a first class global and has only one accessing function
+ // and this function is main (which we know is not recursive we can make
+ // this global a local variable) we replace the global with a local alloca
+ // in this function.
+ //
+ // NOTE: It doesn't make sense to promote non single-value types since we
+ // are just replacing static memory to stack memory.
+ //
+ // If the global is in different address space, don't bring it to stack.
+ if (!GS.HasMultipleAccessingFunctions &&
+ GS.AccessingFunction && !GS.HasNonInstructionUser &&
+ GV->getType()->getElementType()->isSingleValueType() &&
+ GS.AccessingFunction->getName() == "main" &&
+ GS.AccessingFunction->hasExternalLinkage() &&
+ GV->getType()->getAddressSpace() == 0) {
+ DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV);
+ Instruction& FirstI = const_cast<Instruction&>(*GS.AccessingFunction
+ ->getEntryBlock().begin());
+ const Type* ElemTy = GV->getType()->getElementType();
+ // FIXME: Pass Global's alignment when globals have alignment
+ AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI);
+ if (!isa<UndefValue>(GV->getInitializer()))
+ new StoreInst(GV->getInitializer(), Alloca, &FirstI);
+
+ GV->replaceAllUsesWith(Alloca);
+ GV->eraseFromParent();
+ ++NumLocalized;
+ return true;
+ }
+
+ // If the global is never loaded (but may be stored to), it is dead.
+ // Delete it now.
+ if (!GS.isLoaded) {
+ DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
+
+ // Delete any stores we can find to the global. We may not be able to
+ // make it completely dead though.
+ bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
+
+ // If the global is dead now, delete it.
+ if (GV->use_empty()) {
GV->eraseFromParent();
- ++NumLocalized;
- return true;
+ ++NumDeleted;
+ Changed = true;
}
-
- // If the global is never loaded (but may be stored to), it is dead.
- // Delete it now.
- if (!GS.isLoaded) {
- DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV);
-
- // Delete any stores we can find to the global. We may not be able to
- // make it completely dead though.
- bool Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer());
-
- // If the global is dead now, delete it.
- if (GV->use_empty()) {
- GV->eraseFromParent();
- ++NumDeleted;
- Changed = true;
- }
- return Changed;
+ return Changed;
- } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
- DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
- GV->setConstant(true);
+ } else if (GS.StoredType <= GlobalStatus::isInitializerStored) {
+ DEBUG(dbgs() << "MARKING CONSTANT: " << *GV);
+ GV->setConstant(true);
- // Clean up any obviously simplifiable users now.
- CleanupConstantGlobalUsers(GV, GV->getInitializer());
+ // Clean up any obviously simplifiable users now.
+ CleanupConstantGlobalUsers(GV, GV->getInitializer());
+
+ // If the global is dead now, just nuke it.
+ if (GV->use_empty()) {
+ DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
+ << "all users and delete global!\n");
+ GV->eraseFromParent();
+ ++NumDeleted;
+ }
- // If the global is dead now, just nuke it.
- if (GV->use_empty()) {
- DEBUG(dbgs() << " *** Marking constant allowed us to simplify "
- << "all users and delete global!\n");
- GV->eraseFromParent();
- ++NumDeleted;
+ ++NumMarked;
+ return true;
+ } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
+ if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
+ if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
+ GVI = FirstNewGV; // Don't skip the newly produced globals!
+ return true;
+ }
+ } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
+ // If the initial value for the global was an undef value, and if only
+ // one other value was stored into it, we can just change the
+ // initializer to be the stored value, then delete all stores to the
+ // global. This allows us to mark it constant.
+ if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
+ if (isa<UndefValue>(GV->getInitializer())) {
+ // Change the initial value here.
+ GV->setInitializer(SOVConstant);
+
+ // Clean up any obviously simplifiable users now.
+ CleanupConstantGlobalUsers(GV, GV->getInitializer());
+
+ if (GV->use_empty()) {
+ DEBUG(dbgs() << " *** Substituting initializer allowed us to "
+ << "simplify all users and delete global!\n");
+ GV->eraseFromParent();
+ ++NumDeleted;
+ } else {
+ GVI = GV;
+ }
+ ++NumSubstitute;
+ return true;
}
- ++NumMarked;
+ // Try to optimize globals based on the knowledge that only one value
+ // (besides its initializer) is ever stored to the global.
+ if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
+ getAnalysisIfAvailable<TargetData>()))
return true;
- } else if (!GV->getInitializer()->getType()->isSingleValueType()) {
- if (TargetData *TD = getAnalysisIfAvailable<TargetData>())
- if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) {
- GVI = FirstNewGV; // Don't skip the newly produced globals!
- return true;
- }
- } else if (GS.StoredType == GlobalStatus::isStoredOnce) {
- // If the initial value for the global was an undef value, and if only
- // one other value was stored into it, we can just change the
- // initializer to be the stored value, then delete all stores to the
- // global. This allows us to mark it constant.
- if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
- if (isa<UndefValue>(GV->getInitializer())) {
- // Change the initial value here.
- GV->setInitializer(SOVConstant);
-
- // Clean up any obviously simplifiable users now.
- CleanupConstantGlobalUsers(GV, GV->getInitializer());
-
- if (GV->use_empty()) {
- DEBUG(dbgs() << " *** Substituting initializer allowed us to "
- << "simplify all users and delete global!\n");
- GV->eraseFromParent();
- ++NumDeleted;
- } else {
- GVI = GV;
- }
- ++NumSubstitute;
- return true;
- }
- // Try to optimize globals based on the knowledge that only one value
- // (besides its initializer) is ever stored to the global.
- if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GVI,
- getAnalysisIfAvailable<TargetData>()))
+ // Otherwise, if the global was not a boolean, we can shrink it to be a
+ // boolean.
+ if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
+ if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
+ ++NumShrunkToBool;
return true;
-
- // Otherwise, if the global was not a boolean, we can shrink it to be a
- // boolean.
- if (Constant *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue))
- if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
- ++NumShrunkToBool;
- return true;
- }
- }
+ }
}
+
return false;
}
if (New && New != CE)
GV->setInitializer(New);
}
- // Do more involved optimizations if the global is internal.
- if (!GV->isConstant() && GV->hasLocalLinkage() &&
- GV->hasInitializer())
- Changed |= ProcessInternalGlobal(GV, GVI);
+
+ Changed |= ProcessGlobal(GV, GVI);
}
return Changed;
}
/// FindGlobalCtors - Find the llvm.globalctors list, verifying that all
/// initializers have an init priority of 65535.
GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) {
- for (Module::global_iterator I = M.global_begin(), E = M.global_end();
- I != E; ++I)
- if (I->getName() == "llvm.global_ctors") {
- // Found it, verify it's an array of { int, void()* }.
- const ArrayType *ATy =dyn_cast<ArrayType>(I->getType()->getElementType());
- if (!ATy) return 0;
- const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
- if (!STy || STy->getNumElements() != 2 ||
- !STy->getElementType(0)->isIntegerTy(32)) return 0;
- const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
- if (!PFTy) return 0;
- const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
- if (!FTy || !FTy->getReturnType()->isVoidTy() ||
- FTy->isVarArg() || FTy->getNumParams() != 0)
- return 0;
-
- // Verify that the initializer is simple enough for us to handle.
- if (!I->hasDefinitiveInitializer()) return 0;
- ConstantArray *CA = dyn_cast<ConstantArray>(I->getInitializer());
- if (!CA) return 0;
- for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
- if (ConstantStruct *CS = dyn_cast<ConstantStruct>(*i)) {
- if (isa<ConstantPointerNull>(CS->getOperand(1)))
- continue;
+ GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
+ if (GV == 0) return 0;
+
+ // Found it, verify it's an array of { int, void()* }.
+ const ArrayType *ATy =dyn_cast<ArrayType>(GV->getType()->getElementType());
+ if (!ATy) return 0;
+ const StructType *STy = dyn_cast<StructType>(ATy->getElementType());
+ if (!STy || STy->getNumElements() != 2 ||
+ !STy->getElementType(0)->isIntegerTy(32)) return 0;
+ const PointerType *PFTy = dyn_cast<PointerType>(STy->getElementType(1));
+ if (!PFTy) return 0;
+ const FunctionType *FTy = dyn_cast<FunctionType>(PFTy->getElementType());
+ if (!FTy || !FTy->getReturnType()->isVoidTy() ||
+ FTy->isVarArg() || FTy->getNumParams() != 0)
+ return 0;
- // Must have a function or null ptr.
- if (!isa<Function>(CS->getOperand(1)))
- return 0;
-
- // Init priority must be standard.
- ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
- if (!CI || CI->getZExtValue() != 65535)
- return 0;
- } else {
- return 0;
- }
-
- return I;
- }
- return 0;
+ // Verify that the initializer is simple enough for us to handle. We are
+ // only allowed to optimize the initializer if it is unique.
+ if (!GV->hasUniqueInitializer()) return 0;
+
+ ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
+ if (!CA) return 0;
+
+ for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) {
+ ConstantStruct *CS = dyn_cast<ConstantStruct>(*i);
+ if (CS == 0) return 0;
+
+ if (isa<ConstantPointerNull>(CS->getOperand(1)))
+ continue;
+
+ // Must have a function or null ptr.
+ if (!isa<Function>(CS->getOperand(1)))
+ return 0;
+
+ // Init priority must be standard.
+ ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
+ if (!CI || CI->getZExtValue() != 65535)
+ return 0;
+ }
+
+ return GV;
}
/// ParseGlobalCtors - Given a llvm.global_ctors list that we can understand,
/// InstallGlobalCtors - Given a specified llvm.global_ctors list, install the
/// specified array, returning the new global to use.
-static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
+static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL,
const std::vector<Function*> &Ctors) {
// If we made a change, reassemble the initializer list.
std::vector<Constant*> CSVals;
CSVals.push_back(ConstantInt::get(Type::getInt32Ty(GCL->getContext()),65535));
CSVals.push_back(0);
-
+
// Create the new init list.
std::vector<Constant*> CAList;
for (unsigned i = 0, e = Ctors.size(); i != e; ++i) {
}
CAList.push_back(ConstantStruct::get(GCL->getContext(), CSVals, false));
}
-
+
// Create the array initializer.
const Type *StructTy =
cast<ArrayType>(GCL->getType()->getElementType())->getElementType();
- Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
+ Constant *CA = ConstantArray::get(ArrayType::get(StructTy,
CAList.size()), CAList);
-
+
// If we didn't change the number of elements, don't create a new GV.
if (CA->getType() == GCL->getInitializer()->getType()) {
GCL->setInitializer(CA);
return GCL;
}
-
+
// Create the new global and insert it next to the existing list.
GlobalVariable *NGV = new GlobalVariable(CA->getType(), GCL->isConstant(),
GCL->getLinkage(), CA, "",
GCL->isThreadLocal());
GCL->getParent()->getGlobalList().insert(GCL, NGV);
NGV->takeName(GCL);
-
+
// Nuke the old list, replacing any uses with the new one.
if (!GCL->use_empty()) {
Constant *V = NGV;
GCL->replaceAllUsesWith(V);
}
GCL->eraseFromParent();
-
+
if (Ctors.size())
return NGV;
else
}
-static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues,
- Value *V) {
+static Constant *getVal(DenseMap<Value*, Constant*> &ComputedValues, Value *V) {
if (Constant *CV = dyn_cast<Constant>(V)) return CV;
Constant *R = ComputedValues[V];
assert(R && "Reference to an uncomputed value!");
return R;
}
+static inline bool
+isSimpleEnoughValueToCommit(Constant *C,
+ SmallPtrSet<Constant*, 8> &SimpleConstants);
+
+
+/// isSimpleEnoughValueToCommit - Return true if the specified constant can be
+/// handled by the code generator. We don't want to generate something like:
+/// void *X = &X/42;
+/// because the code generator doesn't have a relocation that can handle that.
+///
+/// This function should be called if C was not found (but just got inserted)
+/// in SimpleConstants to avoid having to rescan the same constants all the
+/// time.
+static bool isSimpleEnoughValueToCommitHelper(Constant *C,
+ SmallPtrSet<Constant*, 8> &SimpleConstants) {
+ // Simple integer, undef, constant aggregate zero, global addresses, etc are
+ // all supported.
+ if (C->getNumOperands() == 0 || isa<BlockAddress>(C) ||
+ isa<GlobalValue>(C))
+ return true;
+
+ // Aggregate values are safe if all their elements are.
+ if (isa<ConstantArray>(C) || isa<ConstantStruct>(C) ||
+ isa<ConstantVector>(C)) {
+ for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
+ Constant *Op = cast<Constant>(C->getOperand(i));
+ if (!isSimpleEnoughValueToCommit(Op, SimpleConstants))
+ return false;
+ }
+ return true;
+ }
+
+ // We don't know exactly what relocations are allowed in constant expressions,
+ // so we allow &global+constantoffset, which is safe and uniformly supported
+ // across targets.
+ ConstantExpr *CE = cast<ConstantExpr>(C);
+ switch (CE->getOpcode()) {
+ case Instruction::BitCast:
+ case Instruction::IntToPtr:
+ case Instruction::PtrToInt:
+ // These casts are always fine if the casted value is.
+ return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
+
+ // GEP is fine if it is simple + constant offset.
+ case Instruction::GetElementPtr:
+ for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
+ if (!isa<ConstantInt>(CE->getOperand(i)))
+ return false;
+ return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
+
+ case Instruction::Add:
+ // We allow simple+cst.
+ if (!isa<ConstantInt>(CE->getOperand(1)))
+ return false;
+ return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants);
+ }
+ return false;
+}
+
+static inline bool
+isSimpleEnoughValueToCommit(Constant *C,
+ SmallPtrSet<Constant*, 8> &SimpleConstants) {
+ // If we already checked this constant, we win.
+ if (!SimpleConstants.insert(C)) return true;
+ // Check the constant.
+ return isSimpleEnoughValueToCommitHelper(C, SimpleConstants);
+}
+
+
/// isSimpleEnoughPointerToCommit - Return true if this constant is simple
-/// enough for us to understand. In particular, if it is a cast of something,
-/// we punt. We basically just support direct accesses to globals and GEP's of
+/// enough for us to understand. In particular, if it is a cast to anything
+/// other than from one pointer type to another pointer type, we punt.
+/// We basically just support direct accesses to globals and GEP's of
/// globals. This should be kept up to date with CommitValueTo.
static bool isSimpleEnoughPointerToCommit(Constant *C) {
// Conservatively, avoid aggregate types. This is because we don't
return false;
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C))
- // Do not allow weak/linkonce/dllimport/dllexport linkage or
+ // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
// external globals.
- return GV->hasDefinitiveInitializer();
+ return GV->hasUniqueInitializer();
- if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
// Handle a constantexpr gep.
if (CE->getOpcode() == Instruction::GetElementPtr &&
isa<GlobalVariable>(CE->getOperand(0)) &&
cast<GEPOperator>(CE)->isInBounds()) {
GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
- // Do not allow weak/linkonce/dllimport/dllexport linkage or
+ // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
// external globals.
- if (!GV->hasDefinitiveInitializer())
+ if (!GV->hasUniqueInitializer())
return false;
// The first index must be zero.
return false;
return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE);
+
+ // A constantexpr bitcast from a pointer to another pointer is a no-op,
+ // and we know how to evaluate it by moving the bitcast from the pointer
+ // operand to the value operand.
+ } else if (CE->getOpcode() == Instruction::BitCast &&
+ isa<GlobalVariable>(CE->getOperand(0))) {
+ // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or
+ // external globals.
+ return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer();
}
+ }
+
return false;
}
assert(Val->getType() == Init->getType() && "Type mismatch!");
return Val;
}
-
+
std::vector<Constant*> Elts;
if (const StructType *STy = dyn_cast<StructType>(Init->getType())) {
llvm_unreachable("This code is out of sync with "
" ConstantFoldLoadThroughGEPConstantExpr");
}
-
+
// Replace the element that we are supposed to.
ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
unsigned Idx = CU->getZExtValue();
assert(Idx < STy->getNumElements() && "Struct index out of range!");
Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
-
+
// Return the modified struct.
return ConstantStruct::get(Init->getContext(), &Elts[0], Elts.size(),
STy->isPacked());
NumElts = ATy->getNumElements();
else
NumElts = cast<VectorType>(InitTy)->getNumElements();
-
-
+
+
// Break up the array into elements.
if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) {
for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i)
" ConstantFoldLoadThroughGEPConstantExpr");
Elts.assign(NumElts, UndefValue::get(InitTy->getElementType()));
}
-
+
assert(CI->getZExtValue() < NumElts);
Elts[CI->getZExtValue()] =
EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
-
+
if (Init->getType()->isArrayTy())
return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
- else
- return ConstantVector::get(&Elts[0], Elts.size());
- }
+ return ConstantVector::get(Elts);
+ }
}
/// CommitValueTo - We have decided that Addr (which satisfies the predicate
// is the most up-to-date.
DenseMap<Constant*, Constant*>::const_iterator I = Memory.find(P);
if (I != Memory.end()) return I->second;
-
+
// Access it.
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
if (GV->hasDefinitiveInitializer())
return GV->getInitializer();
return 0;
}
-
+
// Handle a constantexpr getelementptr.
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P))
if (CE->getOpcode() == Instruction::GetElementPtr &&
const SmallVectorImpl<Constant*> &ActualArgs,
std::vector<Function*> &CallStack,
DenseMap<Constant*, Constant*> &MutatedMemory,
- std::vector<GlobalVariable*> &AllocaTmps) {
+ std::vector<GlobalVariable*> &AllocaTmps,
+ SmallPtrSet<Constant*, 8> &SimpleConstants,
+ const TargetData *TD) {
// Check to see if this function is already executing (recursion). If so,
// bail out. TODO: we might want to accept limited recursion.
if (std::find(CallStack.begin(), CallStack.end(), F) != CallStack.end())
return false;
-
+
CallStack.push_back(F);
-
+
/// Values - As we compute SSA register values, we store their contents here.
DenseMap<Value*, Constant*> Values;
-
+
// Initialize arguments to the incoming values specified.
unsigned ArgNo = 0;
for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E;
/// we can only evaluate any one basic block at most once. This set keeps
/// track of what we have executed so we can detect recursive cases etc.
SmallPtrSet<BasicBlock*, 32> ExecutedBlocks;
-
+
// CurInst - The current instruction we're evaluating.
BasicBlock::iterator CurInst = F->begin()->begin();
-
+
// This is the main evaluation loop.
while (1) {
Constant *InstResult = 0;
-
+
if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
if (SI->isVolatile()) return false; // no volatile accesses.
Constant *Ptr = getVal(Values, SI->getOperand(1));
if (!isSimpleEnoughPointerToCommit(Ptr))
// If this is too complex for us to commit, reject it.
return false;
+
Constant *Val = getVal(Values, SI->getOperand(0));
+
+ // If this might be too difficult for the backend to handle (e.g. the addr
+ // of one global variable divided by another) then we can't commit it.
+ if (!isSimpleEnoughValueToCommit(Val, SimpleConstants))
+ return false;
+
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
+ if (CE->getOpcode() == Instruction::BitCast) {
+ // If we're evaluating a store through a bitcast, then we need
+ // to pull the bitcast off the pointer type and push it onto the
+ // stored value.
+ Ptr = CE->getOperand(0);
+
+ const Type *NewTy=cast<PointerType>(Ptr->getType())->getElementType();
+
+ // In order to push the bitcast onto the stored value, a bitcast
+ // from NewTy to Val's type must be legal. If it's not, we can try
+ // introspecting NewTy to find a legal conversion.
+ while (!Val->getType()->canLosslesslyBitCastTo(NewTy)) {
+ // If NewTy is a struct, we can convert the pointer to the struct
+ // into a pointer to its first member.
+ // FIXME: This could be extended to support arrays as well.
+ if (const StructType *STy = dyn_cast<StructType>(NewTy)) {
+ NewTy = STy->getTypeAtIndex(0U);
+
+ const IntegerType *IdxTy =IntegerType::get(NewTy->getContext(), 32);
+ Constant *IdxZero = ConstantInt::get(IdxTy, 0, false);
+ Constant * const IdxList[] = {IdxZero, IdxZero};
+
+ Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList, 2);
+
+ // If we can't improve the situation by introspecting NewTy,
+ // we have to give up.
+ } else {
+ return 0;
+ }
+ }
+
+ // If we found compatible types, go ahead and push the bitcast
+ // onto the stored value.
+ Val = ConstantExpr::getBitCast(Val, NewTy);
+ }
+
MutatedMemory[Ptr] = Val;
} else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) {
InstResult = ConstantExpr::get(BO->getOpcode(),
GlobalValue::InternalLinkage,
UndefValue::get(Ty),
AI->getName()));
- InstResult = AllocaTmps.back();
+ InstResult = AllocaTmps.back();
} else if (CallInst *CI = dyn_cast<CallInst>(CurInst)) {
// Debug info can safely be ignored here.
} else {
if (Callee->getFunctionType()->isVarArg())
return false;
-
+
Constant *RetVal;
// Execute the call, if successful, use the return value.
if (!EvaluateFunction(Callee, RetVal, Formals, CallStack,
- MutatedMemory, AllocaTmps))
+ MutatedMemory, AllocaTmps, SimpleConstants, TD))
return false;
InstResult = RetVal;
}
dyn_cast<ConstantInt>(getVal(Values, BI->getCondition()));
if (!Cond) return false; // Cannot determine.
- NewBB = BI->getSuccessor(!Cond->getZExtValue());
+ NewBB = BI->getSuccessor(!Cond->getZExtValue());
}
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) {
ConstantInt *Val =
} else if (ReturnInst *RI = dyn_cast<ReturnInst>(CurInst)) {
if (RI->getNumOperands())
RetVal = getVal(Values, RI->getOperand(0));
-
+
CallStack.pop_back(); // return from fn.
return true; // We succeeded at evaluating this ctor!
} else {
// invoke, unwind, unreachable.
return false; // Cannot handle this terminator.
}
-
+
// Okay, we succeeded in evaluating this control flow. See if we have
// executed the new block before. If so, we have a looping function,
// which we cannot evaluate in reasonable time.
if (!ExecutedBlocks.insert(NewBB))
return false; // looped!
-
+
// Okay, we have never been in this block before. Check to see if there
// are any PHI nodes. If so, evaluate them with information about where
// we came from.
// Did not know how to evaluate this!
return false;
}
-
- if (!CurInst->use_empty())
+
+ if (!CurInst->use_empty()) {
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(InstResult))
+ InstResult = ConstantFoldConstantExpression(CE, TD);
+
Values[CurInst] = InstResult;
-
+ }
+
// Advance program counter.
++CurInst;
}
/// EvaluateStaticConstructor - Evaluate static constructors in the function, if
/// we can. Return true if we can, false otherwise.
-static bool EvaluateStaticConstructor(Function *F) {
+static bool EvaluateStaticConstructor(Function *F, const TargetData *TD) {
/// MutatedMemory - For each store we execute, we update this map. Loads
/// check this to get the most up-to-date value. If evaluation is successful,
/// this state is committed to the process.
/// to represent its body. This vector is needed so we can delete the
/// temporary globals when we are done.
std::vector<GlobalVariable*> AllocaTmps;
-
+
/// CallStack - This is used to detect recursion. In pathological situations
/// we could hit exponential behavior, but at least there is nothing
/// unbounded.
std::vector<Function*> CallStack;
+ /// SimpleConstants - These are constants we have checked and know to be
+ /// simple enough to live in a static initializer of a global.
+ SmallPtrSet<Constant*, 8> SimpleConstants;
+
// Call the function.
Constant *RetValDummy;
bool EvalSuccess = EvaluateFunction(F, RetValDummy,
SmallVector<Constant*, 0>(), CallStack,
- MutatedMemory, AllocaTmps);
+ MutatedMemory, AllocaTmps,
+ SimpleConstants, TD);
+
if (EvalSuccess) {
// We succeeded at evaluation: commit the result.
DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"
E = MutatedMemory.end(); I != E; ++I)
CommitValueTo(I->second, I->first);
}
-
+
// At this point, we are done interpreting. If we created any 'alloca'
// temporaries, release them now.
while (!AllocaTmps.empty()) {
GlobalVariable *Tmp = AllocaTmps.back();
AllocaTmps.pop_back();
-
+
// If there are still users of the alloca, the program is doing something
// silly, e.g. storing the address of the alloca somewhere and using it
// later. Since this is undefined, we'll just make it be null.
Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType()));
delete Tmp;
}
-
+
return EvalSuccess;
}
std::vector<Function*> Ctors = ParseGlobalCtors(GCL);
bool MadeChange = false;
if (Ctors.empty()) return false;
-
+
+ const TargetData *TD = getAnalysisIfAvailable<TargetData>();
// Loop over global ctors, optimizing them when we can.
for (unsigned i = 0; i != Ctors.size(); ++i) {
Function *F = Ctors[i];
}
break;
}
-
+
// We cannot simplify external ctor functions.
if (F->empty()) continue;
-
+
// If we can evaluate the ctor at compile time, do.
- if (EvaluateStaticConstructor(F)) {
+ if (EvaluateStaticConstructor(F, TD)) {
Ctors.erase(Ctors.begin()+i);
MadeChange = true;
--i;
continue;
}
}
-
+
if (!MadeChange) return false;
-
+
GCL = InstallGlobalCtors(GCL, Ctors);
return true;
}
return Changed;
}
+static Function *FindCXAAtExit(Module &M) {
+ Function *Fn = M.getFunction("__cxa_atexit");
+
+ if (!Fn)
+ return 0;
+
+ const FunctionType *FTy = Fn->getFunctionType();
+
+ // Checking that the function has the right return type, the right number of
+ // parameters and that they all have pointer types should be enough.
+ if (!FTy->getReturnType()->isIntegerTy() ||
+ FTy->getNumParams() != 3 ||
+ !FTy->getParamType(0)->isPointerTy() ||
+ !FTy->getParamType(1)->isPointerTy() ||
+ !FTy->getParamType(2)->isPointerTy())
+ return 0;
+
+ return Fn;
+}
+
+/// cxxDtorIsEmpty - Returns whether the given function is an empty C++
+/// destructor and can therefore be eliminated.
+/// Note that we assume that other optimization passes have already simplified
+/// the code so we only look for a function with a single basic block, where
+/// the only allowed instructions are 'ret' or 'call' to empty C++ dtor.
+static bool cxxDtorIsEmpty(const Function &Fn,
+ SmallPtrSet<const Function *, 8> &CalledFunctions) {
+ // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
+ // nounwind, but that doesn't seem worth doing.
+ if (Fn.isDeclaration())
+ return false;
+
+ if (++Fn.begin() != Fn.end())
+ return false;
+
+ const BasicBlock &EntryBlock = Fn.getEntryBlock();
+ for (BasicBlock::const_iterator I = EntryBlock.begin(), E = EntryBlock.end();
+ I != E; ++I) {
+ if (const CallInst *CI = dyn_cast<CallInst>(I)) {
+ // Ignore debug intrinsics.
+ if (isa<DbgInfoIntrinsic>(CI))
+ continue;
+
+ const Function *CalledFn = CI->getCalledFunction();
+
+ if (!CalledFn)
+ return false;
+
+ SmallPtrSet<const Function *, 8> NewCalledFunctions(CalledFunctions);
+
+ // Don't treat recursive functions as empty.
+ if (!NewCalledFunctions.insert(CalledFn))
+ return false;
+
+ if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions))
+ return false;
+ } else if (isa<ReturnInst>(*I))
+ return true;
+ else
+ return false;
+ }
+
+ return false;
+}
+
+bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
+ /// Itanium C++ ABI p3.3.5:
+ ///
+ /// After constructing a global (or local static) object, that will require
+ /// destruction on exit, a termination function is registered as follows:
+ ///
+ /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
+ ///
+ /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
+ /// call f(p) when DSO d is unloaded, before all such termination calls
+ /// registered before this one. It returns zero if registration is
+ /// successful, nonzero on failure.
+
+ // This pass will look for calls to __cxa_atexit where the function is trivial
+ // and remove them.
+ bool Changed = false;
+
+ for (Function::use_iterator I = CXAAtExitFn->use_begin(),
+ E = CXAAtExitFn->use_end(); I != E;) {
+ // We're only interested in calls. Theoretically, we could handle invoke
+ // instructions as well, but neither llvm-gcc nor clang generate invokes
+ // to __cxa_atexit.
+ CallInst *CI = dyn_cast<CallInst>(*I++);
+ if (!CI)
+ continue;
+
+ Function *DtorFn =
+ dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
+ if (!DtorFn)
+ continue;
+
+ SmallPtrSet<const Function *, 8> CalledFunctions;
+ if (!cxxDtorIsEmpty(*DtorFn, CalledFunctions))
+ continue;
+
+ // Just remove the call.
+ CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
+ CI->eraseFromParent();
+
+ ++NumCXXDtorsRemoved;
+
+ Changed |= true;
+ }
+
+ return Changed;
+}
+
bool GlobalOpt::runOnModule(Module &M) {
bool Changed = false;
-
+
// Try to find the llvm.globalctors list.
GlobalVariable *GlobalCtors = FindGlobalCtors(M);
+ Function *CXAAtExitFn = FindCXAAtExit(M);
+
bool LocalChange = true;
while (LocalChange) {
LocalChange = false;
-
+
// Delete functions that are trivially dead, ccc -> fastcc
LocalChange |= OptimizeFunctions(M);
-
+
// Optimize global_ctors list.
if (GlobalCtors)
LocalChange |= OptimizeGlobalCtorsList(GlobalCtors);
-
+
// Optimize non-address-taken globals.
LocalChange |= OptimizeGlobalVars(M);
// Resolve aliases, when possible.
LocalChange |= OptimizeGlobalAliases(M);
+
+ // Try to remove trivial global destructors.
+ if (CXAAtExitFn)
+ LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
+
Changed |= LocalChange;
}
-
+
// TODO: Move all global ctors functions to the end of the module for code
// layout.
-
+
return Changed;
}
projects / oota-llvm.git / blobdiff