X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FIPO%2FGlobalOpt.cpp;h=c753e2a85dbde92fa36195d21889281e3c023117;hb=8e47daf2858e980210f3e1f007036b24da342c29;hp=0283568f2c2e504ae31a139dbc0330767f4c1026;hpb=10c6d12a9fd4dab411091f64db4db69670b88850;p=oota-llvm.git diff --git a/lib/Transforms/IPO/GlobalOpt.cpp b/lib/Transforms/IPO/GlobalOpt.cpp index 0283568f2c2..c753e2a85db 100644 --- a/lib/Transforms/IPO/GlobalOpt.cpp +++ b/lib/Transforms/IPO/GlobalOpt.cpp @@ -15,28 +15,29 @@ #define DEBUG_TYPE "globalopt" #include "llvm/Transforms/IPO.h" -#include "llvm/CallingConv.h" -#include "llvm/Constants.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Instructions.h" -#include "llvm/IntrinsicInst.h" -#include "llvm/Module.h" -#include "llvm/Operator.h" -#include "llvm/Pass.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/ADT/SmallPtrSet.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/Statistic.h" #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/MemoryBuiltins.h" -#include "llvm/Target/TargetData.h" +#include "llvm/IR/CallingConv.h" +#include "llvm/IR/Constants.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/DerivedTypes.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/Module.h" +#include "llvm/IR/Operator.h" +#include "llvm/Pass.h" #include "llvm/Support/CallSite.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/GetElementPtrTypeIterator.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/ADT/DenseMap.h" -#include "llvm/ADT/SmallPtrSet.h" -#include "llvm/ADT/SmallVector.h" -#include "llvm/ADT/Statistic.h" -#include "llvm/ADT/STLExtras.h" +#include "llvm/Target/TargetLibraryInfo.h" #include using namespace llvm; @@ -61,6 +62,7 @@ namespace { struct GlobalStatus; struct GlobalOpt : public ModulePass { virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.addRequired(); } static char ID; // Pass identification, replacement for typeid GlobalOpt() : ModulePass(ID) { @@ -80,11 +82,17 @@ namespace { const SmallPtrSet &PHIUsers, const GlobalStatus &GS); bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn); + + DataLayout *TD; + TargetLibraryInfo *TLI; }; } char GlobalOpt::ID = 0; -INITIALIZE_PASS(GlobalOpt, "globalopt", +INITIALIZE_PASS_BEGIN(GlobalOpt, "globalopt", + "Global Variable Optimizer", false, false) +INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) +INITIALIZE_PASS_END(GlobalOpt, "globalopt", "Global Variable Optimizer", false, false) ModulePass *llvm::createGlobalOptimizerPass() { return new GlobalOpt(); } @@ -140,21 +148,30 @@ struct GlobalStatus { /// an instruction (e.g. a constant expr or GV initializer). bool HasNonInstructionUser; - /// HasPHIUser - Set to true if this global has a user that is a PHI node. - bool HasPHIUser; + /// AtomicOrdering - Set to the strongest atomic ordering requirement. + AtomicOrdering Ordering; GlobalStatus() : isCompared(false), isLoaded(false), StoredType(NotStored), StoredOnceValue(0), AccessingFunction(0), - HasMultipleAccessingFunctions(false), HasNonInstructionUser(false), - HasPHIUser(false) {} + HasMultipleAccessingFunctions(false), + HasNonInstructionUser(false), Ordering(NotAtomic) {} }; } -// SafeToDestroyConstant - It is safe to destroy a constant iff it is only used -// by constants itself. Note that constants cannot be cyclic, so this test is -// pretty easy to implement recursively. -// +/// StrongerOrdering - Return the stronger of the two ordering. If the two +/// orderings are acquire and release, then return AcquireRelease. +/// +static AtomicOrdering StrongerOrdering(AtomicOrdering X, AtomicOrdering Y) { + if (X == Acquire && Y == Release) return AcquireRelease; + if (Y == Acquire && X == Release) return AcquireRelease; + return (AtomicOrdering)std::max(X, Y); +} + +/// SafeToDestroyConstant - It is safe to destroy a constant iff it is only used +/// by constants itself. Note that constants cannot be cyclic, so this test is +/// pretty easy to implement recursively. +/// static bool SafeToDestroyConstant(const Constant *C) { if (isa(C)) return false; @@ -179,11 +196,11 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, const User *U = *UI; if (const ConstantExpr *CE = dyn_cast(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(CE->getType())) return true; - + if (AnalyzeGlobal(CE, GS, PHIUsers)) return true; } else if (const Instruction *I = dyn_cast(U)) { if (!GS.HasMultipleAccessingFunctions) { @@ -195,12 +212,17 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, } if (const LoadInst *LI = dyn_cast(I)) { GS.isLoaded = true; - if (LI->isVolatile()) return true; // Don't hack on volatile loads. + // Don't hack on volatile loads. + if (LI->isVolatile()) return true; + GS.Ordering = StrongerOrdering(GS.Ordering, LI->getOrdering()); } else if (const StoreInst *SI = dyn_cast(I)) { // Don't allow a store OF the address, only stores TO the address. if (SI->getOperand(0) == V) return true; - if (SI->isVolatile()) return true; // Don't hack on volatile stores. + // Don't hack on volatile stores. + if (SI->isVolatile()) return true; + + GS.Ordering = StrongerOrdering(GS.Ordering, SI->getOrdering()); // If this is a direct store to the global (i.e., the global is a scalar // value, not an aggregate), keep more specific information about @@ -209,6 +231,14 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, if (const GlobalVariable *GV = dyn_cast( SI->getOperand(1))) { Value *StoredVal = SI->getOperand(0); + + if (Constant *C = dyn_cast(StoredVal)) { + if (C->isThreadDependent()) { + // The stored value changes between threads; don't track it. + return true; + } + } + if (StoredVal == GV->getInitializer()) { if (GS.StoredType < GlobalStatus::isInitializerStored) GS.StoredType = GlobalStatus::isInitializerStored; @@ -229,6 +259,8 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, GS.StoredType = GlobalStatus::isStored; } } + } else if (isa(I)) { + if (AnalyzeGlobal(I, GS, PHIUsers)) return true; } else if (isa(I)) { if (AnalyzeGlobal(I, GS, PHIUsers)) return true; } else if (isa(I)) { @@ -238,7 +270,6 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, // have to be careful about infinite recursion. if (PHIUsers.insert(PN)) // Not already visited. if (AnalyzeGlobal(I, GS, PHIUsers)) return true; - GS.HasPHIUser = true; } else if (isa(I)) { GS.isCompared = true; } else if (const MemTransferInst *MTI = dyn_cast(I)) { @@ -269,43 +300,175 @@ static bool AnalyzeGlobal(const Value *V, GlobalStatus &GS, return false; } -static Constant *getAggregateConstantElement(Constant *Agg, Constant *Idx) { - ConstantInt *CI = dyn_cast(Idx); - if (!CI) return 0; - unsigned IdxV = CI->getZExtValue(); - - if (ConstantStruct *CS = dyn_cast(Agg)) { - if (IdxV < CS->getNumOperands()) return CS->getOperand(IdxV); - } else if (ConstantArray *CA = dyn_cast(Agg)) { - if (IdxV < CA->getNumOperands()) return CA->getOperand(IdxV); - } else if (ConstantVector *CP = dyn_cast(Agg)) { - if (IdxV < CP->getNumOperands()) return CP->getOperand(IdxV); - } else if (isa(Agg)) { - if (StructType *STy = dyn_cast(Agg->getType())) { - if (IdxV < STy->getNumElements()) - return Constant::getNullValue(STy->getElementType(IdxV)); - } else if (SequentialType *STy = - dyn_cast(Agg->getType())) { - return Constant::getNullValue(STy->getElementType()); +/// isLeakCheckerRoot - Is this global variable possibly used by a leak checker +/// as a root? If so, we might not really want to eliminate the stores to it. +static bool isLeakCheckerRoot(GlobalVariable *GV) { + // A global variable is a root if it is a pointer, or could plausibly contain + // a pointer. There are two challenges; one is that we could have a struct + // the has an inner member which is a pointer. We recurse through the type to + // detect these (up to a point). The other is that we may actually be a union + // of a pointer and another type, and so our LLVM type is an integer which + // gets converted into a pointer, or our type is an [i8 x #] with a pointer + // potentially contained here. + + if (GV->hasPrivateLinkage()) + return false; + + SmallVector Types; + Types.push_back(cast(GV->getType())->getElementType()); + + unsigned Limit = 20; + do { + Type *Ty = Types.pop_back_val(); + switch (Ty->getTypeID()) { + default: break; + case Type::PointerTyID: return true; + case Type::ArrayTyID: + case Type::VectorTyID: { + SequentialType *STy = cast(Ty); + Types.push_back(STy->getElementType()); + break; + } + case Type::StructTyID: { + StructType *STy = cast(Ty); + if (STy->isOpaque()) return true; + for (StructType::element_iterator I = STy->element_begin(), + E = STy->element_end(); I != E; ++I) { + Type *InnerTy = *I; + if (isa(InnerTy)) return true; + if (isa(InnerTy)) + Types.push_back(InnerTy); + } + break; + } } - } else if (isa(Agg)) { - if (StructType *STy = dyn_cast(Agg->getType())) { - if (IdxV < STy->getNumElements()) - return UndefValue::get(STy->getElementType(IdxV)); - } else if (SequentialType *STy = - dyn_cast(Agg->getType())) { - return UndefValue::get(STy->getElementType()); + if (--Limit == 0) return true; + } while (!Types.empty()); + return false; +} + +/// Given a value that is stored to a global but never read, determine whether +/// it's safe to remove the store and the chain of computation that feeds the +/// store. +static bool IsSafeComputationToRemove(Value *V, const TargetLibraryInfo *TLI) { + do { + if (isa(V)) + return true; + if (!V->hasOneUse()) + return false; + if (isa(V) || isa(V) || isa(V) || + isa(V)) + return false; + if (isAllocationFn(V, TLI)) + return true; + + Instruction *I = cast(V); + if (I->mayHaveSideEffects()) + return false; + if (GetElementPtrInst *GEP = dyn_cast(I)) { + if (!GEP->hasAllConstantIndices()) + return false; + } else if (I->getNumOperands() != 1) { + return false; } - } - return 0; + + V = I->getOperand(0); + } while (1); } +/// CleanupPointerRootUsers - This GV is a pointer root. Loop over all users +/// of the global and clean up any that obviously don't assign the global a +/// value that isn't dynamically allocated. +/// +static bool CleanupPointerRootUsers(GlobalVariable *GV, + const TargetLibraryInfo *TLI) { + // A brief explanation of leak checkers. The goal is to find bugs where + // pointers are forgotten, causing an accumulating growth in memory + // usage over time. The common strategy for leak checkers is to whitelist the + // memory pointed to by globals at exit. This is popular because it also + // solves another problem where the main thread of a C++ program may shut down + // before other threads that are still expecting to use those globals. To + // handle that case, we expect the program may create a singleton and never + // destroy it. + + bool Changed = false; + + // If Dead[n].first is the only use of a malloc result, we can delete its + // chain of computation and the store to the global in Dead[n].second. + SmallVector, 32> Dead; + + // Constants can't be pointers to dynamically allocated memory. + for (Value::use_iterator UI = GV->use_begin(), E = GV->use_end(); + UI != E;) { + User *U = *UI++; + if (StoreInst *SI = dyn_cast(U)) { + Value *V = SI->getValueOperand(); + if (isa(V)) { + Changed = true; + SI->eraseFromParent(); + } else if (Instruction *I = dyn_cast(V)) { + if (I->hasOneUse()) + Dead.push_back(std::make_pair(I, SI)); + } + } else if (MemSetInst *MSI = dyn_cast(U)) { + if (isa(MSI->getValue())) { + Changed = true; + MSI->eraseFromParent(); + } else if (Instruction *I = dyn_cast(MSI->getValue())) { + if (I->hasOneUse()) + Dead.push_back(std::make_pair(I, MSI)); + } + } else if (MemTransferInst *MTI = dyn_cast(U)) { + GlobalVariable *MemSrc = dyn_cast(MTI->getSource()); + if (MemSrc && MemSrc->isConstant()) { + Changed = true; + MTI->eraseFromParent(); + } else if (Instruction *I = dyn_cast(MemSrc)) { + if (I->hasOneUse()) + Dead.push_back(std::make_pair(I, MTI)); + } + } else if (ConstantExpr *CE = dyn_cast(U)) { + if (CE->use_empty()) { + CE->destroyConstant(); + Changed = true; + } + } else if (Constant *C = dyn_cast(U)) { + if (SafeToDestroyConstant(C)) { + C->destroyConstant(); + // This could have invalidated UI, start over from scratch. + Dead.clear(); + CleanupPointerRootUsers(GV, TLI); + return true; + } + } + } + + for (int i = 0, e = Dead.size(); i != e; ++i) { + if (IsSafeComputationToRemove(Dead[i].first, TLI)) { + Dead[i].second->eraseFromParent(); + Instruction *I = Dead[i].first; + do { + if (isAllocationFn(I, TLI)) + break; + Instruction *J = dyn_cast(I->getOperand(0)); + if (!J) + break; + I->eraseFromParent(); + I = J; + } while (1); + I->eraseFromParent(); + } + } + + return Changed; +} /// CleanupConstantGlobalUsers - We just marked GV constant. Loop over all /// users of the global, cleaning up the obvious ones. This is largely just a /// quick scan over the use list to clean up the easy and obvious cruft. This /// returns true if it made a change. -static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { +static bool CleanupConstantGlobalUsers(Value *V, Constant *Init, + DataLayout *TD, TargetLibraryInfo *TLI) { bool Changed = false; for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;) { User *U = *UI++; @@ -326,11 +489,11 @@ static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { Constant *SubInit = 0; if (Init) SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); - Changed |= CleanupConstantGlobalUsers(CE, SubInit); + Changed |= CleanupConstantGlobalUsers(CE, SubInit, TD, TLI); } else if (CE->getOpcode() == Instruction::BitCast && CE->getType()->isPointerTy()) { // Pointer cast, delete any stores and memsets to the global. - Changed |= CleanupConstantGlobalUsers(CE, 0); + Changed |= CleanupConstantGlobalUsers(CE, 0, TD, TLI); } if (CE->use_empty()) { @@ -344,11 +507,17 @@ static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { Constant *SubInit = 0; if (!isa(GEP->getOperand(0))) { ConstantExpr *CE = - dyn_cast_or_null(ConstantFoldInstruction(GEP)); + dyn_cast_or_null(ConstantFoldInstruction(GEP, TD, TLI)); if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr) SubInit = ConstantFoldLoadThroughGEPConstantExpr(Init, CE); + + // If the initializer is an all-null value and we have an inbounds GEP, + // we already know what the result of any load from that GEP is. + // TODO: Handle splats. + if (Init && isa(Init) && GEP->isInBounds()) + SubInit = Constant::getNullValue(GEP->getType()->getElementType()); } - Changed |= CleanupConstantGlobalUsers(GEP, SubInit); + Changed |= CleanupConstantGlobalUsers(GEP, SubInit, TD, TLI); if (GEP->use_empty()) { GEP->eraseFromParent(); @@ -366,7 +535,7 @@ static bool CleanupConstantGlobalUsers(Value *V, Constant *Init) { if (SafeToDestroyConstant(C)) { C->destroyConstant(); // This could have invalidated UI, start over from scratch. - CleanupConstantGlobalUsers(V, Init); + CleanupConstantGlobalUsers(V, Init, TD, TLI); return true; } } @@ -491,7 +660,7 @@ static bool GlobalUsersSafeToSRA(GlobalValue *GV) { /// behavior of the program in a more fine-grained way. We have determined that /// this transformation is safe already. We return the first global variable we /// insert so that the caller can reprocess it. -static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { +static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &TD) { // Make sure this global only has simple uses that we can SRA. if (!GlobalUsersSafeToSRA(GV)) return 0; @@ -512,13 +681,12 @@ static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { NewGlobals.reserve(STy->getNumElements()); const StructLayout &Layout = *TD.getStructLayout(STy); for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { - Constant *In = getAggregateConstantElement(Init, - ConstantInt::get(Type::getInt32Ty(STy->getContext()), i)); + Constant *In = Init->getAggregateElement(i); assert(In && "Couldn't get element of initializer?"); GlobalVariable *NGV = new GlobalVariable(STy->getElementType(i), false, GlobalVariable::InternalLinkage, In, GV->getName()+"."+Twine(i), - GV->isThreadLocal(), + GV->getThreadLocalMode(), GV->getType()->getAddressSpace()); Globals.insert(GV, NGV); NewGlobals.push_back(NGV); @@ -545,14 +713,13 @@ static GlobalVariable *SRAGlobal(GlobalVariable *GV, const TargetData &TD) { uint64_t EltSize = TD.getTypeAllocSize(STy->getElementType()); unsigned EltAlign = TD.getABITypeAlignment(STy->getElementType()); for (unsigned i = 0, e = NumElements; i != e; ++i) { - Constant *In = getAggregateConstantElement(Init, - ConstantInt::get(Type::getInt32Ty(Init->getContext()), i)); + Constant *In = Init->getAggregateElement(i); assert(In && "Couldn't get element of initializer?"); GlobalVariable *NGV = new GlobalVariable(STy->getElementType(), false, GlobalVariable::InternalLinkage, In, GV->getName()+"."+Twine(i), - GV->isThreadLocal(), + GV->getThreadLocalMode(), GV->getType()->getAddressSpace()); Globals.insert(GV, NGV); NewGlobals.push_back(NGV); @@ -768,7 +935,9 @@ static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { /// value stored into it. If there are uses of the loaded value that would trap /// if the loaded value is dynamically null, then we know that they cannot be /// reachable with a null optimize away the load. -static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { +static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV, + DataLayout *TD, + TargetLibraryInfo *TLI) { bool Changed = false; // Keep track of whether we are able to remove all the uses of the global @@ -797,7 +966,9 @@ static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { // If we get here we could have other crazy uses that are transitively // loaded. assert((isa(GlobalUser) || isa(GlobalUser) || - isa(GlobalUser) || isa(GlobalUser)) && + isa(GlobalUser) || isa(GlobalUser) || + isa(GlobalUser) || + isa(GlobalUser)) && "Only expect load and stores!"); } } @@ -810,23 +981,29 @@ static bool OptimizeAwayTrappingUsesOfLoads(GlobalVariable *GV, Constant *LV) { // If we nuked all of the loads, then none of the stores are needed either, // nor is the global. if (AllNonStoreUsesGone) { - DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n"); - CleanupConstantGlobalUsers(GV, 0); + if (isLeakCheckerRoot(GV)) { + Changed |= CleanupPointerRootUsers(GV, TLI); + } else { + Changed = true; + CleanupConstantGlobalUsers(GV, 0, TD, TLI); + } if (GV->use_empty()) { + DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n"); + Changed = true; GV->eraseFromParent(); ++NumDeleted; } - Changed = true; } return Changed; } /// ConstantPropUsersOf - Walk the use list of V, constant folding all of the /// instructions that are foldable. -static void ConstantPropUsersOf(Value *V) { +static void ConstantPropUsersOf(Value *V, + DataLayout *TD, TargetLibraryInfo *TLI) { for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ) if (Instruction *I = dyn_cast(*UI++)) - if (Constant *NewC = ConstantFoldInstruction(I)) { + if (Constant *NewC = ConstantFoldInstruction(I, TD, TLI)) { I->replaceAllUsesWith(NewC); // Advance UI to the next non-I use to avoid invalidating it! @@ -846,7 +1023,8 @@ static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, CallInst *CI, Type *AllocTy, ConstantInt *NElements, - TargetData* TD) { + DataLayout *TD, + TargetLibraryInfo *TLI) { DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI << '\n'); Type *GlobalType; @@ -864,7 +1042,7 @@ static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, UndefValue::get(GlobalType), GV->getName()+".body", GV, - GV->isThreadLocal()); + GV->getThreadLocalMode()); // 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 @@ -897,14 +1075,15 @@ static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, new GlobalVariable(Type::getInt1Ty(GV->getContext()), false, GlobalValue::InternalLinkage, ConstantInt::getFalse(GV->getContext()), - GV->getName()+".init", GV->isThreadLocal()); + GV->getName()+".init", GV->getThreadLocalMode()); bool InitBoolUsed = false; // Loop over all uses of GV, processing them in turn. while (!GV->use_empty()) { if (StoreInst *SI = dyn_cast(GV->use_back())) { // The global is initialized when the store to it occurs. - new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, SI); + new StoreInst(ConstantInt::getTrue(GV->getContext()), InitBool, false, 0, + SI->getOrdering(), SI->getSynchScope(), SI); SI->eraseFromParent(); continue; } @@ -919,7 +1098,10 @@ static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, ICmpInst *ICI = cast(LoadUse.getUser()); // Replace the cmp X, 0 with a use of the bool value. - Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", ICI); + // Sink the load to where the compare was, if atomic rules allow us to. + Value *LV = new LoadInst(InitBool, InitBool->getName()+".val", false, 0, + LI->getOrdering(), LI->getSynchScope(), + LI->isUnordered() ? (Instruction*)ICI : LI); InitBoolUsed = true; switch (ICI->getPredicate()) { default: llvm_unreachable("Unknown ICmp Predicate!"); @@ -960,9 +1142,9 @@ static GlobalVariable *OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, // To further other optimizations, loop over all users of NewGV and try to // constant prop them. This will promote GEP instructions with constant // indices into GEP constant-exprs, which will allow global-opt to hack on it. - ConstantPropUsersOf(NewGV); + ConstantPropUsersOf(NewGV, TD, TLI); if (RepValue != NewGV) - ConstantPropUsersOf(RepValue); + ConstantPropUsersOf(RepValue, TD, TLI); return NewGV; } @@ -1201,7 +1383,6 @@ static Value *GetHeapSROAValue(Value *V, unsigned FieldNo, PHIsToRewrite.push_back(std::make_pair(PN, FieldNo)); } else { llvm_unreachable("Unknown usable value"); - Result = 0; } return FieldVals[FieldNo] = Result; @@ -1291,9 +1472,10 @@ static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load, /// PerformHeapAllocSRoA - CI is an allocation of an array of structures. Break /// it up into multiple allocations of arrays of the fields. static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI, - Value* NElems, TargetData *TD) { + Value *NElems, DataLayout *TD, + const TargetLibraryInfo *TLI) { DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI << '\n'); - Type* MAT = getMallocAllocatedType(CI); + Type *MAT = getMallocAllocatedType(CI, TLI); StructType *STy = cast(MAT); // There is guaranteed to be at least one use of the malloc (storing @@ -1316,7 +1498,7 @@ static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI, PFieldTy, false, GlobalValue::InternalLinkage, Constant::getNullValue(PFieldTy), GV->getName() + ".f" + Twine(FieldNo), GV, - GV->isThreadLocal()); + GV->getThreadLocalMode()); FieldGlobals.push_back(NGV); unsigned TypeSize = TD->getTypeAllocSize(FieldTy); @@ -1374,8 +1556,7 @@ static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI, 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, - Constant::getNullValue(GVVal->getType()), - "tmp"); + Constant::getNullValue(GVVal->getType())); BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it", OrigBB->getParent()); BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next", @@ -1481,8 +1662,10 @@ static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI, static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI, Type *AllocTy, + AtomicOrdering Ordering, Module::global_iterator &GVI, - TargetData *TD) { + DataLayout *TD, + TargetLibraryInfo *TLI) { if (!TD) return false; @@ -1501,7 +1684,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // We can't optimize this if the malloc itself is used in a complex way, // for example, being stored into multiple globals. This allows the - // malloc to be stored into the specified global, loaded setcc'd, and + // malloc to be stored into the specified global, loaded icmp'd, and // GEP'd. These are all things we could transform to using the global // for. SmallPtrSet PHIs; @@ -1513,7 +1696,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // This eliminates dynamic allocation, avoids an indirection accessing the // data, and exposes the resultant global to further GlobalOpt. // We cannot optimize the malloc if we cannot determine malloc array size. - Value *NElems = getMallocArraySize(CI, TD, true); + Value *NElems = getMallocArraySize(CI, TD, TLI, true); if (!NElems) return false; @@ -1522,7 +1705,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // (2048 bytes currently), as we don't want to introduce a 16M global or // something. if (NElements->getZExtValue() * TD->getTypeAllocSize(AllocTy) < 2048) { - GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD); + GVI = OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, TD, TLI); return true; } @@ -1530,6 +1713,9 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // into multiple malloc'd arrays, one for each field. This is basically // SRoA for malloc'd memory. + if (Ordering != NotAtomic) + return false; + // If this is an allocation of a fixed size array of structs, analyze as a // variable size array. malloc [100 x struct],1 -> malloc struct, 100 if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1)) @@ -1547,7 +1733,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // If this is a fixed size array, transform the Malloc to be an alloc of // structs. malloc [100 x struct],1 -> malloc struct, 100 - if (ArrayType *AT = dyn_cast(getMallocAllocatedType(CI))) { + if (ArrayType *AT = dyn_cast(getMallocAllocatedType(CI, TLI))) { Type *IntPtrTy = TD->getIntPtrType(CI->getContext()); unsigned TypeSize = TD->getStructLayout(AllocSTy)->getSizeInBytes(); Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize); @@ -1558,11 +1744,14 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI); CI->replaceAllUsesWith(Cast); CI->eraseFromParent(); - CI = dyn_cast(Malloc) ? - extractMallocCallFromBitCast(Malloc) : cast(Malloc); + if (BitCastInst *BCI = dyn_cast(Malloc)) + CI = cast(BCI->getOperand(0)); + else + CI = cast(Malloc); } - GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, true),TD); + GVI = PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, TD, TLI, true), + TD, TLI); return true; } @@ -1572,8 +1761,9 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, // OptimizeOnceStoredGlobal - Try to optimize globals based on the knowledge // that only one value (besides its initializer) is ever stored to the global. static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, + AtomicOrdering Ordering, Module::global_iterator &GVI, - TargetData *TD) { + DataLayout *TD, TargetLibraryInfo *TLI) { // Ignore no-op GEPs and bitcasts. StoredOnceVal = StoredOnceVal->stripPointerCasts(); @@ -1588,12 +1778,13 @@ static bool OptimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType()); // Optimize away any trapping uses of the loaded value. - if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC)) + if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, TD, TLI)) return true; - } else if (CallInst *CI = extractMallocCall(StoredOnceVal)) { - Type* MallocType = getMallocAllocatedType(CI); - if (MallocType && TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, - GVI, TD)) + } else if (CallInst *CI = extractMallocCall(StoredOnceVal, TLI)) { + Type *MallocType = getMallocAllocatedType(CI, TLI); + if (MallocType && + TryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType, Ordering, GVI, + TD, TLI)) return true; } } @@ -1634,7 +1825,8 @@ static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { GlobalValue::InternalLinkage, ConstantInt::getFalse(GV->getContext()), GV->getName()+".b", - GV->isThreadLocal()); + GV->getThreadLocalMode(), + GV->getType()->getAddressSpace()); GV->getParent()->getGlobalList().insert(GV, NewGV); Constant *InitVal = GV->getInitializer(); @@ -1669,7 +1861,8 @@ static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { if (LoadInst *LI = dyn_cast(StoredVal)) { assert(LI->getOperand(0) == GV && "Not a copy!"); // Insert a new load, to preserve the saved value. - StoreVal = new LoadInst(NewGV, LI->getName()+".b", LI); + StoreVal = new LoadInst(NewGV, LI->getName()+".b", false, 0, + LI->getOrdering(), LI->getSynchScope(), LI); } else { assert((isa(StoredVal) || isa(StoredVal)) && "This is not a form that we understand!"); @@ -1677,11 +1870,13 @@ static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { assert(isa(StoreVal) && "Not a load of NewGV!"); } } - new StoreInst(StoreVal, NewGV, SI); + new StoreInst(StoreVal, NewGV, false, 0, + SI->getOrdering(), SI->getSynchScope(), SI); } else { // Change the load into a load of bool then a select. LoadInst *LI = cast(UI); - LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", LI); + LoadInst *NLI = new LoadInst(NewGV, LI->getName()+".b", false, 0, + LI->getOrdering(), LI->getSynchScope(), LI); Value *NSI; if (IsOneZero) NSI = new ZExtInst(NLI, LI->getType(), "", LI); @@ -1698,11 +1893,11 @@ static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { } -/// ProcessInternalGlobal - Analyze the specified global variable and optimize -/// it if possible. If we make a change, return true. +/// ProcessGlobal - Analyze the specified global variable and optimize it if +/// possible. If we make a change, return true. bool GlobalOpt::ProcessGlobal(GlobalVariable *GV, Module::global_iterator &GVI) { - if (!GV->hasLocalLinkage()) + if (!GV->isDiscardableIfUnused()) return false; // Do more involved optimizations if the global is internal. @@ -1715,6 +1910,9 @@ bool GlobalOpt::ProcessGlobal(GlobalVariable *GV, return true; } + if (!GV->hasLocalLinkage()) + return false; + SmallPtrSet PHIUsers; GlobalStatus GS; @@ -1736,7 +1934,7 @@ bool GlobalOpt::ProcessGlobal(GlobalVariable *GV, /// it if possible. If we make a change, return true. bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, Module::global_iterator &GVI, - const SmallPtrSet &PHIUsers, + const SmallPtrSet &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 @@ -1754,11 +1952,11 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, GS.AccessingFunction->hasExternalLinkage() && GV->getType()->getAddressSpace() == 0) { DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV); - Instruction& FirstI = const_cast(*GS.AccessingFunction + Instruction &FirstI = const_cast(*GS.AccessingFunction ->getEntryBlock().begin()); - Type* ElemTy = GV->getType()->getElementType(); + Type *ElemTy = GV->getType()->getElementType(); // FIXME: Pass Global's alignment when globals have alignment - AllocaInst* Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI); + AllocaInst *Alloca = new AllocaInst(ElemTy, NULL, GV->getName(), &FirstI); if (!isa(GV->getInitializer())) new StoreInst(GV->getInitializer(), Alloca, &FirstI); @@ -1773,9 +1971,15 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, 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()); + bool Changed; + if (isLeakCheckerRoot(GV)) { + // Delete any constant stores to the global. + Changed = CleanupPointerRootUsers(GV, TLI); + } else { + // Delete any stores we can find to the global. We may not be able to + // make it completely dead though. + Changed = CleanupConstantGlobalUsers(GV, GV->getInitializer(), TD, TLI); + } // If the global is dead now, delete it. if (GV->use_empty()) { @@ -1786,11 +1990,11 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, return Changed; } else if (GS.StoredType <= GlobalStatus::isInitializerStored) { - DEBUG(dbgs() << "MARKING CONSTANT: " << *GV); + DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n"); GV->setConstant(true); // Clean up any obviously simplifiable users now. - CleanupConstantGlobalUsers(GV, GV->getInitializer()); + CleanupConstantGlobalUsers(GV, GV->getInitializer(), TD, TLI); // If the global is dead now, just nuke it. if (GV->use_empty()) { @@ -1803,7 +2007,7 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, ++NumMarked; return true; } else if (!GV->getInitializer()->getType()->isSingleValueType()) { - if (TargetData *TD = getAnalysisIfAvailable()) + if (DataLayout *TD = getAnalysisIfAvailable()) if (GlobalVariable *FirstNewGV = SRAGlobal(GV, *TD)) { GVI = FirstNewGV; // Don't skip the newly produced globals! return true; @@ -1819,11 +2023,11 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, GV->setInitializer(SOVConstant); // Clean up any obviously simplifiable users now. - CleanupConstantGlobalUsers(GV, GV->getInitializer()); + CleanupConstantGlobalUsers(GV, GV->getInitializer(), TD, TLI); if (GV->use_empty()) { DEBUG(dbgs() << " *** Substituting initializer allowed us to " - << "simplify all users and delete global!\n"); + << "simplify all users and delete global!\n"); GV->eraseFromParent(); ++NumDeleted; } else { @@ -1835,8 +2039,8 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, // 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())) + if (OptimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, GVI, + TD, TLI)) return true; // Otherwise, if the global was not a boolean, we can shrink it to be a @@ -1855,28 +2059,33 @@ bool GlobalOpt::ProcessInternalGlobal(GlobalVariable *GV, /// function, changing them to FastCC. static void ChangeCalleesToFastCall(Function *F) { for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ + if (isa(*UI)) + continue; CallSite User(cast(*UI)); User.setCallingConv(CallingConv::Fast); } } -static AttrListPtr StripNest(const AttrListPtr &Attrs) { +static AttributeSet StripNest(LLVMContext &C, const AttributeSet &Attrs) { for (unsigned i = 0, e = Attrs.getNumSlots(); i != e; ++i) { - if ((Attrs.getSlot(i).Attrs & Attribute::Nest) == 0) + unsigned Index = Attrs.getSlotIndex(i); + if (!Attrs.getSlotAttributes(i).hasAttribute(Index, Attribute::Nest)) continue; // There can be only one. - return Attrs.removeAttr(Attrs.getSlot(i).Index, Attribute::Nest); + return Attrs.removeAttribute(C, Index, Attribute::Nest); } return Attrs; } static void RemoveNestAttribute(Function *F) { - F->setAttributes(StripNest(F->getAttributes())); + F->setAttributes(StripNest(F->getContext(), F->getAttributes())); for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); UI != E;++UI){ + if (isa(*UI)) + continue; CallSite User(cast(*UI)); - User.setAttributes(StripNest(User.getAttributes())); + User.setAttributes(StripNest(F->getContext(), User.getAttributes())); } } @@ -1889,7 +2098,7 @@ bool GlobalOpt::OptimizeFunctions(Module &M) { if (!F->hasName() && !F->isDeclaration()) F->setLinkage(GlobalValue::InternalLinkage); F->removeDeadConstantUsers(); - if (F->use_empty() && (F->hasLocalLinkage() || F->hasLinkOnceLinkage())) { + if (F->isDefTriviallyDead()) { F->eraseFromParent(); Changed = true; ++NumFnDeleted; @@ -1929,8 +2138,7 @@ bool GlobalOpt::OptimizeGlobalVars(Module &M) { // Simplify the initializer. if (GV->hasInitializer()) if (ConstantExpr *CE = dyn_cast(GV->getInitializer())) { - TargetData *TD = getAnalysisIfAvailable(); - Constant *New = ConstantFoldConstantExpression(CE, TD); + Constant *New = ConstantFoldConstantExpression(CE, TD, TLI); if (New && New != CE) GV->setInitializer(New); } @@ -1945,7 +2153,7 @@ bool GlobalOpt::OptimizeGlobalVars(Module &M) { GlobalVariable *GlobalOpt::FindGlobalCtors(Module &M) { GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors"); if (GV == 0) 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; @@ -2031,7 +2239,7 @@ static GlobalVariable *InstallGlobalCtors(GlobalVariable *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->getThreadLocalMode()); GCL->getParent()->getGlobalList().insert(GCL, NGV); NGV->takeName(GCL); @@ -2051,16 +2259,10 @@ static GlobalVariable *InstallGlobalCtors(GlobalVariable *GCL, } -static Constant *getVal(DenseMap &ComputedValues, Value *V) { - if (Constant *CV = dyn_cast(V)) return CV; - Constant *R = ComputedValues[V]; - assert(R && "Reference to an uncomputed value!"); - return R; -} - -static inline bool +static inline bool isSimpleEnoughValueToCommit(Constant *C, - SmallPtrSet &SimpleConstants); + SmallPtrSet &SimpleConstants, + const DataLayout *TD); /// isSimpleEnoughValueToCommit - Return true if the specified constant can be @@ -2072,58 +2274,67 @@ isSimpleEnoughValueToCommit(Constant *C, /// in SimpleConstants to avoid having to rescan the same constants all the /// time. static bool isSimpleEnoughValueToCommitHelper(Constant *C, - SmallPtrSet &SimpleConstants) { + SmallPtrSet &SimpleConstants, + const DataLayout *TD) { // Simple integer, undef, constant aggregate zero, global addresses, etc are // all supported. if (C->getNumOperands() == 0 || isa(C) || isa(C)) return true; - + // Aggregate values are safe if all their elements are. if (isa(C) || isa(C) || isa(C)) { for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) { Constant *Op = cast(C->getOperand(i)); - if (!isSimpleEnoughValueToCommit(Op, SimpleConstants)) + if (!isSimpleEnoughValueToCommit(Op, SimpleConstants, TD)) 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(C); switch (CE->getOpcode()) { case Instruction::BitCast: + // Bitcast is fine if the casted value is fine. + return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD); + case Instruction::IntToPtr: case Instruction::PtrToInt: - // These casts are always fine if the casted value is. - return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants); - + // int <=> ptr is fine if the int type is the same size as the + // pointer type. + if (!TD || TD->getTypeSizeInBits(CE->getType()) != + TD->getTypeSizeInBits(CE->getOperand(0)->getType())) + return false; + return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD); + // GEP is fine if it is simple + constant offset. case Instruction::GetElementPtr: for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) if (!isa(CE->getOperand(i))) return false; - return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants); - + return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD); + case Instruction::Add: // We allow simple+cst. if (!isa(CE->getOperand(1))) return false; - return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants); + return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, TD); } return false; } -static inline bool +static inline bool isSimpleEnoughValueToCommit(Constant *C, - SmallPtrSet &SimpleConstants) { + SmallPtrSet &SimpleConstants, + const DataLayout *TD) { // If we already checked this constant, we win. if (!SimpleConstants.insert(C)) return true; // Check the constant. - return isSimpleEnoughValueToCommitHelper(C, SimpleConstants); + return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, TD); } @@ -2164,7 +2375,7 @@ static bool isSimpleEnoughPointerToCommit(Constant *C) { 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. @@ -2175,7 +2386,7 @@ static bool isSimpleEnoughPointerToCommit(Constant *C) { return cast(CE->getOperand(0))->hasUniqueInitializer(); } } - + return false; } @@ -2190,23 +2401,11 @@ static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, return Val; } - std::vector Elts; + SmallVector Elts; if (StructType *STy = dyn_cast(Init->getType())) { - // Break up the constant into its elements. - if (ConstantStruct *CS = dyn_cast(Init)) { - for (User::op_iterator i = CS->op_begin(), e = CS->op_end(); i != e; ++i) - Elts.push_back(cast(*i)); - } else if (isa(Init)) { - for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) - Elts.push_back(Constant::getNullValue(STy->getElementType(i))); - } else if (isa(Init)) { - for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) - Elts.push_back(UndefValue::get(STy->getElementType(i))); - } else { - llvm_unreachable("This code is out of sync with " - " ConstantFoldLoadThroughGEPConstantExpr"); - } + for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) + Elts.push_back(Init->getAggregateElement(i)); // Replace the element that we are supposed to. ConstantInt *CU = cast(Addr->getOperand(OpNo)); @@ -2217,7 +2416,7 @@ static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, // Return the modified struct. return ConstantStruct::get(STy, Elts); } - + ConstantInt *CI = cast(Addr->getOperand(OpNo)); SequentialType *InitTy = cast(Init->getType()); @@ -2225,22 +2424,11 @@ static Constant *EvaluateStoreInto(Constant *Init, Constant *Val, if (ArrayType *ATy = dyn_cast(InitTy)) NumElts = ATy->getNumElements(); else - NumElts = cast(InitTy)->getNumElements(); + NumElts = InitTy->getVectorNumElements(); // Break up the array into elements. - if (ConstantArray *CA = dyn_cast(Init)) { - for (User::op_iterator i = CA->op_begin(), e = CA->op_end(); i != e; ++i) - Elts.push_back(cast(*i)); - } else if (ConstantVector *CV = dyn_cast(Init)) { - for (User::op_iterator i = CV->op_begin(), e = CV->op_end(); i != e; ++i) - Elts.push_back(cast(*i)); - } else if (isa(Init)) { - Elts.assign(NumElts, Constant::getNullValue(InitTy->getElementType())); - } else { - assert(isa(Init) && "This code is out of sync with " - " ConstantFoldLoadThroughGEPConstantExpr"); - Elts.assign(NumElts, UndefValue::get(InitTy->getElementType())); - } + for (uint64_t i = 0, e = NumElts; i != e; ++i) + Elts.push_back(Init->getAggregateElement(i)); assert(CI->getZExtValue() < NumElts); Elts[CI->getZExtValue()] = @@ -2265,15 +2453,109 @@ static void CommitValueTo(Constant *Val, Constant *Addr) { GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2)); } +namespace { + +/// Evaluator - This class evaluates LLVM IR, producing the Constant +/// representing each SSA instruction. Changes to global variables are stored +/// in a mapping that can be iterated over after the evaluation is complete. +/// Once an evaluation call fails, the evaluation object should not be reused. +class Evaluator { +public: + Evaluator(const DataLayout *TD, const TargetLibraryInfo *TLI) + : TD(TD), TLI(TLI) { + ValueStack.push_back(new DenseMap); + } + + ~Evaluator() { + DeleteContainerPointers(ValueStack); + 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. + if (!Tmp->use_empty()) + Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); + delete Tmp; + } + } + + /// EvaluateFunction - Evaluate a call to function F, returning true if + /// successful, false if we can't evaluate it. ActualArgs contains the formal + /// arguments for the function. + bool EvaluateFunction(Function *F, Constant *&RetVal, + const SmallVectorImpl &ActualArgs); + + /// EvaluateBlock - Evaluate all instructions in block BB, returning true if + /// successful, false if we can't evaluate it. NewBB returns the next BB that + /// control flows into, or null upon return. + bool EvaluateBlock(BasicBlock::iterator CurInst, BasicBlock *&NextBB); + + Constant *getVal(Value *V) { + if (Constant *CV = dyn_cast(V)) return CV; + Constant *R = ValueStack.back()->lookup(V); + assert(R && "Reference to an uncomputed value!"); + return R; + } + + void setVal(Value *V, Constant *C) { + ValueStack.back()->operator[](V) = C; + } + + const DenseMap &getMutatedMemory() const { + return MutatedMemory; + } + + const SmallPtrSet &getInvariants() const { + return Invariants; + } + +private: + Constant *ComputeLoadResult(Constant *P); + + /// ValueStack - As we compute SSA register values, we store their contents + /// here. The back of the vector contains the current function and the stack + /// contains the values in the calling frames. + SmallVector*, 4> ValueStack; + + /// CallStack - This is used to detect recursion. In pathological situations + /// we could hit exponential behavior, but at least there is nothing + /// unbounded. + SmallVector CallStack; + + /// 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. + DenseMap MutatedMemory; + + /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable + /// to represent its body. This vector is needed so we can delete the + /// temporary globals when we are done. + SmallVector AllocaTmps; + + /// Invariants - These global variables have been marked invariant by the + /// static constructor. + SmallPtrSet Invariants; + + /// SimpleConstants - These are constants we have checked and know to be + /// simple enough to live in a static initializer of a global. + SmallPtrSet SimpleConstants; + + const DataLayout *TD; + const TargetLibraryInfo *TLI; +}; + +} // anonymous namespace + /// ComputeLoadResult - Return the value that would be computed by a load from /// P after the stores reflected by 'memory' have been performed. If we can't /// decide, return null. -static Constant *ComputeLoadResult(Constant *P, - const DenseMap &Memory) { +Constant *Evaluator::ComputeLoadResult(Constant *P) { // If this memory location has been recently stored, use the stored value: it // is the most up-to-date. - DenseMap::const_iterator I = Memory.find(P); - if (I != Memory.end()) return I->second; + DenseMap::const_iterator I = MutatedMemory.find(P); + if (I != MutatedMemory.end()) return I->second; // Access it. if (GlobalVariable *GV = dyn_cast(P)) { @@ -2294,67 +2576,54 @@ static Constant *ComputeLoadResult(Constant *P, return 0; // don't know how to evaluate. } -/// EvaluateFunction - Evaluate a call to function F, returning true if -/// successful, false if we can't evaluate it. ActualArgs contains the formal -/// arguments for the function. -static bool EvaluateFunction(Function *F, Constant *&RetVal, - const SmallVectorImpl &ActualArgs, - std::vector &CallStack, - DenseMap &MutatedMemory, - std::vector &AllocaTmps, - SmallPtrSet &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 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; - ++AI, ++ArgNo) - Values[AI] = ActualArgs[ArgNo]; - - /// ExecutedBlocks - We only handle non-looping, non-recursive code. As such, - /// 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 ExecutedBlocks; - - // CurInst - The current instruction we're evaluating. - BasicBlock::iterator CurInst = F->begin()->begin(); - +/// EvaluateBlock - Evaluate all instructions in block BB, returning true if +/// successful, false if we can't evaluate it. NewBB returns the next BB that +/// control flows into, or null upon return. +bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst, + BasicBlock *&NextBB) { // This is the main evaluation loop. while (1) { Constant *InstResult = 0; + DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n"); + if (StoreInst *SI = dyn_cast(CurInst)) { - if (SI->isVolatile()) return false; // no volatile accesses. - Constant *Ptr = getVal(Values, SI->getOperand(1)); - if (!isSimpleEnoughPointerToCommit(Ptr)) + if (!SI->isSimple()) { + DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n"); + return false; // no volatile/atomic accesses. + } + Constant *Ptr = getVal(SI->getOperand(1)); + if (ConstantExpr *CE = dyn_cast(Ptr)) { + DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr); + Ptr = ConstantFoldConstantExpression(CE, TD, TLI); + DEBUG(dbgs() << "; To: " << *Ptr << "\n"); + } + if (!isSimpleEnoughPointerToCommit(Ptr)) { // If this is too complex for us to commit, reject it. + DEBUG(dbgs() << "Pointer is too complex for us to evaluate store."); return false; - - Constant *Val = getVal(Values, SI->getOperand(0)); + } + + Constant *Val = getVal(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)) + if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, TD)) { + DEBUG(dbgs() << "Store value is too complex to evaluate store. " << *Val + << "\n"); return false; - - if (ConstantExpr *CE = dyn_cast(Ptr)) + } + + if (ConstantExpr *CE = dyn_cast(Ptr)) { if (CE->getOpcode() == Instruction::BitCast) { + DEBUG(dbgs() << "Attempting to resolve bitcast on constant ptr.\n"); // 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); - - Type *NewTy=cast(Ptr->getType())->getElementType(); - + + Type *NewTy = cast(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. @@ -2365,178 +2634,271 @@ static bool EvaluateFunction(Function *F, Constant *&RetVal, if (StructType *STy = dyn_cast(NewTy)) { NewTy = STy->getTypeAtIndex(0U); - IntegerType *IdxTy =IntegerType::get(NewTy->getContext(), 32); + IntegerType *IdxTy = IntegerType::get(NewTy->getContext(), 32); Constant *IdxZero = ConstantInt::get(IdxTy, 0, false); Constant * const IdxList[] = {IdxZero, IdxZero}; Ptr = ConstantExpr::getGetElementPtr(Ptr, IdxList); - + if (ConstantExpr *CE = dyn_cast(Ptr)) + Ptr = ConstantFoldConstantExpression(CE, TD, TLI); + // If we can't improve the situation by introspecting NewTy, // we have to give up. } else { - return 0; + DEBUG(dbgs() << "Failed to bitcast constant ptr, can not " + "evaluate.\n"); + return false; } } - + // If we found compatible types, go ahead and push the bitcast // onto the stored value. Val = ConstantExpr::getBitCast(Val, NewTy); + + DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n"); } - + } + MutatedMemory[Ptr] = Val; } else if (BinaryOperator *BO = dyn_cast(CurInst)) { InstResult = ConstantExpr::get(BO->getOpcode(), - getVal(Values, BO->getOperand(0)), - getVal(Values, BO->getOperand(1))); + getVal(BO->getOperand(0)), + getVal(BO->getOperand(1))); + DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " << *InstResult + << "\n"); } else if (CmpInst *CI = dyn_cast(CurInst)) { InstResult = ConstantExpr::getCompare(CI->getPredicate(), - getVal(Values, CI->getOperand(0)), - getVal(Values, CI->getOperand(1))); + getVal(CI->getOperand(0)), + getVal(CI->getOperand(1))); + DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult + << "\n"); } else if (CastInst *CI = dyn_cast(CurInst)) { InstResult = ConstantExpr::getCast(CI->getOpcode(), - getVal(Values, CI->getOperand(0)), + getVal(CI->getOperand(0)), CI->getType()); + DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult + << "\n"); } else if (SelectInst *SI = dyn_cast(CurInst)) { - InstResult = ConstantExpr::getSelect(getVal(Values, SI->getOperand(0)), - getVal(Values, SI->getOperand(1)), - getVal(Values, SI->getOperand(2))); + InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)), + getVal(SI->getOperand(1)), + getVal(SI->getOperand(2))); + DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult + << "\n"); } else if (GetElementPtrInst *GEP = dyn_cast(CurInst)) { - Constant *P = getVal(Values, GEP->getOperand(0)); + Constant *P = getVal(GEP->getOperand(0)); SmallVector GEPOps; for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e; ++i) - GEPOps.push_back(getVal(Values, *i)); + GEPOps.push_back(getVal(*i)); InstResult = ConstantExpr::getGetElementPtr(P, GEPOps, cast(GEP)->isInBounds()); + DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult + << "\n"); } else if (LoadInst *LI = dyn_cast(CurInst)) { - if (LI->isVolatile()) return false; // no volatile accesses. - InstResult = ComputeLoadResult(getVal(Values, LI->getOperand(0)), - MutatedMemory); - if (InstResult == 0) return false; // Could not evaluate load. + + if (!LI->isSimple()) { + DEBUG(dbgs() << "Found a Load! Not a simple load, can not evaluate.\n"); + return false; // no volatile/atomic accesses. + } + + Constant *Ptr = getVal(LI->getOperand(0)); + if (ConstantExpr *CE = dyn_cast(Ptr)) { + Ptr = ConstantFoldConstantExpression(CE, TD, TLI); + DEBUG(dbgs() << "Found a constant pointer expression, constant " + "folding: " << *Ptr << "\n"); + } + InstResult = ComputeLoadResult(Ptr); + if (InstResult == 0) { + DEBUG(dbgs() << "Failed to compute load result. Can not evaluate load." + "\n"); + return false; // Could not evaluate load. + } + + DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n"); } else if (AllocaInst *AI = dyn_cast(CurInst)) { - if (AI->isArrayAllocation()) return false; // Cannot handle array allocs. + if (AI->isArrayAllocation()) { + DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n"); + return false; // Cannot handle array allocs. + } Type *Ty = AI->getType()->getElementType(); AllocaTmps.push_back(new GlobalVariable(Ty, false, GlobalValue::InternalLinkage, UndefValue::get(Ty), AI->getName())); InstResult = AllocaTmps.back(); - } else if (CallInst *CI = dyn_cast(CurInst)) { + DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n"); + } else if (isa(CurInst) || isa(CurInst)) { + CallSite CS(CurInst); // Debug info can safely be ignored here. - if (isa(CI)) { + if (isa(CS.getInstruction())) { + DEBUG(dbgs() << "Ignoring debug info.\n"); ++CurInst; continue; } // Cannot handle inline asm. - if (isa(CI->getCalledValue())) return false; - - if (MemSetInst *MSI = dyn_cast(CI)) { - if (MSI->isVolatile()) return false; - Constant *Ptr = getVal(Values, MSI->getDest()); - Constant *Val = getVal(Values, MSI->getValue()); - Constant *DestVal = ComputeLoadResult(getVal(Values, Ptr), - MutatedMemory); - if (Val->isNullValue() && DestVal && DestVal->isNullValue()) { - // This memset is a no-op. + if (isa(CS.getCalledValue())) { + DEBUG(dbgs() << "Found inline asm, can not evaluate.\n"); + return false; + } + + if (IntrinsicInst *II = dyn_cast(CS.getInstruction())) { + if (MemSetInst *MSI = dyn_cast(II)) { + if (MSI->isVolatile()) { + DEBUG(dbgs() << "Can not optimize a volatile memset " << + "intrinsic.\n"); + return false; + } + Constant *Ptr = getVal(MSI->getDest()); + Constant *Val = getVal(MSI->getValue()); + Constant *DestVal = ComputeLoadResult(getVal(Ptr)); + if (Val->isNullValue() && DestVal && DestVal->isNullValue()) { + // This memset is a no-op. + DEBUG(dbgs() << "Ignoring no-op memset.\n"); + ++CurInst; + continue; + } + } + + if (II->getIntrinsicID() == Intrinsic::lifetime_start || + II->getIntrinsicID() == Intrinsic::lifetime_end) { + DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n"); + ++CurInst; + continue; + } + + if (II->getIntrinsicID() == Intrinsic::invariant_start) { + // We don't insert an entry into Values, as it doesn't have a + // meaningful return value. + if (!II->use_empty()) { + DEBUG(dbgs() << "Found unused invariant_start. Cant evaluate.\n"); + return false; + } + ConstantInt *Size = cast(II->getArgOperand(0)); + Value *PtrArg = getVal(II->getArgOperand(1)); + Value *Ptr = PtrArg->stripPointerCasts(); + if (GlobalVariable *GV = dyn_cast(Ptr)) { + Type *ElemTy = cast(GV->getType())->getElementType(); + if (!Size->isAllOnesValue() && + Size->getValue().getLimitedValue() >= + TD->getTypeStoreSize(ElemTy)) { + Invariants.insert(GV); + DEBUG(dbgs() << "Found a global var that is an invariant: " << *GV + << "\n"); + } else { + DEBUG(dbgs() << "Found a global var, but can not treat it as an " + "invariant.\n"); + } + } + // Continue even if we do nothing. ++CurInst; continue; } + + DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n"); return false; } // Resolve function pointers. - Function *Callee = dyn_cast(getVal(Values, - CI->getCalledValue())); - if (!Callee) return false; // Cannot resolve. + Function *Callee = dyn_cast(getVal(CS.getCalledValue())); + if (!Callee || Callee->mayBeOverridden()) { + DEBUG(dbgs() << "Can not resolve function pointer.\n"); + return false; // Cannot resolve. + } SmallVector Formals; - CallSite CS(CI); - for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); - i != e; ++i) - Formals.push_back(getVal(Values, *i)); + for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i) + Formals.push_back(getVal(*i)); if (Callee->isDeclaration()) { // If this is a function we can constant fold, do it. - if (Constant *C = ConstantFoldCall(Callee, Formals)) { + if (Constant *C = ConstantFoldCall(Callee, Formals, TLI)) { InstResult = C; + DEBUG(dbgs() << "Constant folded function call. Result: " << + *InstResult << "\n"); } else { + DEBUG(dbgs() << "Can not constant fold function call.\n"); return false; } } else { - if (Callee->getFunctionType()->isVarArg()) + if (Callee->getFunctionType()->isVarArg()) { + DEBUG(dbgs() << "Can not constant fold vararg function call.\n"); return false; + } - Constant *RetVal; + Constant *RetVal = 0; // Execute the call, if successful, use the return value. - if (!EvaluateFunction(Callee, RetVal, Formals, CallStack, - MutatedMemory, AllocaTmps, SimpleConstants, TD)) + ValueStack.push_back(new DenseMap); + if (!EvaluateFunction(Callee, RetVal, Formals)) { + DEBUG(dbgs() << "Failed to evaluate function.\n"); return false; + } + delete ValueStack.pop_back_val(); InstResult = RetVal; + + if (InstResult != NULL) { + DEBUG(dbgs() << "Successfully evaluated function. Result: " << + InstResult << "\n\n"); + } else { + DEBUG(dbgs() << "Successfully evaluated function. Result: 0\n\n"); + } } } else if (isa(CurInst)) { - BasicBlock *NewBB = 0; + DEBUG(dbgs() << "Found a terminator instruction.\n"); + if (BranchInst *BI = dyn_cast(CurInst)) { if (BI->isUnconditional()) { - NewBB = BI->getSuccessor(0); + NextBB = BI->getSuccessor(0); } else { ConstantInt *Cond = - dyn_cast(getVal(Values, BI->getCondition())); + dyn_cast(getVal(BI->getCondition())); if (!Cond) return false; // Cannot determine. - NewBB = BI->getSuccessor(!Cond->getZExtValue()); + NextBB = BI->getSuccessor(!Cond->getZExtValue()); } } else if (SwitchInst *SI = dyn_cast(CurInst)) { ConstantInt *Val = - dyn_cast(getVal(Values, SI->getCondition())); + dyn_cast(getVal(SI->getCondition())); if (!Val) return false; // Cannot determine. - NewBB = SI->getSuccessor(SI->findCaseValue(Val)); + NextBB = SI->findCaseValue(Val).getCaseSuccessor(); } else if (IndirectBrInst *IBI = dyn_cast(CurInst)) { - Value *Val = getVal(Values, IBI->getAddress())->stripPointerCasts(); + Value *Val = getVal(IBI->getAddress())->stripPointerCasts(); if (BlockAddress *BA = dyn_cast(Val)) - NewBB = BA->getBasicBlock(); + NextBB = BA->getBasicBlock(); else return false; // Cannot determine. - } else if (ReturnInst *RI = dyn_cast(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 if (isa(CurInst)) { + NextBB = 0; } else { - // invoke, unwind, unreachable. + // invoke, unwind, resume, unreachable. + DEBUG(dbgs() << "Can not handle terminator."); 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. - BasicBlock *OldBB = CurInst->getParent(); - CurInst = NewBB->begin(); - PHINode *PN; - for (; (PN = dyn_cast(CurInst)); ++CurInst) - Values[PN] = getVal(Values, PN->getIncomingValueForBlock(OldBB)); - - // Do NOT increment CurInst. We know that the terminator had no value. - continue; + // We succeeded at evaluating this block! + DEBUG(dbgs() << "Successfully evaluated block.\n"); + return true; } else { // Did not know how to evaluate this! + DEBUG(dbgs() << "Failed to evaluate block due to unhandled instruction." + "\n"); return false; } if (!CurInst->use_empty()) { if (ConstantExpr *CE = dyn_cast(InstResult)) - InstResult = ConstantFoldConstantExpression(CE, TD); - - Values[CurInst] = InstResult; + InstResult = ConstantFoldConstantExpression(CE, TD, TLI); + + setVal(CurInst, InstResult); + } + + // If we just processed an invoke, we finished evaluating the block. + if (InvokeInst *II = dyn_cast(CurInst)) { + NextBB = II->getNormalDest(); + DEBUG(dbgs() << "Found an invoke instruction. Finished Block.\n\n"); + return true; } // Advance program counter. @@ -2544,64 +2906,98 @@ static bool EvaluateFunction(Function *F, Constant *&RetVal, } } -/// EvaluateStaticConstructor - Evaluate static constructors in the function, if -/// we can. Return true if we can, false otherwise. -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. - DenseMap MutatedMemory; +/// EvaluateFunction - Evaluate a call to function F, returning true if +/// successful, false if we can't evaluate it. ActualArgs contains the formal +/// arguments for the function. +bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal, + const SmallVectorImpl &ActualArgs) { + // 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; - /// AllocaTmps - To 'execute' an alloca, we create a temporary global variable - /// to represent its body. This vector is needed so we can delete the - /// temporary globals when we are done. - std::vector AllocaTmps; + CallStack.push_back(F); - /// CallStack - This is used to detect recursion. In pathological situations - /// we could hit exponential behavior, but at least there is nothing - /// unbounded. - std::vector CallStack; + // Initialize arguments to the incoming values specified. + unsigned ArgNo = 0; + for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; + ++AI, ++ArgNo) + setVal(AI, ActualArgs[ArgNo]); - /// SimpleConstants - These are constants we have checked and know to be - /// simple enough to live in a static initializer of a global. - SmallPtrSet SimpleConstants; - + // ExecutedBlocks - We only handle non-looping, non-recursive code. As such, + // 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 ExecutedBlocks; + + // CurBB - The current basic block we're evaluating. + BasicBlock *CurBB = F->begin(); + + BasicBlock::iterator CurInst = CurBB->begin(); + + while (1) { + BasicBlock *NextBB = 0; // Initialized to avoid compiler warnings. + DEBUG(dbgs() << "Trying to evaluate BB: " << *CurBB << "\n"); + + if (!EvaluateBlock(CurInst, NextBB)) + return false; + + if (NextBB == 0) { + // Successfully running until there's no next block means that we found + // the return. Fill it the return value and pop the call stack. + ReturnInst *RI = cast(CurBB->getTerminator()); + if (RI->getNumOperands()) + RetVal = getVal(RI->getOperand(0)); + CallStack.pop_back(); + return true; + } + + // 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(NextBB)) + 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. + PHINode *PN = 0; + for (CurInst = NextBB->begin(); + (PN = dyn_cast(CurInst)); ++CurInst) + setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB))); + + // Advance to the next block. + CurBB = NextBB; + } +} + +/// EvaluateStaticConstructor - Evaluate static constructors in the function, if +/// we can. Return true if we can, false otherwise. +static bool EvaluateStaticConstructor(Function *F, const DataLayout *TD, + const TargetLibraryInfo *TLI) { // Call the function. + Evaluator Eval(TD, TLI); Constant *RetValDummy; - bool EvalSuccess = EvaluateFunction(F, RetValDummy, - SmallVector(), CallStack, - MutatedMemory, AllocaTmps, - SimpleConstants, TD); - + bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy, + SmallVector()); + if (EvalSuccess) { // We succeeded at evaluation: commit the result. DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" - << F->getName() << "' to " << MutatedMemory.size() + << F->getName() << "' to " << Eval.getMutatedMemory().size() << " stores.\n"); - for (DenseMap::iterator I = MutatedMemory.begin(), - E = MutatedMemory.end(); I != E; ++I) + for (DenseMap::const_iterator I = + Eval.getMutatedMemory().begin(), E = Eval.getMutatedMemory().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. - if (!Tmp->use_empty()) - Tmp->replaceAllUsesWith(Constant::getNullValue(Tmp->getType())); - delete Tmp; + for (SmallPtrSet::const_iterator I = + Eval.getInvariants().begin(), E = Eval.getInvariants().end(); + I != E; ++I) + (*I)->setConstant(true); } return EvalSuccess; } - - /// OptimizeGlobalCtorsList - Simplify and evaluation global ctors if possible. /// Return true if anything changed. bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { @@ -2609,7 +3005,6 @@ bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { bool MadeChange = false; if (Ctors.empty()) return false; - const TargetData *TD = getAnalysisIfAvailable(); // Loop over global ctors, optimizing them when we can. for (unsigned i = 0; i != Ctors.size(); ++i) { Function *F = Ctors[i]; @@ -2622,12 +3017,13 @@ bool GlobalOpt::OptimizeGlobalCtorsList(GlobalVariable *&GCL) { } break; } + DEBUG(dbgs() << "Optimizing Global Constructor: " << *F << "\n"); // We cannot simplify external ctor functions. if (F->empty()) continue; // If we can evaluate the ctor at compile time, do. - if (EvaluateStaticConstructor(F, TD)) { + if (EvaluateStaticConstructor(F, TD, TLI)) { Ctors.erase(Ctors.begin()+i); MadeChange = true; --i; @@ -2699,15 +3095,18 @@ bool GlobalOpt::OptimizeGlobalAliases(Module &M) { return Changed; } -static Function *FindCXAAtExit(Module &M) { - Function *Fn = M.getFunction("__cxa_atexit"); - +static Function *FindCXAAtExit(Module &M, TargetLibraryInfo *TLI) { + if (!TLI->has(LibFunc::cxa_atexit)) + return 0; + + Function *Fn = M.getFunction(TLI->getName(LibFunc::cxa_atexit)); + if (!Fn) return 0; - + FunctionType *FTy = Fn->getFunctionType(); - - // Checking that the function has the right return type, the right number of + + // 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 || @@ -2723,7 +3122,8 @@ static Function *FindCXAAtExit(Module &M) { /// 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. +/// the only allowed instructions are 'ret', 'call' to an empty C++ dtor and +/// other side-effect free instructions. static bool cxxDtorIsEmpty(const Function &Fn, SmallPtrSet &CalledFunctions) { // FIXME: We could eliminate C++ destructors if they're readonly/readnone and @@ -2756,9 +3156,9 @@ static bool cxxDtorIsEmpty(const Function &Fn, if (!cxxDtorIsEmpty(*CalledFn, NewCalledFunctions)) return false; } else if (isa(*I)) - return true; - else - return false; + return true; // We're done. + else if (I->mayHaveSideEffects()) + return false; // Destructor with side effects, bail. } return false; @@ -2781,7 +3181,7 @@ bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) { // and remove them. bool Changed = false; - for (Function::use_iterator I = CXAAtExitFn->use_begin(), + 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 @@ -2790,7 +3190,7 @@ bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) { if (!CI) continue; - Function *DtorFn = + Function *DtorFn = dyn_cast(CI->getArgOperand(0)->stripPointerCasts()); if (!DtorFn) continue; @@ -2814,10 +3214,13 @@ bool GlobalOpt::OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) { bool GlobalOpt::runOnModule(Module &M) { bool Changed = false; + TD = getAnalysisIfAvailable(); + TLI = &getAnalysis(); + // Try to find the llvm.globalctors list. GlobalVariable *GlobalCtors = FindGlobalCtors(M); - Function *CXAAtExitFn = FindCXAAtExit(M); + Function *CXAAtExitFn = FindCXAAtExit(M, TLI); bool LocalChange = true; while (LocalChange) {