X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FLazyValueInfo.cpp;h=281ff89bfe46f869a6ba4d654bf97f717c201c87;hb=5ade584a968527f80fc230404656038fde88f2d1;hp=3148acaac4e67e85f71213cb637698a614d2c804;hpb=5553a3a5108f2942672cfac6f7bcf433acaf89ff;p=oota-llvm.git diff --git a/lib/Analysis/LazyValueInfo.cpp b/lib/Analysis/LazyValueInfo.cpp index 3148acaac4e..281ff89bfe4 100644 --- a/lib/Analysis/LazyValueInfo.cpp +++ b/lib/Analysis/LazyValueInfo.cpp @@ -12,23 +12,35 @@ // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "lazy-value-info" #include "llvm/Analysis/LazyValueInfo.h" -#include "llvm/Constants.h" -#include "llvm/Instructions.h" +#include "llvm/ADT/DenseSet.h" +#include "llvm/ADT/STLExtras.h" #include "llvm/Analysis/ConstantFolding.h" -#include "llvm/Target/TargetData.h" -#include "llvm/Support/CFG.h" +#include "llvm/Analysis/ValueTracking.h" +#include "llvm/IR/CFG.h" +#include "llvm/IR/ConstantRange.h" +#include "llvm/IR/Constants.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/PatternMatch.h" +#include "llvm/IR/ValueHandle.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/ADT/DenseMap.h" -#include "llvm/ADT/PointerIntPair.h" -#include "llvm/ADT/STLExtras.h" +#include "llvm/Target/TargetLibraryInfo.h" +#include +#include using namespace llvm; +using namespace PatternMatch; + +#define DEBUG_TYPE "lazy-value-info" char LazyValueInfo::ID = 0; -static RegisterPass -X("lazy-value-info", "Lazy Value Information Analysis", false, true); +INITIALIZE_PASS_BEGIN(LazyValueInfo, "lazy-value-info", + "Lazy Value Information Analysis", false, true) +INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) +INITIALIZE_PASS_END(LazyValueInfo, "lazy-value-info", + "Lazy Value Information Analysis", false, true) namespace llvm { FunctionPass *createLazyValueInfoPass() { return new LazyValueInfo(); } @@ -48,89 +60,130 @@ namespace llvm { namespace { class LVILatticeVal { enum LatticeValueTy { - /// undefined - This LLVM Value has no known value yet. + /// undefined - This Value has no known value yet. undefined, - /// constant - This LLVM Value has a specific constant value. - constant, - /// notconstant - This LLVM value is known to not have the specified value. + /// constant - This Value has a specific constant value. + constant, + /// notconstant - This Value is known to not have the specified value. notconstant, - - /// overdefined - This instruction is not known to be constant, and we know + + /// constantrange - The Value falls within this range. + constantrange, + + /// overdefined - This value is not known to be constant, and we know that /// it has a value. overdefined }; /// Val: This stores the current lattice value along with the Constant* for /// the constant if this is a 'constant' or 'notconstant' value. - PointerIntPair Val; + LatticeValueTy Tag; + Constant *Val; + ConstantRange Range; public: - LVILatticeVal() : Val(0, undefined) {} + LVILatticeVal() : Tag(undefined), Val(nullptr), Range(1, true) {} static LVILatticeVal get(Constant *C) { LVILatticeVal Res; - Res.markConstant(C); + if (!isa(C)) + Res.markConstant(C); return Res; } static LVILatticeVal getNot(Constant *C) { LVILatticeVal Res; - Res.markNotConstant(C); + if (!isa(C)) + Res.markNotConstant(C); + return Res; + } + static LVILatticeVal getRange(ConstantRange CR) { + LVILatticeVal Res; + Res.markConstantRange(CR); return Res; } - bool isUndefined() const { return Val.getInt() == undefined; } - bool isConstant() const { return Val.getInt() == constant; } - bool isNotConstant() const { return Val.getInt() == notconstant; } - bool isOverdefined() const { return Val.getInt() == overdefined; } + bool isUndefined() const { return Tag == undefined; } + bool isConstant() const { return Tag == constant; } + bool isNotConstant() const { return Tag == notconstant; } + bool isConstantRange() const { return Tag == constantrange; } + bool isOverdefined() const { return Tag == overdefined; } Constant *getConstant() const { assert(isConstant() && "Cannot get the constant of a non-constant!"); - return Val.getPointer(); + return Val; } Constant *getNotConstant() const { assert(isNotConstant() && "Cannot get the constant of a non-notconstant!"); - return Val.getPointer(); + return Val; + } + + ConstantRange getConstantRange() const { + assert(isConstantRange() && + "Cannot get the constant-range of a non-constant-range!"); + return Range; } /// markOverdefined - Return true if this is a change in status. bool markOverdefined() { if (isOverdefined()) return false; - Val.setInt(overdefined); + Tag = overdefined; return true; } /// markConstant - Return true if this is a change in status. bool markConstant(Constant *V) { - if (isConstant()) { - assert(getConstant() == V && "Marking constant with different value"); + assert(V && "Marking constant with NULL"); + if (ConstantInt *CI = dyn_cast(V)) + return markConstantRange(ConstantRange(CI->getValue())); + if (isa(V)) return false; - } - + + assert((!isConstant() || getConstant() == V) && + "Marking constant with different value"); assert(isUndefined()); - Val.setInt(constant); - assert(V && "Marking constant with NULL"); - Val.setPointer(V); + Tag = constant; + Val = V; return true; } /// markNotConstant - Return true if this is a change in status. bool markNotConstant(Constant *V) { - if (isNotConstant()) { - assert(getNotConstant() == V && "Marking !constant with different value"); + assert(V && "Marking constant with NULL"); + if (ConstantInt *CI = dyn_cast(V)) + return markConstantRange(ConstantRange(CI->getValue()+1, CI->getValue())); + if (isa(V)) return false; + + assert((!isConstant() || getConstant() != V) && + "Marking constant !constant with same value"); + assert((!isNotConstant() || getNotConstant() == V) && + "Marking !constant with different value"); + assert(isUndefined() || isConstant()); + Tag = notconstant; + Val = V; + return true; + } + + /// markConstantRange - Return true if this is a change in status. + bool markConstantRange(const ConstantRange NewR) { + if (isConstantRange()) { + if (NewR.isEmptySet()) + return markOverdefined(); + + bool changed = Range != NewR; + Range = NewR; + return changed; } - if (isConstant()) - assert(getConstant() != V && "Marking not constant with different value"); - else - assert(isUndefined()); - - Val.setInt(notconstant); - assert(V && "Marking constant with NULL"); - Val.setPointer(V); + assert(isUndefined()); + if (NewR.isEmptySet()) + return markOverdefined(); + + Tag = constantrange; + Range = NewR; return true; } @@ -140,49 +193,88 @@ public: if (RHS.isUndefined() || isOverdefined()) return false; if (RHS.isOverdefined()) return markOverdefined(); - if (RHS.isNotConstant()) { - if (isNotConstant()) { - if (getNotConstant() != RHS.getNotConstant() || - isa(getNotConstant()) || - isa(RHS.getNotConstant())) - return markOverdefined(); - return false; + if (isUndefined()) { + Tag = RHS.Tag; + Val = RHS.Val; + Range = RHS.Range; + return true; + } + + if (isConstant()) { + if (RHS.isConstant()) { + if (Val == RHS.Val) + return false; + return markOverdefined(); } - if (isConstant()) { - if (getConstant() == RHS.getNotConstant() || - isa(RHS.getNotConstant()) || - isa(getConstant())) + + if (RHS.isNotConstant()) { + if (Val == RHS.Val) return markOverdefined(); - return markNotConstant(RHS.getNotConstant()); + + // Unless we can prove that the two Constants are different, we must + // move to overdefined. + // FIXME: use DataLayout/TargetLibraryInfo for smarter constant folding. + if (ConstantInt *Res = dyn_cast( + ConstantFoldCompareInstOperands(CmpInst::ICMP_NE, + getConstant(), + RHS.getNotConstant()))) + if (Res->isOne()) + return markNotConstant(RHS.getNotConstant()); + + return markOverdefined(); } - - assert(isUndefined() && "Unexpected lattice"); - return markNotConstant(RHS.getNotConstant()); + + // RHS is a ConstantRange, LHS is a non-integer Constant. + + // FIXME: consider the case where RHS is a range [1, 0) and LHS is + // a function. The correct result is to pick up RHS. + + return markOverdefined(); } - - // RHS must be a constant, we must be undef, constant, or notconstant. - if (isUndefined()) - return markConstant(RHS.getConstant()); - - if (isConstant()) { - if (getConstant() != RHS.getConstant()) + + if (isNotConstant()) { + if (RHS.isConstant()) { + if (Val == RHS.Val) + return markOverdefined(); + + // Unless we can prove that the two Constants are different, we must + // move to overdefined. + // FIXME: use DataLayout/TargetLibraryInfo for smarter constant folding. + if (ConstantInt *Res = dyn_cast( + ConstantFoldCompareInstOperands(CmpInst::ICMP_NE, + getNotConstant(), + RHS.getConstant()))) + if (Res->isOne()) + return false; + return markOverdefined(); - return false; + } + + if (RHS.isNotConstant()) { + if (Val == RHS.Val) + return false; + return markOverdefined(); + } + + return markOverdefined(); } - // If we are known "!=4" and RHS is "==5", stay at "!=4". - if (getNotConstant() == RHS.getConstant() || - isa(getNotConstant()) || - isa(RHS.getConstant())) + assert(isConstantRange() && "New LVILattice type?"); + if (!RHS.isConstantRange()) return markOverdefined(); - return false; + + ConstantRange NewR = Range.unionWith(RHS.getConstantRange()); + if (NewR.isFullSet()) + return markOverdefined(); + return markConstantRange(NewR); } - }; } // end anonymous namespace. namespace llvm { +raw_ostream &operator<<(raw_ostream &OS, const LVILatticeVal &Val) + LLVM_ATTRIBUTE_USED; raw_ostream &operator<<(raw_ostream &OS, const LVILatticeVal &Val) { if (Val.isUndefined()) return OS << "undefined"; @@ -191,6 +283,9 @@ raw_ostream &operator<<(raw_ostream &OS, const LVILatticeVal &Val) { if (Val.isNotConstant()) return OS << "notconstant<" << *Val.getNotConstant() << '>'; + else if (Val.isConstantRange()) + return OS << "constantrange<" << Val.getConstantRange().getLower() << ", " + << Val.getConstantRange().getUpper() << '>'; return OS << "constant<" << *Val.getConstant() << '>'; } } @@ -200,25 +295,96 @@ raw_ostream &operator<<(raw_ostream &OS, const LVILatticeVal &Val) { //===----------------------------------------------------------------------===// namespace { + /// LVIValueHandle - A callback value handle updates the cache when + /// values are erased. + class LazyValueInfoCache; + struct LVIValueHandle : public CallbackVH { + LazyValueInfoCache *Parent; + + LVIValueHandle(Value *V, LazyValueInfoCache *P) + : CallbackVH(V), Parent(P) { } + + void deleted() override; + void allUsesReplacedWith(Value *V) override { + deleted(); + } + }; +} + +namespace { /// LazyValueInfoCache - This is the cache kept by LazyValueInfo which /// maintains information about queries across the clients' queries. class LazyValueInfoCache { - public: - /// BlockCacheEntryTy - This is a computed lattice value at the end of the - /// specified basic block for a Value* that depends on context. - typedef std::pair BlockCacheEntryTy; - /// ValueCacheEntryTy - This is all of the cached block information for /// exactly one Value*. The entries are sorted by the BasicBlock* of the /// entries, allowing us to do a lookup with a binary search. - typedef std::vector ValueCacheEntryTy; + typedef std::map, LVILatticeVal> ValueCacheEntryTy; - private: /// ValueCache - This is all of the cached information for all values, /// mapped from Value* to key information. - DenseMap ValueCache; - public: + std::map ValueCache; + /// OverDefinedCache - This tracks, on a per-block basis, the set of + /// values that are over-defined at the end of that block. This is required + /// for cache updating. + typedef std::pair, Value*> OverDefinedPairTy; + DenseSet OverDefinedCache; + + /// SeenBlocks - Keep track of all blocks that we have ever seen, so we + /// don't spend time removing unused blocks from our caches. + DenseSet > SeenBlocks; + + /// BlockValueStack - This stack holds the state of the value solver + /// during a query. It basically emulates the callstack of the naive + /// recursive value lookup process. + std::stack > BlockValueStack; + + friend struct LVIValueHandle; + + /// OverDefinedCacheUpdater - A helper object that ensures that the + /// OverDefinedCache is updated whenever solveBlockValue returns. + struct OverDefinedCacheUpdater { + LazyValueInfoCache *Parent; + Value *Val; + BasicBlock *BB; + LVILatticeVal &BBLV; + + OverDefinedCacheUpdater(Value *V, BasicBlock *B, LVILatticeVal &LV, + LazyValueInfoCache *P) + : Parent(P), Val(V), BB(B), BBLV(LV) { } + + bool markResult(bool changed) { + if (changed && BBLV.isOverdefined()) + Parent->OverDefinedCache.insert(std::make_pair(BB, Val)); + return changed; + } + }; + + + + LVILatticeVal getBlockValue(Value *Val, BasicBlock *BB); + bool getEdgeValue(Value *V, BasicBlock *F, BasicBlock *T, + LVILatticeVal &Result); + bool hasBlockValue(Value *Val, BasicBlock *BB); + + // These methods process one work item and may add more. A false value + // returned means that the work item was not completely processed and must + // be revisited after going through the new items. + bool solveBlockValue(Value *Val, BasicBlock *BB); + bool solveBlockValueNonLocal(LVILatticeVal &BBLV, + Value *Val, BasicBlock *BB); + bool solveBlockValuePHINode(LVILatticeVal &BBLV, + PHINode *PN, BasicBlock *BB); + bool solveBlockValueConstantRange(LVILatticeVal &BBLV, + Instruction *BBI, BasicBlock *BB); + + void solve(); + + ValueCacheEntryTy &lookup(Value *V) { + return ValueCache[LVIValueHandle(V, this)]; + } + + public: /// getValueInBlock - This is the query interface to determine the lattice /// value for the specified Value* at the end of the specified block. LVILatticeVal getValueInBlock(Value *V, BasicBlock *BB); @@ -226,124 +392,138 @@ namespace { /// getValueOnEdge - This is the query interface to determine the lattice /// value for the specified Value* that is true on the specified edge. LVILatticeVal getValueOnEdge(Value *V, BasicBlock *FromBB,BasicBlock *ToBB); - }; -} // end anonymous namespace - -namespace { - struct BlockCacheEntryComparator { - static int Compare(const void *LHSv, const void *RHSv) { - const LazyValueInfoCache::BlockCacheEntryTy *LHS = - static_cast(LHSv); - const LazyValueInfoCache::BlockCacheEntryTy *RHS = - static_cast(RHSv); - if (LHS->first < RHS->first) - return -1; - if (LHS->first > RHS->first) - return 1; - return 0; - } - bool operator()(const LazyValueInfoCache::BlockCacheEntryTy &LHS, - const LazyValueInfoCache::BlockCacheEntryTy &RHS) const { - return LHS.first < RHS.first; + /// threadEdge - This is the update interface to inform the cache that an + /// edge from PredBB to OldSucc has been threaded to be from PredBB to + /// NewSucc. + void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc); + + /// eraseBlock - This is part of the update interface to inform the cache + /// that a block has been deleted. + void eraseBlock(BasicBlock *BB); + + /// clear - Empty the cache. + void clear() { + SeenBlocks.clear(); + ValueCache.clear(); + OverDefinedCache.clear(); } }; +} // end anonymous namespace + +void LVIValueHandle::deleted() { + typedef std::pair, Value*> OverDefinedPairTy; + + SmallVector ToErase; + for (DenseSet::iterator + I = Parent->OverDefinedCache.begin(), + E = Parent->OverDefinedCache.end(); + I != E; ++I) { + if (I->second == getValPtr()) + ToErase.push_back(*I); + } + + for (SmallVectorImpl::iterator I = ToErase.begin(), + E = ToErase.end(); I != E; ++I) + Parent->OverDefinedCache.erase(*I); + + // This erasure deallocates *this, so it MUST happen after we're done + // using any and all members of *this. + Parent->ValueCache.erase(*this); } -//===----------------------------------------------------------------------===// -// LVIQuery Impl -//===----------------------------------------------------------------------===// +void LazyValueInfoCache::eraseBlock(BasicBlock *BB) { + // Shortcut if we have never seen this block. + DenseSet >::iterator I = SeenBlocks.find(BB); + if (I == SeenBlocks.end()) + return; + SeenBlocks.erase(I); -namespace { - /// LVIQuery - This is a transient object that exists while a query is - /// being performed. - /// - /// TODO: Reuse LVIQuery instead of recreating it for every query, this avoids - /// reallocation of the densemap on every query. - class LVIQuery { - typedef LazyValueInfoCache::BlockCacheEntryTy BlockCacheEntryTy; - typedef LazyValueInfoCache::ValueCacheEntryTy ValueCacheEntryTy; - - /// This is the current value being queried for. - Value *Val; - - /// This is all of the cached information about this value. - ValueCacheEntryTy &Cache; - - /// NewBlocks - This is a mpping of the new BasicBlocks which have been - /// added to cache but that are not in sorted order. - DenseMap NewBlockInfo; - public: - - LVIQuery(Value *V, ValueCacheEntryTy &VC) : Val(V), Cache(VC) { - } + SmallVector ToErase; + for (DenseSet::iterator I = OverDefinedCache.begin(), + E = OverDefinedCache.end(); I != E; ++I) { + if (I->first == BB) + ToErase.push_back(*I); + } - ~LVIQuery() { - // When the query is done, insert the newly discovered facts into the - // cache in sorted order. - if (NewBlockInfo.empty()) return; + for (SmallVectorImpl::iterator I = ToErase.begin(), + E = ToErase.end(); I != E; ++I) + OverDefinedCache.erase(*I); - // Grow the cache to exactly fit the new data. - Cache.reserve(Cache.size() + NewBlockInfo.size()); - - // If we only have one new entry, insert it instead of doing a full-on - // sort. - if (NewBlockInfo.size() == 1) { - BlockCacheEntryTy Entry = *NewBlockInfo.begin(); - ValueCacheEntryTy::iterator I = - std::lower_bound(Cache.begin(), Cache.end(), Entry, - BlockCacheEntryComparator()); - assert((I == Cache.end() || I->first != Entry.first) && - "Entry already in map!"); - - Cache.insert(I, Entry); - return; - } - - // TODO: If we only have two new elements, INSERT them both. - - Cache.insert(Cache.end(), NewBlockInfo.begin(), NewBlockInfo.end()); - array_pod_sort(Cache.begin(), Cache.end(), - BlockCacheEntryComparator::Compare); - + for (std::map::iterator + I = ValueCache.begin(), E = ValueCache.end(); I != E; ++I) + I->second.erase(BB); +} + +void LazyValueInfoCache::solve() { + while (!BlockValueStack.empty()) { + std::pair &e = BlockValueStack.top(); + if (solveBlockValue(e.second, e.first)) { + assert(BlockValueStack.top() == e); + BlockValueStack.pop(); } + } +} - LVILatticeVal getBlockValue(BasicBlock *BB); - LVILatticeVal getEdgeValue(BasicBlock *FromBB, BasicBlock *ToBB); +bool LazyValueInfoCache::hasBlockValue(Value *Val, BasicBlock *BB) { + // If already a constant, there is nothing to compute. + if (isa(Val)) + return true; - private: - LVILatticeVal &getCachedEntryForBlock(BasicBlock *BB); - }; -} // end anonymous namespace + LVIValueHandle ValHandle(Val, this); + std::map::iterator I = + ValueCache.find(ValHandle); + if (I == ValueCache.end()) return false; + return I->second.count(BB); +} -/// getCachedEntryForBlock - See if we already have a value for this block. If -/// so, return it, otherwise create a new entry in the NewBlockInfo map to use. -LVILatticeVal &LVIQuery::getCachedEntryForBlock(BasicBlock *BB) { - - // Do a binary search to see if we already have an entry for this block in - // the cache set. If so, find it. - if (!Cache.empty()) { - ValueCacheEntryTy::iterator Entry = - std::lower_bound(Cache.begin(), Cache.end(), - BlockCacheEntryTy(BB, LVILatticeVal()), - BlockCacheEntryComparator()); - if (Entry != Cache.end() && Entry->first == BB) - return Entry->second; - } - - // Otherwise, check to see if it's in NewBlockInfo or create a new entry if - // not. - return NewBlockInfo[BB]; +LVILatticeVal LazyValueInfoCache::getBlockValue(Value *Val, BasicBlock *BB) { + // If already a constant, there is nothing to compute. + if (Constant *VC = dyn_cast(Val)) + return LVILatticeVal::get(VC); + + SeenBlocks.insert(BB); + return lookup(Val)[BB]; } -LVILatticeVal LVIQuery::getBlockValue(BasicBlock *BB) { - // See if we already have a value for this block. - LVILatticeVal &BBLV = getCachedEntryForBlock(BB); +bool LazyValueInfoCache::solveBlockValue(Value *Val, BasicBlock *BB) { + if (isa(Val)) + return true; + + ValueCacheEntryTy &Cache = lookup(Val); + SeenBlocks.insert(BB); + LVILatticeVal &BBLV = Cache[BB]; - // If we've already computed this block's value, return it. - if (!BBLV.isUndefined()) { - DEBUG(errs() << " reuse BB '" << BB->getName() << "' val=" << BBLV <<'\n'); - return BBLV; + // OverDefinedCacheUpdater is a helper object that will update + // the OverDefinedCache for us when this method exits. Make sure to + // call markResult on it as we exist, passing a bool to indicate if the + // cache needs updating, i.e. if we have solve a new value or not. + OverDefinedCacheUpdater ODCacheUpdater(Val, BB, BBLV, this); + + // Once this BB is encountered, Val's value for this BB will not be Undefined + // any longer. When we encounter this BB again, if Val's value is Overdefined, + // we need to compute its value again. + // + // For example, considering this control flow, + // BB1->BB2, BB1->BB3, BB2->BB3, BB2->BB4 + // + // Suppose we have "icmp slt %v, 0" in BB1, and "icmp sgt %v, 0" in BB3. At + // the very beginning, when analyzing edge BB2->BB3, we don't know %v's value + // in BB2, and the data flow algorithm tries to compute BB2's predecessors, so + // then we know %v has negative value on edge BB1->BB2. And then we return to + // check BB2 again, and at this moment BB2 has Overdefined value for %v in + // BB2. So we should have to follow data flow propagation algorithm to get the + // value on edge BB1->BB2 propagated to BB2, and finally %v on BB2 has a + // constant range describing a negative value. + + if (!BBLV.isUndefined() && !BBLV.isOverdefined()) { + DEBUG(dbgs() << " reuse BB '" << BB->getName() << "' val=" << BBLV <<'\n'); + + // Since we're reusing a cached value here, we don't need to update the + // OverDefinedCahce. The cache will have been properly updated + // whenever the cached value was inserted. + ODCacheUpdater.markResult(false); + return true; } // Otherwise, this is the first time we're seeing this block. Reset the @@ -351,60 +531,271 @@ LVILatticeVal LVIQuery::getBlockValue(BasicBlock *BB) { // conservatively correct. BBLV.markOverdefined(); - // If V is live into BB, see if our predecessors know anything about it. Instruction *BBI = dyn_cast(Val); - if (BBI == 0 || BBI->getParent() != BB) { - LVILatticeVal Result; // Start Undefined. - unsigned NumPreds = 0; - - // Loop over all of our predecessors, merging what we know from them into - // result. - for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { - Result.mergeIn(getEdgeValue(*PI, BB)); - - // If we hit overdefined, exit early. The BlockVals entry is already set - // to overdefined. - if (Result.isOverdefined()) { - DEBUG(errs() << " compute BB '" << BB->getName() - << "' - overdefined because of pred.\n"); - return Result; + if (!BBI || BBI->getParent() != BB) { + return ODCacheUpdater.markResult(solveBlockValueNonLocal(BBLV, Val, BB)); + } + + if (PHINode *PN = dyn_cast(BBI)) { + return ODCacheUpdater.markResult(solveBlockValuePHINode(BBLV, PN, BB)); + } + + if (AllocaInst *AI = dyn_cast(BBI)) { + BBLV = LVILatticeVal::getNot(ConstantPointerNull::get(AI->getType())); + return ODCacheUpdater.markResult(true); + } + + // We can only analyze the definitions of certain classes of instructions + // (integral binops and casts at the moment), so bail if this isn't one. + LVILatticeVal Result; + if ((!isa(BBI) && !isa(BBI)) || + !BBI->getType()->isIntegerTy()) { + DEBUG(dbgs() << " compute BB '" << BB->getName() + << "' - overdefined because inst def found.\n"); + BBLV.markOverdefined(); + return ODCacheUpdater.markResult(true); + } + + // FIXME: We're currently limited to binops with a constant RHS. This should + // be improved. + BinaryOperator *BO = dyn_cast(BBI); + if (BO && !isa(BO->getOperand(1))) { + DEBUG(dbgs() << " compute BB '" << BB->getName() + << "' - overdefined because inst def found.\n"); + + BBLV.markOverdefined(); + return ODCacheUpdater.markResult(true); + } + + return ODCacheUpdater.markResult(solveBlockValueConstantRange(BBLV, BBI, BB)); +} + +static bool InstructionDereferencesPointer(Instruction *I, Value *Ptr) { + if (LoadInst *L = dyn_cast(I)) { + return L->getPointerAddressSpace() == 0 && + GetUnderlyingObject(L->getPointerOperand()) == Ptr; + } + if (StoreInst *S = dyn_cast(I)) { + return S->getPointerAddressSpace() == 0 && + GetUnderlyingObject(S->getPointerOperand()) == Ptr; + } + if (MemIntrinsic *MI = dyn_cast(I)) { + if (MI->isVolatile()) return false; + + // FIXME: check whether it has a valuerange that excludes zero? + ConstantInt *Len = dyn_cast(MI->getLength()); + if (!Len || Len->isZero()) return false; + + if (MI->getDestAddressSpace() == 0) + if (GetUnderlyingObject(MI->getRawDest()) == Ptr) + return true; + if (MemTransferInst *MTI = dyn_cast(MI)) + if (MTI->getSourceAddressSpace() == 0) + if (GetUnderlyingObject(MTI->getRawSource()) == Ptr) + return true; + } + return false; +} + +bool LazyValueInfoCache::solveBlockValueNonLocal(LVILatticeVal &BBLV, + Value *Val, BasicBlock *BB) { + LVILatticeVal Result; // Start Undefined. + + // If this is a pointer, and there's a load from that pointer in this BB, + // then we know that the pointer can't be NULL. + bool NotNull = false; + if (Val->getType()->isPointerTy()) { + if (isKnownNonNull(Val)) { + NotNull = true; + } else { + Value *UnderlyingVal = GetUnderlyingObject(Val); + // If 'GetUnderlyingObject' didn't converge, skip it. It won't converge + // inside InstructionDereferencesPointer either. + if (UnderlyingVal == GetUnderlyingObject(UnderlyingVal, nullptr, 1)) { + for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); + BI != BE; ++BI) { + if (InstructionDereferencesPointer(BI, UnderlyingVal)) { + NotNull = true; + break; + } + } } - ++NumPreds; } - - // If this is the entry block, we must be asking about an argument. The - // value is overdefined. - if (NumPreds == 0 && BB == &BB->getParent()->front()) { - assert(isa(Val) && "Unknown live-in to the entry block"); + } + + // If this is the entry block, we must be asking about an argument. The + // value is overdefined. + if (BB == &BB->getParent()->getEntryBlock()) { + assert(isa(Val) && "Unknown live-in to the entry block"); + if (NotNull) { + PointerType *PTy = cast(Val->getType()); + Result = LVILatticeVal::getNot(ConstantPointerNull::get(PTy)); + } else { Result.markOverdefined(); - return Result; } - - // Return the merged value, which is more precise than 'overdefined'. - assert(!Result.isOverdefined()); - return getCachedEntryForBlock(BB) = Result; + BBLV = Result; + return true; + } + + // Loop over all of our predecessors, merging what we know from them into + // result. + bool EdgesMissing = false; + for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { + LVILatticeVal EdgeResult; + EdgesMissing |= !getEdgeValue(Val, *PI, BB, EdgeResult); + if (EdgesMissing) + continue; + + Result.mergeIn(EdgeResult); + + // If we hit overdefined, exit early. The BlockVals entry is already set + // to overdefined. + if (Result.isOverdefined()) { + DEBUG(dbgs() << " compute BB '" << BB->getName() + << "' - overdefined because of pred.\n"); + // If we previously determined that this is a pointer that can't be null + // then return that rather than giving up entirely. + if (NotNull) { + PointerType *PTy = cast(Val->getType()); + Result = LVILatticeVal::getNot(ConstantPointerNull::get(PTy)); + } + + BBLV = Result; + return true; + } } + if (EdgesMissing) + return false; + + // Return the merged value, which is more precise than 'overdefined'. + assert(!Result.isOverdefined()); + BBLV = Result; + return true; +} - // If this value is defined by an instruction in this block, we have to - // process it here somehow or return overdefined. - if (PHINode *PN = dyn_cast(BBI)) { - (void)PN; - // TODO: PHI Translation in preds. - } else { - +bool LazyValueInfoCache::solveBlockValuePHINode(LVILatticeVal &BBLV, + PHINode *PN, BasicBlock *BB) { + LVILatticeVal Result; // Start Undefined. + + // Loop over all of our predecessors, merging what we know from them into + // result. + bool EdgesMissing = false; + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { + BasicBlock *PhiBB = PN->getIncomingBlock(i); + Value *PhiVal = PN->getIncomingValue(i); + LVILatticeVal EdgeResult; + EdgesMissing |= !getEdgeValue(PhiVal, PhiBB, BB, EdgeResult); + if (EdgesMissing) + continue; + + Result.mergeIn(EdgeResult); + + // If we hit overdefined, exit early. The BlockVals entry is already set + // to overdefined. + if (Result.isOverdefined()) { + DEBUG(dbgs() << " compute BB '" << BB->getName() + << "' - overdefined because of pred.\n"); + + BBLV = Result; + return true; + } + } + if (EdgesMissing) + return false; + + // Return the merged value, which is more precise than 'overdefined'. + assert(!Result.isOverdefined() && "Possible PHI in entry block?"); + BBLV = Result; + return true; +} + +bool LazyValueInfoCache::solveBlockValueConstantRange(LVILatticeVal &BBLV, + Instruction *BBI, + BasicBlock *BB) { + // Figure out the range of the LHS. If that fails, bail. + if (!hasBlockValue(BBI->getOperand(0), BB)) { + BlockValueStack.push(std::make_pair(BB, BBI->getOperand(0))); + return false; + } + + LVILatticeVal LHSVal = getBlockValue(BBI->getOperand(0), BB); + if (!LHSVal.isConstantRange()) { + BBLV.markOverdefined(); + return true; } - DEBUG(errs() << " compute BB '" << BB->getName() - << "' - overdefined because inst def found.\n"); + ConstantRange LHSRange = LHSVal.getConstantRange(); + ConstantRange RHSRange(1); + IntegerType *ResultTy = cast(BBI->getType()); + if (isa(BBI)) { + if (ConstantInt *RHS = dyn_cast(BBI->getOperand(1))) { + RHSRange = ConstantRange(RHS->getValue()); + } else { + BBLV.markOverdefined(); + return true; + } + } + // NOTE: We're currently limited by the set of operations that ConstantRange + // can evaluate symbolically. Enhancing that set will allows us to analyze + // more definitions. LVILatticeVal Result; - Result.markOverdefined(); - return getCachedEntryForBlock(BB) = Result; + switch (BBI->getOpcode()) { + case Instruction::Add: + Result.markConstantRange(LHSRange.add(RHSRange)); + break; + case Instruction::Sub: + Result.markConstantRange(LHSRange.sub(RHSRange)); + break; + case Instruction::Mul: + Result.markConstantRange(LHSRange.multiply(RHSRange)); + break; + case Instruction::UDiv: + Result.markConstantRange(LHSRange.udiv(RHSRange)); + break; + case Instruction::Shl: + Result.markConstantRange(LHSRange.shl(RHSRange)); + break; + case Instruction::LShr: + Result.markConstantRange(LHSRange.lshr(RHSRange)); + break; + case Instruction::Trunc: + Result.markConstantRange(LHSRange.truncate(ResultTy->getBitWidth())); + break; + case Instruction::SExt: + Result.markConstantRange(LHSRange.signExtend(ResultTy->getBitWidth())); + break; + case Instruction::ZExt: + Result.markConstantRange(LHSRange.zeroExtend(ResultTy->getBitWidth())); + break; + case Instruction::BitCast: + Result.markConstantRange(LHSRange); + break; + case Instruction::And: + Result.markConstantRange(LHSRange.binaryAnd(RHSRange)); + break; + case Instruction::Or: + Result.markConstantRange(LHSRange.binaryOr(RHSRange)); + break; + + // Unhandled instructions are overdefined. + default: + DEBUG(dbgs() << " compute BB '" << BB->getName() + << "' - overdefined because inst def found.\n"); + Result.markOverdefined(); + break; + } + + BBLV = Result; + return true; } - -/// getEdgeValue - This method -LVILatticeVal LVIQuery::getEdgeValue(BasicBlock *BBFrom, BasicBlock *BBTo) { +/// \brief Compute the value of Val on the edge BBFrom -> BBTo. Returns false if +/// Val is not constrained on the edge. +static bool getEdgeValueLocal(Value *Val, BasicBlock *BBFrom, + BasicBlock *BBTo, LVILatticeVal &Result) { + // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we + // know that v != 0. if (BranchInst *BI = dyn_cast(BBFrom->getTerminator())) { // If this is a conditional branch and only one successor goes to BBTo, then // we maybe able to infer something from the condition. @@ -416,78 +807,217 @@ LVILatticeVal LVIQuery::getEdgeValue(BasicBlock *BBFrom, BasicBlock *BBTo) { // If V is the condition of the branch itself, then we know exactly what // it is. - if (BI->getCondition() == Val) - return LVILatticeVal::get(ConstantInt::get( - Type::getInt1Ty(Val->getContext()), isTrueDest)); + if (BI->getCondition() == Val) { + Result = LVILatticeVal::get(ConstantInt::get( + Type::getInt1Ty(Val->getContext()), isTrueDest)); + return true; + } // If the condition of the branch is an equality comparison, we may be // able to infer the value. - if (ICmpInst *ICI = dyn_cast(BI->getCondition())) - if (ICI->isEquality() && ICI->getOperand(0) == Val && - isa(ICI->getOperand(1))) { + ICmpInst *ICI = dyn_cast(BI->getCondition()); + if (ICI && isa(ICI->getOperand(1))) { + if (ICI->isEquality() && ICI->getOperand(0) == Val) { // We know that V has the RHS constant if this is a true SETEQ or // false SETNE. if (isTrueDest == (ICI->getPredicate() == ICmpInst::ICMP_EQ)) - return LVILatticeVal::get(cast(ICI->getOperand(1))); - return LVILatticeVal::getNot(cast(ICI->getOperand(1))); + Result = LVILatticeVal::get(cast(ICI->getOperand(1))); + else + Result = LVILatticeVal::getNot(cast(ICI->getOperand(1))); + return true; + } + + // Recognize the range checking idiom that InstCombine produces. + // (X-C1) u< C2 --> [C1, C1+C2) + ConstantInt *NegOffset = nullptr; + if (ICI->getPredicate() == ICmpInst::ICMP_ULT) + match(ICI->getOperand(0), m_Add(m_Specific(Val), + m_ConstantInt(NegOffset))); + + ConstantInt *CI = dyn_cast(ICI->getOperand(1)); + if (CI && (ICI->getOperand(0) == Val || NegOffset)) { + // Calculate the range of values that would satisfy the comparison. + ConstantRange CmpRange(CI->getValue()); + ConstantRange TrueValues = + ConstantRange::makeICmpRegion(ICI->getPredicate(), CmpRange); + + if (NegOffset) // Apply the offset from above. + TrueValues = TrueValues.subtract(NegOffset->getValue()); + + // If we're interested in the false dest, invert the condition. + if (!isTrueDest) TrueValues = TrueValues.inverse(); + + Result = LVILatticeVal::getRange(TrueValues); + return true; } + } } } - - // TODO: Info from switch. - - // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we - // know that v != 0. - - // Otherwise see if the value is known in the block. - return getBlockValue(BBFrom); + + // If the edge was formed by a switch on the value, then we may know exactly + // what it is. + if (SwitchInst *SI = dyn_cast(BBFrom->getTerminator())) { + if (SI->getCondition() != Val) + return false; + + bool DefaultCase = SI->getDefaultDest() == BBTo; + unsigned BitWidth = Val->getType()->getIntegerBitWidth(); + ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/); + + for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); + i != e; ++i) { + ConstantRange EdgeVal(i.getCaseValue()->getValue()); + if (DefaultCase) { + // It is possible that the default destination is the destination of + // some cases. There is no need to perform difference for those cases. + if (i.getCaseSuccessor() != BBTo) + EdgesVals = EdgesVals.difference(EdgeVal); + } else if (i.getCaseSuccessor() == BBTo) + EdgesVals = EdgesVals.unionWith(EdgeVal); + } + Result = LVILatticeVal::getRange(EdgesVals); + return true; + } + return false; } +/// \brief Compute the value of Val on the edge BBFrom -> BBTo, or the value at +/// the basic block if the edge does not constraint Val. +bool LazyValueInfoCache::getEdgeValue(Value *Val, BasicBlock *BBFrom, + BasicBlock *BBTo, LVILatticeVal &Result) { + // If already a constant, there is nothing to compute. + if (Constant *VC = dyn_cast(Val)) { + Result = LVILatticeVal::get(VC); + return true; + } -//===----------------------------------------------------------------------===// -// LazyValueInfoCache Impl -//===----------------------------------------------------------------------===// + if (getEdgeValueLocal(Val, BBFrom, BBTo, Result)) { + if (!Result.isConstantRange() || + Result.getConstantRange().getSingleElement()) + return true; + + // FIXME: this check should be moved to the beginning of the function when + // LVI better supports recursive values. Even for the single value case, we + // can intersect to detect dead code (an empty range). + if (!hasBlockValue(Val, BBFrom)) { + BlockValueStack.push(std::make_pair(BBFrom, Val)); + return false; + } + + // Try to intersect ranges of the BB and the constraint on the edge. + LVILatticeVal InBlock = getBlockValue(Val, BBFrom); + if (!InBlock.isConstantRange()) + return true; + + ConstantRange Range = + Result.getConstantRange().intersectWith(InBlock.getConstantRange()); + Result = LVILatticeVal::getRange(Range); + return true; + } + + if (!hasBlockValue(Val, BBFrom)) { + BlockValueStack.push(std::make_pair(BBFrom, Val)); + return false; + } + + // if we couldn't compute the value on the edge, use the value from the BB + Result = getBlockValue(Val, BBFrom); + return true; +} LVILatticeVal LazyValueInfoCache::getValueInBlock(Value *V, BasicBlock *BB) { - // If already a constant, there is nothing to compute. - if (Constant *VC = dyn_cast(V)) - return LVILatticeVal::get(VC); - - DEBUG(errs() << "LVI Getting block end value " << *V << " at '" + DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '" << BB->getName() << "'\n"); - LVILatticeVal Result = LVIQuery(V, ValueCache[V]).getBlockValue(BB); - - DEBUG(errs() << " Result = " << Result << "\n"); + BlockValueStack.push(std::make_pair(BB, V)); + solve(); + LVILatticeVal Result = getBlockValue(V, BB); + + DEBUG(dbgs() << " Result = " << Result << "\n"); return Result; } LVILatticeVal LazyValueInfoCache:: getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB) { - // If already a constant, there is nothing to compute. - if (Constant *VC = dyn_cast(V)) - return LVILatticeVal::get(VC); - - DEBUG(errs() << "LVI Getting edge value " << *V << " from '" + DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '" << FromBB->getName() << "' to '" << ToBB->getName() << "'\n"); - LVILatticeVal Result = - LVIQuery(V, ValueCache[V]).getEdgeValue(FromBB, ToBB); - - DEBUG(errs() << " Result = " << Result << "\n"); + LVILatticeVal Result; + if (!getEdgeValue(V, FromBB, ToBB, Result)) { + solve(); + bool WasFastQuery = getEdgeValue(V, FromBB, ToBB, Result); + (void)WasFastQuery; + assert(WasFastQuery && "More work to do after problem solved?"); + } + + DEBUG(dbgs() << " Result = " << Result << "\n"); return Result; } +void LazyValueInfoCache::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, + BasicBlock *NewSucc) { + // When an edge in the graph has been threaded, values that we could not + // determine a value for before (i.e. were marked overdefined) may be possible + // to solve now. We do NOT try to proactively update these values. Instead, + // we clear their entries from the cache, and allow lazy updating to recompute + // them when needed. + + // The updating process is fairly simple: we need to dropped cached info + // for all values that were marked overdefined in OldSucc, and for those same + // values in any successor of OldSucc (except NewSucc) in which they were + // also marked overdefined. + std::vector worklist; + worklist.push_back(OldSucc); + + DenseSet ClearSet; + for (DenseSet::iterator I = OverDefinedCache.begin(), + E = OverDefinedCache.end(); I != E; ++I) { + if (I->first == OldSucc) + ClearSet.insert(I->second); + } + + // Use a worklist to perform a depth-first search of OldSucc's successors. + // NOTE: We do not need a visited list since any blocks we have already + // visited will have had their overdefined markers cleared already, and we + // thus won't loop to their successors. + while (!worklist.empty()) { + BasicBlock *ToUpdate = worklist.back(); + worklist.pop_back(); + + // Skip blocks only accessible through NewSucc. + if (ToUpdate == NewSucc) continue; + + bool changed = false; + for (DenseSet::iterator I = ClearSet.begin(), E = ClearSet.end(); + I != E; ++I) { + // If a value was marked overdefined in OldSucc, and is here too... + DenseSet::iterator OI = + OverDefinedCache.find(std::make_pair(ToUpdate, *I)); + if (OI == OverDefinedCache.end()) continue; + + // Remove it from the caches. + ValueCacheEntryTy &Entry = ValueCache[LVIValueHandle(*I, this)]; + ValueCacheEntryTy::iterator CI = Entry.find(ToUpdate); + + assert(CI != Entry.end() && "Couldn't find entry to update?"); + Entry.erase(CI); + OverDefinedCache.erase(OI); + + // If we removed anything, then we potentially need to update + // blocks successors too. + changed = true; + } + + if (!changed) continue; + + worklist.insert(worklist.end(), succ_begin(ToUpdate), succ_end(ToUpdate)); + } +} + //===----------------------------------------------------------------------===// // LazyValueInfo Impl //===----------------------------------------------------------------------===// -bool LazyValueInfo::runOnFunction(Function &F) { - TD = getAnalysisIfAvailable(); - // Fully lazy. - return false; -} - /// getCache - This lazily constructs the LazyValueInfoCache. static LazyValueInfoCache &getCache(void *&PImpl) { if (!PImpl) @@ -495,11 +1025,28 @@ static LazyValueInfoCache &getCache(void *&PImpl) { return *static_cast(PImpl); } +bool LazyValueInfo::runOnFunction(Function &F) { + if (PImpl) + getCache(PImpl).clear(); + + DataLayoutPass *DLP = getAnalysisIfAvailable(); + DL = DLP ? &DLP->getDataLayout() : nullptr; + TLI = &getAnalysis(); + + // Fully lazy. + return false; +} + +void LazyValueInfo::getAnalysisUsage(AnalysisUsage &AU) const { + AU.setPreservesAll(); + AU.addRequired(); +} + void LazyValueInfo::releaseMemory() { // If the cache was allocated, free it. if (PImpl) { delete &getCache(PImpl); - PImpl = 0; + PImpl = nullptr; } } @@ -508,7 +1055,12 @@ Constant *LazyValueInfo::getConstant(Value *V, BasicBlock *BB) { if (Result.isConstant()) return Result.getConstant(); - return 0; + if (Result.isConstantRange()) { + ConstantRange CR = Result.getConstantRange(); + if (const APInt *SingleVal = CR.getSingleElement()) + return ConstantInt::get(V->getContext(), *SingleVal); + } + return nullptr; } /// getConstantOnEdge - Determine whether the specified value is known to be a @@ -519,7 +1071,12 @@ Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB, if (Result.isConstant()) return Result.getConstant(); - return 0; + if (Result.isConstantRange()) { + ConstantRange CR = Result.getConstantRange(); + if (const APInt *SingleVal = CR.getSingleElement()) + return ConstantInt::get(V->getContext(), *SingleVal); + } + return nullptr; } /// getPredicateOnEdge - Determine whether the specified value comparison @@ -531,27 +1088,59 @@ LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C, LVILatticeVal Result = getCache(PImpl).getValueOnEdge(V, FromBB, ToBB); // If we know the value is a constant, evaluate the conditional. - Constant *Res = 0; + Constant *Res = nullptr; if (Result.isConstant()) { - Res = ConstantFoldCompareInstOperands(Pred, Result.getConstant(), C, TD); - if (ConstantInt *ResCI = dyn_cast_or_null(Res)) + Res = ConstantFoldCompareInstOperands(Pred, Result.getConstant(), C, DL, + TLI); + if (ConstantInt *ResCI = dyn_cast(Res)) return ResCI->isZero() ? False : True; return Unknown; } + if (Result.isConstantRange()) { + ConstantInt *CI = dyn_cast(C); + if (!CI) return Unknown; + + ConstantRange CR = Result.getConstantRange(); + if (Pred == ICmpInst::ICMP_EQ) { + if (!CR.contains(CI->getValue())) + return False; + + if (CR.isSingleElement() && CR.contains(CI->getValue())) + return True; + } else if (Pred == ICmpInst::ICMP_NE) { + if (!CR.contains(CI->getValue())) + return True; + + if (CR.isSingleElement() && CR.contains(CI->getValue())) + return False; + } + + // Handle more complex predicates. + ConstantRange TrueValues = + ICmpInst::makeConstantRange((ICmpInst::Predicate)Pred, CI->getValue()); + if (TrueValues.contains(CR)) + return True; + if (TrueValues.inverse().contains(CR)) + return False; + return Unknown; + } + if (Result.isNotConstant()) { // If this is an equality comparison, we can try to fold it knowing that // "V != C1". if (Pred == ICmpInst::ICMP_EQ) { // !C1 == C -> false iff C1 == C. Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, - Result.getNotConstant(), C, TD); + Result.getNotConstant(), C, DL, + TLI); if (Res->isNullValue()) return False; } else if (Pred == ICmpInst::ICMP_NE) { // !C1 != C -> true iff C1 == C. Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, - Result.getNotConstant(), C, TD); + Result.getNotConstant(), C, DL, + TLI); if (Res->isNullValue()) return True; } @@ -561,4 +1150,11 @@ LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C, return Unknown; } +void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, + BasicBlock *NewSucc) { + if (PImpl) getCache(PImpl).threadEdge(PredBB, OldSucc, NewSucc); +} +void LazyValueInfo::eraseBlock(BasicBlock *BB) { + if (PImpl) getCache(PImpl).eraseBlock(BB); +}