X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FPredicateSimplifier.cpp;h=24707bd4d86722a6296c7a529359967bc1f7b3bf;hb=333c40096561218bc3597cf153c0a3895274414c;hp=8a677c4461cded940dd7f7a6ac0e26b407957327;hpb=438e08e35796ca2e4d2be71185b34e11c2d12c80;p=oota-llvm.git diff --git a/lib/Transforms/Scalar/PredicateSimplifier.cpp b/lib/Transforms/Scalar/PredicateSimplifier.cpp index 8a677c4461c..24707bd4d86 100644 --- a/lib/Transforms/Scalar/PredicateSimplifier.cpp +++ b/lib/Transforms/Scalar/PredicateSimplifier.cpp @@ -2,8 +2,8 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by Nick Lewycky and is distributed under the -// University of Illinois Open Source License. See LICENSE.TXT for details. +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // @@ -22,16 +22,16 @@ // //===----------------------------------------------------------------------===// // -// This pass focusses on four properties; equals, not equals, less-than -// and less-than-or-equals-to. The greater-than forms are also held just -// to allow walking from a lesser node to a greater one. These properties +// The InequalityGraph focusses on four properties; equals, not equals, +// less-than and less-than-or-equals-to. The greater-than forms are also held +// just to allow walking from a lesser node to a greater one. These properties // are stored in a lattice; LE can become LT or EQ, NE can become LT or GT. // // These relationships define a graph between values of the same type. Each // Value is stored in a map table that retrieves the associated Node. This -// is how EQ relationships are stored; the map contains pointers to the -// same node. The node contains a most canonical Value* form and the list of -// known relationships. +// is how EQ relationships are stored; the map contains pointers from equal +// Value to the same node. The node contains a most canonical Value* form +// and the list of known relationships with other nodes. // // If two nodes are known to be inequal, then they will contain pointers to // each other with an "NE" relationship. If node getNode(%x) is less than @@ -41,25 +41,25 @@ // %a < %b < %c < %d // // with four nodes representing the properties. The InequalityGraph provides -// queries (such as "isEqual") and mutators (such as "addEqual"). To implement -// "isLess(%a, %c)", we start with getNode(%c) and walk downwards until -// we reach %a or the leaf node. Note that the graph is directed and acyclic, -// but may contain joins, meaning that this walk is not a linear time -// algorithm. +// querying with "isRelatedBy" and mutators "addEquality" and "addInequality". +// To find a relationship, we start with one of the nodes any binary search +// through its list to find where the relationships with the second node start. +// Then we iterate through those to find the first relationship that dominates +// our context node. // // To create these properties, we wait until a branch or switch instruction // implies that a particular value is true (or false). The VRPSolver is // responsible for analyzing the variable and seeing what new inferences // can be made from each property. For example: // -// %P = seteq int* %ptr, null -// %a = or bool %P, %Q -// br bool %a label %cond_true, label %cond_false +// %P = icmp ne i32* %ptr, null +// %a = and i1 %P, %Q +// br i1 %a label %cond_true, label %cond_false // // For the true branch, the VRPSolver will start with %a EQ true and look at // the definition of %a and find that it can infer that %P and %Q are both // true. From %P being true, it can infer that %ptr NE null. For the false -// branch it can't infer anything from the "or" instruction. +// branch it can't infer anything from the "and" instruction. // // Besides branches, we can also infer properties from instruction that may // have undefined behaviour in certain cases. For example, the dividend of @@ -67,6 +67,17 @@ // that the dividend is not equal to zero. // //===----------------------------------------------------------------------===// +// +// The ValueRanges class stores the known integer bounds of a Value. When we +// encounter i8 %a u< %b, the ValueRanges stores that %a = [1, 255] and +// %b = [0, 254]. +// +// It never stores an empty range, because that means that the code is +// unreachable. It never stores a single-element range since that's an equality +// relationship and better stored in the InequalityGraph, nor an empty range +// since that is better stored in UnreachableBlocks. +// +//===----------------------------------------------------------------------===// #define DEBUG_TYPE "predsimplify" #include "llvm/Transforms/Scalar.h" @@ -74,29 +85,504 @@ #include "llvm/DerivedTypes.h" #include "llvm/Instructions.h" #include "llvm/Pass.h" +#include "llvm/ADT/DepthFirstIterator.h" #include "llvm/ADT/SetOperations.h" -#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/SetVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include "llvm/Analysis/Dominators.h" -#include "llvm/Analysis/ET-Forest.h" #include "llvm/Assembly/Writer.h" #include "llvm/Support/CFG.h" #include "llvm/Support/Compiler.h" +#include "llvm/Support/ConstantRange.h" #include "llvm/Support/Debug.h" #include "llvm/Support/InstVisitor.h" +#include "llvm/Target/TargetData.h" #include "llvm/Transforms/Utils/Local.h" #include #include -#include -#include +#include using namespace llvm; STATISTIC(NumVarsReplaced, "Number of argument substitutions"); STATISTIC(NumInstruction , "Number of instructions removed"); STATISTIC(NumSimple , "Number of simple replacements"); +STATISTIC(NumBlocks , "Number of blocks marked unreachable"); +STATISTIC(NumSnuggle , "Number of comparisons snuggled"); + +static const ConstantRange empty(1, false); namespace { + class DomTreeDFS { + public: + class Node { + friend class DomTreeDFS; + public: + typedef std::vector::iterator iterator; + typedef std::vector::const_iterator const_iterator; + + unsigned getDFSNumIn() const { return DFSin; } + unsigned getDFSNumOut() const { return DFSout; } + + BasicBlock *getBlock() const { return BB; } + + iterator begin() { return Children.begin(); } + iterator end() { return Children.end(); } + + const_iterator begin() const { return Children.begin(); } + const_iterator end() const { return Children.end(); } + + bool dominates(const Node *N) const { + return DFSin <= N->DFSin && DFSout >= N->DFSout; + } + + bool DominatedBy(const Node *N) const { + return N->dominates(this); + } + + /// Sorts by the number of descendants. With this, you can iterate + /// through a sorted list and the first matching entry is the most + /// specific match for your basic block. The order provided is stable; + /// DomTreeDFS::Nodes with the same number of descendants are sorted by + /// DFS in number. + bool operator<(const Node &N) const { + unsigned spread = DFSout - DFSin; + unsigned N_spread = N.DFSout - N.DFSin; + if (spread == N_spread) return DFSin < N.DFSin; + return spread < N_spread; + } + bool operator>(const Node &N) const { return N < *this; } + + private: + unsigned DFSin, DFSout; + BasicBlock *BB; + + std::vector Children; + }; + + // XXX: this may be slow. Instead of using "new" for each node, consider + // putting them in a vector to keep them contiguous. + explicit DomTreeDFS(DominatorTree *DT) { + std::stack > S; + + Entry = new Node; + Entry->BB = DT->getRootNode()->getBlock(); + S.push(std::make_pair(Entry, DT->getRootNode())); + + NodeMap[Entry->BB] = Entry; + + while (!S.empty()) { + std::pair &Pair = S.top(); + Node *N = Pair.first; + DomTreeNode *DTNode = Pair.second; + S.pop(); + + for (DomTreeNode::iterator I = DTNode->begin(), E = DTNode->end(); + I != E; ++I) { + Node *NewNode = new Node; + NewNode->BB = (*I)->getBlock(); + N->Children.push_back(NewNode); + S.push(std::make_pair(NewNode, *I)); + + NodeMap[NewNode->BB] = NewNode; + } + } + + renumber(); + +#ifndef NDEBUG + DEBUG(dump()); +#endif + } + +#ifndef NDEBUG + virtual +#endif + ~DomTreeDFS() { + std::stack S; + + S.push(Entry); + while (!S.empty()) { + Node *N = S.top(); S.pop(); + + for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) + S.push(*I); + + delete N; + } + } + + /// getRootNode - This returns the entry node for the CFG of the function. + Node *getRootNode() const { return Entry; } + + /// getNodeForBlock - return the node for the specified basic block. + Node *getNodeForBlock(BasicBlock *BB) const { + if (!NodeMap.count(BB)) return 0; + return const_cast(this)->NodeMap[BB]; + } + + /// dominates - returns true if the basic block for I1 dominates that of + /// the basic block for I2. If the instructions belong to the same basic + /// block, the instruction first instruction sequentially in the block is + /// considered dominating. + bool dominates(Instruction *I1, Instruction *I2) { + BasicBlock *BB1 = I1->getParent(), + *BB2 = I2->getParent(); + if (BB1 == BB2) { + if (isa(I1)) return false; + if (isa(I2)) return true; + if ( isa(I1) && !isa(I2)) return true; + if (!isa(I1) && isa(I2)) return false; + + for (BasicBlock::const_iterator I = BB2->begin(), E = BB2->end(); + I != E; ++I) { + if (&*I == I1) return true; + else if (&*I == I2) return false; + } + assert(!"Instructions not found in parent BasicBlock?"); + } else { + Node *Node1 = getNodeForBlock(BB1), + *Node2 = getNodeForBlock(BB2); + return Node1 && Node2 && Node1->dominates(Node2); + } + return false; // Not reached + } + + private: + /// renumber - calculates the depth first search numberings and applies + /// them onto the nodes. + void renumber() { + std::stack > S; + unsigned n = 0; + + Entry->DFSin = ++n; + S.push(std::make_pair(Entry, Entry->begin())); + + while (!S.empty()) { + std::pair &Pair = S.top(); + Node *N = Pair.first; + Node::iterator &I = Pair.second; + + if (I == N->end()) { + N->DFSout = ++n; + S.pop(); + } else { + Node *Next = *I++; + Next->DFSin = ++n; + S.push(std::make_pair(Next, Next->begin())); + } + } + } + +#ifndef NDEBUG + virtual void dump() const { + dump(*cerr.stream()); + } + + void dump(std::ostream &os) const { + os << "Predicate simplifier DomTreeDFS: \n"; + dump(Entry, 0, os); + os << "\n\n"; + } + + void dump(Node *N, int depth, std::ostream &os) const { + ++depth; + for (int i = 0; i < depth; ++i) { os << " "; } + os << "[" << depth << "] "; + + os << N->getBlock()->getName() << " (" << N->getDFSNumIn() + << ", " << N->getDFSNumOut() << ")\n"; + + for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) + dump(*I, depth, os); + } +#endif + + Node *Entry; + std::map NodeMap; + }; + + // SLT SGT ULT UGT EQ + // 0 1 0 1 0 -- GT 10 + // 0 1 0 1 1 -- GE 11 + // 0 1 1 0 0 -- SGTULT 12 + // 0 1 1 0 1 -- SGEULE 13 + // 0 1 1 1 0 -- SGT 14 + // 0 1 1 1 1 -- SGE 15 + // 1 0 0 1 0 -- SLTUGT 18 + // 1 0 0 1 1 -- SLEUGE 19 + // 1 0 1 0 0 -- LT 20 + // 1 0 1 0 1 -- LE 21 + // 1 0 1 1 0 -- SLT 22 + // 1 0 1 1 1 -- SLE 23 + // 1 1 0 1 0 -- UGT 26 + // 1 1 0 1 1 -- UGE 27 + // 1 1 1 0 0 -- ULT 28 + // 1 1 1 0 1 -- ULE 29 + // 1 1 1 1 0 -- NE 30 + enum LatticeBits { + EQ_BIT = 1, UGT_BIT = 2, ULT_BIT = 4, SGT_BIT = 8, SLT_BIT = 16 + }; + enum LatticeVal { + GT = SGT_BIT | UGT_BIT, + GE = GT | EQ_BIT, + LT = SLT_BIT | ULT_BIT, + LE = LT | EQ_BIT, + NE = SLT_BIT | SGT_BIT | ULT_BIT | UGT_BIT, + SGTULT = SGT_BIT | ULT_BIT, + SGEULE = SGTULT | EQ_BIT, + SLTUGT = SLT_BIT | UGT_BIT, + SLEUGE = SLTUGT | EQ_BIT, + ULT = SLT_BIT | SGT_BIT | ULT_BIT, + UGT = SLT_BIT | SGT_BIT | UGT_BIT, + SLT = SLT_BIT | ULT_BIT | UGT_BIT, + SGT = SGT_BIT | ULT_BIT | UGT_BIT, + SLE = SLT | EQ_BIT, + SGE = SGT | EQ_BIT, + ULE = ULT | EQ_BIT, + UGE = UGT | EQ_BIT + }; + +#ifndef NDEBUG + /// validPredicate - determines whether a given value is actually a lattice + /// value. Only used in assertions or debugging. + static bool validPredicate(LatticeVal LV) { + switch (LV) { + case GT: case GE: case LT: case LE: case NE: + case SGTULT: case SGT: case SGEULE: + case SLTUGT: case SLT: case SLEUGE: + case ULT: case UGT: + case SLE: case SGE: case ULE: case UGE: + return true; + default: + return false; + } + } +#endif + + /// reversePredicate - reverse the direction of the inequality + static LatticeVal reversePredicate(LatticeVal LV) { + unsigned reverse = LV ^ (SLT_BIT|SGT_BIT|ULT_BIT|UGT_BIT); //preserve EQ_BIT + + if ((reverse & (SLT_BIT|SGT_BIT)) == 0) + reverse |= (SLT_BIT|SGT_BIT); + + if ((reverse & (ULT_BIT|UGT_BIT)) == 0) + reverse |= (ULT_BIT|UGT_BIT); + + LatticeVal Rev = static_cast(reverse); + assert(validPredicate(Rev) && "Failed reversing predicate."); + return Rev; + } + + /// ValueNumbering stores the scope-specific value numbers for a given Value. + class VISIBILITY_HIDDEN ValueNumbering { + + /// VNPair is a tuple of {Value, index number, DomTreeDFS::Node}. It + /// includes the comparison operators necessary to allow you to store it + /// in a sorted vector. + class VISIBILITY_HIDDEN VNPair { + public: + Value *V; + unsigned index; + DomTreeDFS::Node *Subtree; + + VNPair(Value *V, unsigned index, DomTreeDFS::Node *Subtree) + : V(V), index(index), Subtree(Subtree) {} + + bool operator==(const VNPair &RHS) const { + return V == RHS.V && Subtree == RHS.Subtree; + } + + bool operator<(const VNPair &RHS) const { + if (V != RHS.V) return V < RHS.V; + return *Subtree < *RHS.Subtree; + } + + bool operator<(Value *RHS) const { + return V < RHS; + } + + bool operator>(Value *RHS) const { + return V > RHS; + } + + friend bool operator<(Value *RHS, const VNPair &pair) { + return pair.operator>(RHS); + } + }; + + typedef std::vector VNMapType; + VNMapType VNMap; + + /// The canonical choice for value number at index. + std::vector Values; + + DomTreeDFS *DTDFS; + + public: +#ifndef NDEBUG + virtual ~ValueNumbering() {} + virtual void dump() { + dump(*cerr.stream()); + } + + void dump(std::ostream &os) { + for (unsigned i = 1; i <= Values.size(); ++i) { + os << i << " = "; + WriteAsOperand(os, Values[i-1]); + os << " {"; + for (unsigned j = 0; j < VNMap.size(); ++j) { + if (VNMap[j].index == i) { + WriteAsOperand(os, VNMap[j].V); + os << " (" << VNMap[j].Subtree->getDFSNumIn() << ") "; + } + } + os << "}\n"; + } + } +#endif + + /// compare - returns true if V1 is a better canonical value than V2. + bool compare(Value *V1, Value *V2) const { + if (isa(V1)) + return !isa(V2); + else if (isa(V2)) + return false; + else if (isa(V1)) + return !isa(V2); + else if (isa(V2)) + return false; + + Instruction *I1 = dyn_cast(V1); + Instruction *I2 = dyn_cast(V2); + + if (!I1 || !I2) + return V1->getNumUses() < V2->getNumUses(); + + return DTDFS->dominates(I1, I2); + } + + ValueNumbering(DomTreeDFS *DTDFS) : DTDFS(DTDFS) {} + + /// valueNumber - finds the value number for V under the Subtree. If + /// there is no value number, returns zero. + unsigned valueNumber(Value *V, DomTreeDFS::Node *Subtree) { + if (!(isa(V) || isa(V) || isa(V)) + || V->getType() == Type::VoidTy) return 0; + + VNMapType::iterator E = VNMap.end(); + VNPair pair(V, 0, Subtree); + VNMapType::iterator I = std::lower_bound(VNMap.begin(), E, pair); + while (I != E && I->V == V) { + if (I->Subtree->dominates(Subtree)) + return I->index; + ++I; + } + return 0; + } + + /// getOrInsertVN - always returns a value number, creating it if necessary. + unsigned getOrInsertVN(Value *V, DomTreeDFS::Node *Subtree) { + if (unsigned n = valueNumber(V, Subtree)) + return n; + else + return newVN(V); + } + + /// newVN - creates a new value number. Value V must not already have a + /// value number assigned. + unsigned newVN(Value *V) { + assert((isa(V) || isa(V) || isa(V)) && + "Bad Value for value numbering."); + assert(V->getType() != Type::VoidTy && "Won't value number a void value"); + + Values.push_back(V); + + VNPair pair = VNPair(V, Values.size(), DTDFS->getRootNode()); + VNMapType::iterator I = std::lower_bound(VNMap.begin(), VNMap.end(), pair); + assert((I == VNMap.end() || value(I->index) != V) && + "Attempt to create a duplicate value number."); + VNMap.insert(I, pair); + + return Values.size(); + } + + /// value - returns the Value associated with a value number. + Value *value(unsigned index) const { + assert(index != 0 && "Zero index is reserved for not found."); + assert(index <= Values.size() && "Index out of range."); + return Values[index-1]; + } + + /// canonicalize - return a Value that is equal to V under Subtree. + Value *canonicalize(Value *V, DomTreeDFS::Node *Subtree) { + if (isa(V)) return V; + + if (unsigned n = valueNumber(V, Subtree)) + return value(n); + else + return V; + } + + /// addEquality - adds that value V belongs to the set of equivalent + /// values defined by value number n under Subtree. + void addEquality(unsigned n, Value *V, DomTreeDFS::Node *Subtree) { + assert(canonicalize(value(n), Subtree) == value(n) && + "Node's 'canonical' choice isn't best within this subtree."); + + // Suppose that we are given "%x -> node #1 (%y)". The problem is that + // we may already have "%z -> node #2 (%x)" somewhere above us in the + // graph. We need to find those edges and add "%z -> node #1 (%y)" + // to keep the lookups canonical. + + std::vector ToRepoint(1, V); + + if (unsigned Conflict = valueNumber(V, Subtree)) { + for (VNMapType::iterator I = VNMap.begin(), E = VNMap.end(); + I != E; ++I) { + if (I->index == Conflict && I->Subtree->dominates(Subtree)) + ToRepoint.push_back(I->V); + } + } + + for (std::vector::iterator VI = ToRepoint.begin(), + VE = ToRepoint.end(); VI != VE; ++VI) { + Value *V = *VI; + + VNPair pair(V, n, Subtree); + VNMapType::iterator B = VNMap.begin(), E = VNMap.end(); + VNMapType::iterator I = std::lower_bound(B, E, pair); + if (I != E && I->V == V && I->Subtree == Subtree) + I->index = n; // Update best choice + else + VNMap.insert(I, pair); // New Value + + // XXX: we currently don't have to worry about updating values with + // more specific Subtrees, but we will need to for PHI node support. + +#ifndef NDEBUG + Value *V_n = value(n); + if (isa(V) && isa(V_n)) { + assert(V == V_n && "Constant equals different constant?"); + } +#endif + } + } + + /// remove - removes all references to value V. + void remove(Value *V) { + VNMapType::iterator B = VNMap.begin(), E = VNMap.end(); + VNPair pair(V, 0, DTDFS->getRootNode()); + VNMapType::iterator J = std::upper_bound(B, E, pair); + VNMapType::iterator I = J; + + while (I != B && (I == E || I->V == V)) --I; + + VNMap.erase(I, J); + } + }; + /// The InequalityGraph stores the relationships between values. /// Each Value in the graph is assigned to a Node. Nodes are pointer /// comparable for equality. The caller is expected to maintain the logical @@ -105,410 +591,778 @@ namespace { /// The InequalityGraph class may invalidate Node*s after any mutator call. /// @brief The InequalityGraph stores the relationships between values. class VISIBILITY_HIDDEN InequalityGraph { + ValueNumbering &VN; + DomTreeDFS::Node *TreeRoot; + + InequalityGraph(); // DO NOT IMPLEMENT + InequalityGraph(InequalityGraph &); // DO NOT IMPLEMENT public: + InequalityGraph(ValueNumbering &VN, DomTreeDFS::Node *TreeRoot) + : VN(VN), TreeRoot(TreeRoot) {} + class Node; - // LT GT EQ - // 0 0 0 -- invalid (false) - // 0 0 1 -- invalid (EQ) - // 0 1 0 -- GT - // 0 1 1 -- GE - // 1 0 0 -- LT - // 1 0 1 -- LE - // 1 1 0 -- NE - // 1 1 1 -- invalid (true) - enum LatticeBits { - EQ_BIT = 1, GT_BIT = 2, LT_BIT = 4 - }; - enum LatticeVal { - GT = GT_BIT, GE = GT_BIT | EQ_BIT, - LT = LT_BIT, LE = LT_BIT | EQ_BIT, - NE = GT_BIT | LT_BIT - }; + /// An Edge is contained inside a Node making one end of the edge implicit + /// and contains a pointer to the other end. The edge contains a lattice + /// value specifying the relationship and an DomTreeDFS::Node specifying + /// the root in the dominator tree to which this edge applies. + class VISIBILITY_HIDDEN Edge { + public: + Edge(unsigned T, LatticeVal V, DomTreeDFS::Node *ST) + : To(T), LV(V), Subtree(ST) {} - static bool validPredicate(LatticeVal LV) { - return LV > 1 && LV < 7; - } + unsigned To; + LatticeVal LV; + DomTreeDFS::Node *Subtree; - private: - typedef std::map NodeMapType; - NodeMapType Nodes; + bool operator<(const Edge &edge) const { + if (To != edge.To) return To < edge.To; + return *Subtree < *edge.Subtree; + } - const InequalityGraph *ConcreteIG; + bool operator<(unsigned to) const { + return To < to; + } + + bool operator>(unsigned to) const { + return To > to; + } + + friend bool operator<(unsigned to, const Edge &edge) { + return edge.operator>(to); + } + }; - public: /// A single node in the InequalityGraph. This stores the canonical Value /// for the node, as well as the relationships with the neighbours. /// - /// Because the lists are intended to be used for traversal, it is invalid - /// for the node to list itself in LessEqual or GreaterEqual lists. The - /// fact that a node is equal to itself is implied, and may be checked - /// with pointer comparison. /// @brief A single node in the InequalityGraph. class VISIBILITY_HIDDEN Node { friend class InequalityGraph; - Value *Canonical; - - typedef SmallVector, 4> RelationsType; + typedef SmallVector RelationsType; RelationsType Relations; + + // TODO: can this idea improve performance? + //friend class std::vector; + //Node(Node &N) { RelationsType.swap(N.RelationsType); } + public: typedef RelationsType::iterator iterator; typedef RelationsType::const_iterator const_iterator; +#ifndef NDEBUG + virtual ~Node() {} + virtual void dump() const { + dump(*cerr.stream()); + } private: - /// Updates the lattice value for a given node. Create a new entry if - /// one doesn't exist, otherwise it merges the values. The new lattice - /// value must not be inconsistent with any previously existing value. - void update(Node *N, LatticeVal R) { - iterator I = find(N); - if (I == end()) { - Relations.push_back(std::make_pair(N, R)); - } else { - I->second = static_cast(I->second & R); - assert(validPredicate(I->second) && - "Invalid union of lattice values."); + void dump(std::ostream &os) const { + static const std::string names[32] = + { "000000", "000001", "000002", "000003", "000004", "000005", + "000006", "000007", "000008", "000009", " >", " >=", + " s>u<", "s>=u<=", " s>", " s>=", "000016", "000017", + " s", "s<=u>=", " <", " <=", " s<", " s<=", + "000024", "000025", " u>", " u>=", " u<", " u<=", + " !=", "000031" }; + for (Node::const_iterator NI = begin(), NE = end(); NI != NE; ++NI) { + os << names[NI->LV] << " " << NI->To + << " (" << NI->Subtree->getDFSNumIn() << "), "; } } + public: +#endif - void assign(Node *N, LatticeVal R) { - iterator I = find(N); - if (I != end()) I->second = R; - - Relations.push_back(std::make_pair(N, R)); + iterator begin() { return Relations.begin(); } + iterator end() { return Relations.end(); } + const_iterator begin() const { return Relations.begin(); } + const_iterator end() const { return Relations.end(); } + + iterator find(unsigned n, DomTreeDFS::Node *Subtree) { + iterator E = end(); + for (iterator I = std::lower_bound(begin(), E, n); + I != E && I->To == n; ++I) { + if (Subtree->DominatedBy(I->Subtree)) + return I; + } + return E; } - public: - iterator begin() { return Relations.begin(); } - iterator end() { return Relations.end(); } - iterator find(Node *N) { - iterator I = begin(); - for (iterator E = end(); I != E; ++I) - if (I->first == N) break; - return I; + const_iterator find(unsigned n, DomTreeDFS::Node *Subtree) const { + const_iterator E = end(); + for (const_iterator I = std::lower_bound(begin(), E, n); + I != E && I->To == n; ++I) { + if (Subtree->DominatedBy(I->Subtree)) + return I; + } + return E; } - const_iterator begin() const { return Relations.begin(); } - const_iterator end() const { return Relations.end(); } - const_iterator find(Node *N) const { - const_iterator I = begin(); - for (const_iterator E = end(); I != E; ++I) - if (I->first == N) break; - return I; - } + /// update - updates the lattice value for a given node, creating a new + /// entry if one doesn't exist. The new lattice value must not be + /// inconsistent with any previously existing value. + void update(unsigned n, LatticeVal R, DomTreeDFS::Node *Subtree) { + assert(validPredicate(R) && "Invalid predicate."); + + Edge edge(n, R, Subtree); + iterator B = begin(), E = end(); + iterator I = std::lower_bound(B, E, edge); + + iterator J = I; + while (J != E && J->To == n) { + if (Subtree->DominatedBy(J->Subtree)) + break; + ++J; + } - unsigned findIndex(Node *N) { - unsigned i = 0; - iterator I = begin(); - for (iterator E = end(); I != E; ++I, ++i) - if (I->first == N) return i; - return (unsigned)-1; - } + if (J != E && J->To == n) { + edge.LV = static_cast(J->LV & R); + assert(validPredicate(edge.LV) && "Invalid union of lattice values."); - void erase(iterator i) { Relations.erase(i); } + if (edge.LV == J->LV) + return; // This update adds nothing new. + } - Value *getValue() const { return Canonical; } - void setValue(Value *V) { Canonical = V; } + if (I != B) { + // We also have to tighten any edge beneath our update. + for (iterator K = I - 1; K->To == n; --K) { + if (K->Subtree->DominatedBy(Subtree)) { + LatticeVal LV = static_cast(K->LV & edge.LV); + assert(validPredicate(LV) && "Invalid union of lattice values"); + K->LV = LV; + } + if (K == B) break; + } + } - void addNotEqual(Node *N) { update(N, NE); } - void addLess(Node *N) { update(N, LT); } - void addLessEqual(Node *N) { update(N, LE); } - void addGreater(Node *N) { update(N, GT); } - void addGreaterEqual(Node *N) { update(N, GE); } + // Insert new edge at Subtree if it isn't already there. + if (I == E || I->To != n || Subtree != I->Subtree) + Relations.insert(I, edge); + } }; - InequalityGraph() : ConcreteIG(NULL) {} + private: - InequalityGraph(const InequalityGraph &_IG) { -#if 0 // disable COW - if (_IG.ConcreteIG) ConcreteIG = _IG.ConcreteIG; - else ConcreteIG = &_IG; -#else - ConcreteIG = &_IG; - materialize(); -#endif + std::vector Nodes; + + public: + /// node - returns the node object at a given value number. The pointer + /// returned may be invalidated on the next call to node(). + Node *node(unsigned index) { + assert(VN.value(index)); // This triggers the necessary checks. + if (Nodes.size() < index) Nodes.resize(index); + return &Nodes[index-1]; } - ~InequalityGraph(); + /// isRelatedBy - true iff n1 op n2 + bool isRelatedBy(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree, + LatticeVal LV) { + if (n1 == n2) return LV & EQ_BIT; - private: - void materialize(); + Node *N1 = node(n1); + Node::iterator I = N1->find(n2, Subtree), E = N1->end(); + if (I != E) return (I->LV & LV) == I->LV; - public: - /// If the Value is in the graph, return the canonical form. Otherwise, - /// return the original Value. - Value *canonicalize(Value *V) const { - if (const Node *N = getNode(V)) - return N->getValue(); - else - return V; + return false; } - /// Returns the node currently representing Value V, or null if no such - /// node exists. - Node *getNode(Value *V) { - materialize(); + // The add* methods assume that your input is logically valid and may + // assertion-fail or infinitely loop if you attempt a contradiction. - NodeMapType::const_iterator I = Nodes.find(V); - return (I != Nodes.end()) ? I->second : 0; - } + /// addInequality - Sets n1 op n2. + /// It is also an error to call this on an inequality that is already true. + void addInequality(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree, + LatticeVal LV1) { + assert(n1 != n2 && "A node can't be inequal to itself."); + + if (LV1 != NE) + assert(!isRelatedBy(n1, n2, Subtree, reversePredicate(LV1)) && + "Contradictory inequality."); + + // Suppose we're adding %n1 < %n2. Find all the %a < %n1 and + // add %a < %n2 too. This keeps the graph fully connected. + if (LV1 != NE) { + // Break up the relationship into signed and unsigned comparison parts. + // If the signed parts of %a op1 %n1 match that of %n1 op2 %n2, and + // op1 and op2 aren't NE, then add %a op3 %n2. The new relationship + // should have the EQ_BIT iff it's set for both op1 and op2. + + unsigned LV1_s = LV1 & (SLT_BIT|SGT_BIT); + unsigned LV1_u = LV1 & (ULT_BIT|UGT_BIT); + + for (Node::iterator I = node(n1)->begin(), E = node(n1)->end(); I != E; ++I) { + if (I->LV != NE && I->To != n2) { + + DomTreeDFS::Node *Local_Subtree = NULL; + if (Subtree->DominatedBy(I->Subtree)) + Local_Subtree = Subtree; + else if (I->Subtree->DominatedBy(Subtree)) + Local_Subtree = I->Subtree; + + if (Local_Subtree) { + unsigned new_relationship = 0; + LatticeVal ILV = reversePredicate(I->LV); + unsigned ILV_s = ILV & (SLT_BIT|SGT_BIT); + unsigned ILV_u = ILV & (ULT_BIT|UGT_BIT); + + if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s) + new_relationship |= ILV_s; + if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u) + new_relationship |= ILV_u; + + if (new_relationship) { + if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0) + new_relationship |= (SLT_BIT|SGT_BIT); + if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0) + new_relationship |= (ULT_BIT|UGT_BIT); + if ((LV1 & EQ_BIT) && (ILV & EQ_BIT)) + new_relationship |= EQ_BIT; + + LatticeVal NewLV = static_cast(new_relationship); + + node(I->To)->update(n2, NewLV, Local_Subtree); + node(n2)->update(I->To, reversePredicate(NewLV), Local_Subtree); + } + } + } + } - const Node *getNode(Value *V) const { - if (ConcreteIG) return ConcreteIG->getNode(V); + for (Node::iterator I = node(n2)->begin(), E = node(n2)->end(); I != E; ++I) { + if (I->LV != NE && I->To != n1) { + DomTreeDFS::Node *Local_Subtree = NULL; + if (Subtree->DominatedBy(I->Subtree)) + Local_Subtree = Subtree; + else if (I->Subtree->DominatedBy(Subtree)) + Local_Subtree = I->Subtree; + + if (Local_Subtree) { + unsigned new_relationship = 0; + unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT); + unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT); + + if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s) + new_relationship |= ILV_s; + + if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u) + new_relationship |= ILV_u; + + if (new_relationship) { + if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0) + new_relationship |= (SLT_BIT|SGT_BIT); + if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0) + new_relationship |= (ULT_BIT|UGT_BIT); + if ((LV1 & EQ_BIT) && (I->LV & EQ_BIT)) + new_relationship |= EQ_BIT; + + LatticeVal NewLV = static_cast(new_relationship); + + node(n1)->update(I->To, NewLV, Local_Subtree); + node(I->To)->update(n1, reversePredicate(NewLV), Local_Subtree); + } + } + } + } + } - NodeMapType::const_iterator I = Nodes.find(V); - return (I != Nodes.end()) ? I->second : 0; + node(n1)->update(n2, LV1, Subtree); + node(n2)->update(n1, reversePredicate(LV1), Subtree); } - Node *getOrInsertNode(Value *V) { - if (Node *N = getNode(V)) - return N; - else - return newNode(V); + /// remove - removes a node from the graph by removing all references to + /// and from it. + void remove(unsigned n) { + Node *N = node(n); + for (Node::iterator NI = N->begin(), NE = N->end(); NI != NE; ++NI) { + Node::iterator Iter = node(NI->To)->find(n, TreeRoot); + do { + node(NI->To)->Relations.erase(Iter); + Iter = node(NI->To)->find(n, TreeRoot); + } while (Iter != node(NI->To)->end()); + } + N->Relations.clear(); } - Node *newNode(Value *V) { - //DOUT << "new node: " << *V << "\n"; - materialize(); - Node *&N = Nodes[V]; - assert(N == 0 && "Node already exists for value."); - N = new Node(); - N->setValue(V); - return N; +#ifndef NDEBUG + virtual ~InequalityGraph() {} + virtual void dump() { + dump(*cerr.stream()); } - /// Returns true iff the nodes are provably inequal. - bool isNotEqual(const Node *N1, const Node *N2) const { - if (N1 == N2) return false; - for (Node::const_iterator I = N1->begin(), E = N1->end(); I != E; ++I) { - if (I->first == N2) - return (I->second & EQ_BIT) == 0; + void dump(std::ostream &os) { + for (unsigned i = 1; i <= Nodes.size(); ++i) { + os << i << " = {"; + node(i)->dump(os); + os << "}\n"; } - return isLess(N1, N2) || isGreater(N1, N2); } +#endif + }; + + class VRPSolver; - /// Returns true iff N1 is provably less than N2. - bool isLess(const Node *N1, const Node *N2) const { - if (N1 == N2) return false; - for (Node::const_iterator I = N2->begin(), E = N2->end(); I != E; ++I) { - if (I->first == N1) - return I->second == LT; + /// ValueRanges tracks the known integer ranges and anti-ranges of the nodes + /// in the InequalityGraph. + class VISIBILITY_HIDDEN ValueRanges { + ValueNumbering &VN; + TargetData *TD; + + class VISIBILITY_HIDDEN ScopedRange { + typedef std::vector > + RangeListType; + RangeListType RangeList; + + static bool swo(const std::pair &LHS, + const std::pair &RHS) { + return *LHS.first < *RHS.first; } - for (Node::const_iterator I = N2->begin(), E = N2->end(); I != E; ++I) { - if ((I->second & (LT_BIT | GT_BIT)) == LT_BIT) - if (isLess(N1, I->first)) return true; + + public: +#ifndef NDEBUG + virtual ~ScopedRange() {} + virtual void dump() const { + dump(*cerr.stream()); } - return false; + + void dump(std::ostream &os) const { + os << "{"; + for (const_iterator I = begin(), E = end(); I != E; ++I) { + os << &I->second << " (" << I->first->getDFSNumIn() << "), "; + } + os << "}"; + } +#endif + + typedef RangeListType::iterator iterator; + typedef RangeListType::const_iterator const_iterator; + + iterator begin() { return RangeList.begin(); } + iterator end() { return RangeList.end(); } + const_iterator begin() const { return RangeList.begin(); } + const_iterator end() const { return RangeList.end(); } + + iterator find(DomTreeDFS::Node *Subtree) { + iterator E = end(); + iterator I = std::lower_bound(begin(), E, + std::make_pair(Subtree, empty), swo); + + while (I != E && !I->first->dominates(Subtree)) ++I; + return I; + } + + const_iterator find(DomTreeDFS::Node *Subtree) const { + const_iterator E = end(); + const_iterator I = std::lower_bound(begin(), E, + std::make_pair(Subtree, empty), swo); + + while (I != E && !I->first->dominates(Subtree)) ++I; + return I; + } + + void update(const ConstantRange &CR, DomTreeDFS::Node *Subtree) { + assert(!CR.isEmptySet() && "Empty ConstantRange."); + assert(!CR.isSingleElement() && "Refusing to store single element."); + + iterator E = end(); + iterator I = + std::lower_bound(begin(), E, std::make_pair(Subtree, empty), swo); + + if (I != end() && I->first == Subtree) { + ConstantRange CR2 = I->second.maximalIntersectWith(CR); + assert(!CR2.isEmptySet() && !CR2.isSingleElement() && + "Invalid union of ranges."); + I->second = CR2; + } else + RangeList.insert(I, std::make_pair(Subtree, CR)); + } + }; + + std::vector Ranges; + + void update(unsigned n, const ConstantRange &CR, DomTreeDFS::Node *Subtree){ + if (CR.isFullSet()) return; + if (Ranges.size() < n) Ranges.resize(n); + Ranges[n-1].update(CR, Subtree); } - /// Returns true iff N1 is provably less than or equal to N2. - bool isLessEqual(const Node *N1, const Node *N2) const { - if (N1 == N2) return true; - for (Node::const_iterator I = N2->begin(), E = N2->end(); I != E; ++I) { - if (I->first == N1) - return (I->second & (LT_BIT | GT_BIT)) == LT_BIT; + /// create - Creates a ConstantRange that matches the given LatticeVal + /// relation with a given integer. + ConstantRange create(LatticeVal LV, const ConstantRange &CR) { + assert(!CR.isEmptySet() && "Can't deal with empty set."); + + if (LV == NE) + return makeConstantRange(ICmpInst::ICMP_NE, CR); + + unsigned LV_s = LV & (SGT_BIT|SLT_BIT); + unsigned LV_u = LV & (UGT_BIT|ULT_BIT); + bool hasEQ = LV & EQ_BIT; + + ConstantRange Range(CR.getBitWidth()); + + if (LV_s == SGT_BIT) { + Range = Range.maximalIntersectWith(makeConstantRange( + hasEQ ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SGT, CR)); + } else if (LV_s == SLT_BIT) { + Range = Range.maximalIntersectWith(makeConstantRange( + hasEQ ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_SLT, CR)); } - for (Node::const_iterator I = N2->begin(), E = N2->end(); I != E; ++I) { - if ((I->second & (LT_BIT | GT_BIT)) == LT_BIT) - if (isLessEqual(N1, I->first)) return true; + + if (LV_u == UGT_BIT) { + Range = Range.maximalIntersectWith(makeConstantRange( + hasEQ ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_UGT, CR)); + } else if (LV_u == ULT_BIT) { + Range = Range.maximalIntersectWith(makeConstantRange( + hasEQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT, CR)); } - return false; + + return Range; } - /// Returns true iff N1 is provably greater than N2. - bool isGreater(const Node *N1, const Node *N2) const { - return isLess(N2, N1); + /// makeConstantRange - Creates a ConstantRange representing the set of all + /// value that match the ICmpInst::Predicate with any of the values in CR. + ConstantRange makeConstantRange(ICmpInst::Predicate ICmpOpcode, + const ConstantRange &CR) { + uint32_t W = CR.getBitWidth(); + switch (ICmpOpcode) { + default: assert(!"Invalid ICmp opcode to makeConstantRange()"); + case ICmpInst::ICMP_EQ: + return ConstantRange(CR.getLower(), CR.getUpper()); + case ICmpInst::ICMP_NE: + if (CR.isSingleElement()) + return ConstantRange(CR.getUpper(), CR.getLower()); + return ConstantRange(W); + case ICmpInst::ICMP_ULT: + return ConstantRange(APInt::getMinValue(W), CR.getUnsignedMax()); + case ICmpInst::ICMP_SLT: + return ConstantRange(APInt::getSignedMinValue(W), CR.getSignedMax()); + case ICmpInst::ICMP_ULE: { + APInt UMax(CR.getUnsignedMax()); + if (UMax.isMaxValue()) + return ConstantRange(W); + return ConstantRange(APInt::getMinValue(W), UMax + 1); + } + case ICmpInst::ICMP_SLE: { + APInt SMax(CR.getSignedMax()); + if (SMax.isMaxSignedValue() || (SMax+1).isMaxSignedValue()) + return ConstantRange(W); + return ConstantRange(APInt::getSignedMinValue(W), SMax + 1); + } + case ICmpInst::ICMP_UGT: + return ConstantRange(CR.getUnsignedMin() + 1, APInt::getNullValue(W)); + case ICmpInst::ICMP_SGT: + return ConstantRange(CR.getSignedMin() + 1, + APInt::getSignedMinValue(W)); + case ICmpInst::ICMP_UGE: { + APInt UMin(CR.getUnsignedMin()); + if (UMin.isMinValue()) + return ConstantRange(W); + return ConstantRange(UMin, APInt::getNullValue(W)); + } + case ICmpInst::ICMP_SGE: { + APInt SMin(CR.getSignedMin()); + if (SMin.isMinSignedValue()) + return ConstantRange(W); + return ConstantRange(SMin, APInt::getSignedMinValue(W)); + } + } } - /// Returns true iff N1 is provably greater than or equal to N2. - bool isGreaterEqual(const Node *N1, const Node *N2) const { - return isLessEqual(N2, N1); +#ifndef NDEBUG + bool isCanonical(Value *V, DomTreeDFS::Node *Subtree) { + return V == VN.canonicalize(V, Subtree); } +#endif - // The add* methods assume that your input is logically valid and may - // assertion-fail or infinitely loop if you attempt a contradiction. + public: + + ValueRanges(ValueNumbering &VN, TargetData *TD) : VN(VN), TD(TD) {} - void addEqual(Node *N, Value *V) { - materialize(); - Nodes[V] = N; +#ifndef NDEBUG + virtual ~ValueRanges() {} + + virtual void dump() const { + dump(*cerr.stream()); } - void addNotEqual(Node *N1, Node *N2) { - assert(N1 != N2 && "A node can't be inequal to itself."); - materialize(); - N1->addNotEqual(N2); - N2->addNotEqual(N1); + void dump(std::ostream &os) const { + for (unsigned i = 0, e = Ranges.size(); i != e; ++i) { + os << (i+1) << " = "; + Ranges[i].dump(os); + os << "\n"; + } } +#endif - /// N1 is less than N2. - void addLess(Node *N1, Node *N2) { - assert(N1 != N2 && !isLess(N2, N1) && "Attempt to create < cycle."); - materialize(); - N2->addLess(N1); - N1->addGreater(N2); + /// range - looks up the ConstantRange associated with a value number. + ConstantRange range(unsigned n, DomTreeDFS::Node *Subtree) { + assert(VN.value(n)); // performs range checks + + if (n <= Ranges.size()) { + ScopedRange::iterator I = Ranges[n-1].find(Subtree); + if (I != Ranges[n-1].end()) return I->second; + } + + Value *V = VN.value(n); + ConstantRange CR = range(V); + return CR; } - /// N1 is less than or equal to N2. - void addLessEqual(Node *N1, Node *N2) { - assert(N1 != N2 && "Nodes are equal. Use mergeNodes instead."); - assert(!isGreater(N1, N2) && "Impossible: Adding x <= y when x > y."); - materialize(); - N2->addLessEqual(N1); - N1->addGreaterEqual(N2); + /// range - determine a range from a Value without performing any lookups. + ConstantRange range(Value *V) const { + if (ConstantInt *C = dyn_cast(V)) + return ConstantRange(C->getValue()); + else if (isa(V)) + return ConstantRange(APInt::getNullValue(typeToWidth(V->getType()))); + else + return ConstantRange(typeToWidth(V->getType())); } - /// Find the transitive closure starting at a node walking down the edges - /// of type Val. Type Inserter must be an inserter that accepts Node *. - template - void transitiveClosure(Node *N, LatticeVal Val, Inserter insert) { - for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I) { - if (I->second == Val) { - *insert = I->first; - transitiveClosure(I->first, Val, insert); - } + // typeToWidth - returns the number of bits necessary to store a value of + // this type, or zero if unknown. + uint32_t typeToWidth(const Type *Ty) const { + if (TD) + return TD->getTypeSizeInBits(Ty); + else + return Ty->getPrimitiveSizeInBits(); + } + + static bool isRelatedBy(const ConstantRange &CR1, const ConstantRange &CR2, + LatticeVal LV) { + switch (LV) { + default: assert(!"Impossible lattice value!"); + case NE: + return CR1.maximalIntersectWith(CR2).isEmptySet(); + case ULT: + return CR1.getUnsignedMax().ult(CR2.getUnsignedMin()); + case ULE: + return CR1.getUnsignedMax().ule(CR2.getUnsignedMin()); + case UGT: + return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax()); + case UGE: + return CR1.getUnsignedMin().uge(CR2.getUnsignedMax()); + case SLT: + return CR1.getSignedMax().slt(CR2.getSignedMin()); + case SLE: + return CR1.getSignedMax().sle(CR2.getSignedMin()); + case SGT: + return CR1.getSignedMin().sgt(CR2.getSignedMax()); + case SGE: + return CR1.getSignedMin().sge(CR2.getSignedMax()); + case LT: + return CR1.getUnsignedMax().ult(CR2.getUnsignedMin()) && + CR1.getSignedMax().slt(CR2.getUnsignedMin()); + case LE: + return CR1.getUnsignedMax().ule(CR2.getUnsignedMin()) && + CR1.getSignedMax().sle(CR2.getUnsignedMin()); + case GT: + return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax()) && + CR1.getSignedMin().sgt(CR2.getSignedMax()); + case GE: + return CR1.getUnsignedMin().uge(CR2.getUnsignedMax()) && + CR1.getSignedMin().sge(CR2.getSignedMax()); + case SLTUGT: + return CR1.getSignedMax().slt(CR2.getSignedMin()) && + CR1.getUnsignedMin().ugt(CR2.getUnsignedMax()); + case SLEUGE: + return CR1.getSignedMax().sle(CR2.getSignedMin()) && + CR1.getUnsignedMin().uge(CR2.getUnsignedMax()); + case SGTULT: + return CR1.getSignedMin().sgt(CR2.getSignedMax()) && + CR1.getUnsignedMax().ult(CR2.getUnsignedMin()); + case SGEULE: + return CR1.getSignedMin().sge(CR2.getSignedMax()) && + CR1.getUnsignedMax().ule(CR2.getUnsignedMin()); } } - /// Kills off all the nodes in Kill by replicating their properties into - /// node N. The elements of Kill must be unique. After merging, N's new - /// canonical value is NewCanonical. Type C must be a container of Node *. - template - void mergeNodes(Node *N, C &Kill, Value *NewCanonical); + bool isRelatedBy(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree, + LatticeVal LV) { + ConstantRange CR1 = range(n1, Subtree); + ConstantRange CR2 = range(n2, Subtree); - /// Removes a Value from the graph, but does not delete any nodes. As this - /// method does not delete Nodes, V may not be the canonical choice for - /// any node. - void remove(Value *V) { - materialize(); + // True iff all values in CR1 are LV to all values in CR2. + return isRelatedBy(CR1, CR2, LV); + } + + void addToWorklist(Value *V, Constant *C, ICmpInst::Predicate Pred, + VRPSolver *VRP); + void markBlock(VRPSolver *VRP); + + void mergeInto(Value **I, unsigned n, unsigned New, + DomTreeDFS::Node *Subtree, VRPSolver *VRP) { + ConstantRange CR_New = range(New, Subtree); + ConstantRange Merged = CR_New; + + for (; n != 0; ++I, --n) { + unsigned i = VN.valueNumber(*I, Subtree); + ConstantRange CR_Kill = i ? range(i, Subtree) : range(*I); + if (CR_Kill.isFullSet()) continue; + Merged = Merged.maximalIntersectWith(CR_Kill); + } - for (NodeMapType::iterator I = Nodes.begin(), E = Nodes.end(); I != E;) { - NodeMapType::iterator J = I++; - assert(J->second->getValue() != V && "Can't delete canonical choice."); - if (J->first == V) Nodes.erase(J); + if (Merged.isFullSet() || Merged == CR_New) return; + + applyRange(New, Merged, Subtree, VRP); + } + + void applyRange(unsigned n, const ConstantRange &CR, + DomTreeDFS::Node *Subtree, VRPSolver *VRP) { + ConstantRange Merged = CR.maximalIntersectWith(range(n, Subtree)); + if (Merged.isEmptySet()) { + markBlock(VRP); + return; } - } -#ifndef NDEBUG - void debug(std::ostream &os) const { - std::set VisitedNodes; - for (NodeMapType::const_iterator I = Nodes.begin(), E = Nodes.end(); - I != E; ++I) { - Node *N = I->second; - os << *I->first << " == " << *N->getValue() << "\n"; - if (VisitedNodes.insert(N).second) { - os << *N->getValue() << ":\n"; - for (Node::const_iterator NI = N->begin(), NE = N->end(); - NI != NE; ++NI) { - static const std::string names[8] = - { "00", "01", " <", "<=", " >", ">=", "!=", "07" }; - os << " " << names[NI->second] << " " - << *NI->first->getValue() << "\n"; + if (const APInt *I = Merged.getSingleElement()) { + Value *V = VN.value(n); // XXX: redesign worklist. + const Type *Ty = V->getType(); + if (Ty->isInteger()) { + addToWorklist(V, ConstantInt::get(*I), ICmpInst::ICMP_EQ, VRP); + return; + } else if (const PointerType *PTy = dyn_cast(Ty)) { + assert(*I == 0 && "Pointer is null but not zero?"); + addToWorklist(V, ConstantPointerNull::get(PTy), + ICmpInst::ICMP_EQ, VRP); + return; } } - } - } -#endif - }; - - InequalityGraph::~InequalityGraph() { - if (ConcreteIG) return; - std::vector Remove; - for (NodeMapType::iterator I = Nodes.begin(), E = Nodes.end(); - I != E; ++I) { - if (I->first == I->second->getValue()) - Remove.push_back(I->second); - } - for (std::vector::iterator I = Remove.begin(), E = Remove.end(); - I != E; ++I) { - delete *I; + update(n, Merged, Subtree); } - } - template - void InequalityGraph::mergeNodes(Node *N, C &Kill, Value *NewCanonical) { - materialize(); + void addNotEquals(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree, + VRPSolver *VRP) { + ConstantRange CR1 = range(n1, Subtree); + ConstantRange CR2 = range(n2, Subtree); + + uint32_t W = CR1.getBitWidth(); + + if (const APInt *I = CR1.getSingleElement()) { + if (CR2.isFullSet()) { + ConstantRange NewCR2(CR1.getUpper(), CR1.getLower()); + applyRange(n2, NewCR2, Subtree, VRP); + } else if (*I == CR2.getLower()) { + APInt NewLower(CR2.getLower() + 1), + NewUpper(CR2.getUpper()); + if (NewLower == NewUpper) + NewLower = NewUpper = APInt::getMinValue(W); + + ConstantRange NewCR2(NewLower, NewUpper); + applyRange(n2, NewCR2, Subtree, VRP); + } else if (*I == CR2.getUpper() - 1) { + APInt NewLower(CR2.getLower()), + NewUpper(CR2.getUpper() - 1); + if (NewLower == NewUpper) + NewLower = NewUpper = APInt::getMinValue(W); + + ConstantRange NewCR2(NewLower, NewUpper); + applyRange(n2, NewCR2, Subtree, VRP); + } + } - // Merge the relationships from the members of Kill into N. - for (typename C::iterator KI = Kill.begin(), KE = Kill.end(); - KI != KE; ++KI) { + if (const APInt *I = CR2.getSingleElement()) { + if (CR1.isFullSet()) { + ConstantRange NewCR1(CR2.getUpper(), CR2.getLower()); + applyRange(n1, NewCR1, Subtree, VRP); + } else if (*I == CR1.getLower()) { + APInt NewLower(CR1.getLower() + 1), + NewUpper(CR1.getUpper()); + if (NewLower == NewUpper) + NewLower = NewUpper = APInt::getMinValue(W); + + ConstantRange NewCR1(NewLower, NewUpper); + applyRange(n1, NewCR1, Subtree, VRP); + } else if (*I == CR1.getUpper() - 1) { + APInt NewLower(CR1.getLower()), + NewUpper(CR1.getUpper() - 1); + if (NewLower == NewUpper) + NewLower = NewUpper = APInt::getMinValue(W); + + ConstantRange NewCR1(NewLower, NewUpper); + applyRange(n1, NewCR1, Subtree, VRP); + } + } + } - for (Node::iterator I = (*KI)->begin(), E = (*KI)->end(); I != E; ++I) { - if (I->first == N) continue; + void addInequality(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree, + LatticeVal LV, VRPSolver *VRP) { + assert(!isRelatedBy(n1, n2, Subtree, LV) && "Asked to do useless work."); - Node::iterator NI = N->find(I->first); - if (NI == N->end()) { - N->Relations.push_back(std::make_pair(I->first, I->second)); - } else { - unsigned char LV = NI->second & I->second; - if (LV == EQ_BIT) { + if (LV == NE) { + addNotEquals(n1, n2, Subtree, VRP); + return; + } - assert(std::find(Kill.begin(), Kill.end(), I->first) != Kill.end() - && "Lost EQ property."); - N->erase(NI); - } else { - NI->second = static_cast(LV); - assert(InequalityGraph::validPredicate(NI->second) && - "Invalid union of lattice values."); - } - } + ConstantRange CR1 = range(n1, Subtree); + ConstantRange CR2 = range(n2, Subtree); - // All edges are reciprocal; every Node that Kill points to also - // contains a pointer to Kill. Replace those with pointers with N. - unsigned iter = I->first->findIndex(*KI); - assert(iter != (unsigned)-1 && "Edge not reciprocal."); - I->first->assign(N, (I->first->begin()+iter)->second); - I->first->erase(I->first->begin()+iter); + if (!CR1.isSingleElement()) { + ConstantRange NewCR1 = CR1.maximalIntersectWith(create(LV, CR2)); + if (NewCR1 != CR1) + applyRange(n1, NewCR1, Subtree, VRP); } - // Removing references from N to Kill. - Node::iterator NI = N->find(*KI); - if (NI != N->end()) { - N->erase(NI); // breaks reciprocity until Kill is deleted. + if (!CR2.isSingleElement()) { + ConstantRange NewCR2 = CR2.maximalIntersectWith( + create(reversePredicate(LV), CR1)); + if (NewCR2 != CR2) + applyRange(n2, NewCR2, Subtree, VRP); } } + }; - N->setValue(NewCanonical); + /// UnreachableBlocks keeps tracks of blocks that are for one reason or + /// another discovered to be unreachable. This is used to cull the graph when + /// analyzing instructions, and to mark blocks with the "unreachable" + /// terminator instruction after the function has executed. + class VISIBILITY_HIDDEN UnreachableBlocks { + private: + std::vector DeadBlocks; - // Update value mapping to point to the merged node. - for (NodeMapType::iterator I = Nodes.begin(), E = Nodes.end(); - I != E; ++I) { - if (std::find(Kill.begin(), Kill.end(), I->second) != Kill.end()) - I->second = N; - } + public: + /// mark - mark a block as dead + void mark(BasicBlock *BB) { + std::vector::iterator E = DeadBlocks.end(); + std::vector::iterator I = + std::lower_bound(DeadBlocks.begin(), E, BB); - for (typename C::iterator KI = Kill.begin(), KE = Kill.end(); - KI != KE; ++KI) { - delete *KI; + if (I == E || *I != BB) DeadBlocks.insert(I, BB); } - } - void InequalityGraph::materialize() { - if (!ConcreteIG) return; - const InequalityGraph *IG = ConcreteIG; - ConcreteIG = NULL; + /// isDead - returns whether a block is known to be dead already + bool isDead(BasicBlock *BB) { + std::vector::iterator E = DeadBlocks.end(); + std::vector::iterator I = + std::lower_bound(DeadBlocks.begin(), E, BB); - for (NodeMapType::const_iterator I = IG->Nodes.begin(), - E = IG->Nodes.end(); I != E; ++I) { - if (I->first == I->second->getValue()) { - Node *N = newNode(I->first); - N->Relations.reserve(N->Relations.size()); - } + return I != E && *I == BB; } - for (NodeMapType::const_iterator I = IG->Nodes.begin(), - E = IG->Nodes.end(); I != E; ++I) { - if (I->first != I->second->getValue()) { - Nodes[I->first] = getNode(I->second->getValue()); - } else { - Node *Old = I->second; - Node *N = getNode(I->first); - for (Node::const_iterator NI = Old->begin(), NE = Old->end(); - NI != NE; ++NI) { - N->assign(getNode(NI->first->getValue()), NI->second); + + /// kill - replace the dead blocks' terminator with an UnreachableInst. + bool kill() { + bool modified = false; + for (std::vector::iterator I = DeadBlocks.begin(), + E = DeadBlocks.end(); I != E; ++I) { + BasicBlock *BB = *I; + + DOUT << "unreachable block: " << BB->getName() << "\n"; + + for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); + SI != SE; ++SI) { + BasicBlock *Succ = *SI; + Succ->removePredecessor(BB); } + + TerminatorInst *TI = BB->getTerminator(); + TI->replaceAllUsesWith(UndefValue::get(TI->getType())); + TI->eraseFromParent(); + new UnreachableInst(BB); + ++NumBlocks; + modified = true; } + DeadBlocks.clear(); + return modified; } - } + }; /// VRPSolver keeps track of how changes to one variable affect other /// variables, and forwards changes along to the InequalityGraph. It @@ -516,588 +1370,979 @@ namespace { /// @brief VRPSolver calculates inferences from a new relationship. class VISIBILITY_HIDDEN VRPSolver { private: - std::deque WorkList; - - InequalityGraph &IG; - const InequalityGraph &cIG; - ETForest *Forest; - ETNode *Top; + friend class ValueRanges; - typedef InequalityGraph::Node Node; + struct Operation { + Value *LHS, *RHS; + ICmpInst::Predicate Op; - /// Returns true if V1 is a better canonical value than V2. - bool compare(Value *V1, Value *V2) const { - if (isa(V1)) - return !isa(V2); - else if (isa(V2)) - return false; - else if (isa(V1)) - return !isa(V2); - else if (isa(V2)) - return false; + BasicBlock *ContextBB; // XXX use a DomTreeDFS::Node instead + Instruction *ContextInst; + }; + std::deque WorkList; - Instruction *I1 = dyn_cast(V1); - Instruction *I2 = dyn_cast(V2); + ValueNumbering &VN; + InequalityGraph &IG; + UnreachableBlocks &UB; + ValueRanges &VR; + DomTreeDFS *DTDFS; + DomTreeDFS::Node *Top; + BasicBlock *TopBB; + Instruction *TopInst; + bool &modified; - if (!I1 || !I2) return false; + typedef InequalityGraph::Node Node; - BasicBlock *BB1 = I1->getParent(), - *BB2 = I2->getParent(); - if (BB1 == BB2) { - for (BasicBlock::const_iterator I = BB1->begin(), E = BB1->end(); - I != E; ++I) { - if (&*I == I1) return true; - if (&*I == I2) return false; + // below - true if the Instruction is dominated by the current context + // block or instruction + bool below(Instruction *I) { + BasicBlock *BB = I->getParent(); + if (TopInst && TopInst->getParent() == BB) { + if (isa(TopInst)) return false; + if (isa(I)) return true; + if ( isa(TopInst) && !isa(I)) return true; + if (!isa(TopInst) && isa(I)) return false; + + for (BasicBlock::const_iterator Iter = BB->begin(), E = BB->end(); + Iter != E; ++Iter) { + if (&*Iter == TopInst) return true; + else if (&*Iter == I) return false; } assert(!"Instructions not found in parent BasicBlock?"); } else { - return Forest->properlyDominates(BB1, BB2); + DomTreeDFS::Node *Node = DTDFS->getNodeForBlock(BB); + if (!Node) return false; + return Top->dominates(Node); } - return false; + return false; // Not reached } - void addToWorklist(Instruction *I) { - //DOUT << "addToWorklist: " << *I << "\n"; - - if (!isa(I) && !isa(I)) return; + // aboveOrBelow - true if the Instruction either dominates or is dominated + // by the current context block or instruction + bool aboveOrBelow(Instruction *I) { + BasicBlock *BB = I->getParent(); + DomTreeDFS::Node *Node = DTDFS->getNodeForBlock(BB); + if (!Node) return false; - const Type *Ty = I->getType(); - if (Ty == Type::VoidTy || Ty->isFPOrFPVector()) return; + return Top == Node || Top->dominates(Node) || Node->dominates(Top); + } - if (isInstructionTriviallyDead(I)) return; + bool makeEqual(Value *V1, Value *V2) { + DOUT << "makeEqual(" << *V1 << ", " << *V2 << ")\n"; + DOUT << "context is "; + if (TopInst) DOUT << "I: " << *TopInst << "\n"; + else DOUT << "BB: " << TopBB->getName() + << "(" << Top->getDFSNumIn() << ")\n"; - WorkList.push_back(I); - } + assert(V1->getType() == V2->getType() && + "Can't make two values with different types equal."); - void addRecursive(Value *V) { - //DOUT << "addRecursive: " << *V << "\n"; + if (V1 == V2) return true; - Instruction *I = dyn_cast(V); - if (I) - addToWorklist(I); - else if (!isa(V)) - return; + if (isa(V1) && isa(V2)) + return false; - //DOUT << "addRecursive uses...\n"; - for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); - UI != UE; ++UI) { - // Use must be either be dominated by Top, or dominate Top. - if (Instruction *Inst = dyn_cast(*UI)) { - ETNode *INode = Forest->getNodeForBlock(Inst->getParent()); - if (INode->DominatedBy(Top) || Top->DominatedBy(INode)) - addToWorklist(Inst); - } - } + unsigned n1 = VN.valueNumber(V1, Top), n2 = VN.valueNumber(V2, Top); - if (I) { - //DOUT << "addRecursive ops...\n"; - for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); - OI != OE; ++OI) { - if (Instruction *Inst = dyn_cast(*OI)) - addToWorklist(Inst); - } + if (n1 && n2) { + if (n1 == n2) return true; + if (IG.isRelatedBy(n1, n2, Top, NE)) return false; } - //DOUT << "exit addRecursive (" << *V << ").\n"; - } - - public: - VRPSolver(InequalityGraph &IG, ETForest *Forest, BasicBlock *TopBB) - : IG(IG), cIG(IG), Forest(Forest), Top(Forest->getNodeForBlock(TopBB)) {} - bool isEqual(Value *V1, Value *V2) const { - if (V1 == V2) return true; - if (const Node *N1 = cIG.getNode(V1)) - return N1 == cIG.getNode(V2); - return false; - } + if (n1) assert(V1 == VN.value(n1) && "Value isn't canonical."); + if (n2) assert(V2 == VN.value(n2) && "Value isn't canonical."); - bool isNotEqual(Value *V1, Value *V2) const { - if (V1 == V2) return false; - if (const Node *N1 = cIG.getNode(V1)) - if (const Node *N2 = cIG.getNode(V2)) - return cIG.isNotEqual(N1, N2); - return false; - } + assert(!VN.compare(V2, V1) && "Please order parameters to makeEqual."); - bool isLess(Value *V1, Value *V2) const { - if (V1 == V2) return false; - if (const Node *N1 = cIG.getNode(V1)) - if (const Node *N2 = cIG.getNode(V2)) - return cIG.isLess(N1, N2); - return false; - } + assert(!isa(V2) && "Tried to remove a constant."); - bool isLessEqual(Value *V1, Value *V2) const { - if (V1 == V2) return true; - if (const Node *N1 = cIG.getNode(V1)) - if (const Node *N2 = cIG.getNode(V2)) - return cIG.isLessEqual(N1, N2); - return false; - } + SetVector Remove; + if (n2) Remove.insert(n2); - bool isGreater(Value *V1, Value *V2) const { - if (V1 == V2) return false; - if (const Node *N1 = cIG.getNode(V1)) - if (const Node *N2 = cIG.getNode(V2)) - return cIG.isGreater(N1, N2); - return false; - } + if (n1 && n2) { + // Suppose we're being told that %x == %y, and %x <= %z and %y >= %z. + // We can't just merge %x and %y because the relationship with %z would + // be EQ and that's invalid. What we're doing is looking for any nodes + // %z such that %x <= %z and %y >= %z, and vice versa. - bool isGreaterEqual(Value *V1, Value *V2) const { - if (V1 == V2) return true; - if (const Node *N1 = IG.getNode(V1)) - if (const Node *N2 = IG.getNode(V2)) - return cIG.isGreaterEqual(N1, N2); - return false; - } + Node::iterator end = IG.node(n2)->end(); - // All of the add* functions return true if the InequalityGraph represents - // the property, and false if there is a logical contradiction. On false, - // you may no longer perform any queries on the InequalityGraph. + // Find the intersection between N1 and N2 which is dominated by + // Top. If we find %x where N1 <= %x <= N2 (or >=) then add %x to + // Remove. + for (Node::iterator I = IG.node(n1)->begin(), E = IG.node(n1)->end(); + I != E; ++I) { + if (!(I->LV & EQ_BIT) || !Top->DominatedBy(I->Subtree)) continue; + + unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT); + unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT); + Node::iterator NI = IG.node(n2)->find(I->To, Top); + if (NI != end) { + LatticeVal NILV = reversePredicate(NI->LV); + unsigned NILV_s = NILV & (SLT_BIT|SGT_BIT); + unsigned NILV_u = NILV & (ULT_BIT|UGT_BIT); + + if ((ILV_s != (SLT_BIT|SGT_BIT) && ILV_s == NILV_s) || + (ILV_u != (ULT_BIT|UGT_BIT) && ILV_u == NILV_u)) + Remove.insert(I->To); + } + } - bool addEqual(Value *V1, Value *V2) { - //DOUT << "addEqual(" << *V1 << ", " << *V2 << ")\n"; - if (isEqual(V1, V2)) return true; + // See if one of the nodes about to be removed is actually a better + // canonical choice than n1. + unsigned orig_n1 = n1; + SetVector::iterator DontRemove = Remove.end(); + for (SetVector::iterator I = Remove.begin()+1 /* skip n2 */, + E = Remove.end(); I != E; ++I) { + unsigned n = *I; + Value *V = VN.value(n); + if (VN.compare(V, V1)) { + V1 = V; + n1 = n; + DontRemove = I; + } + } + if (DontRemove != Remove.end()) { + unsigned n = *DontRemove; + Remove.remove(n); + Remove.insert(orig_n1); + } + } - const Node *cN1 = cIG.getNode(V1), *cN2 = cIG.getNode(V2); + // We'd like to allow makeEqual on two values to perform a simple + // substitution without creating nodes in the IG whenever possible. + // + // The first iteration through this loop operates on V2 before going + // through the Remove list and operating on those too. If all of the + // iterations performed simple replacements then we exit early. + bool mergeIGNode = false; + unsigned i = 0; + for (Value *R = V2; i == 0 || i < Remove.size(); ++i) { + if (i) R = VN.value(Remove[i]); // skip n2. + + // Try to replace the whole instruction. If we can, we're done. + Instruction *I2 = dyn_cast(R); + if (I2 && below(I2)) { + std::vector ToNotify; + for (Value::use_iterator UI = I2->use_begin(), UE = I2->use_end(); + UI != UE;) { + Use &TheUse = UI.getUse(); + ++UI; + Instruction *I = cast(TheUse.getUser()); + ToNotify.push_back(I); + } - if (cN1 && cN2 && cIG.isNotEqual(cN1, cN2)) - return false; + DOUT << "Simply removing " << *I2 + << ", replacing with " << *V1 << "\n"; + I2->replaceAllUsesWith(V1); + // leave it dead; it'll get erased later. + ++NumInstruction; + modified = true; - if (compare(V2, V1)) { std::swap(V1, V2); std::swap(cN1, cN2); } + for (std::vector::iterator II = ToNotify.begin(), + IE = ToNotify.end(); II != IE; ++II) { + opsToDef(*II); + } - if (cN1) { - if (ConstantBool *CB = dyn_cast(V1)) { - Node *N1 = IG.getNode(V1); - - // When "addEqual" is performed and the new value is a ConstantBool, - // iterate through the NE set and fix them up to be EQ of the - // opposite bool. + continue; + } - for (Node::iterator I = N1->begin(), E = N1->end(); I != E; ++I) - if ((I->second & 1) == 0) { - assert(N1 != I->first && "Node related to itself?"); - addEqual(I->first->getValue(), - ConstantBool::get(!CB->getValue())); + // Otherwise, replace all dominated uses. + for (Value::use_iterator UI = R->use_begin(), UE = R->use_end(); + UI != UE;) { + Use &TheUse = UI.getUse(); + ++UI; + if (Instruction *I = dyn_cast(TheUse.getUser())) { + if (below(I)) { + TheUse.set(V1); + modified = true; + ++NumVarsReplaced; + opsToDef(I); } + } + } + + // If that killed the instruction, stop here. + if (I2 && isInstructionTriviallyDead(I2)) { + DOUT << "Killed all uses of " << *I2 + << ", replacing with " << *V1 << "\n"; + continue; } + + // If we make it to here, then we will need to create a node for N1. + // Otherwise, we can skip out early! + mergeIGNode = true; } - if (!cN2) { - if (Instruction *I2 = dyn_cast(V2)) { - ETNode *Node_I2 = Forest->getNodeForBlock(I2->getParent()); - if (Top != Node_I2 && Node_I2->DominatedBy(Top)) { - Value *V = V1; - if (cN1 && compare(V1, cN1->getValue())) V = cN1->getValue(); - //DOUT << "Simply removing " << *I2 - // << ", replacing with " << *V << "\n"; - I2->replaceAllUsesWith(V); - // leave it dead; it'll get erased later. - ++NumSimple; - addRecursive(V1); - return true; + if (!isa(V1)) { + if (Remove.empty()) { + VR.mergeInto(&V2, 1, VN.getOrInsertVN(V1, Top), Top, this); + } else { + std::vector RemoveVals; + RemoveVals.reserve(Remove.size()); + + for (SetVector::iterator I = Remove.begin(), + E = Remove.end(); I != E; ++I) { + Value *V = VN.value(*I); + if (!V->use_empty()) + RemoveVals.push_back(V); } + VR.mergeInto(&RemoveVals[0], RemoveVals.size(), + VN.getOrInsertVN(V1, Top), Top, this); } } - Node *N1 = IG.getNode(V1), *N2 = IG.getNode(V2); + if (mergeIGNode) { + // Create N1. + if (!n1) n1 = VN.getOrInsertVN(V1, Top); + IG.node(n1); // Ensure that IG.Nodes won't get resized - if ( N1 && !N2) { - IG.addEqual(N1, V2); - if (compare(V1, N1->getValue())) N1->setValue(V1); - } - if (!N1 && N2) { - IG.addEqual(N2, V1); - if (compare(V1, N2->getValue())) N2->setValue(V1); - } - if ( N1 && N2) { - // Suppose we're being told that %x == %y, and %x <= %z and %y >= %z. - // We can't just merge %x and %y because the relationship with %z would - // be EQ and that's invalid; they need to be the same Node. - // - // What we're doing is looking for any chain of nodes reaching %z such - // that %x <= %z and %y >= %z, and vice versa. The cool part is that - // every node in between is also equal because of the squeeze principle. - - std::vector N1_GE, N2_LE, N1_LE, N2_GE; - IG.transitiveClosure(N1, InequalityGraph::GE, back_inserter(N1_GE)); - std::sort(N1_GE.begin(), N1_GE.end()); - N1_GE.erase(std::unique(N1_GE.begin(), N1_GE.end()), N1_GE.end()); - IG.transitiveClosure(N2, InequalityGraph::LE, back_inserter(N2_LE)); - std::sort(N1_LE.begin(), N1_LE.end()); - N1_LE.erase(std::unique(N1_LE.begin(), N1_LE.end()), N1_LE.end()); - IG.transitiveClosure(N1, InequalityGraph::LE, back_inserter(N1_LE)); - std::sort(N2_GE.begin(), N2_GE.end()); - N2_GE.erase(std::unique(N2_GE.begin(), N2_GE.end()), N2_GE.end()); - std::unique(N2_GE.begin(), N2_GE.end()); - IG.transitiveClosure(N2, InequalityGraph::GE, back_inserter(N2_GE)); - std::sort(N2_LE.begin(), N2_LE.end()); - N2_LE.erase(std::unique(N2_LE.begin(), N2_LE.end()), N2_LE.end()); - - std::vector Set1, Set2; - std::set_intersection(N1_GE.begin(), N1_GE.end(), - N2_LE.begin(), N2_LE.end(), - back_inserter(Set1)); - std::set_intersection(N1_LE.begin(), N1_LE.end(), - N2_GE.begin(), N2_GE.end(), - back_inserter(Set2)); - - std::vector Equal; - std::set_union(Set1.begin(), Set1.end(), Set2.begin(), Set2.end(), - back_inserter(Equal)); - - Value *Best = N1->getValue(); - if (compare(N2->getValue(), Best)) Best = N2->getValue(); - - for (std::vector::iterator I = Equal.begin(), E = Equal.end(); + // Migrate relationships from removed nodes to N1. + for (SetVector::iterator I = Remove.begin(), E = Remove.end(); I != E; ++I) { - Value *V = (*I)->getValue(); - if (compare(V, Best)) Best = V; + unsigned n = *I; + for (Node::iterator NI = IG.node(n)->begin(), NE = IG.node(n)->end(); + NI != NE; ++NI) { + if (NI->Subtree->DominatedBy(Top)) { + if (NI->To == n1) { + assert((NI->LV & EQ_BIT) && "Node inequal to itself."); + continue; + } + if (Remove.count(NI->To)) + continue; + + IG.node(NI->To)->update(n1, reversePredicate(NI->LV), Top); + IG.node(n1)->update(NI->To, NI->LV, Top); + } + } + } + + // Point V2 (and all items in Remove) to N1. + if (!n2) + VN.addEquality(n1, V2, Top); + else { + for (SetVector::iterator I = Remove.begin(), + E = Remove.end(); I != E; ++I) { + VN.addEquality(n1, VN.value(*I), Top); + } } - Equal.push_back(N2); - IG.mergeNodes(N1, Equal, Best); + // If !Remove.empty() then V2 = Remove[0]->getValue(). + // Even when Remove is empty, we still want to process V2. + i = 0; + for (Value *R = V2; i == 0 || i < Remove.size(); ++i) { + if (i) R = VN.value(Remove[i]); // skip n2. + + if (Instruction *I2 = dyn_cast(R)) { + if (aboveOrBelow(I2)) + defToOps(I2); + } + for (Value::use_iterator UI = V2->use_begin(), UE = V2->use_end(); + UI != UE;) { + Use &TheUse = UI.getUse(); + ++UI; + if (Instruction *I = dyn_cast(TheUse.getUser())) { + if (aboveOrBelow(I)) + opsToDef(I); + } + } + } } - if (!N1 && !N2) IG.addEqual(IG.newNode(V1), V2); - addRecursive(V1); - addRecursive(V2); + // re-opsToDef all dominated users of V1. + if (Instruction *I = dyn_cast(V1)) { + for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); + UI != UE;) { + Use &TheUse = UI.getUse(); + ++UI; + Value *V = TheUse.getUser(); + if (!V->use_empty()) { + Instruction *Inst = cast(V); + if (aboveOrBelow(Inst)) + opsToDef(Inst); + } + } + } return true; } - bool addNotEqual(Value *V1, Value *V2) { - //DOUT << "addNotEqual(" << *V1 << ", " << *V2 << ")\n"); - if (isNotEqual(V1, V2)) return true; - - // Never permit %x NE true/false. - if (ConstantBool *B1 = dyn_cast(V1)) { - return addEqual(ConstantBool::get(!B1->getValue()), V2); - } else if (ConstantBool *B2 = dyn_cast(V2)) { - return addEqual(V1, ConstantBool::get(!B2->getValue())); + /// cmpInstToLattice - converts an CmpInst::Predicate to lattice value + /// Requires that the lattice value be valid; does not accept ICMP_EQ. + static LatticeVal cmpInstToLattice(ICmpInst::Predicate Pred) { + switch (Pred) { + case ICmpInst::ICMP_EQ: + assert(!"No matching lattice value."); + return static_cast(EQ_BIT); + default: + assert(!"Invalid 'icmp' predicate."); + case ICmpInst::ICMP_NE: + return NE; + case ICmpInst::ICMP_UGT: + return UGT; + case ICmpInst::ICMP_UGE: + return UGE; + case ICmpInst::ICMP_ULT: + return ULT; + case ICmpInst::ICMP_ULE: + return ULE; + case ICmpInst::ICMP_SGT: + return SGT; + case ICmpInst::ICMP_SGE: + return SGE; + case ICmpInst::ICMP_SLT: + return SLT; + case ICmpInst::ICMP_SLE: + return SLE; } + } + + public: + VRPSolver(ValueNumbering &VN, InequalityGraph &IG, UnreachableBlocks &UB, + ValueRanges &VR, DomTreeDFS *DTDFS, bool &modified, + BasicBlock *TopBB) + : VN(VN), + IG(IG), + UB(UB), + VR(VR), + DTDFS(DTDFS), + Top(DTDFS->getNodeForBlock(TopBB)), + TopBB(TopBB), + TopInst(NULL), + modified(modified) + { + assert(Top && "VRPSolver created for unreachable basic block."); + } - Node *N1 = IG.getOrInsertNode(V1), - *N2 = IG.getOrInsertNode(V2); + VRPSolver(ValueNumbering &VN, InequalityGraph &IG, UnreachableBlocks &UB, + ValueRanges &VR, DomTreeDFS *DTDFS, bool &modified, + Instruction *TopInst) + : VN(VN), + IG(IG), + UB(UB), + VR(VR), + DTDFS(DTDFS), + Top(DTDFS->getNodeForBlock(TopInst->getParent())), + TopBB(TopInst->getParent()), + TopInst(TopInst), + modified(modified) + { + assert(Top && "VRPSolver created for unreachable basic block."); + assert(Top->getBlock() == TopInst->getParent() && "Context mismatch."); + } - if (N1 == N2) return false; + bool isRelatedBy(Value *V1, Value *V2, ICmpInst::Predicate Pred) const { + if (Constant *C1 = dyn_cast(V1)) + if (Constant *C2 = dyn_cast(V2)) + return ConstantExpr::getCompare(Pred, C1, C2) == + ConstantInt::getTrue(); + + unsigned n1 = VN.valueNumber(V1, Top); + unsigned n2 = VN.valueNumber(V2, Top); + + if (n1 && n2) { + if (n1 == n2) return Pred == ICmpInst::ICMP_EQ || + Pred == ICmpInst::ICMP_ULE || + Pred == ICmpInst::ICMP_UGE || + Pred == ICmpInst::ICMP_SLE || + Pred == ICmpInst::ICMP_SGE; + if (Pred == ICmpInst::ICMP_EQ) return false; + if (IG.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true; + if (VR.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true; + } - IG.addNotEqual(N1, N2); + if ((n1 && !n2 && isa(V2)) || + (n2 && !n1 && isa(V1))) { + ConstantRange CR1 = n1 ? VR.range(n1, Top) : VR.range(V1); + ConstantRange CR2 = n2 ? VR.range(n2, Top) : VR.range(V2); - addRecursive(V1); - addRecursive(V2); + if (Pred == ICmpInst::ICMP_EQ) + return CR1.isSingleElement() && + CR1.getSingleElement() == CR2.getSingleElement(); - return true; + return VR.isRelatedBy(CR1, CR2, cmpInstToLattice(Pred)); + } + if (Pred == ICmpInst::ICMP_EQ) return V1 == V2; + return false; } - /// Set V1 less than V2. - bool addLess(Value *V1, Value *V2) { - if (isLess(V1, V2)) return true; - if (isGreaterEqual(V1, V2)) return false; + /// add - adds a new property to the work queue + void add(Value *V1, Value *V2, ICmpInst::Predicate Pred, + Instruction *I = NULL) { + DOUT << "adding " << *V1 << " " << Pred << " " << *V2; + if (I) DOUT << " context: " << *I; + else DOUT << " default context (" << Top->getDFSNumIn() << ")"; + DOUT << "\n"; + + assert(V1->getType() == V2->getType() && + "Can't relate two values with different types."); + + WorkList.push_back(Operation()); + Operation &O = WorkList.back(); + O.LHS = V1, O.RHS = V2, O.Op = Pred, O.ContextInst = I; + O.ContextBB = I ? I->getParent() : TopBB; + } - Node *N1 = IG.getOrInsertNode(V1), *N2 = IG.getOrInsertNode(V2); + /// defToOps - Given an instruction definition that we've learned something + /// new about, find any new relationships between its operands. + void defToOps(Instruction *I) { + Instruction *NewContext = below(I) ? I : TopInst; + Value *Canonical = VN.canonicalize(I, Top); - if (N1 == N2) return false; + if (BinaryOperator *BO = dyn_cast(I)) { + const Type *Ty = BO->getType(); + assert(!Ty->isFPOrFPVector() && "Float in work queue!"); - IG.addLess(N1, N2); + Value *Op0 = VN.canonicalize(BO->getOperand(0), Top); + Value *Op1 = VN.canonicalize(BO->getOperand(1), Top); - addRecursive(V1); - addRecursive(V2); + // TODO: "and i32 -1, %x" EQ %y then %x EQ %y. - return true; + switch (BO->getOpcode()) { + case Instruction::And: { + // "and i32 %a, %b" EQ -1 then %a EQ -1 and %b EQ -1 + ConstantInt *CI = ConstantInt::getAllOnesValue(Ty); + if (Canonical == CI) { + add(CI, Op0, ICmpInst::ICMP_EQ, NewContext); + add(CI, Op1, ICmpInst::ICMP_EQ, NewContext); + } + } break; + case Instruction::Or: { + // "or i32 %a, %b" EQ 0 then %a EQ 0 and %b EQ 0 + Constant *Zero = Constant::getNullValue(Ty); + if (Canonical == Zero) { + add(Zero, Op0, ICmpInst::ICMP_EQ, NewContext); + add(Zero, Op1, ICmpInst::ICMP_EQ, NewContext); + } + } break; + case Instruction::Xor: { + // "xor i32 %c, %a" EQ %b then %a EQ %c ^ %b + // "xor i32 %c, %a" EQ %c then %a EQ 0 + // "xor i32 %c, %a" NE %c then %a NE 0 + // Repeat the above, with order of operands reversed. + Value *LHS = Op0; + Value *RHS = Op1; + if (!isa(LHS)) std::swap(LHS, RHS); + + if (ConstantInt *CI = dyn_cast(Canonical)) { + if (ConstantInt *Arg = dyn_cast(LHS)) { + add(RHS, ConstantInt::get(CI->getValue() ^ Arg->getValue()), + ICmpInst::ICMP_EQ, NewContext); + } + } + if (Canonical == LHS) { + if (isa(Canonical)) + add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_EQ, + NewContext); + } else if (isRelatedBy(LHS, Canonical, ICmpInst::ICMP_NE)) { + add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_NE, + NewContext); + } + } break; + default: + break; + } + } else if (ICmpInst *IC = dyn_cast(I)) { + // "icmp ult i32 %a, %y" EQ true then %a u< y + // etc. + + if (Canonical == ConstantInt::getTrue()) { + add(IC->getOperand(0), IC->getOperand(1), IC->getPredicate(), + NewContext); + } else if (Canonical == ConstantInt::getFalse()) { + add(IC->getOperand(0), IC->getOperand(1), + ICmpInst::getInversePredicate(IC->getPredicate()), NewContext); + } + } else if (SelectInst *SI = dyn_cast(I)) { + if (I->getType()->isFPOrFPVector()) return; + + // Given: "%a = select i1 %x, i32 %b, i32 %c" + // %a EQ %b and %b NE %c then %x EQ true + // %a EQ %c and %b NE %c then %x EQ false + + Value *True = SI->getTrueValue(); + Value *False = SI->getFalseValue(); + if (isRelatedBy(True, False, ICmpInst::ICMP_NE)) { + if (Canonical == VN.canonicalize(True, Top) || + isRelatedBy(Canonical, False, ICmpInst::ICMP_NE)) + add(SI->getCondition(), ConstantInt::getTrue(), + ICmpInst::ICMP_EQ, NewContext); + else if (Canonical == VN.canonicalize(False, Top) || + isRelatedBy(Canonical, True, ICmpInst::ICMP_NE)) + add(SI->getCondition(), ConstantInt::getFalse(), + ICmpInst::ICMP_EQ, NewContext); + } + } else if (GetElementPtrInst *GEPI = dyn_cast(I)) { + for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(), + OE = GEPI->idx_end(); OI != OE; ++OI) { + ConstantInt *Op = dyn_cast(VN.canonicalize(*OI, Top)); + if (!Op || !Op->isZero()) return; + } + // TODO: The GEPI indices are all zero. Copy from definition to operand, + // jumping the type plane as needed. + if (isRelatedBy(GEPI, Constant::getNullValue(GEPI->getType()), + ICmpInst::ICMP_NE)) { + Value *Ptr = GEPI->getPointerOperand(); + add(Ptr, Constant::getNullValue(Ptr->getType()), ICmpInst::ICMP_NE, + NewContext); + } + } else if (CastInst *CI = dyn_cast(I)) { + const Type *SrcTy = CI->getSrcTy(); + + unsigned ci = VN.getOrInsertVN(CI, Top); + uint32_t W = VR.typeToWidth(SrcTy); + if (!W) return; + ConstantRange CR = VR.range(ci, Top); + + if (CR.isFullSet()) return; + + switch (CI->getOpcode()) { + default: break; + case Instruction::ZExt: + case Instruction::SExt: + VR.applyRange(VN.getOrInsertVN(CI->getOperand(0), Top), + CR.truncate(W), Top, this); + break; + case Instruction::BitCast: + VR.applyRange(VN.getOrInsertVN(CI->getOperand(0), Top), + CR, Top, this); + break; + } + } } - /// Set V1 less than or equal to V2. - bool addLessEqual(Value *V1, Value *V2) { - if (isLessEqual(V1, V2)) return true; - if (V1 == V2) return true; + /// opsToDef - A new relationship was discovered involving one of this + /// instruction's operands. Find any new relationship involving the + /// definition, or another operand. + void opsToDef(Instruction *I) { + Instruction *NewContext = below(I) ? I : TopInst; + + if (BinaryOperator *BO = dyn_cast(I)) { + Value *Op0 = VN.canonicalize(BO->getOperand(0), Top); + Value *Op1 = VN.canonicalize(BO->getOperand(1), Top); + + if (ConstantInt *CI0 = dyn_cast(Op0)) + if (ConstantInt *CI1 = dyn_cast(Op1)) { + add(BO, ConstantExpr::get(BO->getOpcode(), CI0, CI1), + ICmpInst::ICMP_EQ, NewContext); + return; + } - if (isLessEqual(V2, V1)) - return addEqual(V1, V2); + // "%y = and i1 true, %x" then %x EQ %y + // "%y = or i1 false, %x" then %x EQ %y + // "%x = add i32 %y, 0" then %x EQ %y + // "%x = mul i32 %y, 0" then %x EQ 0 + + Instruction::BinaryOps Opcode = BO->getOpcode(); + const Type *Ty = BO->getType(); + assert(!Ty->isFPOrFPVector() && "Float in work queue!"); + + Constant *Zero = Constant::getNullValue(Ty); + Constant *One = ConstantInt::get(Ty, 1); + ConstantInt *AllOnes = ConstantInt::getAllOnesValue(Ty); + + switch (Opcode) { + default: break; + case Instruction::LShr: + case Instruction::AShr: + case Instruction::Shl: + if (Op1 == Zero) { + add(BO, Op0, ICmpInst::ICMP_EQ, NewContext); + return; + } + break; + case Instruction::Sub: + if (Op1 == Zero) { + add(BO, Op0, ICmpInst::ICMP_EQ, NewContext); + return; + } + if (ConstantInt *CI0 = dyn_cast(Op0)) { + unsigned n_ci0 = VN.getOrInsertVN(Op1, Top); + ConstantRange CR = VR.range(n_ci0, Top); + if (!CR.isFullSet()) { + CR.subtract(CI0->getValue()); + unsigned n_bo = VN.getOrInsertVN(BO, Top); + VR.applyRange(n_bo, CR, Top, this); + return; + } + } + if (ConstantInt *CI1 = dyn_cast(Op1)) { + unsigned n_ci1 = VN.getOrInsertVN(Op0, Top); + ConstantRange CR = VR.range(n_ci1, Top); + if (!CR.isFullSet()) { + CR.subtract(CI1->getValue()); + unsigned n_bo = VN.getOrInsertVN(BO, Top); + VR.applyRange(n_bo, CR, Top, this); + return; + } + } + break; + case Instruction::Or: + if (Op0 == AllOnes || Op1 == AllOnes) { + add(BO, AllOnes, ICmpInst::ICMP_EQ, NewContext); + return; + } + if (Op0 == Zero) { + add(BO, Op1, ICmpInst::ICMP_EQ, NewContext); + return; + } else if (Op1 == Zero) { + add(BO, Op0, ICmpInst::ICMP_EQ, NewContext); + return; + } + break; + case Instruction::Add: + if (ConstantInt *CI0 = dyn_cast(Op0)) { + unsigned n_ci0 = VN.getOrInsertVN(Op1, Top); + ConstantRange CR = VR.range(n_ci0, Top); + if (!CR.isFullSet()) { + CR.subtract(-CI0->getValue()); + unsigned n_bo = VN.getOrInsertVN(BO, Top); + VR.applyRange(n_bo, CR, Top, this); + return; + } + } + if (ConstantInt *CI1 = dyn_cast(Op1)) { + unsigned n_ci1 = VN.getOrInsertVN(Op0, Top); + ConstantRange CR = VR.range(n_ci1, Top); + if (!CR.isFullSet()) { + CR.subtract(-CI1->getValue()); + unsigned n_bo = VN.getOrInsertVN(BO, Top); + VR.applyRange(n_bo, CR, Top, this); + return; + } + } + // fall-through + case Instruction::Xor: + if (Op0 == Zero) { + add(BO, Op1, ICmpInst::ICMP_EQ, NewContext); + return; + } else if (Op1 == Zero) { + add(BO, Op0, ICmpInst::ICMP_EQ, NewContext); + return; + } + break; + case Instruction::And: + if (Op0 == AllOnes) { + add(BO, Op1, ICmpInst::ICMP_EQ, NewContext); + return; + } else if (Op1 == AllOnes) { + add(BO, Op0, ICmpInst::ICMP_EQ, NewContext); + return; + } + if (Op0 == Zero || Op1 == Zero) { + add(BO, Zero, ICmpInst::ICMP_EQ, NewContext); + return; + } + break; + case Instruction::Mul: + if (Op0 == Zero || Op1 == Zero) { + add(BO, Zero, ICmpInst::ICMP_EQ, NewContext); + return; + } + if (Op0 == One) { + add(BO, Op1, ICmpInst::ICMP_EQ, NewContext); + return; + } else if (Op1 == One) { + add(BO, Op0, ICmpInst::ICMP_EQ, NewContext); + return; + } + break; + } - if (isGreater(V1, V2)) return false; + // "%x = add i32 %y, %z" and %x EQ %y then %z EQ 0 + // "%x = add i32 %y, %z" and %x EQ %z then %y EQ 0 + // "%x = shl i32 %y, %z" and %x EQ %y and %y NE 0 then %z EQ 0 + // "%x = udiv i32 %y, %z" and %x EQ %y and %y NE 0 then %z EQ 1 + + Value *Known = Op0, *Unknown = Op1, + *TheBO = VN.canonicalize(BO, Top); + if (Known != TheBO) std::swap(Known, Unknown); + if (Known == TheBO) { + switch (Opcode) { + default: break; + case Instruction::LShr: + case Instruction::AShr: + case Instruction::Shl: + if (!isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) break; + // otherwise, fall-through. + case Instruction::Sub: + if (Unknown == Op0) break; + // otherwise, fall-through. + case Instruction::Xor: + case Instruction::Add: + add(Unknown, Zero, ICmpInst::ICMP_EQ, NewContext); + break; + case Instruction::UDiv: + case Instruction::SDiv: + if (Unknown == Op1) break; + if (isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) + add(Unknown, One, ICmpInst::ICMP_EQ, NewContext); + break; + } + } - Node *N1 = IG.getOrInsertNode(V1), - *N2 = IG.getOrInsertNode(V2); + // TODO: "%a = add i32 %b, 1" and %b > %z then %a >= %z. + + } else if (ICmpInst *IC = dyn_cast(I)) { + // "%a = icmp ult i32 %b, %c" and %b u< %c then %a EQ true + // "%a = icmp ult i32 %b, %c" and %b u>= %c then %a EQ false + // etc. + + Value *Op0 = VN.canonicalize(IC->getOperand(0), Top); + Value *Op1 = VN.canonicalize(IC->getOperand(1), Top); + + ICmpInst::Predicate Pred = IC->getPredicate(); + if (isRelatedBy(Op0, Op1, Pred)) + add(IC, ConstantInt::getTrue(), ICmpInst::ICMP_EQ, NewContext); + else if (isRelatedBy(Op0, Op1, ICmpInst::getInversePredicate(Pred))) + add(IC, ConstantInt::getFalse(), ICmpInst::ICMP_EQ, NewContext); + + } else if (SelectInst *SI = dyn_cast(I)) { + if (I->getType()->isFPOrFPVector()) return; + + // Given: "%a = select i1 %x, i32 %b, i32 %c" + // %x EQ true then %a EQ %b + // %x EQ false then %a EQ %c + // %b EQ %c then %a EQ %b + + Value *Canonical = VN.canonicalize(SI->getCondition(), Top); + if (Canonical == ConstantInt::getTrue()) { + add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext); + } else if (Canonical == ConstantInt::getFalse()) { + add(SI, SI->getFalseValue(), ICmpInst::ICMP_EQ, NewContext); + } else if (VN.canonicalize(SI->getTrueValue(), Top) == + VN.canonicalize(SI->getFalseValue(), Top)) { + add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext); + } + } else if (CastInst *CI = dyn_cast(I)) { + const Type *DestTy = CI->getDestTy(); + if (DestTy->isFPOrFPVector()) return; - if (N1 == N2) return true; + Value *Op = VN.canonicalize(CI->getOperand(0), Top); + Instruction::CastOps Opcode = CI->getOpcode(); - IG.addLessEqual(N1, N2); + if (Constant *C = dyn_cast(Op)) { + add(CI, ConstantExpr::getCast(Opcode, C, DestTy), + ICmpInst::ICMP_EQ, NewContext); + } - addRecursive(V1); - addRecursive(V2); + uint32_t W = VR.typeToWidth(DestTy); + unsigned ci = VN.getOrInsertVN(CI, Top); + ConstantRange CR = VR.range(VN.getOrInsertVN(Op, Top), Top); - return true; + if (!CR.isFullSet()) { + switch (Opcode) { + default: break; + case Instruction::ZExt: + VR.applyRange(ci, CR.zeroExtend(W), Top, this); + break; + case Instruction::SExt: + VR.applyRange(ci, CR.signExtend(W), Top, this); + break; + case Instruction::Trunc: { + ConstantRange Result = CR.truncate(W); + if (!Result.isFullSet()) + VR.applyRange(ci, Result, Top, this); + } break; + case Instruction::BitCast: + VR.applyRange(ci, CR, Top, this); + break; + // TODO: other casts? + } + } + } else if (GetElementPtrInst *GEPI = dyn_cast(I)) { + for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(), + OE = GEPI->idx_end(); OI != OE; ++OI) { + ConstantInt *Op = dyn_cast(VN.canonicalize(*OI, Top)); + if (!Op || !Op->isZero()) return; + } + // TODO: The GEPI indices are all zero. Copy from operand to definition, + // jumping the type plane as needed. + Value *Ptr = GEPI->getPointerOperand(); + if (isRelatedBy(Ptr, Constant::getNullValue(Ptr->getType()), + ICmpInst::ICMP_NE)) { + add(GEPI, Constant::getNullValue(GEPI->getType()), ICmpInst::ICMP_NE, + NewContext); + } + } } + /// solve - process the work queue void solve() { - DOUT << "WorkList entry, size: " << WorkList.size() << "\n"; + //DOUT << "WorkList entry, size: " << WorkList.size() << "\n"; while (!WorkList.empty()) { - DOUT << "WorkList size: " << WorkList.size() << "\n"; - - Instruction *I = WorkList.front(); - WorkList.pop_front(); - - Value *Canonical = cIG.canonicalize(I); - const Type *Ty = I->getType(); - - //DOUT << "solving: " << *I << "\n"; - //DEBUG(IG.debug(*cerr.stream())); - - if (BinaryOperator *BO = dyn_cast(I)) { - Value *Op0 = cIG.canonicalize(BO->getOperand(0)), - *Op1 = cIG.canonicalize(BO->getOperand(1)); - - ConstantIntegral *CI1 = dyn_cast(Op0), - *CI2 = dyn_cast(Op1); + //DOUT << "WorkList size: " << WorkList.size() << "\n"; + + Operation &O = WorkList.front(); + TopInst = O.ContextInst; + TopBB = O.ContextBB; + Top = DTDFS->getNodeForBlock(TopBB); // XXX move this into Context + + O.LHS = VN.canonicalize(O.LHS, Top); + O.RHS = VN.canonicalize(O.RHS, Top); + + assert(O.LHS == VN.canonicalize(O.LHS, Top) && "Canonicalize isn't."); + assert(O.RHS == VN.canonicalize(O.RHS, Top) && "Canonicalize isn't."); + + DOUT << "solving " << *O.LHS << " " << O.Op << " " << *O.RHS; + if (O.ContextInst) DOUT << " context inst: " << *O.ContextInst; + else DOUT << " context block: " << O.ContextBB->getName(); + DOUT << "\n"; + + DEBUG(VN.dump()); + DEBUG(IG.dump()); + DEBUG(VR.dump()); + + // If they're both Constant, skip it. Check for contradiction and mark + // the BB as unreachable if so. + if (Constant *CI_L = dyn_cast(O.LHS)) { + if (Constant *CI_R = dyn_cast(O.RHS)) { + if (ConstantExpr::getCompare(O.Op, CI_L, CI_R) == + ConstantInt::getFalse()) + UB.mark(TopBB); + + WorkList.pop_front(); + continue; + } + } - if (CI1 && CI2) - addEqual(BO, ConstantExpr::get(BO->getOpcode(), CI1, CI2)); + if (VN.compare(O.LHS, O.RHS)) { + std::swap(O.LHS, O.RHS); + O.Op = ICmpInst::getSwappedPredicate(O.Op); + } - switch (BO->getOpcode()) { - case Instruction::SetEQ: - // "seteq int %a, %b" EQ true then %a EQ %b - // "seteq int %a, %b" EQ false then %a NE %b - if (Canonical == ConstantBool::getTrue()) - addEqual(Op0, Op1); - else if (Canonical == ConstantBool::getFalse()) - addNotEqual(Op0, Op1); + if (O.Op == ICmpInst::ICMP_EQ) { + if (!makeEqual(O.RHS, O.LHS)) + UB.mark(TopBB); + } else { + LatticeVal LV = cmpInstToLattice(O.Op); - // %a EQ %b then "seteq int %a, %b" EQ true - // %a NE %b then "seteq int %a, %b" EQ false - if (isEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getTrue()); - else if (isNotEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getFalse()); + if ((LV & EQ_BIT) && + isRelatedBy(O.LHS, O.RHS, ICmpInst::getSwappedPredicate(O.Op))) { + if (!makeEqual(O.RHS, O.LHS)) + UB.mark(TopBB); + } else { + if (isRelatedBy(O.LHS, O.RHS, ICmpInst::getInversePredicate(O.Op))){ + UB.mark(TopBB); + WorkList.pop_front(); + continue; + } - break; - case Instruction::SetNE: - // "setne int %a, %b" EQ true then %a NE %b - // "setne int %a, %b" EQ false then %a EQ %b - if (Canonical == ConstantBool::getTrue()) - addNotEqual(Op0, Op1); - else if (Canonical == ConstantBool::getFalse()) - addEqual(Op0, Op1); - - // %a EQ %b then "setne int %a, %b" EQ false - // %a NE %b then "setne int %a, %b" EQ true - if (isEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getFalse()); - else if (isNotEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getTrue()); + unsigned n1 = VN.getOrInsertVN(O.LHS, Top); + unsigned n2 = VN.getOrInsertVN(O.RHS, Top); - break; - case Instruction::SetLT: - // "setlt int %a, %b" EQ true then %a LT %b - // "setlt int %a, %b" EQ false then %b LE %a - if (Canonical == ConstantBool::getTrue()) - addLess(Op0, Op1); - else if (Canonical == ConstantBool::getFalse()) - addLessEqual(Op1, Op0); - - // %a LT %b then "setlt int %a, %b" EQ true - // %a GE %b then "setlt int %a, %b" EQ false - if (isLess(Op0, Op1)) - addEqual(BO, ConstantBool::getTrue()); - else if (isGreaterEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getFalse()); + if (n1 == n2) { + if (O.Op != ICmpInst::ICMP_UGE && O.Op != ICmpInst::ICMP_ULE && + O.Op != ICmpInst::ICMP_SGE && O.Op != ICmpInst::ICMP_SLE) + UB.mark(TopBB); - break; - case Instruction::SetLE: - // "setle int %a, %b" EQ true then %a LE %b - // "setle int %a, %b" EQ false then %b LT %a - if (Canonical == ConstantBool::getTrue()) - addLessEqual(Op0, Op1); - else if (Canonical == ConstantBool::getFalse()) - addLess(Op1, Op0); - - // %a LE %b then "setle int %a, %b" EQ true - // %a GT %b then "setle int %a, %b" EQ false - if (isLessEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getTrue()); - else if (isGreater(Op0, Op1)) - addEqual(BO, ConstantBool::getFalse()); + WorkList.pop_front(); + continue; + } - break; - case Instruction::SetGT: - // "setgt int %a, %b" EQ true then %b LT %a - // "setgt int %a, %b" EQ false then %a LE %b - if (Canonical == ConstantBool::getTrue()) - addLess(Op1, Op0); - else if (Canonical == ConstantBool::getFalse()) - addLessEqual(Op0, Op1); - - // %a GT %b then "setgt int %a, %b" EQ true - // %a LE %b then "setgt int %a, %b" EQ false - if (isGreater(Op0, Op1)) - addEqual(BO, ConstantBool::getTrue()); - else if (isLessEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getFalse()); + if (VR.isRelatedBy(n1, n2, Top, LV) || + IG.isRelatedBy(n1, n2, Top, LV)) { + WorkList.pop_front(); + continue; + } - break; - case Instruction::SetGE: - // "setge int %a, %b" EQ true then %b LE %a - // "setge int %a, %b" EQ false then %a LT %b - if (Canonical == ConstantBool::getTrue()) - addLessEqual(Op1, Op0); - else if (Canonical == ConstantBool::getFalse()) - addLess(Op0, Op1); - - // %a GE %b then "setge int %a, %b" EQ true - // %a LT %b then "setlt int %a, %b" EQ false - if (isGreaterEqual(Op0, Op1)) - addEqual(BO, ConstantBool::getTrue()); - else if (isLess(Op0, Op1)) - addEqual(BO, ConstantBool::getFalse()); + VR.addInequality(n1, n2, Top, LV, this); + if ((!isa(O.RHS) && !isa(O.LHS)) || + LV == NE) + IG.addInequality(n1, n2, Top, LV); - break; - case Instruction::And: { - // "and int %a, %b" EQ -1 then %a EQ -1 and %b EQ -1 - // "and bool %a, %b" EQ true then %a EQ true and %b EQ true - ConstantIntegral *CI = ConstantIntegral::getAllOnesValue(Ty); - if (Canonical == CI) { - addEqual(CI, Op0); - addEqual(CI, Op1); - } - } break; - case Instruction::Or: { - // "or int %a, %b" EQ 0 then %a EQ 0 and %b EQ 0 - // "or bool %a, %b" EQ false then %a EQ false and %b EQ false - Constant *Zero = Constant::getNullValue(Ty); - if (Canonical == Zero) { - addEqual(Zero, Op0); - addEqual(Zero, Op1); - } - } break; - case Instruction::Xor: { - // "xor bool true, %a" EQ true then %a EQ false - // "xor bool true, %a" EQ false then %a EQ true - // "xor bool false, %a" EQ true then %a EQ true - // "xor bool false, %a" EQ false then %a EQ false - // "xor int %c, %a" EQ %c then %a EQ 0 - // "xor int %c, %a" NE %c then %a NE 0 - // 1. Repeat all of the above, with order of operands reversed. - Value *LHS = Op0, *RHS = Op1; - if (!isa(LHS)) std::swap(LHS, RHS); - - if (ConstantBool *CB = dyn_cast(Canonical)) { - if (ConstantBool *A = dyn_cast(LHS)) - addEqual(RHS, ConstantBool::get(A->getValue() ^ - CB->getValue())); + if (Instruction *I1 = dyn_cast(O.LHS)) { + if (aboveOrBelow(I1)) + defToOps(I1); + } + if (isa(O.LHS) || isa(O.LHS)) { + for (Value::use_iterator UI = O.LHS->use_begin(), + UE = O.LHS->use_end(); UI != UE;) { + Use &TheUse = UI.getUse(); + ++UI; + Instruction *I = cast(TheUse.getUser()); + if (aboveOrBelow(I)) + opsToDef(I); } - if (Canonical == LHS) { - if (isa(Canonical)) - addEqual(RHS, Constant::getNullValue(Ty)); - } else if (isNotEqual(LHS, Canonical)) { - addNotEqual(RHS, Constant::getNullValue(Ty)); + } + if (Instruction *I2 = dyn_cast(O.RHS)) { + if (aboveOrBelow(I2)) + defToOps(I2); + } + if (isa(O.RHS) || isa(O.RHS)) { + for (Value::use_iterator UI = O.RHS->use_begin(), + UE = O.RHS->use_end(); UI != UE;) { + Use &TheUse = UI.getUse(); + ++UI; + Instruction *I = cast(TheUse.getUser()); + if (aboveOrBelow(I)) + opsToDef(I); } - } break; - default: - break; - } - - // "%x = add int %y, %z" and %x EQ %y then %z EQ 0 - // "%x = mul int %y, %z" and %x EQ %y then %z EQ 1 - // 1. Repeat all of the above, with order of operands reversed. - // "%x = fdiv float %y, %z" and %x EQ %y then %z EQ 1 - Value *Known = Op0, *Unknown = Op1; - if (Known != BO) std::swap(Known, Unknown); - if (Known == BO) { - switch (BO->getOpcode()) { - default: break; - case Instruction::Xor: - case Instruction::Or: - case Instruction::Add: - case Instruction::Sub: - if (!Ty->isFloatingPoint()) - addEqual(Unknown, Constant::getNullValue(Ty)); - break; - case Instruction::UDiv: - case Instruction::SDiv: - case Instruction::FDiv: - if (Unknown == Op0) break; // otherwise, fallthrough - case Instruction::And: - case Instruction::Mul: - Constant *One = NULL; - if (isa(Unknown)) - One = ConstantInt::get(Ty, 1); - else if (isa(Unknown)) - One = ConstantFP::get(Ty, 1); - else if (isa(Unknown)) - One = ConstantBool::getTrue(); - - if (One) addEqual(Unknown, One); - break; } } - } else if (SelectInst *SI = dyn_cast(I)) { - // Given: "%a = select bool %x, int %b, int %c" - // %a EQ %b then %x EQ true - // %a EQ %c then %x EQ false - if (isEqual(I, SI->getTrueValue()) || - isNotEqual(I, SI->getFalseValue())) - addEqual(SI->getCondition(), ConstantBool::getTrue()); - else if (isEqual(I, SI->getFalseValue()) || - isNotEqual(I, SI->getTrueValue())) - addEqual(SI->getCondition(), ConstantBool::getFalse()); - - // %x EQ true then %a EQ %b - // %x EQ false then %a NE %b - if (isEqual(SI->getCondition(), ConstantBool::getTrue())) - addEqual(SI, SI->getTrueValue()); - else if (isEqual(SI->getCondition(), ConstantBool::getFalse())) - addEqual(SI, SI->getFalseValue()); } + WorkList.pop_front(); } } }; + void ValueRanges::addToWorklist(Value *V, Constant *C, + ICmpInst::Predicate Pred, VRPSolver *VRP) { + VRP->add(V, C, Pred, VRP->TopInst); + } + + void ValueRanges::markBlock(VRPSolver *VRP) { + VRP->UB.mark(VRP->TopBB); + } + /// PredicateSimplifier - This class is a simplifier that replaces /// one equivalent variable with another. It also tracks what /// can't be equal and will solve setcc instructions when possible. /// @brief Root of the predicate simplifier optimization. class VISIBILITY_HIDDEN PredicateSimplifier : public FunctionPass { - DominatorTree *DT; - ETForest *Forest; + DomTreeDFS *DTDFS; bool modified; + ValueNumbering *VN; + InequalityGraph *IG; + UnreachableBlocks UB; + ValueRanges *VR; - class State { - public: - BasicBlock *ToVisit; - InequalityGraph *IG; - - State(BasicBlock *BB, InequalityGraph *IG) : ToVisit(BB), IG(IG) {} - }; - - std::vector WorkList; + std::vector WorkList; public: + static char ID; // Pass identification, replacement for typeid + PredicateSimplifier() : FunctionPass(&ID) {} + bool runOnFunction(Function &F); virtual void getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequiredID(BreakCriticalEdgesID); AU.addRequired(); - AU.addRequired(); - AU.setPreservesCFG(); - AU.addPreservedID(BreakCriticalEdgesID); + AU.addRequired(); + AU.addPreserved(); } private: - /// Forwards - Adds new properties into PropertySet and uses them to + /// Forwards - Adds new properties to VRPSolver and uses them to /// simplify instructions. Because new properties sometimes apply to /// a transition from one BasicBlock to another, this will use the /// PredicateSimplifier::proceedToSuccessor(s) interface to enter the - /// basic block with the new PropertySet. + /// basic block. /// @brief Performs abstract execution of the program. class VISIBILITY_HIDDEN Forwards : public InstVisitor { friend class InstVisitor; PredicateSimplifier *PS; + DomTreeDFS::Node *DTNode; public: + ValueNumbering &VN; InequalityGraph &IG; + UnreachableBlocks &UB; + ValueRanges &VR; - Forwards(PredicateSimplifier *PS, InequalityGraph &IG) - : PS(PS), IG(IG) {} + Forwards(PredicateSimplifier *PS, DomTreeDFS::Node *DTNode) + : PS(PS), DTNode(DTNode), VN(*PS->VN), IG(*PS->IG), UB(PS->UB), + VR(*PS->VR) {} void visitTerminatorInst(TerminatorInst &TI); void visitBranchInst(BranchInst &BI); @@ -1107,53 +2352,66 @@ namespace { void visitLoadInst(LoadInst &LI); void visitStoreInst(StoreInst &SI); + void visitSExtInst(SExtInst &SI); + void visitZExtInst(ZExtInst &ZI); + void visitBinaryOperator(BinaryOperator &BO); + void visitICmpInst(ICmpInst &IC); }; - + // Used by terminator instructions to proceed from the current basic // block to the next. Verifies that "current" dominates "next", // then calls visitBasicBlock. - void proceedToSuccessors(const InequalityGraph &IG, BasicBlock *BBCurrent) { - DominatorTree::Node *Current = DT->getNode(BBCurrent); - for (DominatorTree::Node::iterator I = Current->begin(), + void proceedToSuccessors(DomTreeDFS::Node *Current) { + for (DomTreeDFS::Node::iterator I = Current->begin(), E = Current->end(); I != E; ++I) { - //visitBasicBlock((*I)->getBlock(), IG); - WorkList.push_back(State((*I)->getBlock(), new InequalityGraph(IG))); + WorkList.push_back(*I); } } - void proceedToSuccessor(InequalityGraph *NextIG, BasicBlock *Next) { - //visitBasicBlock(Next, NextIG); - WorkList.push_back(State(Next, NextIG)); + void proceedToSuccessor(DomTreeDFS::Node *Next) { + WorkList.push_back(Next); } // Visits each instruction in the basic block. - void visitBasicBlock(BasicBlock *BB, InequalityGraph &IG) { - DOUT << "Entering Basic Block: " << BB->getName() << "\n"; + void visitBasicBlock(DomTreeDFS::Node *Node) { + BasicBlock *BB = Node->getBlock(); + DOUT << "Entering Basic Block: " << BB->getName() + << " (" << Node->getDFSNumIn() << ")\n"; for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { - visitInstruction(I++, IG); + visitInstruction(I++, Node); } } - // Tries to simplify each Instruction and add new properties to - // the PropertySet. - void visitInstruction(Instruction *I, InequalityGraph &IG) { + // Tries to simplify each Instruction and add new properties. + void visitInstruction(Instruction *I, DomTreeDFS::Node *DT) { DOUT << "Considering instruction " << *I << "\n"; - DEBUG(IG.debug(*cerr.stream())); + DEBUG(VN->dump()); + DEBUG(IG->dump()); + DEBUG(VR->dump()); - // Sometimes instructions are made dead due to earlier analysis. + // Sometimes instructions are killed in earlier analysis. if (isInstructionTriviallyDead(I)) { + ++NumSimple; + modified = true; + if (unsigned n = VN->valueNumber(I, DTDFS->getRootNode())) + if (VN->value(n) == I) IG->remove(n); + VN->remove(I); I->eraseFromParent(); return; } +#ifndef NDEBUG // Try to replace the whole instruction. - Value *V = IG.canonicalize(I); + Value *V = VN->canonicalize(I, DT); + assert(V == I && "Late instruction canonicalization."); if (V != I) { modified = true; ++NumInstruction; DOUT << "Removing " << *I << ", replacing with " << *V << "\n"; - IG.remove(I); + if (unsigned n = VN->valueNumber(I, DTDFS->getRootNode())) + if (VN->value(n) == I) IG->remove(n); + VN->remove(I); I->replaceAllUsesWith(V); I->eraseFromParent(); return; @@ -1162,7 +2420,8 @@ namespace { // Try to substitute operands. for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { Value *Oper = I->getOperand(i); - Value *V = IG.canonicalize(Oper); + Value *V = VN->canonicalize(Oper, DT); + assert(V == Oper && "Late operand canonicalization."); if (V != Oper) { modified = true; ++NumVarsReplaced; @@ -1171,76 +2430,90 @@ namespace { DOUT << " into " << *I; } } +#endif - //DOUT << "push (%" << I->getParent()->getName() << ")\n"; - Forwards visit(this, IG); + std::string name = I->getParent()->getName(); + DOUT << "push (%" << name << ")\n"; + Forwards visit(this, DT); visit.visit(*I); - //DOUT << "pop (%" << I->getParent()->getName() << ")\n"; + DOUT << "pop (%" << name << ")\n"; } }; bool PredicateSimplifier::runOnFunction(Function &F) { - DT = &getAnalysis(); - Forest = &getAnalysis(); + DominatorTree *DT = &getAnalysis(); + DTDFS = new DomTreeDFS(DT); + TargetData *TD = &getAnalysis(); DOUT << "Entering Function: " << F.getName() << "\n"; modified = false; - WorkList.push_back(State(DT->getRoot(), new InequalityGraph())); + DomTreeDFS::Node *Root = DTDFS->getRootNode(); + VN = new ValueNumbering(DTDFS); + IG = new InequalityGraph(*VN, Root); + VR = new ValueRanges(*VN, TD); + WorkList.push_back(Root); do { - State S = WorkList.back(); + DomTreeDFS::Node *DTNode = WorkList.back(); WorkList.pop_back(); - visitBasicBlock(S.ToVisit, *S.IG); - delete S.IG; + if (!UB.isDead(DTNode->getBlock())) visitBasicBlock(DTNode); } while (!WorkList.empty()); - //DEBUG(F.viewCFG()); + delete DTDFS; + delete VR; + delete IG; + delete VN; + + modified |= UB.kill(); return modified; } void PredicateSimplifier::Forwards::visitTerminatorInst(TerminatorInst &TI) { - PS->proceedToSuccessors(IG, TI.getParent()); + PS->proceedToSuccessors(DTNode); } void PredicateSimplifier::Forwards::visitBranchInst(BranchInst &BI) { - BasicBlock *BB = BI.getParent(); - if (BI.isUnconditional()) { - PS->proceedToSuccessors(IG, BB); + PS->proceedToSuccessors(DTNode); return; } Value *Condition = BI.getCondition(); - BasicBlock *TrueDest = BI.getSuccessor(0), - *FalseDest = BI.getSuccessor(1); + BasicBlock *TrueDest = BI.getSuccessor(0); + BasicBlock *FalseDest = BI.getSuccessor(1); - if (isa(Condition) || TrueDest == FalseDest) { - PS->proceedToSuccessors(IG, BB); + if (isa(Condition) || TrueDest == FalseDest) { + PS->proceedToSuccessors(DTNode); return; } - DominatorTree::Node *Node = PS->DT->getNode(BB); - for (DominatorTree::Node::iterator I = Node->begin(), E = Node->end(); + for (DomTreeDFS::Node::iterator I = DTNode->begin(), E = DTNode->end(); I != E; ++I) { BasicBlock *Dest = (*I)->getBlock(); - InequalityGraph *DestProperties = new InequalityGraph(IG); - VRPSolver Solver(*DestProperties, PS->Forest, Dest); + DOUT << "Branch thinking about %" << Dest->getName() + << "(" << PS->DTDFS->getNodeForBlock(Dest)->getDFSNumIn() << ")\n"; if (Dest == TrueDest) { - DOUT << "(" << BB->getName() << ") true set:\n"; - if (!Solver.addEqual(ConstantBool::getTrue(), Condition)) continue; - Solver.solve(); - DEBUG(DestProperties->debug(*cerr.stream())); + DOUT << "(" << DTNode->getBlock()->getName() << ") true set:\n"; + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, Dest); + VRP.add(ConstantInt::getTrue(), Condition, ICmpInst::ICMP_EQ); + VRP.solve(); + DEBUG(VN.dump()); + DEBUG(IG.dump()); + DEBUG(VR.dump()); } else if (Dest == FalseDest) { - DOUT << "(" << BB->getName() << ") false set:\n"; - if (!Solver.addEqual(ConstantBool::getFalse(), Condition)) continue; - Solver.solve(); - DEBUG(DestProperties->debug(*cerr.stream())); + DOUT << "(" << DTNode->getBlock()->getName() << ") false set:\n"; + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, Dest); + VRP.add(ConstantInt::getFalse(), Condition, ICmpInst::ICMP_EQ); + VRP.solve(); + DEBUG(VN.dump()); + DEBUG(IG.dump()); + DEBUG(VR.dump()); } - PS->proceedToSuccessor(DestProperties, Dest); + PS->proceedToSuccessor(*I); } } @@ -1249,41 +2522,40 @@ namespace { // Set the EQProperty in each of the cases BBs, and the NEProperties // in the default BB. - // InequalityGraph DefaultProperties(IG); - DominatorTree::Node *Node = PS->DT->getNode(SI.getParent()); - for (DominatorTree::Node::iterator I = Node->begin(), E = Node->end(); + for (DomTreeDFS::Node::iterator I = DTNode->begin(), E = DTNode->end(); I != E; ++I) { BasicBlock *BB = (*I)->getBlock(); + DOUT << "Switch thinking about BB %" << BB->getName() + << "(" << PS->DTDFS->getNodeForBlock(BB)->getDFSNumIn() << ")\n"; - InequalityGraph *BBProperties = new InequalityGraph(IG); - VRPSolver Solver(*BBProperties, PS->Forest, BB); + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, BB); if (BB == SI.getDefaultDest()) { for (unsigned i = 1, e = SI.getNumCases(); i < e; ++i) if (SI.getSuccessor(i) != BB) - if (!Solver.addNotEqual(Condition, SI.getCaseValue(i))) continue; - Solver.solve(); + VRP.add(Condition, SI.getCaseValue(i), ICmpInst::ICMP_NE); + VRP.solve(); } else if (ConstantInt *CI = SI.findCaseDest(BB)) { - if (!Solver.addEqual(Condition, CI)) continue; - Solver.solve(); + VRP.add(Condition, CI, ICmpInst::ICMP_EQ); + VRP.solve(); } - PS->proceedToSuccessor(BBProperties, BB); + PS->proceedToSuccessor(*I); } } void PredicateSimplifier::Forwards::visitAllocaInst(AllocaInst &AI) { - VRPSolver VRP(IG, PS->Forest, AI.getParent()); - VRP.addNotEqual(Constant::getNullValue(AI.getType()), &AI); + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &AI); + VRP.add(Constant::getNullValue(AI.getType()), &AI, ICmpInst::ICMP_NE); VRP.solve(); } void PredicateSimplifier::Forwards::visitLoadInst(LoadInst &LI) { Value *Ptr = LI.getPointerOperand(); - // avoid "load uint* null" -> null NE null. + // avoid "load i8* null" -> null NE null. if (isa(Ptr)) return; - VRPSolver VRP(IG, PS->Forest, LI.getParent()); - VRP.addNotEqual(Constant::getNullValue(Ptr->getType()), Ptr); + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &LI); + VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE); VRP.solve(); } @@ -1291,8 +2563,28 @@ namespace { Value *Ptr = SI.getPointerOperand(); if (isa(Ptr)) return; - VRPSolver VRP(IG, PS->Forest, SI.getParent()); - VRP.addNotEqual(Constant::getNullValue(Ptr->getType()), Ptr); + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &SI); + VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE); + VRP.solve(); + } + + void PredicateSimplifier::Forwards::visitSExtInst(SExtInst &SI) { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &SI); + uint32_t SrcBitWidth = cast(SI.getSrcTy())->getBitWidth(); + uint32_t DstBitWidth = cast(SI.getDestTy())->getBitWidth(); + APInt Min(APInt::getHighBitsSet(DstBitWidth, DstBitWidth-SrcBitWidth+1)); + APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth-1)); + VRP.add(ConstantInt::get(Min), &SI, ICmpInst::ICMP_SLE); + VRP.add(ConstantInt::get(Max), &SI, ICmpInst::ICMP_SGE); + VRP.solve(); + } + + void PredicateSimplifier::Forwards::visitZExtInst(ZExtInst &ZI) { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &ZI); + uint32_t SrcBitWidth = cast(ZI.getSrcTy())->getBitWidth(); + uint32_t DstBitWidth = cast(ZI.getDestTy())->getBitWidth(); + APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth)); + VRP.add(ConstantInt::get(Max), &ZI, ICmpInst::ICMP_UGE); VRP.solve(); } @@ -1300,28 +2592,130 @@ namespace { Instruction::BinaryOps ops = BO.getOpcode(); switch (ops) { - case Instruction::URem: - case Instruction::SRem: - case Instruction::FRem: - case Instruction::UDiv: - case Instruction::SDiv: - case Instruction::FDiv: { - Value *Divisor = BO.getOperand(1); - VRPSolver VRP(IG, PS->Forest, BO.getParent()); - VRP.addNotEqual(Constant::getNullValue(Divisor->getType()), Divisor); - VRP.solve(); - break; - } - default: - break; + default: break; + case Instruction::URem: + case Instruction::SRem: + case Instruction::UDiv: + case Instruction::SDiv: { + Value *Divisor = BO.getOperand(1); + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO); + VRP.add(Constant::getNullValue(Divisor->getType()), Divisor, + ICmpInst::ICMP_NE); + VRP.solve(); + break; + } + } + + switch (ops) { + default: break; + case Instruction::Shl: { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO); + VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE); + VRP.solve(); + } break; + case Instruction::AShr: { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO); + VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_SLE); + VRP.solve(); + } break; + case Instruction::LShr: + case Instruction::UDiv: { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO); + VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE); + VRP.solve(); + } break; + case Instruction::URem: { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO); + VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE); + VRP.solve(); + } break; + case Instruction::And: { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO); + VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE); + VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE); + VRP.solve(); + } break; + case Instruction::Or: { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO); + VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE); + VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_UGE); + VRP.solve(); + } break; } } + void PredicateSimplifier::Forwards::visitICmpInst(ICmpInst &IC) { + // If possible, squeeze the ICmp predicate into something simpler. + // Eg., if x = [0, 4) and we're being asked icmp uge %x, 3 then change + // the predicate to eq. + + // XXX: once we do full PHI handling, modifying the instruction in the + // Forwards visitor will cause missed optimizations. + + ICmpInst::Predicate Pred = IC.getPredicate(); + + switch (Pred) { + default: break; + case ICmpInst::ICMP_ULE: Pred = ICmpInst::ICMP_ULT; break; + case ICmpInst::ICMP_UGE: Pred = ICmpInst::ICMP_UGT; break; + case ICmpInst::ICMP_SLE: Pred = ICmpInst::ICMP_SLT; break; + case ICmpInst::ICMP_SGE: Pred = ICmpInst::ICMP_SGT; break; + } + if (Pred != IC.getPredicate()) { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &IC); + if (VRP.isRelatedBy(IC.getOperand(1), IC.getOperand(0), + ICmpInst::ICMP_NE)) { + ++NumSnuggle; + PS->modified = true; + IC.setPredicate(Pred); + } + } + + Pred = IC.getPredicate(); + + if (ConstantInt *Op1 = dyn_cast(IC.getOperand(1))) { + ConstantInt *NextVal = 0; + switch (Pred) { + default: break; + case ICmpInst::ICMP_SLT: + case ICmpInst::ICMP_ULT: + if (Op1->getValue() != 0) + NextVal = ConstantInt::get(Op1->getValue()-1); + break; + case ICmpInst::ICMP_SGT: + case ICmpInst::ICMP_UGT: + if (!Op1->getValue().isAllOnesValue()) + NextVal = ConstantInt::get(Op1->getValue()+1); + break; + } - RegisterPass X("predsimplify", - "Predicate Simplifier"); + if (NextVal) { + VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &IC); + if (VRP.isRelatedBy(IC.getOperand(0), NextVal, + ICmpInst::getInversePredicate(Pred))) { + ICmpInst *NewIC = new ICmpInst(&IC, ICmpInst::ICMP_EQ, + IC.getOperand(0), NextVal, ""); + NewIC->takeName(&IC); + IC.replaceAllUsesWith(NewIC); + + // XXX: prove this isn't necessary + if (unsigned n = VN.valueNumber(&IC, PS->DTDFS->getRootNode())) + if (VN.value(n) == &IC) IG.remove(n); + VN.remove(&IC); + + IC.eraseFromParent(); + ++NumSnuggle; + PS->modified = true; + } + } + } + } } +char PredicateSimplifier::ID = 0; +static RegisterPass +X("predsimplify", "Predicate Simplifier"); + FunctionPass *llvm::createPredicateSimplifierPass() { return new PredicateSimplifier(); }