1 //===-- PredicateSimplifier.cpp - Path Sensitive Simplifier ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Path-sensitive optimizer. In a branch where x == y, replace uses of
11 // x with y. Permits further optimization, such as the elimination of
12 // the unreachable call:
14 // void test(int *p, int *q)
20 // foo(); // unreachable
23 //===----------------------------------------------------------------------===//
25 // The InequalityGraph focusses on four properties; equals, not equals,
26 // less-than and less-than-or-equals-to. The greater-than forms are also held
27 // just to allow walking from a lesser node to a greater one. These properties
28 // are stored in a lattice; LE can become LT or EQ, NE can become LT or GT.
30 // These relationships define a graph between values of the same type. Each
31 // Value is stored in a map table that retrieves the associated Node. This
32 // is how EQ relationships are stored; the map contains pointers from equal
33 // Value to the same node. The node contains a most canonical Value* form
34 // and the list of known relationships with other nodes.
36 // If two nodes are known to be inequal, then they will contain pointers to
37 // each other with an "NE" relationship. If node getNode(%x) is less than
38 // getNode(%y), then the %x node will contain <%y, GT> and %y will contain
39 // <%x, LT>. This allows us to tie nodes together into a graph like this:
43 // with four nodes representing the properties. The InequalityGraph provides
44 // querying with "isRelatedBy" and mutators "addEquality" and "addInequality".
45 // To find a relationship, we start with one of the nodes any binary search
46 // through its list to find where the relationships with the second node start.
47 // Then we iterate through those to find the first relationship that dominates
50 // To create these properties, we wait until a branch or switch instruction
51 // implies that a particular value is true (or false). The VRPSolver is
52 // responsible for analyzing the variable and seeing what new inferences
53 // can be made from each property. For example:
55 // %P = icmp ne i32* %ptr, null
57 // br i1 %a label %cond_true, label %cond_false
59 // For the true branch, the VRPSolver will start with %a EQ true and look at
60 // the definition of %a and find that it can infer that %P and %Q are both
61 // true. From %P being true, it can infer that %ptr NE null. For the false
62 // branch it can't infer anything from the "and" instruction.
64 // Besides branches, we can also infer properties from instruction that may
65 // have undefined behaviour in certain cases. For example, the dividend of
66 // a division may never be zero. After the division instruction, we may assume
67 // that the dividend is not equal to zero.
69 //===----------------------------------------------------------------------===//
71 // The ValueRanges class stores the known integer bounds of a Value. When we
72 // encounter i8 %a u< %b, the ValueRanges stores that %a = [1, 255] and
75 // It never stores an empty range, because that means that the code is
76 // unreachable. It never stores a single-element range since that's an equality
77 // relationship and better stored in the InequalityGraph, nor an empty range
78 // since that is better stored in UnreachableBlocks.
80 //===----------------------------------------------------------------------===//
82 #define DEBUG_TYPE "predsimplify"
83 #include "llvm/Transforms/Scalar.h"
84 #include "llvm/Constants.h"
85 #include "llvm/DerivedTypes.h"
86 #include "llvm/Instructions.h"
87 #include "llvm/Pass.h"
88 #include "llvm/ADT/DepthFirstIterator.h"
89 #include "llvm/ADT/SetOperations.h"
90 #include "llvm/ADT/SetVector.h"
91 #include "llvm/ADT/Statistic.h"
92 #include "llvm/ADT/STLExtras.h"
93 #include "llvm/Analysis/Dominators.h"
94 #include "llvm/Assembly/Writer.h"
95 #include "llvm/Support/CFG.h"
96 #include "llvm/Support/Compiler.h"
97 #include "llvm/Support/ConstantRange.h"
98 #include "llvm/Support/Debug.h"
99 #include "llvm/Support/InstVisitor.h"
100 #include "llvm/Target/TargetData.h"
101 #include "llvm/Transforms/Utils/Local.h"
105 using namespace llvm;
107 STATISTIC(NumVarsReplaced, "Number of argument substitutions");
108 STATISTIC(NumInstruction , "Number of instructions removed");
109 STATISTIC(NumSimple , "Number of simple replacements");
110 STATISTIC(NumBlocks , "Number of blocks marked unreachable");
111 STATISTIC(NumSnuggle , "Number of comparisons snuggled");
117 friend class DomTreeDFS;
119 typedef std::vector<Node *>::iterator iterator;
120 typedef std::vector<Node *>::const_iterator const_iterator;
122 unsigned getDFSNumIn() const { return DFSin; }
123 unsigned getDFSNumOut() const { return DFSout; }
125 BasicBlock *getBlock() const { return BB; }
127 iterator begin() { return Children.begin(); }
128 iterator end() { return Children.end(); }
130 const_iterator begin() const { return Children.begin(); }
131 const_iterator end() const { return Children.end(); }
133 bool dominates(const Node *N) const {
134 return DFSin <= N->DFSin && DFSout >= N->DFSout;
137 bool DominatedBy(const Node *N) const {
138 return N->dominates(this);
141 /// Sorts by the number of descendants. With this, you can iterate
142 /// through a sorted list and the first matching entry is the most
143 /// specific match for your basic block. The order provided is stable;
144 /// DomTreeDFS::Nodes with the same number of descendants are sorted by
146 bool operator<(const Node &N) const {
147 unsigned spread = DFSout - DFSin;
148 unsigned N_spread = N.DFSout - N.DFSin;
149 if (spread == N_spread) return DFSin < N.DFSin;
150 return spread < N_spread;
152 bool operator>(const Node &N) const { return N < *this; }
155 unsigned DFSin, DFSout;
158 std::vector<Node *> Children;
161 // XXX: this may be slow. Instead of using "new" for each node, consider
162 // putting them in a vector to keep them contiguous.
163 explicit DomTreeDFS(DominatorTree *DT) {
164 std::stack<std::pair<Node *, DomTreeNode *> > S;
167 Entry->BB = DT->getRootNode()->getBlock();
168 S.push(std::make_pair(Entry, DT->getRootNode()));
170 NodeMap[Entry->BB] = Entry;
173 std::pair<Node *, DomTreeNode *> &Pair = S.top();
174 Node *N = Pair.first;
175 DomTreeNode *DTNode = Pair.second;
178 for (DomTreeNode::iterator I = DTNode->begin(), E = DTNode->end();
180 Node *NewNode = new Node;
181 NewNode->BB = (*I)->getBlock();
182 N->Children.push_back(NewNode);
183 S.push(std::make_pair(NewNode, *I));
185 NodeMap[NewNode->BB] = NewNode;
200 std::stack<Node *> S;
204 Node *N = S.top(); S.pop();
206 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I)
213 /// getRootNode - This returns the entry node for the CFG of the function.
214 Node *getRootNode() const { return Entry; }
216 /// getNodeForBlock - return the node for the specified basic block.
217 Node *getNodeForBlock(BasicBlock *BB) const {
218 if (!NodeMap.count(BB)) return 0;
219 return const_cast<DomTreeDFS*>(this)->NodeMap[BB];
222 /// dominates - returns true if the basic block for I1 dominates that of
223 /// the basic block for I2. If the instructions belong to the same basic
224 /// block, the instruction first instruction sequentially in the block is
225 /// considered dominating.
226 bool dominates(Instruction *I1, Instruction *I2) {
227 BasicBlock *BB1 = I1->getParent(),
228 *BB2 = I2->getParent();
230 if (isa<TerminatorInst>(I1)) return false;
231 if (isa<TerminatorInst>(I2)) return true;
232 if ( isa<PHINode>(I1) && !isa<PHINode>(I2)) return true;
233 if (!isa<PHINode>(I1) && isa<PHINode>(I2)) return false;
235 for (BasicBlock::const_iterator I = BB2->begin(), E = BB2->end();
237 if (&*I == I1) return true;
238 else if (&*I == I2) return false;
240 assert(!"Instructions not found in parent BasicBlock?");
242 Node *Node1 = getNodeForBlock(BB1),
243 *Node2 = getNodeForBlock(BB2);
244 return Node1 && Node2 && Node1->dominates(Node2);
246 return false; // Not reached
250 /// renumber - calculates the depth first search numberings and applies
251 /// them onto the nodes.
253 std::stack<std::pair<Node *, Node::iterator> > S;
257 S.push(std::make_pair(Entry, Entry->begin()));
260 std::pair<Node *, Node::iterator> &Pair = S.top();
261 Node *N = Pair.first;
262 Node::iterator &I = Pair.second;
270 S.push(std::make_pair(Next, Next->begin()));
276 virtual void dump() const {
277 dump(*cerr.stream());
280 void dump(std::ostream &os) const {
281 os << "Predicate simplifier DomTreeDFS: \n";
286 void dump(Node *N, int depth, std::ostream &os) const {
288 for (int i = 0; i < depth; ++i) { os << " "; }
289 os << "[" << depth << "] ";
291 os << N->getBlock()->getName() << " (" << N->getDFSNumIn()
292 << ", " << N->getDFSNumOut() << ")\n";
294 for (Node::iterator I = N->begin(), E = N->end(); I != E; ++I)
300 std::map<BasicBlock *, Node *> NodeMap;
303 // SLT SGT ULT UGT EQ
304 // 0 1 0 1 0 -- GT 10
305 // 0 1 0 1 1 -- GE 11
306 // 0 1 1 0 0 -- SGTULT 12
307 // 0 1 1 0 1 -- SGEULE 13
308 // 0 1 1 1 0 -- SGT 14
309 // 0 1 1 1 1 -- SGE 15
310 // 1 0 0 1 0 -- SLTUGT 18
311 // 1 0 0 1 1 -- SLEUGE 19
312 // 1 0 1 0 0 -- LT 20
313 // 1 0 1 0 1 -- LE 21
314 // 1 0 1 1 0 -- SLT 22
315 // 1 0 1 1 1 -- SLE 23
316 // 1 1 0 1 0 -- UGT 26
317 // 1 1 0 1 1 -- UGE 27
318 // 1 1 1 0 0 -- ULT 28
319 // 1 1 1 0 1 -- ULE 29
320 // 1 1 1 1 0 -- NE 30
322 EQ_BIT = 1, UGT_BIT = 2, ULT_BIT = 4, SGT_BIT = 8, SLT_BIT = 16
325 GT = SGT_BIT | UGT_BIT,
327 LT = SLT_BIT | ULT_BIT,
329 NE = SLT_BIT | SGT_BIT | ULT_BIT | UGT_BIT,
330 SGTULT = SGT_BIT | ULT_BIT,
331 SGEULE = SGTULT | EQ_BIT,
332 SLTUGT = SLT_BIT | UGT_BIT,
333 SLEUGE = SLTUGT | EQ_BIT,
334 ULT = SLT_BIT | SGT_BIT | ULT_BIT,
335 UGT = SLT_BIT | SGT_BIT | UGT_BIT,
336 SLT = SLT_BIT | ULT_BIT | UGT_BIT,
337 SGT = SGT_BIT | ULT_BIT | UGT_BIT,
344 /// validPredicate - determines whether a given value is actually a lattice
345 /// value. Only used in assertions or debugging.
346 static bool validPredicate(LatticeVal LV) {
348 case GT: case GE: case LT: case LE: case NE:
349 case SGTULT: case SGT: case SGEULE:
350 case SLTUGT: case SLT: case SLEUGE:
352 case SLE: case SGE: case ULE: case UGE:
359 /// reversePredicate - reverse the direction of the inequality
360 static LatticeVal reversePredicate(LatticeVal LV) {
361 unsigned reverse = LV ^ (SLT_BIT|SGT_BIT|ULT_BIT|UGT_BIT); //preserve EQ_BIT
363 if ((reverse & (SLT_BIT|SGT_BIT)) == 0)
364 reverse |= (SLT_BIT|SGT_BIT);
366 if ((reverse & (ULT_BIT|UGT_BIT)) == 0)
367 reverse |= (ULT_BIT|UGT_BIT);
369 LatticeVal Rev = static_cast<LatticeVal>(reverse);
370 assert(validPredicate(Rev) && "Failed reversing predicate.");
374 /// ValueNumbering stores the scope-specific value numbers for a given Value.
375 class VISIBILITY_HIDDEN ValueNumbering {
377 /// VNPair is a tuple of {Value, index number, DomTreeDFS::Node}. It
378 /// includes the comparison operators necessary to allow you to store it
379 /// in a sorted vector.
380 class VISIBILITY_HIDDEN VNPair {
384 DomTreeDFS::Node *Subtree;
386 VNPair(Value *V, unsigned index, DomTreeDFS::Node *Subtree)
387 : V(V), index(index), Subtree(Subtree) {}
389 bool operator==(const VNPair &RHS) const {
390 return V == RHS.V && Subtree == RHS.Subtree;
393 bool operator<(const VNPair &RHS) const {
394 if (V != RHS.V) return V < RHS.V;
395 return *Subtree < *RHS.Subtree;
398 bool operator<(Value *RHS) const {
402 bool operator>(Value *RHS) const {
406 friend bool operator<(Value *RHS, const VNPair &pair) {
407 return pair.operator>(RHS);
411 typedef std::vector<VNPair> VNMapType;
414 /// The canonical choice for value number at index.
415 std::vector<Value *> Values;
421 virtual ~ValueNumbering() {}
422 virtual void dump() {
423 dump(*cerr.stream());
426 void dump(std::ostream &os) {
427 for (unsigned i = 1; i <= Values.size(); ++i) {
429 WriteAsOperand(os, Values[i-1]);
431 for (unsigned j = 0; j < VNMap.size(); ++j) {
432 if (VNMap[j].index == i) {
433 WriteAsOperand(os, VNMap[j].V);
434 os << " (" << VNMap[j].Subtree->getDFSNumIn() << ") ";
442 /// compare - returns true if V1 is a better canonical value than V2.
443 bool compare(Value *V1, Value *V2) const {
444 if (isa<Constant>(V1))
445 return !isa<Constant>(V2);
446 else if (isa<Constant>(V2))
448 else if (isa<Argument>(V1))
449 return !isa<Argument>(V2);
450 else if (isa<Argument>(V2))
453 Instruction *I1 = dyn_cast<Instruction>(V1);
454 Instruction *I2 = dyn_cast<Instruction>(V2);
457 return V1->getNumUses() < V2->getNumUses();
459 return DTDFS->dominates(I1, I2);
462 ValueNumbering(DomTreeDFS *DTDFS) : DTDFS(DTDFS) {}
464 /// valueNumber - finds the value number for V under the Subtree. If
465 /// there is no value number, returns zero.
466 unsigned valueNumber(Value *V, DomTreeDFS::Node *Subtree) {
467 if (!(isa<Constant>(V) || isa<Argument>(V) || isa<Instruction>(V))
468 || V->getType() == Type::VoidTy) return 0;
470 VNMapType::iterator E = VNMap.end();
471 VNPair pair(V, 0, Subtree);
472 VNMapType::iterator I = std::lower_bound(VNMap.begin(), E, pair);
473 while (I != E && I->V == V) {
474 if (I->Subtree->dominates(Subtree))
481 /// getOrInsertVN - always returns a value number, creating it if necessary.
482 unsigned getOrInsertVN(Value *V, DomTreeDFS::Node *Subtree) {
483 if (unsigned n = valueNumber(V, Subtree))
489 /// newVN - creates a new value number. Value V must not already have a
490 /// value number assigned.
491 unsigned newVN(Value *V) {
492 assert((isa<Constant>(V) || isa<Argument>(V) || isa<Instruction>(V)) &&
493 "Bad Value for value numbering.");
494 assert(V->getType() != Type::VoidTy && "Won't value number a void value");
498 VNPair pair = VNPair(V, Values.size(), DTDFS->getRootNode());
499 VNMapType::iterator I = std::lower_bound(VNMap.begin(), VNMap.end(), pair);
500 assert((I == VNMap.end() || value(I->index) != V) &&
501 "Attempt to create a duplicate value number.");
502 VNMap.insert(I, pair);
504 return Values.size();
507 /// value - returns the Value associated with a value number.
508 Value *value(unsigned index) const {
509 assert(index != 0 && "Zero index is reserved for not found.");
510 assert(index <= Values.size() && "Index out of range.");
511 return Values[index-1];
514 /// canonicalize - return a Value that is equal to V under Subtree.
515 Value *canonicalize(Value *V, DomTreeDFS::Node *Subtree) {
516 if (isa<Constant>(V)) return V;
518 if (unsigned n = valueNumber(V, Subtree))
524 /// addEquality - adds that value V belongs to the set of equivalent
525 /// values defined by value number n under Subtree.
526 void addEquality(unsigned n, Value *V, DomTreeDFS::Node *Subtree) {
527 assert(canonicalize(value(n), Subtree) == value(n) &&
528 "Node's 'canonical' choice isn't best within this subtree.");
530 // Suppose that we are given "%x -> node #1 (%y)". The problem is that
531 // we may already have "%z -> node #2 (%x)" somewhere above us in the
532 // graph. We need to find those edges and add "%z -> node #1 (%y)"
533 // to keep the lookups canonical.
535 std::vector<Value *> ToRepoint(1, V);
537 if (unsigned Conflict = valueNumber(V, Subtree)) {
538 for (VNMapType::iterator I = VNMap.begin(), E = VNMap.end();
540 if (I->index == Conflict && I->Subtree->dominates(Subtree))
541 ToRepoint.push_back(I->V);
545 for (std::vector<Value *>::iterator VI = ToRepoint.begin(),
546 VE = ToRepoint.end(); VI != VE; ++VI) {
549 VNPair pair(V, n, Subtree);
550 VNMapType::iterator B = VNMap.begin(), E = VNMap.end();
551 VNMapType::iterator I = std::lower_bound(B, E, pair);
552 if (I != E && I->V == V && I->Subtree == Subtree)
553 I->index = n; // Update best choice
555 VNMap.insert(I, pair); // New Value
557 // XXX: we currently don't have to worry about updating values with
558 // more specific Subtrees, but we will need to for PHI node support.
561 Value *V_n = value(n);
562 if (isa<Constant>(V) && isa<Constant>(V_n)) {
563 assert(V == V_n && "Constant equals different constant?");
569 /// remove - removes all references to value V.
570 void remove(Value *V) {
571 VNMapType::iterator B = VNMap.begin(), E = VNMap.end();
572 VNPair pair(V, 0, DTDFS->getRootNode());
573 VNMapType::iterator J = std::upper_bound(B, E, pair);
574 VNMapType::iterator I = J;
576 while (I != B && (I == E || I->V == V)) --I;
582 /// The InequalityGraph stores the relationships between values.
583 /// Each Value in the graph is assigned to a Node. Nodes are pointer
584 /// comparable for equality. The caller is expected to maintain the logical
585 /// consistency of the system.
587 /// The InequalityGraph class may invalidate Node*s after any mutator call.
588 /// @brief The InequalityGraph stores the relationships between values.
589 class VISIBILITY_HIDDEN InequalityGraph {
591 DomTreeDFS::Node *TreeRoot;
593 InequalityGraph(); // DO NOT IMPLEMENT
594 InequalityGraph(InequalityGraph &); // DO NOT IMPLEMENT
596 InequalityGraph(ValueNumbering &VN, DomTreeDFS::Node *TreeRoot)
597 : VN(VN), TreeRoot(TreeRoot) {}
601 /// An Edge is contained inside a Node making one end of the edge implicit
602 /// and contains a pointer to the other end. The edge contains a lattice
603 /// value specifying the relationship and an DomTreeDFS::Node specifying
604 /// the root in the dominator tree to which this edge applies.
605 class VISIBILITY_HIDDEN Edge {
607 Edge(unsigned T, LatticeVal V, DomTreeDFS::Node *ST)
608 : To(T), LV(V), Subtree(ST) {}
612 DomTreeDFS::Node *Subtree;
614 bool operator<(const Edge &edge) const {
615 if (To != edge.To) return To < edge.To;
616 return *Subtree < *edge.Subtree;
619 bool operator<(unsigned to) const {
623 bool operator>(unsigned to) const {
627 friend bool operator<(unsigned to, const Edge &edge) {
628 return edge.operator>(to);
632 /// A single node in the InequalityGraph. This stores the canonical Value
633 /// for the node, as well as the relationships with the neighbours.
635 /// @brief A single node in the InequalityGraph.
636 class VISIBILITY_HIDDEN Node {
637 friend class InequalityGraph;
639 typedef SmallVector<Edge, 4> RelationsType;
640 RelationsType Relations;
642 // TODO: can this idea improve performance?
643 //friend class std::vector<Node>;
644 //Node(Node &N) { RelationsType.swap(N.RelationsType); }
647 typedef RelationsType::iterator iterator;
648 typedef RelationsType::const_iterator const_iterator;
652 virtual void dump() const {
653 dump(*cerr.stream());
656 void dump(std::ostream &os) const {
657 static const std::string names[32] =
658 { "000000", "000001", "000002", "000003", "000004", "000005",
659 "000006", "000007", "000008", "000009", " >", " >=",
660 " s>u<", "s>=u<=", " s>", " s>=", "000016", "000017",
661 " s<u>", "s<=u>=", " <", " <=", " s<", " s<=",
662 "000024", "000025", " u>", " u>=", " u<", " u<=",
664 for (Node::const_iterator NI = begin(), NE = end(); NI != NE; ++NI) {
665 os << names[NI->LV] << " " << NI->To
666 << " (" << NI->Subtree->getDFSNumIn() << "), ";
672 iterator begin() { return Relations.begin(); }
673 iterator end() { return Relations.end(); }
674 const_iterator begin() const { return Relations.begin(); }
675 const_iterator end() const { return Relations.end(); }
677 iterator find(unsigned n, DomTreeDFS::Node *Subtree) {
679 for (iterator I = std::lower_bound(begin(), E, n);
680 I != E && I->To == n; ++I) {
681 if (Subtree->DominatedBy(I->Subtree))
687 const_iterator find(unsigned n, DomTreeDFS::Node *Subtree) const {
688 const_iterator E = end();
689 for (const_iterator I = std::lower_bound(begin(), E, n);
690 I != E && I->To == n; ++I) {
691 if (Subtree->DominatedBy(I->Subtree))
697 /// update - updates the lattice value for a given node, creating a new
698 /// entry if one doesn't exist. The new lattice value must not be
699 /// inconsistent with any previously existing value.
700 void update(unsigned n, LatticeVal R, DomTreeDFS::Node *Subtree) {
701 assert(validPredicate(R) && "Invalid predicate.");
703 Edge edge(n, R, Subtree);
704 iterator B = begin(), E = end();
705 iterator I = std::lower_bound(B, E, edge);
708 while (J != E && J->To == n) {
709 if (Subtree->DominatedBy(J->Subtree))
714 if (J != E && J->To == n) {
715 edge.LV = static_cast<LatticeVal>(J->LV & R);
716 assert(validPredicate(edge.LV) && "Invalid union of lattice values.");
718 if (edge.LV == J->LV)
719 return; // This update adds nothing new.
723 // We also have to tighten any edge beneath our update.
724 for (iterator K = I - 1; K->To == n; --K) {
725 if (K->Subtree->DominatedBy(Subtree)) {
726 LatticeVal LV = static_cast<LatticeVal>(K->LV & edge.LV);
727 assert(validPredicate(LV) && "Invalid union of lattice values");
734 // Insert new edge at Subtree if it isn't already there.
735 if (I == E || I->To != n || Subtree != I->Subtree)
736 Relations.insert(I, edge);
742 std::vector<Node> Nodes;
745 /// node - returns the node object at a given value number. The pointer
746 /// returned may be invalidated on the next call to node().
747 Node *node(unsigned index) {
748 assert(VN.value(index)); // This triggers the necessary checks.
749 if (Nodes.size() < index) Nodes.resize(index);
750 return &Nodes[index-1];
753 /// isRelatedBy - true iff n1 op n2
754 bool isRelatedBy(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
756 if (n1 == n2) return LV & EQ_BIT;
759 Node::iterator I = N1->find(n2, Subtree), E = N1->end();
760 if (I != E) return (I->LV & LV) == I->LV;
765 // The add* methods assume that your input is logically valid and may
766 // assertion-fail or infinitely loop if you attempt a contradiction.
768 /// addInequality - Sets n1 op n2.
769 /// It is also an error to call this on an inequality that is already true.
770 void addInequality(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
772 assert(n1 != n2 && "A node can't be inequal to itself.");
775 assert(!isRelatedBy(n1, n2, Subtree, reversePredicate(LV1)) &&
776 "Contradictory inequality.");
778 // Suppose we're adding %n1 < %n2. Find all the %a < %n1 and
779 // add %a < %n2 too. This keeps the graph fully connected.
781 // Break up the relationship into signed and unsigned comparison parts.
782 // If the signed parts of %a op1 %n1 match that of %n1 op2 %n2, and
783 // op1 and op2 aren't NE, then add %a op3 %n2. The new relationship
784 // should have the EQ_BIT iff it's set for both op1 and op2.
786 unsigned LV1_s = LV1 & (SLT_BIT|SGT_BIT);
787 unsigned LV1_u = LV1 & (ULT_BIT|UGT_BIT);
789 for (Node::iterator I = node(n1)->begin(), E = node(n1)->end(); I != E; ++I) {
790 if (I->LV != NE && I->To != n2) {
792 DomTreeDFS::Node *Local_Subtree = NULL;
793 if (Subtree->DominatedBy(I->Subtree))
794 Local_Subtree = Subtree;
795 else if (I->Subtree->DominatedBy(Subtree))
796 Local_Subtree = I->Subtree;
799 unsigned new_relationship = 0;
800 LatticeVal ILV = reversePredicate(I->LV);
801 unsigned ILV_s = ILV & (SLT_BIT|SGT_BIT);
802 unsigned ILV_u = ILV & (ULT_BIT|UGT_BIT);
804 if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s)
805 new_relationship |= ILV_s;
806 if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u)
807 new_relationship |= ILV_u;
809 if (new_relationship) {
810 if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0)
811 new_relationship |= (SLT_BIT|SGT_BIT);
812 if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0)
813 new_relationship |= (ULT_BIT|UGT_BIT);
814 if ((LV1 & EQ_BIT) && (ILV & EQ_BIT))
815 new_relationship |= EQ_BIT;
817 LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
819 node(I->To)->update(n2, NewLV, Local_Subtree);
820 node(n2)->update(I->To, reversePredicate(NewLV), Local_Subtree);
826 for (Node::iterator I = node(n2)->begin(), E = node(n2)->end(); I != E; ++I) {
827 if (I->LV != NE && I->To != n1) {
828 DomTreeDFS::Node *Local_Subtree = NULL;
829 if (Subtree->DominatedBy(I->Subtree))
830 Local_Subtree = Subtree;
831 else if (I->Subtree->DominatedBy(Subtree))
832 Local_Subtree = I->Subtree;
835 unsigned new_relationship = 0;
836 unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT);
837 unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT);
839 if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s)
840 new_relationship |= ILV_s;
842 if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u)
843 new_relationship |= ILV_u;
845 if (new_relationship) {
846 if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0)
847 new_relationship |= (SLT_BIT|SGT_BIT);
848 if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0)
849 new_relationship |= (ULT_BIT|UGT_BIT);
850 if ((LV1 & EQ_BIT) && (I->LV & EQ_BIT))
851 new_relationship |= EQ_BIT;
853 LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
855 node(n1)->update(I->To, NewLV, Local_Subtree);
856 node(I->To)->update(n1, reversePredicate(NewLV), Local_Subtree);
863 node(n1)->update(n2, LV1, Subtree);
864 node(n2)->update(n1, reversePredicate(LV1), Subtree);
867 /// remove - removes a node from the graph by removing all references to
869 void remove(unsigned n) {
871 for (Node::iterator NI = N->begin(), NE = N->end(); NI != NE; ++NI) {
872 Node::iterator Iter = node(NI->To)->find(n, TreeRoot);
874 node(NI->To)->Relations.erase(Iter);
875 Iter = node(NI->To)->find(n, TreeRoot);
876 } while (Iter != node(NI->To)->end());
878 N->Relations.clear();
882 virtual ~InequalityGraph() {}
883 virtual void dump() {
884 dump(*cerr.stream());
887 void dump(std::ostream &os) {
888 for (unsigned i = 1; i <= Nodes.size(); ++i) {
899 /// ValueRanges tracks the known integer ranges and anti-ranges of the nodes
900 /// in the InequalityGraph.
901 class VISIBILITY_HIDDEN ValueRanges {
905 class VISIBILITY_HIDDEN ScopedRange {
906 typedef std::vector<std::pair<DomTreeDFS::Node *, ConstantRange> >
908 RangeListType RangeList;
910 static bool swo(const std::pair<DomTreeDFS::Node *, ConstantRange> &LHS,
911 const std::pair<DomTreeDFS::Node *, ConstantRange> &RHS) {
912 return *LHS.first < *RHS.first;
917 virtual ~ScopedRange() {}
918 virtual void dump() const {
919 dump(*cerr.stream());
922 void dump(std::ostream &os) const {
924 for (const_iterator I = begin(), E = end(); I != E; ++I) {
925 os << I->second << " (" << I->first->getDFSNumIn() << "), ";
931 typedef RangeListType::iterator iterator;
932 typedef RangeListType::const_iterator const_iterator;
934 iterator begin() { return RangeList.begin(); }
935 iterator end() { return RangeList.end(); }
936 const_iterator begin() const { return RangeList.begin(); }
937 const_iterator end() const { return RangeList.end(); }
939 iterator find(DomTreeDFS::Node *Subtree) {
940 static ConstantRange empty(1, false);
942 iterator I = std::lower_bound(begin(), E,
943 std::make_pair(Subtree, empty), swo);
945 while (I != E && !I->first->dominates(Subtree)) ++I;
949 const_iterator find(DomTreeDFS::Node *Subtree) const {
950 static const ConstantRange empty(1, false);
951 const_iterator E = end();
952 const_iterator I = std::lower_bound(begin(), E,
953 std::make_pair(Subtree, empty), swo);
955 while (I != E && !I->first->dominates(Subtree)) ++I;
959 void update(const ConstantRange &CR, DomTreeDFS::Node *Subtree) {
960 assert(!CR.isEmptySet() && "Empty ConstantRange.");
961 assert(!CR.isSingleElement() && "Refusing to store single element.");
963 static ConstantRange empty(1, false);
966 std::lower_bound(begin(), E, std::make_pair(Subtree, empty), swo);
968 if (I != end() && I->first == Subtree) {
969 ConstantRange CR2 = I->second.maximalIntersectWith(CR);
970 assert(!CR2.isEmptySet() && !CR2.isSingleElement() &&
971 "Invalid union of ranges.");
974 RangeList.insert(I, std::make_pair(Subtree, CR));
978 std::vector<ScopedRange> Ranges;
980 void update(unsigned n, const ConstantRange &CR, DomTreeDFS::Node *Subtree){
981 if (CR.isFullSet()) return;
982 if (Ranges.size() < n) Ranges.resize(n);
983 Ranges[n-1].update(CR, Subtree);
986 /// create - Creates a ConstantRange that matches the given LatticeVal
987 /// relation with a given integer.
988 ConstantRange create(LatticeVal LV, const ConstantRange &CR) {
989 assert(!CR.isEmptySet() && "Can't deal with empty set.");
992 return makeConstantRange(ICmpInst::ICMP_NE, CR);
994 unsigned LV_s = LV & (SGT_BIT|SLT_BIT);
995 unsigned LV_u = LV & (UGT_BIT|ULT_BIT);
996 bool hasEQ = LV & EQ_BIT;
998 ConstantRange Range(CR.getBitWidth());
1000 if (LV_s == SGT_BIT) {
1001 Range = Range.maximalIntersectWith(makeConstantRange(
1002 hasEQ ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SGT, CR));
1003 } else if (LV_s == SLT_BIT) {
1004 Range = Range.maximalIntersectWith(makeConstantRange(
1005 hasEQ ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_SLT, CR));
1008 if (LV_u == UGT_BIT) {
1009 Range = Range.maximalIntersectWith(makeConstantRange(
1010 hasEQ ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_UGT, CR));
1011 } else if (LV_u == ULT_BIT) {
1012 Range = Range.maximalIntersectWith(makeConstantRange(
1013 hasEQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT, CR));
1019 /// makeConstantRange - Creates a ConstantRange representing the set of all
1020 /// value that match the ICmpInst::Predicate with any of the values in CR.
1021 ConstantRange makeConstantRange(ICmpInst::Predicate ICmpOpcode,
1022 const ConstantRange &CR) {
1023 uint32_t W = CR.getBitWidth();
1024 switch (ICmpOpcode) {
1025 default: assert(!"Invalid ICmp opcode to makeConstantRange()");
1026 case ICmpInst::ICMP_EQ:
1027 return ConstantRange(CR.getLower(), CR.getUpper());
1028 case ICmpInst::ICMP_NE:
1029 if (CR.isSingleElement())
1030 return ConstantRange(CR.getUpper(), CR.getLower());
1031 return ConstantRange(W);
1032 case ICmpInst::ICMP_ULT:
1033 return ConstantRange(APInt::getMinValue(W), CR.getUnsignedMax());
1034 case ICmpInst::ICMP_SLT:
1035 return ConstantRange(APInt::getSignedMinValue(W), CR.getSignedMax());
1036 case ICmpInst::ICMP_ULE: {
1037 APInt UMax(CR.getUnsignedMax());
1038 if (UMax.isMaxValue())
1039 return ConstantRange(W);
1040 return ConstantRange(APInt::getMinValue(W), UMax + 1);
1042 case ICmpInst::ICMP_SLE: {
1043 APInt SMax(CR.getSignedMax());
1044 if (SMax.isMaxSignedValue() || (SMax+1).isMaxSignedValue())
1045 return ConstantRange(W);
1046 return ConstantRange(APInt::getSignedMinValue(W), SMax + 1);
1048 case ICmpInst::ICMP_UGT:
1049 return ConstantRange(CR.getUnsignedMin() + 1, APInt::getNullValue(W));
1050 case ICmpInst::ICMP_SGT:
1051 return ConstantRange(CR.getSignedMin() + 1,
1052 APInt::getSignedMinValue(W));
1053 case ICmpInst::ICMP_UGE: {
1054 APInt UMin(CR.getUnsignedMin());
1055 if (UMin.isMinValue())
1056 return ConstantRange(W);
1057 return ConstantRange(UMin, APInt::getNullValue(W));
1059 case ICmpInst::ICMP_SGE: {
1060 APInt SMin(CR.getSignedMin());
1061 if (SMin.isMinSignedValue())
1062 return ConstantRange(W);
1063 return ConstantRange(SMin, APInt::getSignedMinValue(W));
1069 bool isCanonical(Value *V, DomTreeDFS::Node *Subtree) {
1070 return V == VN.canonicalize(V, Subtree);
1076 ValueRanges(ValueNumbering &VN, TargetData *TD) : VN(VN), TD(TD) {}
1079 virtual ~ValueRanges() {}
1081 virtual void dump() const {
1082 dump(*cerr.stream());
1085 void dump(std::ostream &os) const {
1086 for (unsigned i = 0, e = Ranges.size(); i != e; ++i) {
1087 os << (i+1) << " = ";
1094 /// range - looks up the ConstantRange associated with a value number.
1095 ConstantRange range(unsigned n, DomTreeDFS::Node *Subtree) {
1096 assert(VN.value(n)); // performs range checks
1098 if (n <= Ranges.size()) {
1099 ScopedRange::iterator I = Ranges[n-1].find(Subtree);
1100 if (I != Ranges[n-1].end()) return I->second;
1103 Value *V = VN.value(n);
1104 ConstantRange CR = range(V);
1108 /// range - determine a range from a Value without performing any lookups.
1109 ConstantRange range(Value *V) const {
1110 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
1111 return ConstantRange(C->getValue());
1112 else if (isa<ConstantPointerNull>(V))
1113 return ConstantRange(APInt::getNullValue(typeToWidth(V->getType())));
1115 return ConstantRange(typeToWidth(V->getType()));
1118 // typeToWidth - returns the number of bits necessary to store a value of
1119 // this type, or zero if unknown.
1120 uint32_t typeToWidth(const Type *Ty) const {
1122 return TD->getTypeSizeInBits(Ty);
1124 return Ty->getPrimitiveSizeInBits();
1127 static bool isRelatedBy(const ConstantRange &CR1, const ConstantRange &CR2,
1130 default: assert(!"Impossible lattice value!");
1132 return CR1.maximalIntersectWith(CR2).isEmptySet();
1134 return CR1.getUnsignedMax().ult(CR2.getUnsignedMin());
1136 return CR1.getUnsignedMax().ule(CR2.getUnsignedMin());
1138 return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax());
1140 return CR1.getUnsignedMin().uge(CR2.getUnsignedMax());
1142 return CR1.getSignedMax().slt(CR2.getSignedMin());
1144 return CR1.getSignedMax().sle(CR2.getSignedMin());
1146 return CR1.getSignedMin().sgt(CR2.getSignedMax());
1148 return CR1.getSignedMin().sge(CR2.getSignedMax());
1150 return CR1.getUnsignedMax().ult(CR2.getUnsignedMin()) &&
1151 CR1.getSignedMax().slt(CR2.getUnsignedMin());
1153 return CR1.getUnsignedMax().ule(CR2.getUnsignedMin()) &&
1154 CR1.getSignedMax().sle(CR2.getUnsignedMin());
1156 return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax()) &&
1157 CR1.getSignedMin().sgt(CR2.getSignedMax());
1159 return CR1.getUnsignedMin().uge(CR2.getUnsignedMax()) &&
1160 CR1.getSignedMin().sge(CR2.getSignedMax());
1162 return CR1.getSignedMax().slt(CR2.getSignedMin()) &&
1163 CR1.getUnsignedMin().ugt(CR2.getUnsignedMax());
1165 return CR1.getSignedMax().sle(CR2.getSignedMin()) &&
1166 CR1.getUnsignedMin().uge(CR2.getUnsignedMax());
1168 return CR1.getSignedMin().sgt(CR2.getSignedMax()) &&
1169 CR1.getUnsignedMax().ult(CR2.getUnsignedMin());
1171 return CR1.getSignedMin().sge(CR2.getSignedMax()) &&
1172 CR1.getUnsignedMax().ule(CR2.getUnsignedMin());
1176 bool isRelatedBy(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
1178 ConstantRange CR1 = range(n1, Subtree);
1179 ConstantRange CR2 = range(n2, Subtree);
1181 // True iff all values in CR1 are LV to all values in CR2.
1182 return isRelatedBy(CR1, CR2, LV);
1185 void addToWorklist(Value *V, Constant *C, ICmpInst::Predicate Pred,
1187 void markBlock(VRPSolver *VRP);
1189 void mergeInto(Value **I, unsigned n, unsigned New,
1190 DomTreeDFS::Node *Subtree, VRPSolver *VRP) {
1191 ConstantRange CR_New = range(New, Subtree);
1192 ConstantRange Merged = CR_New;
1194 for (; n != 0; ++I, --n) {
1195 unsigned i = VN.valueNumber(*I, Subtree);
1196 ConstantRange CR_Kill = i ? range(i, Subtree) : range(*I);
1197 if (CR_Kill.isFullSet()) continue;
1198 Merged = Merged.maximalIntersectWith(CR_Kill);
1201 if (Merged.isFullSet() || Merged == CR_New) return;
1203 applyRange(New, Merged, Subtree, VRP);
1206 void applyRange(unsigned n, const ConstantRange &CR,
1207 DomTreeDFS::Node *Subtree, VRPSolver *VRP) {
1208 ConstantRange Merged = CR.maximalIntersectWith(range(n, Subtree));
1209 if (Merged.isEmptySet()) {
1214 if (const APInt *I = Merged.getSingleElement()) {
1215 Value *V = VN.value(n); // XXX: redesign worklist.
1216 const Type *Ty = V->getType();
1217 if (Ty->isInteger()) {
1218 addToWorklist(V, ConstantInt::get(*I), ICmpInst::ICMP_EQ, VRP);
1220 } else if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) {
1221 assert(*I == 0 && "Pointer is null but not zero?");
1222 addToWorklist(V, ConstantPointerNull::get(PTy),
1223 ICmpInst::ICMP_EQ, VRP);
1228 update(n, Merged, Subtree);
1231 void addNotEquals(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
1233 ConstantRange CR1 = range(n1, Subtree);
1234 ConstantRange CR2 = range(n2, Subtree);
1236 uint32_t W = CR1.getBitWidth();
1238 if (const APInt *I = CR1.getSingleElement()) {
1239 if (CR2.isFullSet()) {
1240 ConstantRange NewCR2(CR1.getUpper(), CR1.getLower());
1241 applyRange(n2, NewCR2, Subtree, VRP);
1242 } else if (*I == CR2.getLower()) {
1243 APInt NewLower(CR2.getLower() + 1),
1244 NewUpper(CR2.getUpper());
1245 if (NewLower == NewUpper)
1246 NewLower = NewUpper = APInt::getMinValue(W);
1248 ConstantRange NewCR2(NewLower, NewUpper);
1249 applyRange(n2, NewCR2, Subtree, VRP);
1250 } else if (*I == CR2.getUpper() - 1) {
1251 APInt NewLower(CR2.getLower()),
1252 NewUpper(CR2.getUpper() - 1);
1253 if (NewLower == NewUpper)
1254 NewLower = NewUpper = APInt::getMinValue(W);
1256 ConstantRange NewCR2(NewLower, NewUpper);
1257 applyRange(n2, NewCR2, Subtree, VRP);
1261 if (const APInt *I = CR2.getSingleElement()) {
1262 if (CR1.isFullSet()) {
1263 ConstantRange NewCR1(CR2.getUpper(), CR2.getLower());
1264 applyRange(n1, NewCR1, Subtree, VRP);
1265 } else if (*I == CR1.getLower()) {
1266 APInt NewLower(CR1.getLower() + 1),
1267 NewUpper(CR1.getUpper());
1268 if (NewLower == NewUpper)
1269 NewLower = NewUpper = APInt::getMinValue(W);
1271 ConstantRange NewCR1(NewLower, NewUpper);
1272 applyRange(n1, NewCR1, Subtree, VRP);
1273 } else if (*I == CR1.getUpper() - 1) {
1274 APInt NewLower(CR1.getLower()),
1275 NewUpper(CR1.getUpper() - 1);
1276 if (NewLower == NewUpper)
1277 NewLower = NewUpper = APInt::getMinValue(W);
1279 ConstantRange NewCR1(NewLower, NewUpper);
1280 applyRange(n1, NewCR1, Subtree, VRP);
1285 void addInequality(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
1286 LatticeVal LV, VRPSolver *VRP) {
1287 assert(!isRelatedBy(n1, n2, Subtree, LV) && "Asked to do useless work.");
1290 addNotEquals(n1, n2, Subtree, VRP);
1294 ConstantRange CR1 = range(n1, Subtree);
1295 ConstantRange CR2 = range(n2, Subtree);
1297 if (!CR1.isSingleElement()) {
1298 ConstantRange NewCR1 = CR1.maximalIntersectWith(create(LV, CR2));
1300 applyRange(n1, NewCR1, Subtree, VRP);
1303 if (!CR2.isSingleElement()) {
1304 ConstantRange NewCR2 = CR2.maximalIntersectWith(
1305 create(reversePredicate(LV), CR1));
1307 applyRange(n2, NewCR2, Subtree, VRP);
1312 /// UnreachableBlocks keeps tracks of blocks that are for one reason or
1313 /// another discovered to be unreachable. This is used to cull the graph when
1314 /// analyzing instructions, and to mark blocks with the "unreachable"
1315 /// terminator instruction after the function has executed.
1316 class VISIBILITY_HIDDEN UnreachableBlocks {
1318 std::vector<BasicBlock *> DeadBlocks;
1321 /// mark - mark a block as dead
1322 void mark(BasicBlock *BB) {
1323 std::vector<BasicBlock *>::iterator E = DeadBlocks.end();
1324 std::vector<BasicBlock *>::iterator I =
1325 std::lower_bound(DeadBlocks.begin(), E, BB);
1327 if (I == E || *I != BB) DeadBlocks.insert(I, BB);
1330 /// isDead - returns whether a block is known to be dead already
1331 bool isDead(BasicBlock *BB) {
1332 std::vector<BasicBlock *>::iterator E = DeadBlocks.end();
1333 std::vector<BasicBlock *>::iterator I =
1334 std::lower_bound(DeadBlocks.begin(), E, BB);
1336 return I != E && *I == BB;
1339 /// kill - replace the dead blocks' terminator with an UnreachableInst.
1341 bool modified = false;
1342 for (std::vector<BasicBlock *>::iterator I = DeadBlocks.begin(),
1343 E = DeadBlocks.end(); I != E; ++I) {
1344 BasicBlock *BB = *I;
1346 DOUT << "unreachable block: " << BB->getName() << "\n";
1348 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
1350 BasicBlock *Succ = *SI;
1351 Succ->removePredecessor(BB);
1354 TerminatorInst *TI = BB->getTerminator();
1355 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1356 TI->eraseFromParent();
1357 new UnreachableInst(BB);
1366 /// VRPSolver keeps track of how changes to one variable affect other
1367 /// variables, and forwards changes along to the InequalityGraph. It
1368 /// also maintains the correct choice for "canonical" in the IG.
1369 /// @brief VRPSolver calculates inferences from a new relationship.
1370 class VISIBILITY_HIDDEN VRPSolver {
1372 friend class ValueRanges;
1376 ICmpInst::Predicate Op;
1378 BasicBlock *ContextBB; // XXX use a DomTreeDFS::Node instead
1379 Instruction *ContextInst;
1381 std::deque<Operation> WorkList;
1384 InequalityGraph &IG;
1385 UnreachableBlocks &UB;
1388 DomTreeDFS::Node *Top;
1390 Instruction *TopInst;
1393 typedef InequalityGraph::Node Node;
1395 // below - true if the Instruction is dominated by the current context
1396 // block or instruction
1397 bool below(Instruction *I) {
1398 BasicBlock *BB = I->getParent();
1399 if (TopInst && TopInst->getParent() == BB) {
1400 if (isa<TerminatorInst>(TopInst)) return false;
1401 if (isa<TerminatorInst>(I)) return true;
1402 if ( isa<PHINode>(TopInst) && !isa<PHINode>(I)) return true;
1403 if (!isa<PHINode>(TopInst) && isa<PHINode>(I)) return false;
1405 for (BasicBlock::const_iterator Iter = BB->begin(), E = BB->end();
1406 Iter != E; ++Iter) {
1407 if (&*Iter == TopInst) return true;
1408 else if (&*Iter == I) return false;
1410 assert(!"Instructions not found in parent BasicBlock?");
1412 DomTreeDFS::Node *Node = DTDFS->getNodeForBlock(BB);
1413 if (!Node) return false;
1414 return Top->dominates(Node);
1416 return false; // Not reached
1419 // aboveOrBelow - true if the Instruction either dominates or is dominated
1420 // by the current context block or instruction
1421 bool aboveOrBelow(Instruction *I) {
1422 BasicBlock *BB = I->getParent();
1423 DomTreeDFS::Node *Node = DTDFS->getNodeForBlock(BB);
1424 if (!Node) return false;
1426 return Top == Node || Top->dominates(Node) || Node->dominates(Top);
1429 bool makeEqual(Value *V1, Value *V2) {
1430 DOUT << "makeEqual(" << *V1 << ", " << *V2 << ")\n";
1431 DOUT << "context is ";
1432 if (TopInst) DOUT << "I: " << *TopInst << "\n";
1433 else DOUT << "BB: " << TopBB->getName()
1434 << "(" << Top->getDFSNumIn() << ")\n";
1436 assert(V1->getType() == V2->getType() &&
1437 "Can't make two values with different types equal.");
1439 if (V1 == V2) return true;
1441 if (isa<Constant>(V1) && isa<Constant>(V2))
1444 unsigned n1 = VN.valueNumber(V1, Top), n2 = VN.valueNumber(V2, Top);
1447 if (n1 == n2) return true;
1448 if (IG.isRelatedBy(n1, n2, Top, NE)) return false;
1451 if (n1) assert(V1 == VN.value(n1) && "Value isn't canonical.");
1452 if (n2) assert(V2 == VN.value(n2) && "Value isn't canonical.");
1454 assert(!VN.compare(V2, V1) && "Please order parameters to makeEqual.");
1456 assert(!isa<Constant>(V2) && "Tried to remove a constant.");
1458 SetVector<unsigned> Remove;
1459 if (n2) Remove.insert(n2);
1462 // Suppose we're being told that %x == %y, and %x <= %z and %y >= %z.
1463 // We can't just merge %x and %y because the relationship with %z would
1464 // be EQ and that's invalid. What we're doing is looking for any nodes
1465 // %z such that %x <= %z and %y >= %z, and vice versa.
1467 Node::iterator end = IG.node(n2)->end();
1469 // Find the intersection between N1 and N2 which is dominated by
1470 // Top. If we find %x where N1 <= %x <= N2 (or >=) then add %x to
1472 for (Node::iterator I = IG.node(n1)->begin(), E = IG.node(n1)->end();
1474 if (!(I->LV & EQ_BIT) || !Top->DominatedBy(I->Subtree)) continue;
1476 unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT);
1477 unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT);
1478 Node::iterator NI = IG.node(n2)->find(I->To, Top);
1480 LatticeVal NILV = reversePredicate(NI->LV);
1481 unsigned NILV_s = NILV & (SLT_BIT|SGT_BIT);
1482 unsigned NILV_u = NILV & (ULT_BIT|UGT_BIT);
1484 if ((ILV_s != (SLT_BIT|SGT_BIT) && ILV_s == NILV_s) ||
1485 (ILV_u != (ULT_BIT|UGT_BIT) && ILV_u == NILV_u))
1486 Remove.insert(I->To);
1490 // See if one of the nodes about to be removed is actually a better
1491 // canonical choice than n1.
1492 unsigned orig_n1 = n1;
1493 SetVector<unsigned>::iterator DontRemove = Remove.end();
1494 for (SetVector<unsigned>::iterator I = Remove.begin()+1 /* skip n2 */,
1495 E = Remove.end(); I != E; ++I) {
1497 Value *V = VN.value(n);
1498 if (VN.compare(V, V1)) {
1504 if (DontRemove != Remove.end()) {
1505 unsigned n = *DontRemove;
1507 Remove.insert(orig_n1);
1511 // We'd like to allow makeEqual on two values to perform a simple
1512 // substitution without creating nodes in the IG whenever possible.
1514 // The first iteration through this loop operates on V2 before going
1515 // through the Remove list and operating on those too. If all of the
1516 // iterations performed simple replacements then we exit early.
1517 bool mergeIGNode = false;
1519 for (Value *R = V2; i == 0 || i < Remove.size(); ++i) {
1520 if (i) R = VN.value(Remove[i]); // skip n2.
1522 // Try to replace the whole instruction. If we can, we're done.
1523 Instruction *I2 = dyn_cast<Instruction>(R);
1524 if (I2 && below(I2)) {
1525 std::vector<Instruction *> ToNotify;
1526 for (Value::use_iterator UI = R->use_begin(), UE = R->use_end();
1528 Use &TheUse = UI.getUse();
1530 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser()))
1531 ToNotify.push_back(I);
1534 DOUT << "Simply removing " << *I2
1535 << ", replacing with " << *V1 << "\n";
1536 I2->replaceAllUsesWith(V1);
1537 // leave it dead; it'll get erased later.
1541 for (std::vector<Instruction *>::iterator II = ToNotify.begin(),
1542 IE = ToNotify.end(); II != IE; ++II) {
1549 // Otherwise, replace all dominated uses.
1550 for (Value::use_iterator UI = R->use_begin(), UE = R->use_end();
1552 Use &TheUse = UI.getUse();
1554 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1564 // If that killed the instruction, stop here.
1565 if (I2 && isInstructionTriviallyDead(I2)) {
1566 DOUT << "Killed all uses of " << *I2
1567 << ", replacing with " << *V1 << "\n";
1571 // If we make it to here, then we will need to create a node for N1.
1572 // Otherwise, we can skip out early!
1576 if (!isa<Constant>(V1)) {
1577 if (Remove.empty()) {
1578 VR.mergeInto(&V2, 1, VN.getOrInsertVN(V1, Top), Top, this);
1580 std::vector<Value*> RemoveVals;
1581 RemoveVals.reserve(Remove.size());
1583 for (SetVector<unsigned>::iterator I = Remove.begin(),
1584 E = Remove.end(); I != E; ++I) {
1585 Value *V = VN.value(*I);
1586 if (!V->use_empty())
1587 RemoveVals.push_back(V);
1589 VR.mergeInto(&RemoveVals[0], RemoveVals.size(),
1590 VN.getOrInsertVN(V1, Top), Top, this);
1596 if (!n1) n1 = VN.getOrInsertVN(V1, Top);
1597 IG.node(n1); // Ensure that IG.Nodes won't get resized
1599 // Migrate relationships from removed nodes to N1.
1600 for (SetVector<unsigned>::iterator I = Remove.begin(), E = Remove.end();
1603 for (Node::iterator NI = IG.node(n)->begin(), NE = IG.node(n)->end();
1605 if (NI->Subtree->DominatedBy(Top)) {
1607 assert((NI->LV & EQ_BIT) && "Node inequal to itself.");
1610 if (Remove.count(NI->To))
1613 IG.node(NI->To)->update(n1, reversePredicate(NI->LV), Top);
1614 IG.node(n1)->update(NI->To, NI->LV, Top);
1619 // Point V2 (and all items in Remove) to N1.
1621 VN.addEquality(n1, V2, Top);
1623 for (SetVector<unsigned>::iterator I = Remove.begin(),
1624 E = Remove.end(); I != E; ++I) {
1625 VN.addEquality(n1, VN.value(*I), Top);
1629 // If !Remove.empty() then V2 = Remove[0]->getValue().
1630 // Even when Remove is empty, we still want to process V2.
1632 for (Value *R = V2; i == 0 || i < Remove.size(); ++i) {
1633 if (i) R = VN.value(Remove[i]); // skip n2.
1635 if (Instruction *I2 = dyn_cast<Instruction>(R)) {
1636 if (aboveOrBelow(I2))
1639 for (Value::use_iterator UI = V2->use_begin(), UE = V2->use_end();
1641 Use &TheUse = UI.getUse();
1643 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1644 if (aboveOrBelow(I))
1651 // re-opsToDef all dominated users of V1.
1652 if (Instruction *I = dyn_cast<Instruction>(V1)) {
1653 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1655 Use &TheUse = UI.getUse();
1657 Value *V = TheUse.getUser();
1658 if (!V->use_empty()) {
1659 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
1660 if (aboveOrBelow(Inst))
1670 /// cmpInstToLattice - converts an CmpInst::Predicate to lattice value
1671 /// Requires that the lattice value be valid; does not accept ICMP_EQ.
1672 static LatticeVal cmpInstToLattice(ICmpInst::Predicate Pred) {
1674 case ICmpInst::ICMP_EQ:
1675 assert(!"No matching lattice value.");
1676 return static_cast<LatticeVal>(EQ_BIT);
1678 assert(!"Invalid 'icmp' predicate.");
1679 case ICmpInst::ICMP_NE:
1681 case ICmpInst::ICMP_UGT:
1683 case ICmpInst::ICMP_UGE:
1685 case ICmpInst::ICMP_ULT:
1687 case ICmpInst::ICMP_ULE:
1689 case ICmpInst::ICMP_SGT:
1691 case ICmpInst::ICMP_SGE:
1693 case ICmpInst::ICMP_SLT:
1695 case ICmpInst::ICMP_SLE:
1701 VRPSolver(ValueNumbering &VN, InequalityGraph &IG, UnreachableBlocks &UB,
1702 ValueRanges &VR, DomTreeDFS *DTDFS, bool &modified,
1709 Top(DTDFS->getNodeForBlock(TopBB)),
1714 assert(Top && "VRPSolver created for unreachable basic block.");
1717 VRPSolver(ValueNumbering &VN, InequalityGraph &IG, UnreachableBlocks &UB,
1718 ValueRanges &VR, DomTreeDFS *DTDFS, bool &modified,
1719 Instruction *TopInst)
1725 Top(DTDFS->getNodeForBlock(TopInst->getParent())),
1726 TopBB(TopInst->getParent()),
1730 assert(Top && "VRPSolver created for unreachable basic block.");
1731 assert(Top->getBlock() == TopInst->getParent() && "Context mismatch.");
1734 bool isRelatedBy(Value *V1, Value *V2, ICmpInst::Predicate Pred) const {
1735 if (Constant *C1 = dyn_cast<Constant>(V1))
1736 if (Constant *C2 = dyn_cast<Constant>(V2))
1737 return ConstantExpr::getCompare(Pred, C1, C2) ==
1738 ConstantInt::getTrue();
1740 unsigned n1 = VN.valueNumber(V1, Top);
1741 unsigned n2 = VN.valueNumber(V2, Top);
1744 if (n1 == n2) return Pred == ICmpInst::ICMP_EQ ||
1745 Pred == ICmpInst::ICMP_ULE ||
1746 Pred == ICmpInst::ICMP_UGE ||
1747 Pred == ICmpInst::ICMP_SLE ||
1748 Pred == ICmpInst::ICMP_SGE;
1749 if (Pred == ICmpInst::ICMP_EQ) return false;
1750 if (IG.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true;
1751 if (VR.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true;
1754 if ((n1 && !n2 && isa<Constant>(V2)) ||
1755 (n2 && !n1 && isa<Constant>(V1))) {
1756 ConstantRange CR1 = n1 ? VR.range(n1, Top) : VR.range(V1);
1757 ConstantRange CR2 = n2 ? VR.range(n2, Top) : VR.range(V2);
1759 if (Pred == ICmpInst::ICMP_EQ)
1760 return CR1.isSingleElement() &&
1761 CR1.getSingleElement() == CR2.getSingleElement();
1763 return VR.isRelatedBy(CR1, CR2, cmpInstToLattice(Pred));
1765 if (Pred == ICmpInst::ICMP_EQ) return V1 == V2;
1769 /// add - adds a new property to the work queue
1770 void add(Value *V1, Value *V2, ICmpInst::Predicate Pred,
1771 Instruction *I = NULL) {
1772 DOUT << "adding " << *V1 << " " << Pred << " " << *V2;
1773 if (I) DOUT << " context: " << *I;
1774 else DOUT << " default context (" << Top->getDFSNumIn() << ")";
1777 assert(V1->getType() == V2->getType() &&
1778 "Can't relate two values with different types.");
1780 WorkList.push_back(Operation());
1781 Operation &O = WorkList.back();
1782 O.LHS = V1, O.RHS = V2, O.Op = Pred, O.ContextInst = I;
1783 O.ContextBB = I ? I->getParent() : TopBB;
1786 /// defToOps - Given an instruction definition that we've learned something
1787 /// new about, find any new relationships between its operands.
1788 void defToOps(Instruction *I) {
1789 Instruction *NewContext = below(I) ? I : TopInst;
1790 Value *Canonical = VN.canonicalize(I, Top);
1792 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1793 const Type *Ty = BO->getType();
1794 assert(!Ty->isFPOrFPVector() && "Float in work queue!");
1796 Value *Op0 = VN.canonicalize(BO->getOperand(0), Top);
1797 Value *Op1 = VN.canonicalize(BO->getOperand(1), Top);
1799 // TODO: "and i32 -1, %x" EQ %y then %x EQ %y.
1801 switch (BO->getOpcode()) {
1802 case Instruction::And: {
1803 // "and i32 %a, %b" EQ -1 then %a EQ -1 and %b EQ -1
1804 ConstantInt *CI = ConstantInt::getAllOnesValue(Ty);
1805 if (Canonical == CI) {
1806 add(CI, Op0, ICmpInst::ICMP_EQ, NewContext);
1807 add(CI, Op1, ICmpInst::ICMP_EQ, NewContext);
1810 case Instruction::Or: {
1811 // "or i32 %a, %b" EQ 0 then %a EQ 0 and %b EQ 0
1812 Constant *Zero = Constant::getNullValue(Ty);
1813 if (Canonical == Zero) {
1814 add(Zero, Op0, ICmpInst::ICMP_EQ, NewContext);
1815 add(Zero, Op1, ICmpInst::ICMP_EQ, NewContext);
1818 case Instruction::Xor: {
1819 // "xor i32 %c, %a" EQ %b then %a EQ %c ^ %b
1820 // "xor i32 %c, %a" EQ %c then %a EQ 0
1821 // "xor i32 %c, %a" NE %c then %a NE 0
1822 // Repeat the above, with order of operands reversed.
1825 if (!isa<Constant>(LHS)) std::swap(LHS, RHS);
1827 if (ConstantInt *CI = dyn_cast<ConstantInt>(Canonical)) {
1828 if (ConstantInt *Arg = dyn_cast<ConstantInt>(LHS)) {
1829 add(RHS, ConstantInt::get(CI->getValue() ^ Arg->getValue()),
1830 ICmpInst::ICMP_EQ, NewContext);
1833 if (Canonical == LHS) {
1834 if (isa<ConstantInt>(Canonical))
1835 add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_EQ,
1837 } else if (isRelatedBy(LHS, Canonical, ICmpInst::ICMP_NE)) {
1838 add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_NE,
1845 } else if (ICmpInst *IC = dyn_cast<ICmpInst>(I)) {
1846 // "icmp ult i32 %a, %y" EQ true then %a u< y
1849 if (Canonical == ConstantInt::getTrue()) {
1850 add(IC->getOperand(0), IC->getOperand(1), IC->getPredicate(),
1852 } else if (Canonical == ConstantInt::getFalse()) {
1853 add(IC->getOperand(0), IC->getOperand(1),
1854 ICmpInst::getInversePredicate(IC->getPredicate()), NewContext);
1856 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
1857 if (I->getType()->isFPOrFPVector()) return;
1859 // Given: "%a = select i1 %x, i32 %b, i32 %c"
1860 // %a EQ %b and %b NE %c then %x EQ true
1861 // %a EQ %c and %b NE %c then %x EQ false
1863 Value *True = SI->getTrueValue();
1864 Value *False = SI->getFalseValue();
1865 if (isRelatedBy(True, False, ICmpInst::ICMP_NE)) {
1866 if (Canonical == VN.canonicalize(True, Top) ||
1867 isRelatedBy(Canonical, False, ICmpInst::ICMP_NE))
1868 add(SI->getCondition(), ConstantInt::getTrue(),
1869 ICmpInst::ICMP_EQ, NewContext);
1870 else if (Canonical == VN.canonicalize(False, Top) ||
1871 isRelatedBy(Canonical, True, ICmpInst::ICMP_NE))
1872 add(SI->getCondition(), ConstantInt::getFalse(),
1873 ICmpInst::ICMP_EQ, NewContext);
1875 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1876 for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(),
1877 OE = GEPI->idx_end(); OI != OE; ++OI) {
1878 ConstantInt *Op = dyn_cast<ConstantInt>(VN.canonicalize(*OI, Top));
1879 if (!Op || !Op->isZero()) return;
1881 // TODO: The GEPI indices are all zero. Copy from definition to operand,
1882 // jumping the type plane as needed.
1883 if (isRelatedBy(GEPI, Constant::getNullValue(GEPI->getType()),
1884 ICmpInst::ICMP_NE)) {
1885 Value *Ptr = GEPI->getPointerOperand();
1886 add(Ptr, Constant::getNullValue(Ptr->getType()), ICmpInst::ICMP_NE,
1889 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
1890 const Type *SrcTy = CI->getSrcTy();
1892 unsigned ci = VN.getOrInsertVN(CI, Top);
1893 uint32_t W = VR.typeToWidth(SrcTy);
1895 ConstantRange CR = VR.range(ci, Top);
1897 if (CR.isFullSet()) return;
1899 switch (CI->getOpcode()) {
1901 case Instruction::ZExt:
1902 case Instruction::SExt:
1903 VR.applyRange(VN.getOrInsertVN(CI->getOperand(0), Top),
1904 CR.truncate(W), Top, this);
1906 case Instruction::BitCast:
1907 VR.applyRange(VN.getOrInsertVN(CI->getOperand(0), Top),
1914 /// opsToDef - A new relationship was discovered involving one of this
1915 /// instruction's operands. Find any new relationship involving the
1916 /// definition, or another operand.
1917 void opsToDef(Instruction *I) {
1918 Instruction *NewContext = below(I) ? I : TopInst;
1920 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1921 Value *Op0 = VN.canonicalize(BO->getOperand(0), Top);
1922 Value *Op1 = VN.canonicalize(BO->getOperand(1), Top);
1924 if (ConstantInt *CI0 = dyn_cast<ConstantInt>(Op0))
1925 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Op1)) {
1926 add(BO, ConstantExpr::get(BO->getOpcode(), CI0, CI1),
1927 ICmpInst::ICMP_EQ, NewContext);
1931 // "%y = and i1 true, %x" then %x EQ %y
1932 // "%y = or i1 false, %x" then %x EQ %y
1933 // "%x = add i32 %y, 0" then %x EQ %y
1934 // "%x = mul i32 %y, 0" then %x EQ 0
1936 Instruction::BinaryOps Opcode = BO->getOpcode();
1937 const Type *Ty = BO->getType();
1938 assert(!Ty->isFPOrFPVector() && "Float in work queue!");
1940 Constant *Zero = Constant::getNullValue(Ty);
1941 ConstantInt *AllOnes = ConstantInt::getAllOnesValue(Ty);
1945 case Instruction::LShr:
1946 case Instruction::AShr:
1947 case Instruction::Shl:
1948 case Instruction::Sub:
1950 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1954 case Instruction::Or:
1955 if (Op0 == AllOnes || Op1 == AllOnes) {
1956 add(BO, AllOnes, ICmpInst::ICMP_EQ, NewContext);
1959 case Instruction::Xor:
1960 case Instruction::Add:
1962 add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
1964 } else if (Op1 == Zero) {
1965 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1969 case Instruction::And:
1970 if (Op0 == AllOnes) {
1971 add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
1973 } else if (Op1 == AllOnes) {
1974 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1978 case Instruction::Mul:
1979 if (Op0 == Zero || Op1 == Zero) {
1980 add(BO, Zero, ICmpInst::ICMP_EQ, NewContext);
1986 // "%x = add i32 %y, %z" and %x EQ %y then %z EQ 0
1987 // "%x = add i32 %y, %z" and %x EQ %z then %y EQ 0
1988 // "%x = shl i32 %y, %z" and %x EQ %y and %y NE 0 then %z EQ 0
1989 // "%x = udiv i32 %y, %z" and %x EQ %y then %z EQ 1
1991 Value *Known = Op0, *Unknown = Op1,
1992 *TheBO = VN.canonicalize(BO, Top);
1993 if (Known != TheBO) std::swap(Known, Unknown);
1994 if (Known == TheBO) {
1997 case Instruction::LShr:
1998 case Instruction::AShr:
1999 case Instruction::Shl:
2000 if (!isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) break;
2001 // otherwise, fall-through.
2002 case Instruction::Sub:
2003 if (Unknown == Op0) break;
2004 // otherwise, fall-through.
2005 case Instruction::Xor:
2006 case Instruction::Add:
2007 add(Unknown, Zero, ICmpInst::ICMP_EQ, NewContext);
2009 case Instruction::UDiv:
2010 case Instruction::SDiv:
2011 if (Unknown == Op1) break;
2012 if (isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) {
2013 Constant *One = ConstantInt::get(Ty, 1);
2014 add(Unknown, One, ICmpInst::ICMP_EQ, NewContext);
2020 // TODO: "%a = add i32 %b, 1" and %b > %z then %a >= %z.
2022 } else if (ICmpInst *IC = dyn_cast<ICmpInst>(I)) {
2023 // "%a = icmp ult i32 %b, %c" and %b u< %c then %a EQ true
2024 // "%a = icmp ult i32 %b, %c" and %b u>= %c then %a EQ false
2027 Value *Op0 = VN.canonicalize(IC->getOperand(0), Top);
2028 Value *Op1 = VN.canonicalize(IC->getOperand(1), Top);
2030 ICmpInst::Predicate Pred = IC->getPredicate();
2031 if (isRelatedBy(Op0, Op1, Pred))
2032 add(IC, ConstantInt::getTrue(), ICmpInst::ICMP_EQ, NewContext);
2033 else if (isRelatedBy(Op0, Op1, ICmpInst::getInversePredicate(Pred)))
2034 add(IC, ConstantInt::getFalse(), ICmpInst::ICMP_EQ, NewContext);
2036 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
2037 if (I->getType()->isFPOrFPVector()) return;
2039 // Given: "%a = select i1 %x, i32 %b, i32 %c"
2040 // %x EQ true then %a EQ %b
2041 // %x EQ false then %a EQ %c
2042 // %b EQ %c then %a EQ %b
2044 Value *Canonical = VN.canonicalize(SI->getCondition(), Top);
2045 if (Canonical == ConstantInt::getTrue()) {
2046 add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext);
2047 } else if (Canonical == ConstantInt::getFalse()) {
2048 add(SI, SI->getFalseValue(), ICmpInst::ICMP_EQ, NewContext);
2049 } else if (VN.canonicalize(SI->getTrueValue(), Top) ==
2050 VN.canonicalize(SI->getFalseValue(), Top)) {
2051 add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext);
2053 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2054 const Type *DestTy = CI->getDestTy();
2055 if (DestTy->isFPOrFPVector()) return;
2057 Value *Op = VN.canonicalize(CI->getOperand(0), Top);
2058 Instruction::CastOps Opcode = CI->getOpcode();
2060 if (Constant *C = dyn_cast<Constant>(Op)) {
2061 add(CI, ConstantExpr::getCast(Opcode, C, DestTy),
2062 ICmpInst::ICMP_EQ, NewContext);
2065 uint32_t W = VR.typeToWidth(DestTy);
2066 unsigned ci = VN.getOrInsertVN(CI, Top);
2067 ConstantRange CR = VR.range(VN.getOrInsertVN(Op, Top), Top);
2069 if (!CR.isFullSet()) {
2072 case Instruction::ZExt:
2073 VR.applyRange(ci, CR.zeroExtend(W), Top, this);
2075 case Instruction::SExt:
2076 VR.applyRange(ci, CR.signExtend(W), Top, this);
2078 case Instruction::Trunc: {
2079 ConstantRange Result = CR.truncate(W);
2080 if (!Result.isFullSet())
2081 VR.applyRange(ci, Result, Top, this);
2083 case Instruction::BitCast:
2084 VR.applyRange(ci, CR, Top, this);
2086 // TODO: other casts?
2089 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
2090 for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(),
2091 OE = GEPI->idx_end(); OI != OE; ++OI) {
2092 ConstantInt *Op = dyn_cast<ConstantInt>(VN.canonicalize(*OI, Top));
2093 if (!Op || !Op->isZero()) return;
2095 // TODO: The GEPI indices are all zero. Copy from operand to definition,
2096 // jumping the type plane as needed.
2097 Value *Ptr = GEPI->getPointerOperand();
2098 if (isRelatedBy(Ptr, Constant::getNullValue(Ptr->getType()),
2099 ICmpInst::ICMP_NE)) {
2100 add(GEPI, Constant::getNullValue(GEPI->getType()), ICmpInst::ICMP_NE,
2106 /// solve - process the work queue
2108 //DOUT << "WorkList entry, size: " << WorkList.size() << "\n";
2109 while (!WorkList.empty()) {
2110 //DOUT << "WorkList size: " << WorkList.size() << "\n";
2112 Operation &O = WorkList.front();
2113 TopInst = O.ContextInst;
2114 TopBB = O.ContextBB;
2115 Top = DTDFS->getNodeForBlock(TopBB); // XXX move this into Context
2117 O.LHS = VN.canonicalize(O.LHS, Top);
2118 O.RHS = VN.canonicalize(O.RHS, Top);
2120 assert(O.LHS == VN.canonicalize(O.LHS, Top) && "Canonicalize isn't.");
2121 assert(O.RHS == VN.canonicalize(O.RHS, Top) && "Canonicalize isn't.");
2123 DOUT << "solving " << *O.LHS << " " << O.Op << " " << *O.RHS;
2124 if (O.ContextInst) DOUT << " context inst: " << *O.ContextInst;
2125 else DOUT << " context block: " << O.ContextBB->getName();
2132 // If they're both Constant, skip it. Check for contradiction and mark
2133 // the BB as unreachable if so.
2134 if (Constant *CI_L = dyn_cast<Constant>(O.LHS)) {
2135 if (Constant *CI_R = dyn_cast<Constant>(O.RHS)) {
2136 if (ConstantExpr::getCompare(O.Op, CI_L, CI_R) ==
2137 ConstantInt::getFalse())
2140 WorkList.pop_front();
2145 if (VN.compare(O.LHS, O.RHS)) {
2146 std::swap(O.LHS, O.RHS);
2147 O.Op = ICmpInst::getSwappedPredicate(O.Op);
2150 if (O.Op == ICmpInst::ICMP_EQ) {
2151 if (!makeEqual(O.RHS, O.LHS))
2154 LatticeVal LV = cmpInstToLattice(O.Op);
2156 if ((LV & EQ_BIT) &&
2157 isRelatedBy(O.LHS, O.RHS, ICmpInst::getSwappedPredicate(O.Op))) {
2158 if (!makeEqual(O.RHS, O.LHS))
2161 if (isRelatedBy(O.LHS, O.RHS, ICmpInst::getInversePredicate(O.Op))){
2163 WorkList.pop_front();
2167 unsigned n1 = VN.getOrInsertVN(O.LHS, Top);
2168 unsigned n2 = VN.getOrInsertVN(O.RHS, Top);
2171 if (O.Op != ICmpInst::ICMP_UGE && O.Op != ICmpInst::ICMP_ULE &&
2172 O.Op != ICmpInst::ICMP_SGE && O.Op != ICmpInst::ICMP_SLE)
2175 WorkList.pop_front();
2179 if (VR.isRelatedBy(n1, n2, Top, LV) ||
2180 IG.isRelatedBy(n1, n2, Top, LV)) {
2181 WorkList.pop_front();
2185 VR.addInequality(n1, n2, Top, LV, this);
2186 if ((!isa<ConstantInt>(O.RHS) && !isa<ConstantInt>(O.LHS)) ||
2188 IG.addInequality(n1, n2, Top, LV);
2190 if (Instruction *I1 = dyn_cast<Instruction>(O.LHS)) {
2191 if (aboveOrBelow(I1))
2194 if (isa<Instruction>(O.LHS) || isa<Argument>(O.LHS)) {
2195 for (Value::use_iterator UI = O.LHS->use_begin(),
2196 UE = O.LHS->use_end(); UI != UE;) {
2197 Use &TheUse = UI.getUse();
2199 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
2200 if (aboveOrBelow(I))
2205 if (Instruction *I2 = dyn_cast<Instruction>(O.RHS)) {
2206 if (aboveOrBelow(I2))
2209 if (isa<Instruction>(O.RHS) || isa<Argument>(O.RHS)) {
2210 for (Value::use_iterator UI = O.RHS->use_begin(),
2211 UE = O.RHS->use_end(); UI != UE;) {
2212 Use &TheUse = UI.getUse();
2214 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
2215 if (aboveOrBelow(I))
2222 WorkList.pop_front();
2227 void ValueRanges::addToWorklist(Value *V, Constant *C,
2228 ICmpInst::Predicate Pred, VRPSolver *VRP) {
2229 VRP->add(V, C, Pred, VRP->TopInst);
2232 void ValueRanges::markBlock(VRPSolver *VRP) {
2233 VRP->UB.mark(VRP->TopBB);
2236 /// PredicateSimplifier - This class is a simplifier that replaces
2237 /// one equivalent variable with another. It also tracks what
2238 /// can't be equal and will solve setcc instructions when possible.
2239 /// @brief Root of the predicate simplifier optimization.
2240 class VISIBILITY_HIDDEN PredicateSimplifier : public FunctionPass {
2244 InequalityGraph *IG;
2245 UnreachableBlocks UB;
2248 std::vector<DomTreeDFS::Node *> WorkList;
2251 static char ID; // Pass identification, replacement for typeid
2252 PredicateSimplifier() : FunctionPass((intptr_t)&ID) {}
2254 bool runOnFunction(Function &F);
2256 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
2257 AU.addRequiredID(BreakCriticalEdgesID);
2258 AU.addRequired<DominatorTree>();
2259 AU.addRequired<TargetData>();
2260 AU.addPreserved<TargetData>();
2264 /// Forwards - Adds new properties to VRPSolver and uses them to
2265 /// simplify instructions. Because new properties sometimes apply to
2266 /// a transition from one BasicBlock to another, this will use the
2267 /// PredicateSimplifier::proceedToSuccessor(s) interface to enter the
2269 /// @brief Performs abstract execution of the program.
2270 class VISIBILITY_HIDDEN Forwards : public InstVisitor<Forwards> {
2271 friend class InstVisitor<Forwards>;
2272 PredicateSimplifier *PS;
2273 DomTreeDFS::Node *DTNode;
2277 InequalityGraph &IG;
2278 UnreachableBlocks &UB;
2281 Forwards(PredicateSimplifier *PS, DomTreeDFS::Node *DTNode)
2282 : PS(PS), DTNode(DTNode), VN(*PS->VN), IG(*PS->IG), UB(PS->UB),
2285 void visitTerminatorInst(TerminatorInst &TI);
2286 void visitBranchInst(BranchInst &BI);
2287 void visitSwitchInst(SwitchInst &SI);
2289 void visitAllocaInst(AllocaInst &AI);
2290 void visitLoadInst(LoadInst &LI);
2291 void visitStoreInst(StoreInst &SI);
2293 void visitSExtInst(SExtInst &SI);
2294 void visitZExtInst(ZExtInst &ZI);
2296 void visitBinaryOperator(BinaryOperator &BO);
2297 void visitICmpInst(ICmpInst &IC);
2300 // Used by terminator instructions to proceed from the current basic
2301 // block to the next. Verifies that "current" dominates "next",
2302 // then calls visitBasicBlock.
2303 void proceedToSuccessors(DomTreeDFS::Node *Current) {
2304 for (DomTreeDFS::Node::iterator I = Current->begin(),
2305 E = Current->end(); I != E; ++I) {
2306 WorkList.push_back(*I);
2310 void proceedToSuccessor(DomTreeDFS::Node *Next) {
2311 WorkList.push_back(Next);
2314 // Visits each instruction in the basic block.
2315 void visitBasicBlock(DomTreeDFS::Node *Node) {
2316 BasicBlock *BB = Node->getBlock();
2317 DOUT << "Entering Basic Block: " << BB->getName()
2318 << " (" << Node->getDFSNumIn() << ")\n";
2319 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2320 visitInstruction(I++, Node);
2324 // Tries to simplify each Instruction and add new properties.
2325 void visitInstruction(Instruction *I, DomTreeDFS::Node *DT) {
2326 DOUT << "Considering instruction " << *I << "\n";
2331 // Sometimes instructions are killed in earlier analysis.
2332 if (isInstructionTriviallyDead(I)) {
2335 if (unsigned n = VN->valueNumber(I, DTDFS->getRootNode()))
2336 if (VN->value(n) == I) IG->remove(n);
2338 I->eraseFromParent();
2343 // Try to replace the whole instruction.
2344 Value *V = VN->canonicalize(I, DT);
2345 assert(V == I && "Late instruction canonicalization.");
2349 DOUT << "Removing " << *I << ", replacing with " << *V << "\n";
2350 if (unsigned n = VN->valueNumber(I, DTDFS->getRootNode()))
2351 if (VN->value(n) == I) IG->remove(n);
2353 I->replaceAllUsesWith(V);
2354 I->eraseFromParent();
2358 // Try to substitute operands.
2359 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2360 Value *Oper = I->getOperand(i);
2361 Value *V = VN->canonicalize(Oper, DT);
2362 assert(V == Oper && "Late operand canonicalization.");
2366 DOUT << "Resolving " << *I;
2367 I->setOperand(i, V);
2368 DOUT << " into " << *I;
2373 std::string name = I->getParent()->getName();
2374 DOUT << "push (%" << name << ")\n";
2375 Forwards visit(this, DT);
2377 DOUT << "pop (%" << name << ")\n";
2381 bool PredicateSimplifier::runOnFunction(Function &F) {
2382 DominatorTree *DT = &getAnalysis<DominatorTree>();
2383 DTDFS = new DomTreeDFS(DT);
2384 TargetData *TD = &getAnalysis<TargetData>();
2386 DOUT << "Entering Function: " << F.getName() << "\n";
2389 DomTreeDFS::Node *Root = DTDFS->getRootNode();
2390 VN = new ValueNumbering(DTDFS);
2391 IG = new InequalityGraph(*VN, Root);
2392 VR = new ValueRanges(*VN, TD);
2393 WorkList.push_back(Root);
2396 DomTreeDFS::Node *DTNode = WorkList.back();
2397 WorkList.pop_back();
2398 if (!UB.isDead(DTNode->getBlock())) visitBasicBlock(DTNode);
2399 } while (!WorkList.empty());
2405 modified |= UB.kill();
2410 void PredicateSimplifier::Forwards::visitTerminatorInst(TerminatorInst &TI) {
2411 PS->proceedToSuccessors(DTNode);
2414 void PredicateSimplifier::Forwards::visitBranchInst(BranchInst &BI) {
2415 if (BI.isUnconditional()) {
2416 PS->proceedToSuccessors(DTNode);
2420 Value *Condition = BI.getCondition();
2421 BasicBlock *TrueDest = BI.getSuccessor(0);
2422 BasicBlock *FalseDest = BI.getSuccessor(1);
2424 if (isa<Constant>(Condition) || TrueDest == FalseDest) {
2425 PS->proceedToSuccessors(DTNode);
2429 for (DomTreeDFS::Node::iterator I = DTNode->begin(), E = DTNode->end();
2431 BasicBlock *Dest = (*I)->getBlock();
2432 DOUT << "Branch thinking about %" << Dest->getName()
2433 << "(" << PS->DTDFS->getNodeForBlock(Dest)->getDFSNumIn() << ")\n";
2435 if (Dest == TrueDest) {
2436 DOUT << "(" << DTNode->getBlock()->getName() << ") true set:\n";
2437 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, Dest);
2438 VRP.add(ConstantInt::getTrue(), Condition, ICmpInst::ICMP_EQ);
2443 } else if (Dest == FalseDest) {
2444 DOUT << "(" << DTNode->getBlock()->getName() << ") false set:\n";
2445 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, Dest);
2446 VRP.add(ConstantInt::getFalse(), Condition, ICmpInst::ICMP_EQ);
2453 PS->proceedToSuccessor(*I);
2457 void PredicateSimplifier::Forwards::visitSwitchInst(SwitchInst &SI) {
2458 Value *Condition = SI.getCondition();
2460 // Set the EQProperty in each of the cases BBs, and the NEProperties
2461 // in the default BB.
2463 for (DomTreeDFS::Node::iterator I = DTNode->begin(), E = DTNode->end();
2465 BasicBlock *BB = (*I)->getBlock();
2466 DOUT << "Switch thinking about BB %" << BB->getName()
2467 << "(" << PS->DTDFS->getNodeForBlock(BB)->getDFSNumIn() << ")\n";
2469 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, BB);
2470 if (BB == SI.getDefaultDest()) {
2471 for (unsigned i = 1, e = SI.getNumCases(); i < e; ++i)
2472 if (SI.getSuccessor(i) != BB)
2473 VRP.add(Condition, SI.getCaseValue(i), ICmpInst::ICMP_NE);
2475 } else if (ConstantInt *CI = SI.findCaseDest(BB)) {
2476 VRP.add(Condition, CI, ICmpInst::ICMP_EQ);
2479 PS->proceedToSuccessor(*I);
2483 void PredicateSimplifier::Forwards::visitAllocaInst(AllocaInst &AI) {
2484 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &AI);
2485 VRP.add(Constant::getNullValue(AI.getType()), &AI, ICmpInst::ICMP_NE);
2489 void PredicateSimplifier::Forwards::visitLoadInst(LoadInst &LI) {
2490 Value *Ptr = LI.getPointerOperand();
2491 // avoid "load uint* null" -> null NE null.
2492 if (isa<Constant>(Ptr)) return;
2494 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &LI);
2495 VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
2499 void PredicateSimplifier::Forwards::visitStoreInst(StoreInst &SI) {
2500 Value *Ptr = SI.getPointerOperand();
2501 if (isa<Constant>(Ptr)) return;
2503 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &SI);
2504 VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
2508 void PredicateSimplifier::Forwards::visitSExtInst(SExtInst &SI) {
2509 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &SI);
2510 uint32_t SrcBitWidth = cast<IntegerType>(SI.getSrcTy())->getBitWidth();
2511 uint32_t DstBitWidth = cast<IntegerType>(SI.getDestTy())->getBitWidth();
2512 APInt Min(APInt::getHighBitsSet(DstBitWidth, DstBitWidth-SrcBitWidth+1));
2513 APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth-1));
2514 VRP.add(ConstantInt::get(Min), &SI, ICmpInst::ICMP_SLE);
2515 VRP.add(ConstantInt::get(Max), &SI, ICmpInst::ICMP_SGE);
2519 void PredicateSimplifier::Forwards::visitZExtInst(ZExtInst &ZI) {
2520 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &ZI);
2521 uint32_t SrcBitWidth = cast<IntegerType>(ZI.getSrcTy())->getBitWidth();
2522 uint32_t DstBitWidth = cast<IntegerType>(ZI.getDestTy())->getBitWidth();
2523 APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth));
2524 VRP.add(ConstantInt::get(Max), &ZI, ICmpInst::ICMP_UGE);
2528 void PredicateSimplifier::Forwards::visitBinaryOperator(BinaryOperator &BO) {
2529 Instruction::BinaryOps ops = BO.getOpcode();
2533 case Instruction::URem:
2534 case Instruction::SRem:
2535 case Instruction::UDiv:
2536 case Instruction::SDiv: {
2537 Value *Divisor = BO.getOperand(1);
2538 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
2539 VRP.add(Constant::getNullValue(Divisor->getType()), Divisor,
2548 case Instruction::Shl: {
2549 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
2550 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE);
2553 case Instruction::AShr: {
2554 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
2555 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_SLE);
2558 case Instruction::LShr:
2559 case Instruction::UDiv: {
2560 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
2561 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE);
2564 case Instruction::URem: {
2565 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
2566 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE);
2569 case Instruction::And: {
2570 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
2571 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE);
2572 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE);
2575 case Instruction::Or: {
2576 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
2577 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE);
2578 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_UGE);
2584 void PredicateSimplifier::Forwards::visitICmpInst(ICmpInst &IC) {
2585 // If possible, squeeze the ICmp predicate into something simpler.
2586 // Eg., if x = [0, 4) and we're being asked icmp uge %x, 3 then change
2587 // the predicate to eq.
2589 // XXX: once we do full PHI handling, modifying the instruction in the
2590 // Forwards visitor will cause missed optimizations.
2592 ICmpInst::Predicate Pred = IC.getPredicate();
2596 case ICmpInst::ICMP_ULE: Pred = ICmpInst::ICMP_ULT; break;
2597 case ICmpInst::ICMP_UGE: Pred = ICmpInst::ICMP_UGT; break;
2598 case ICmpInst::ICMP_SLE: Pred = ICmpInst::ICMP_SLT; break;
2599 case ICmpInst::ICMP_SGE: Pred = ICmpInst::ICMP_SGT; break;
2601 if (Pred != IC.getPredicate()) {
2602 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &IC);
2603 if (VRP.isRelatedBy(IC.getOperand(1), IC.getOperand(0),
2604 ICmpInst::ICMP_NE)) {
2606 PS->modified = true;
2607 IC.setPredicate(Pred);
2611 Pred = IC.getPredicate();
2613 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(IC.getOperand(1))) {
2614 ConstantInt *NextVal = 0;
2617 case ICmpInst::ICMP_SLT:
2618 case ICmpInst::ICMP_ULT:
2619 if (Op1->getValue() != 0)
2620 NextVal = ConstantInt::get(Op1->getValue()-1);
2622 case ICmpInst::ICMP_SGT:
2623 case ICmpInst::ICMP_UGT:
2624 if (!Op1->getValue().isAllOnesValue())
2625 NextVal = ConstantInt::get(Op1->getValue()+1);
2630 VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &IC);
2631 if (VRP.isRelatedBy(IC.getOperand(0), NextVal,
2632 ICmpInst::getInversePredicate(Pred))) {
2633 ICmpInst *NewIC = new ICmpInst(ICmpInst::ICMP_EQ, IC.getOperand(0),
2635 NewIC->takeName(&IC);
2636 IC.replaceAllUsesWith(NewIC);
2638 // XXX: prove this isn't necessary
2639 if (unsigned n = VN.valueNumber(&IC, PS->DTDFS->getRootNode()))
2640 if (VN.value(n) == &IC) IG.remove(n);
2643 IC.eraseFromParent();
2645 PS->modified = true;
2652 char PredicateSimplifier::ID = 0;
2653 static RegisterPass<PredicateSimplifier>
2654 X("predsimplify", "Predicate Simplifier");
2656 FunctionPass *llvm::createPredicateSimplifierPass() {
2657 return new PredicateSimplifier();