1 //===-- PredicateSimplifier.cpp - Path Sensitive Simplifier ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nick Lewycky and is distributed under the
6 // University of Illinois Open Source 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
73 // %b = [0, 254]. Because we store these by Value*, you should always
74 // canonicalize through the InequalityGraph first.
76 // It never stores an empty range, because that means that the code is
77 // unreachable. It never stores a single-element range since that's an equality
78 // relationship and better stored in the InequalityGraph, nor an empty range
79 // since that is better stored in UnreachableBlocks.
81 //===----------------------------------------------------------------------===//
83 #define DEBUG_TYPE "predsimplify"
84 #include "llvm/Transforms/Scalar.h"
85 #include "llvm/Constants.h"
86 #include "llvm/DerivedTypes.h"
87 #include "llvm/Instructions.h"
88 #include "llvm/Pass.h"
89 #include "llvm/ADT/DepthFirstIterator.h"
90 #include "llvm/ADT/SetOperations.h"
91 #include "llvm/ADT/SetVector.h"
92 #include "llvm/ADT/Statistic.h"
93 #include "llvm/ADT/STLExtras.h"
94 #include "llvm/Analysis/Dominators.h"
95 #include "llvm/Analysis/ET-Forest.h"
96 #include "llvm/Support/CFG.h"
97 #include "llvm/Support/Compiler.h"
98 #include "llvm/Support/ConstantRange.h"
99 #include "llvm/Support/Debug.h"
100 #include "llvm/Support/InstVisitor.h"
101 #include "llvm/Target/TargetData.h"
102 #include "llvm/Transforms/Utils/Local.h"
106 using namespace llvm;
108 STATISTIC(NumVarsReplaced, "Number of argument substitutions");
109 STATISTIC(NumInstruction , "Number of instructions removed");
110 STATISTIC(NumSimple , "Number of simple replacements");
111 STATISTIC(NumBlocks , "Number of blocks marked unreachable");
112 STATISTIC(NumSnuggle , "Number of comparisons snuggled");
115 // SLT SGT ULT UGT EQ
116 // 0 1 0 1 0 -- GT 10
117 // 0 1 0 1 1 -- GE 11
118 // 0 1 1 0 0 -- SGTULT 12
119 // 0 1 1 0 1 -- SGEULE 13
120 // 0 1 1 1 0 -- SGT 14
121 // 0 1 1 1 1 -- SGE 15
122 // 1 0 0 1 0 -- SLTUGT 18
123 // 1 0 0 1 1 -- SLEUGE 19
124 // 1 0 1 0 0 -- LT 20
125 // 1 0 1 0 1 -- LE 21
126 // 1 0 1 1 0 -- SLT 22
127 // 1 0 1 1 1 -- SLE 23
128 // 1 1 0 1 0 -- UGT 26
129 // 1 1 0 1 1 -- UGE 27
130 // 1 1 1 0 0 -- ULT 28
131 // 1 1 1 0 1 -- ULE 29
132 // 1 1 1 1 0 -- NE 30
134 EQ_BIT = 1, UGT_BIT = 2, ULT_BIT = 4, SGT_BIT = 8, SLT_BIT = 16
137 GT = SGT_BIT | UGT_BIT,
139 LT = SLT_BIT | ULT_BIT,
141 NE = SLT_BIT | SGT_BIT | ULT_BIT | UGT_BIT,
142 SGTULT = SGT_BIT | ULT_BIT,
143 SGEULE = SGTULT | EQ_BIT,
144 SLTUGT = SLT_BIT | UGT_BIT,
145 SLEUGE = SLTUGT | EQ_BIT,
146 ULT = SLT_BIT | SGT_BIT | ULT_BIT,
147 UGT = SLT_BIT | SGT_BIT | UGT_BIT,
148 SLT = SLT_BIT | ULT_BIT | UGT_BIT,
149 SGT = SGT_BIT | ULT_BIT | UGT_BIT,
156 static bool validPredicate(LatticeVal LV) {
158 case GT: case GE: case LT: case LE: case NE:
159 case SGTULT: case SGT: case SGEULE:
160 case SLTUGT: case SLT: case SLEUGE:
162 case SLE: case SGE: case ULE: case UGE:
169 /// reversePredicate - reverse the direction of the inequality
170 static LatticeVal reversePredicate(LatticeVal LV) {
171 unsigned reverse = LV ^ (SLT_BIT|SGT_BIT|ULT_BIT|UGT_BIT); //preserve EQ_BIT
173 if ((reverse & (SLT_BIT|SGT_BIT)) == 0)
174 reverse |= (SLT_BIT|SGT_BIT);
176 if ((reverse & (ULT_BIT|UGT_BIT)) == 0)
177 reverse |= (ULT_BIT|UGT_BIT);
179 LatticeVal Rev = static_cast<LatticeVal>(reverse);
180 assert(validPredicate(Rev) && "Failed reversing predicate.");
184 /// This is a StrictWeakOrdering predicate that sorts ETNodes by how many
185 /// descendants they have. With this, you can iterate through a list sorted
186 /// by this operation and the first matching entry is the most specific
187 /// match for your basic block. The order provided is stable; ETNodes with
188 /// the same number of children are sorted by pointer address.
189 struct VISIBILITY_HIDDEN OrderByDominance {
190 bool operator()(const ETNode *LHS, const ETNode *RHS) const {
191 unsigned LHS_spread = LHS->getDFSNumOut() - LHS->getDFSNumIn();
192 unsigned RHS_spread = RHS->getDFSNumOut() - RHS->getDFSNumIn();
193 if (LHS_spread != RHS_spread) return LHS_spread < RHS_spread;
194 else return LHS < RHS;
198 /// The InequalityGraph stores the relationships between values.
199 /// Each Value in the graph is assigned to a Node. Nodes are pointer
200 /// comparable for equality. The caller is expected to maintain the logical
201 /// consistency of the system.
203 /// The InequalityGraph class may invalidate Node*s after any mutator call.
204 /// @brief The InequalityGraph stores the relationships between values.
205 class VISIBILITY_HIDDEN InequalityGraph {
208 InequalityGraph(); // DO NOT IMPLEMENT
209 InequalityGraph(InequalityGraph &); // DO NOT IMPLEMENT
211 explicit InequalityGraph(ETNode *TreeRoot) : TreeRoot(TreeRoot) {}
215 /// An Edge is contained inside a Node making one end of the edge implicit
216 /// and contains a pointer to the other end. The edge contains a lattice
217 /// value specifying the relationship and an ETNode specifying the root
218 /// in the dominator tree to which this edge applies.
219 class VISIBILITY_HIDDEN Edge {
221 Edge(unsigned T, LatticeVal V, ETNode *ST)
222 : To(T), LV(V), Subtree(ST) {}
228 bool operator<(const Edge &edge) const {
229 if (To != edge.To) return To < edge.To;
230 else return OrderByDominance()(Subtree, edge.Subtree);
232 bool operator<(unsigned to) const {
237 /// A single node in the InequalityGraph. This stores the canonical Value
238 /// for the node, as well as the relationships with the neighbours.
240 /// @brief A single node in the InequalityGraph.
241 class VISIBILITY_HIDDEN Node {
242 friend class InequalityGraph;
244 typedef SmallVector<Edge, 4> RelationsType;
245 RelationsType Relations;
249 // TODO: can this idea improve performance?
250 //friend class std::vector<Node>;
251 //Node(Node &N) { RelationsType.swap(N.RelationsType); }
254 typedef RelationsType::iterator iterator;
255 typedef RelationsType::const_iterator const_iterator;
257 Node(Value *V) : Canonical(V) {}
263 virtual void dump() const {
264 dump(*cerr.stream());
267 void dump(std::ostream &os) const {
268 os << *getValue() << ":\n";
269 for (Node::const_iterator NI = begin(), NE = end(); NI != NE; ++NI) {
270 static const std::string names[32] =
271 { "000000", "000001", "000002", "000003", "000004", "000005",
272 "000006", "000007", "000008", "000009", " >", " >=",
273 " s>u<", "s>=u<=", " s>", " s>=", "000016", "000017",
274 " s<u>", "s<=u>=", " <", " <=", " s<", " s<=",
275 "000024", "000025", " u>", " u>=", " u<", " u<=",
277 os << " " << names[NI->LV] << " " << NI->To
278 << " (" << NI->Subtree->getDFSNumIn() << ")\n";
284 iterator begin() { return Relations.begin(); }
285 iterator end() { return Relations.end(); }
286 const_iterator begin() const { return Relations.begin(); }
287 const_iterator end() const { return Relations.end(); }
289 iterator find(unsigned n, ETNode *Subtree) {
291 for (iterator I = std::lower_bound(begin(), E, n);
292 I != E && I->To == n; ++I) {
293 if (Subtree->DominatedBy(I->Subtree))
299 const_iterator find(unsigned n, ETNode *Subtree) const {
300 const_iterator E = end();
301 for (const_iterator I = std::lower_bound(begin(), E, n);
302 I != E && I->To == n; ++I) {
303 if (Subtree->DominatedBy(I->Subtree))
309 Value *getValue() const
314 /// Updates the lattice value for a given node. Create a new entry if
315 /// one doesn't exist, otherwise it merges the values. The new lattice
316 /// value must not be inconsistent with any previously existing value.
317 void update(unsigned n, LatticeVal R, ETNode *Subtree) {
318 assert(validPredicate(R) && "Invalid predicate.");
319 iterator I = find(n, Subtree);
321 Edge edge(n, R, Subtree);
322 iterator Insert = std::lower_bound(begin(), end(), edge);
323 Relations.insert(Insert, edge);
325 LatticeVal LV = static_cast<LatticeVal>(I->LV & R);
326 assert(validPredicate(LV) && "Invalid union of lattice values.");
328 if (Subtree != I->Subtree) {
329 assert(Subtree->DominatedBy(I->Subtree) &&
330 "Find returned subtree that doesn't apply.");
332 Edge edge(n, R, Subtree);
333 iterator Insert = std::lower_bound(begin(), end(), edge);
334 Relations.insert(Insert, edge); // invalidates I
335 I = find(n, Subtree);
338 // Also, we have to tighten any edge that Subtree dominates.
339 for (iterator B = begin(); I->To == n; --I) {
340 if (I->Subtree->DominatedBy(Subtree)) {
341 LatticeVal LV = static_cast<LatticeVal>(I->LV & R);
342 assert(validPredicate(LV) && "Invalid union of lattice values");
353 struct VISIBILITY_HIDDEN NodeMapEdge {
358 NodeMapEdge(Value *V, unsigned index, ETNode *Subtree)
359 : V(V), index(index), Subtree(Subtree) {}
361 bool operator==(const NodeMapEdge &RHS) const {
363 Subtree == RHS.Subtree;
366 bool operator<(const NodeMapEdge &RHS) const {
367 if (V != RHS.V) return V < RHS.V;
368 return OrderByDominance()(Subtree, RHS.Subtree);
371 bool operator<(Value *RHS) const {
376 typedef std::vector<NodeMapEdge> NodeMapType;
379 std::vector<Node> Nodes;
382 /// node - returns the node object at a given index retrieved from getNode.
383 /// Index zero is reserved and may not be passed in here. The pointer
384 /// returned is valid until the next call to newNode or getOrInsertNode.
385 Node *node(unsigned index) {
386 assert(index != 0 && "Zero index is reserved for not found.");
387 assert(index <= Nodes.size() && "Index out of range.");
388 return &Nodes[index-1];
391 /// Returns the node currently representing Value V, or zero if no such
393 unsigned getNode(Value *V, ETNode *Subtree) {
394 NodeMapType::iterator E = NodeMap.end();
395 NodeMapEdge Edge(V, 0, Subtree);
396 NodeMapType::iterator I = std::lower_bound(NodeMap.begin(), E, Edge);
397 while (I != E && I->V == V) {
398 if (Subtree->DominatedBy(I->Subtree))
405 /// getOrInsertNode - always returns a valid node index, creating a node
406 /// to match the Value if needed.
407 unsigned getOrInsertNode(Value *V, ETNode *Subtree) {
408 if (unsigned n = getNode(V, Subtree))
414 /// newNode - creates a new node for a given Value and returns the index.
415 unsigned newNode(Value *V) {
416 Nodes.push_back(Node(V));
418 NodeMapEdge MapEntry = NodeMapEdge(V, Nodes.size(), TreeRoot);
419 assert(!std::binary_search(NodeMap.begin(), NodeMap.end(), MapEntry) &&
420 "Attempt to create a duplicate Node.");
421 NodeMap.insert(std::lower_bound(NodeMap.begin(), NodeMap.end(),
422 MapEntry), MapEntry);
423 return MapEntry.index;
426 /// If the Value is in the graph, return the canonical form. Otherwise,
427 /// return the original Value.
428 Value *canonicalize(Value *V, ETNode *Subtree) {
429 if (isa<Constant>(V)) return V;
431 if (unsigned n = getNode(V, Subtree))
432 return node(n)->getValue();
437 /// isRelatedBy - true iff n1 op n2
438 bool isRelatedBy(unsigned n1, unsigned n2, ETNode *Subtree, LatticeVal LV) {
439 if (n1 == n2) return LV & EQ_BIT;
442 Node::iterator I = N1->find(n2, Subtree), E = N1->end();
443 if (I != E) return (I->LV & LV) == I->LV;
448 // The add* methods assume that your input is logically valid and may
449 // assertion-fail or infinitely loop if you attempt a contradiction.
451 void addEquality(unsigned n, Value *V, ETNode *Subtree) {
452 assert(canonicalize(node(n)->getValue(), Subtree) == node(n)->getValue()
453 && "Node's 'canonical' choice isn't best within this subtree.");
455 // Suppose that we are given "%x -> node #1 (%y)". The problem is that
456 // we may already have "%z -> node #2 (%x)" somewhere above us in the
457 // graph. We need to find those edges and add "%z -> node #1 (%y)"
458 // to keep the lookups canonical.
460 std::vector<Value *> ToRepoint;
461 ToRepoint.push_back(V);
463 if (unsigned Conflict = getNode(V, Subtree)) {
464 for (NodeMapType::iterator I = NodeMap.begin(), E = NodeMap.end();
466 if (I->index == Conflict && Subtree->DominatedBy(I->Subtree))
467 ToRepoint.push_back(I->V);
471 for (std::vector<Value *>::iterator VI = ToRepoint.begin(),
472 VE = ToRepoint.end(); VI != VE; ++VI) {
475 // XXX: review this code. This may be doing too many insertions.
476 NodeMapEdge Edge(V, n, Subtree);
477 NodeMapType::iterator E = NodeMap.end();
478 NodeMapType::iterator I = std::lower_bound(NodeMap.begin(), E, Edge);
479 if (I == E || I->V != V || I->Subtree != Subtree) {
481 NodeMap.insert(I, Edge);
482 } else if (I != E && I->V == V && I->Subtree == Subtree) {
483 // Update best choice
489 if (isa<Constant>(V)) {
490 if (isa<Constant>(N->getValue())) {
491 assert(V == N->getValue() && "Constant equals different constant?");
498 /// addInequality - Sets n1 op n2.
499 /// It is also an error to call this on an inequality that is already true.
500 void addInequality(unsigned n1, unsigned n2, ETNode *Subtree,
502 assert(n1 != n2 && "A node can't be inequal to itself.");
505 assert(!isRelatedBy(n1, n2, Subtree, reversePredicate(LV1)) &&
506 "Contradictory inequality.");
511 // Suppose we're adding %n1 < %n2. Find all the %a < %n1 and
512 // add %a < %n2 too. This keeps the graph fully connected.
514 // Break up the relationship into signed and unsigned comparison parts.
515 // If the signed parts of %a op1 %n1 match that of %n1 op2 %n2, and
516 // op1 and op2 aren't NE, then add %a op3 %n2. The new relationship
517 // should have the EQ_BIT iff it's set for both op1 and op2.
519 unsigned LV1_s = LV1 & (SLT_BIT|SGT_BIT);
520 unsigned LV1_u = LV1 & (ULT_BIT|UGT_BIT);
522 for (Node::iterator I = N1->begin(), E = N1->end(); I != E; ++I) {
523 if (I->LV != NE && I->To != n2) {
525 ETNode *Local_Subtree = NULL;
526 if (Subtree->DominatedBy(I->Subtree))
527 Local_Subtree = Subtree;
528 else if (I->Subtree->DominatedBy(Subtree))
529 Local_Subtree = I->Subtree;
532 unsigned new_relationship = 0;
533 LatticeVal ILV = reversePredicate(I->LV);
534 unsigned ILV_s = ILV & (SLT_BIT|SGT_BIT);
535 unsigned ILV_u = ILV & (ULT_BIT|UGT_BIT);
537 if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s)
538 new_relationship |= ILV_s;
539 if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u)
540 new_relationship |= ILV_u;
542 if (new_relationship) {
543 if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0)
544 new_relationship |= (SLT_BIT|SGT_BIT);
545 if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0)
546 new_relationship |= (ULT_BIT|UGT_BIT);
547 if ((LV1 & EQ_BIT) && (ILV & EQ_BIT))
548 new_relationship |= EQ_BIT;
550 LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
552 node(I->To)->update(n2, NewLV, Local_Subtree);
553 N2->update(I->To, reversePredicate(NewLV), Local_Subtree);
559 for (Node::iterator I = N2->begin(), E = N2->end(); I != E; ++I) {
560 if (I->LV != NE && I->To != n1) {
561 ETNode *Local_Subtree = NULL;
562 if (Subtree->DominatedBy(I->Subtree))
563 Local_Subtree = Subtree;
564 else if (I->Subtree->DominatedBy(Subtree))
565 Local_Subtree = I->Subtree;
568 unsigned new_relationship = 0;
569 unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT);
570 unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT);
572 if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s)
573 new_relationship |= ILV_s;
575 if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u)
576 new_relationship |= ILV_u;
578 if (new_relationship) {
579 if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0)
580 new_relationship |= (SLT_BIT|SGT_BIT);
581 if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0)
582 new_relationship |= (ULT_BIT|UGT_BIT);
583 if ((LV1 & EQ_BIT) && (I->LV & EQ_BIT))
584 new_relationship |= EQ_BIT;
586 LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
588 N1->update(I->To, NewLV, Local_Subtree);
589 node(I->To)->update(n1, reversePredicate(NewLV), Local_Subtree);
596 N1->update(n2, LV1, Subtree);
597 N2->update(n1, reversePredicate(LV1), Subtree);
600 /// remove - Removes a Value from the graph. If the value is the canonical
601 /// choice for a Node, destroys the Node from the graph deleting all edges
602 /// to and from it. This method does not renumber the nodes.
603 void remove(Value *V) {
604 for (unsigned i = 0; i < NodeMap.size();) {
605 NodeMapType::iterator I = NodeMap.begin()+i;
607 Node *N = node(I->index);
608 if (node(I->index)->getValue() == V) {
609 for (Node::iterator NI = N->begin(), NE = N->end(); NI != NE; ++NI){
610 Node::iterator Iter = node(NI->To)->find(I->index, TreeRoot);
612 node(NI->To)->Relations.erase(Iter);
613 Iter = node(NI->To)->find(I->index, TreeRoot);
614 } while (Iter != node(NI->To)->end());
618 N->Relations.clear();
625 virtual ~InequalityGraph() {}
626 virtual void dump() {
627 dump(*cerr.stream());
630 void dump(std::ostream &os) {
631 std::set<Node *> VisitedNodes;
632 for (NodeMapType::const_iterator I = NodeMap.begin(), E = NodeMap.end();
634 Node *N = node(I->index);
635 os << *I->V << " == " << I->index
636 << "(" << I->Subtree->getDFSNumIn() << ")\n";
637 if (VisitedNodes.insert(N).second) {
638 os << I->index << ". ";
639 if (!N->getValue()) os << "(deleted node)\n";
649 /// ValueRanges tracks the known integer ranges and anti-ranges of the nodes
650 /// in the InequalityGraph.
651 class VISIBILITY_HIDDEN ValueRanges {
653 /// A ScopedRange ties an InequalityGraph node with a ConstantRange under
654 /// the scope of a rooted subtree in the dominator tree.
655 class VISIBILITY_HIDDEN ScopedRange {
657 ScopedRange(Value *V, ConstantRange CR, ETNode *ST)
658 : V(V), CR(CR), Subtree(ST) {}
664 bool operator<(const ScopedRange &range) const {
665 if (V != range.V) return V < range.V;
666 else return OrderByDominance()(Subtree, range.Subtree);
669 bool operator<(const Value *value) const {
676 std::vector<ScopedRange> Ranges;
677 typedef std::vector<ScopedRange>::iterator iterator;
679 // XXX: this is a copy of the code in InequalityGraph::Node. Perhaps a
680 // intrusive domtree-scoped container is in order?
682 iterator begin() { return Ranges.begin(); }
683 iterator end() { return Ranges.end(); }
685 iterator find(Value *V, ETNode *Subtree) {
687 for (iterator I = std::lower_bound(begin(), E, V);
688 I != E && I->V == V; ++I) {
689 if (Subtree->DominatedBy(I->Subtree))
695 void update(Value *V, ConstantRange CR, ETNode *Subtree) {
696 assert(!CR.isEmptySet() && "Empty ConstantRange!");
697 if (CR.isFullSet()) return;
699 iterator I = find(V, Subtree);
701 ScopedRange range(V, CR, Subtree);
702 iterator Insert = std::lower_bound(begin(), end(), range);
703 Ranges.insert(Insert, range);
705 CR = CR.intersectWith(I->CR);
706 assert(!CR.isEmptySet() && "Empty intersection of ConstantRanges!");
709 if (Subtree != I->Subtree) {
710 assert(Subtree->DominatedBy(I->Subtree) &&
711 "Find returned subtree that doesn't apply.");
713 ScopedRange range(V, CR, Subtree);
714 iterator Insert = std::lower_bound(begin(), end(), range);
715 Ranges.insert(Insert, range); // invalidates I
716 I = find(V, Subtree);
719 // Also, we have to tighten any edge that Subtree dominates.
720 for (iterator B = begin(); I->V == V; --I) {
721 if (I->Subtree->DominatedBy(Subtree)) {
722 I->CR = CR.intersectWith(I->CR);
723 assert(!I->CR.isEmptySet() &&
724 "Empty intersection of ConstantRanges!");
732 /// range - Creates a ConstantRange representing the set of all values
733 /// that match the ICmpInst::Predicate with any of the values in CR.
734 ConstantRange range(ICmpInst::Predicate ICmpOpcode,
735 const ConstantRange &CR) {
736 uint32_t W = CR.getBitWidth();
737 switch (ICmpOpcode) {
738 default: assert(!"Invalid ICmp opcode to range()");
739 case ICmpInst::ICMP_EQ:
740 return ConstantRange(CR.getLower(), CR.getUpper());
741 case ICmpInst::ICMP_NE:
742 if (CR.isSingleElement())
743 return ConstantRange(CR.getUpper(), CR.getLower());
744 return ConstantRange(W);
745 case ICmpInst::ICMP_ULT:
746 return ConstantRange(APInt::getMinValue(W), CR.getUnsignedMax());
747 case ICmpInst::ICMP_SLT:
748 return ConstantRange(APInt::getSignedMinValue(W), CR.getSignedMax());
749 case ICmpInst::ICMP_ULE: {
750 APInt UMax(CR.getUnsignedMax());
751 if (UMax == APInt::getMaxValue(W))
752 return ConstantRange(W);
753 return ConstantRange(APInt::getMinValue(W), UMax + 1);
755 case ICmpInst::ICMP_SLE: {
756 APInt SMax(CR.getSignedMax());
757 if (SMax == APInt::getSignedMaxValue(W) ||
758 SMax + 1 == APInt::getSignedMaxValue(W))
759 return ConstantRange(W);
760 return ConstantRange(APInt::getSignedMinValue(W), SMax + 1);
762 case ICmpInst::ICMP_UGT:
763 return ConstantRange(CR.getUnsignedMin() + 1, APInt::getNullValue(W));
764 case ICmpInst::ICMP_SGT:
765 return ConstantRange(CR.getSignedMin() + 1,
766 APInt::getSignedMinValue(W));
767 case ICmpInst::ICMP_UGE: {
768 APInt UMin(CR.getUnsignedMin());
769 if (UMin == APInt::getMinValue(W))
770 return ConstantRange(W);
771 return ConstantRange(UMin, APInt::getNullValue(W));
773 case ICmpInst::ICMP_SGE: {
774 APInt SMin(CR.getSignedMin());
775 if (SMin == APInt::getSignedMinValue(W))
776 return ConstantRange(W);
777 return ConstantRange(SMin, APInt::getSignedMinValue(W));
782 /// create - Creates a ConstantRange that matches the given LatticeVal
783 /// relation with a given integer.
784 ConstantRange create(LatticeVal LV, const ConstantRange &CR) {
785 assert(!CR.isEmptySet() && "Can't deal with empty set.");
788 return range(ICmpInst::ICMP_NE, CR);
790 unsigned LV_s = LV & (SGT_BIT|SLT_BIT);
791 unsigned LV_u = LV & (UGT_BIT|ULT_BIT);
792 bool hasEQ = LV & EQ_BIT;
794 ConstantRange Range(CR.getBitWidth());
796 if (LV_s == SGT_BIT) {
797 Range = Range.intersectWith(range(
798 hasEQ ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SGT, CR));
799 } else if (LV_s == SLT_BIT) {
800 Range = Range.intersectWith(range(
801 hasEQ ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_SLT, CR));
804 if (LV_u == UGT_BIT) {
805 Range = Range.intersectWith(range(
806 hasEQ ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_UGT, CR));
807 } else if (LV_u == ULT_BIT) {
808 Range = Range.intersectWith(range(
809 hasEQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT, CR));
816 bool isCanonical(Value *V, ETNode *Subtree, VRPSolver *VRP);
821 explicit ValueRanges(TargetData *TD) : TD(TD) {}
823 // rangeFromValue - converts a Value into a range. If the value is a
824 // constant it constructs the single element range, otherwise it performs
825 // a lookup. The width W must be retrieved from typeToWidth and may not
827 ConstantRange rangeFromValue(Value *V, ETNode *Subtree, uint32_t W) {
828 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) {
829 return ConstantRange(C->getValue());
830 } else if (isa<ConstantPointerNull>(V)) {
831 return ConstantRange(APInt::getNullValue(W));
833 iterator I = find(V, Subtree);
837 return ConstantRange(W);
840 // typeToWidth - returns the number of bits necessary to store a value of
841 // this type, or zero if unknown.
842 uint32_t typeToWidth(const Type *Ty) const {
844 return TD->getTypeSizeInBits(Ty);
846 if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty))
847 return ITy->getBitWidth();
852 bool isRelatedBy(Value *V1, Value *V2, ETNode *Subtree, LatticeVal LV) {
853 uint32_t W = typeToWidth(V1->getType());
854 if (!W) return false;
856 ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
857 ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
859 // True iff all values in CR1 are LV to all values in CR2.
861 default: assert(!"Impossible lattice value!");
863 return CR1.intersectWith(CR2).isEmptySet();
865 return CR1.getUnsignedMax().ult(CR2.getUnsignedMin());
867 return CR1.getUnsignedMax().ule(CR2.getUnsignedMin());
869 return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax());
871 return CR1.getUnsignedMin().uge(CR2.getUnsignedMax());
873 return CR1.getSignedMax().slt(CR2.getSignedMin());
875 return CR1.getSignedMax().sle(CR2.getSignedMin());
877 return CR1.getSignedMin().sgt(CR2.getSignedMax());
879 return CR1.getSignedMin().sge(CR2.getSignedMax());
881 return CR1.getUnsignedMax().ult(CR2.getUnsignedMin()) &&
882 CR1.getSignedMax().slt(CR2.getUnsignedMin());
884 return CR1.getUnsignedMax().ule(CR2.getUnsignedMin()) &&
885 CR1.getSignedMax().sle(CR2.getUnsignedMin());
887 return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax()) &&
888 CR1.getSignedMin().sgt(CR2.getSignedMax());
890 return CR1.getUnsignedMin().uge(CR2.getUnsignedMax()) &&
891 CR1.getSignedMin().sge(CR2.getSignedMax());
893 return CR1.getSignedMax().slt(CR2.getSignedMin()) &&
894 CR1.getUnsignedMin().ugt(CR2.getUnsignedMax());
896 return CR1.getSignedMax().sle(CR2.getSignedMin()) &&
897 CR1.getUnsignedMin().uge(CR2.getUnsignedMax());
899 return CR1.getSignedMin().sgt(CR2.getSignedMax()) &&
900 CR1.getUnsignedMax().ult(CR2.getUnsignedMin());
902 return CR1.getSignedMin().sge(CR2.getSignedMax()) &&
903 CR1.getUnsignedMax().ule(CR2.getUnsignedMin());
907 void addToWorklist(Value *V, Constant *C, ICmpInst::Predicate Pred,
909 void markBlock(VRPSolver *VRP);
911 void mergeInto(Value **I, unsigned n, Value *New, ETNode *Subtree,
913 assert(isCanonical(New, Subtree, VRP) && "Best choice not canonical?");
915 uint32_t W = typeToWidth(New->getType());
918 ConstantRange CR_New = rangeFromValue(New, Subtree, W);
919 ConstantRange Merged = CR_New;
921 for (; n != 0; ++I, --n) {
922 ConstantRange CR_Kill = rangeFromValue(*I, Subtree, W);
923 if (CR_Kill.isFullSet()) continue;
924 Merged = Merged.intersectWith(CR_Kill);
927 if (Merged.isFullSet() || Merged == CR_New) return;
929 applyRange(New, Merged, Subtree, VRP);
932 void applyRange(Value *V, const ConstantRange &CR, ETNode *Subtree,
934 assert(isCanonical(V, Subtree, VRP) && "Value not canonical.");
936 if (const APInt *I = CR.getSingleElement()) {
937 const Type *Ty = V->getType();
938 if (Ty->isInteger()) {
939 addToWorklist(V, ConstantInt::get(*I), ICmpInst::ICMP_EQ, VRP);
941 } else if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) {
942 assert(*I == 0 && "Pointer is null but not zero?");
943 addToWorklist(V, ConstantPointerNull::get(PTy),
944 ICmpInst::ICMP_EQ, VRP);
949 ConstantRange Merged = CR.intersectWith(
950 rangeFromValue(V, Subtree, CR.getBitWidth()));
951 if (Merged.isEmptySet()) {
956 update(V, Merged, Subtree);
959 void addNotEquals(Value *V1, Value *V2, ETNode *Subtree, VRPSolver *VRP) {
960 uint32_t W = typeToWidth(V1->getType());
963 ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
964 ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
966 if (const APInt *I = CR1.getSingleElement()) {
967 if (CR2.isFullSet()) {
968 ConstantRange NewCR2(CR1.getUpper(), CR1.getLower());
969 applyRange(V2, NewCR2, Subtree, VRP);
970 } else if (*I == CR2.getLower()) {
971 APInt NewLower(CR2.getLower() + 1),
972 NewUpper(CR2.getUpper());
973 if (NewLower == NewUpper)
974 NewLower = NewUpper = APInt::getMinValue(W);
976 ConstantRange NewCR2(NewLower, NewUpper);
977 applyRange(V2, NewCR2, Subtree, VRP);
978 } else if (*I == CR2.getUpper() - 1) {
979 APInt NewLower(CR2.getLower()),
980 NewUpper(CR2.getUpper() - 1);
981 if (NewLower == NewUpper)
982 NewLower = NewUpper = APInt::getMinValue(W);
984 ConstantRange NewCR2(NewLower, NewUpper);
985 applyRange(V2, NewCR2, Subtree, VRP);
989 if (const APInt *I = CR2.getSingleElement()) {
990 if (CR1.isFullSet()) {
991 ConstantRange NewCR1(CR2.getUpper(), CR2.getLower());
992 applyRange(V1, NewCR1, Subtree, VRP);
993 } else if (*I == CR1.getLower()) {
994 APInt NewLower(CR1.getLower() + 1),
995 NewUpper(CR1.getUpper());
996 if (NewLower == NewUpper)
997 NewLower = NewUpper = APInt::getMinValue(W);
999 ConstantRange NewCR1(NewLower, NewUpper);
1000 applyRange(V1, NewCR1, Subtree, VRP);
1001 } else if (*I == CR1.getUpper() - 1) {
1002 APInt NewLower(CR1.getLower()),
1003 NewUpper(CR1.getUpper() - 1);
1004 if (NewLower == NewUpper)
1005 NewLower = NewUpper = APInt::getMinValue(W);
1007 ConstantRange NewCR1(NewLower, NewUpper);
1008 applyRange(V1, NewCR1, Subtree, VRP);
1013 void addInequality(Value *V1, Value *V2, ETNode *Subtree, LatticeVal LV,
1015 assert(!isRelatedBy(V1, V2, Subtree, LV) && "Asked to do useless work.");
1017 assert(isCanonical(V1, Subtree, VRP) && "Value not canonical.");
1018 assert(isCanonical(V2, Subtree, VRP) && "Value not canonical.");
1021 addNotEquals(V1, V2, Subtree, VRP);
1025 uint32_t W = typeToWidth(V1->getType());
1028 ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
1029 ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
1031 if (!CR1.isSingleElement()) {
1032 ConstantRange NewCR1 = CR1.intersectWith(create(LV, CR2));
1034 applyRange(V1, NewCR1, Subtree, VRP);
1037 if (!CR2.isSingleElement()) {
1038 ConstantRange NewCR2 = CR2.intersectWith(create(reversePredicate(LV),
1041 applyRange(V2, NewCR2, Subtree, VRP);
1046 /// UnreachableBlocks keeps tracks of blocks that are for one reason or
1047 /// another discovered to be unreachable. This is used to cull the graph when
1048 /// analyzing instructions, and to mark blocks with the "unreachable"
1049 /// terminator instruction after the function has executed.
1050 class VISIBILITY_HIDDEN UnreachableBlocks {
1052 std::vector<BasicBlock *> DeadBlocks;
1055 /// mark - mark a block as dead
1056 void mark(BasicBlock *BB) {
1057 std::vector<BasicBlock *>::iterator E = DeadBlocks.end();
1058 std::vector<BasicBlock *>::iterator I =
1059 std::lower_bound(DeadBlocks.begin(), E, BB);
1061 if (I == E || *I != BB) DeadBlocks.insert(I, BB);
1064 /// isDead - returns whether a block is known to be dead already
1065 bool isDead(BasicBlock *BB) {
1066 std::vector<BasicBlock *>::iterator E = DeadBlocks.end();
1067 std::vector<BasicBlock *>::iterator I =
1068 std::lower_bound(DeadBlocks.begin(), E, BB);
1070 return I != E && *I == BB;
1073 /// kill - replace the dead blocks' terminator with an UnreachableInst.
1075 bool modified = false;
1076 for (std::vector<BasicBlock *>::iterator I = DeadBlocks.begin(),
1077 E = DeadBlocks.end(); I != E; ++I) {
1078 BasicBlock *BB = *I;
1080 DOUT << "unreachable block: " << BB->getName() << "\n";
1082 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
1084 BasicBlock *Succ = *SI;
1085 Succ->removePredecessor(BB);
1088 TerminatorInst *TI = BB->getTerminator();
1089 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1090 TI->eraseFromParent();
1091 new UnreachableInst(BB);
1100 /// VRPSolver keeps track of how changes to one variable affect other
1101 /// variables, and forwards changes along to the InequalityGraph. It
1102 /// also maintains the correct choice for "canonical" in the IG.
1103 /// @brief VRPSolver calculates inferences from a new relationship.
1104 class VISIBILITY_HIDDEN VRPSolver {
1106 friend class ValueRanges;
1110 ICmpInst::Predicate Op;
1112 BasicBlock *ContextBB;
1113 Instruction *ContextInst;
1115 std::deque<Operation> WorkList;
1117 InequalityGraph &IG;
1118 UnreachableBlocks &UB;
1124 Instruction *TopInst;
1127 typedef InequalityGraph::Node Node;
1129 /// IdomI - Determines whether one Instruction dominates another.
1130 bool IdomI(Instruction *I1, Instruction *I2) const {
1131 BasicBlock *BB1 = I1->getParent(),
1132 *BB2 = I2->getParent();
1134 if (isa<TerminatorInst>(I1)) return false;
1135 if (isa<TerminatorInst>(I2)) return true;
1136 if (isa<PHINode>(I1) && !isa<PHINode>(I2)) return true;
1137 if (!isa<PHINode>(I1) && isa<PHINode>(I2)) return false;
1139 for (BasicBlock::const_iterator I = BB1->begin(), E = BB1->end();
1141 if (&*I == I1) return true;
1142 if (&*I == I2) return false;
1144 assert(!"Instructions not found in parent BasicBlock?");
1146 return Forest->properlyDominates(BB1, BB2);
1151 /// Returns true if V1 is a better canonical value than V2.
1152 bool compare(Value *V1, Value *V2) const {
1153 if (isa<Constant>(V1))
1154 return !isa<Constant>(V2);
1155 else if (isa<Constant>(V2))
1157 else if (isa<Argument>(V1))
1158 return !isa<Argument>(V2);
1159 else if (isa<Argument>(V2))
1162 Instruction *I1 = dyn_cast<Instruction>(V1);
1163 Instruction *I2 = dyn_cast<Instruction>(V2);
1166 return V1->getNumUses() < V2->getNumUses();
1168 return IdomI(I1, I2);
1171 // below - true if the Instruction is dominated by the current context
1172 // block or instruction
1173 bool below(Instruction *I) {
1175 return IdomI(TopInst, I);
1177 ETNode *Node = Forest->getNodeForBlock(I->getParent());
1178 return Node->DominatedBy(Top);
1182 bool makeEqual(Value *V1, Value *V2) {
1183 DOUT << "makeEqual(" << *V1 << ", " << *V2 << ")\n";
1185 assert(V1->getType() == V2->getType() &&
1186 "Can't make two values with different types equal.");
1188 if (V1 == V2) return true;
1190 if (isa<Constant>(V1) && isa<Constant>(V2))
1193 unsigned n1 = IG.getNode(V1, Top), n2 = IG.getNode(V2, Top);
1196 if (n1 == n2) return true;
1197 if (IG.isRelatedBy(n1, n2, Top, NE)) return false;
1200 if (n1) assert(V1 == IG.node(n1)->getValue() && "Value isn't canonical.");
1201 if (n2) assert(V2 == IG.node(n2)->getValue() && "Value isn't canonical.");
1203 assert(!compare(V2, V1) && "Please order parameters to makeEqual.");
1205 assert(!isa<Constant>(V2) && "Tried to remove a constant.");
1207 SetVector<unsigned> Remove;
1208 if (n2) Remove.insert(n2);
1211 // Suppose we're being told that %x == %y, and %x <= %z and %y >= %z.
1212 // We can't just merge %x and %y because the relationship with %z would
1213 // be EQ and that's invalid. What we're doing is looking for any nodes
1214 // %z such that %x <= %z and %y >= %z, and vice versa.
1216 Node *N1 = IG.node(n1);
1217 Node *N2 = IG.node(n2);
1218 Node::iterator end = N2->end();
1220 // Find the intersection between N1 and N2 which is dominated by
1221 // Top. If we find %x where N1 <= %x <= N2 (or >=) then add %x to
1223 for (Node::iterator I = N1->begin(), E = N1->end(); I != E; ++I) {
1224 if (!(I->LV & EQ_BIT) || !Top->DominatedBy(I->Subtree)) continue;
1226 unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT);
1227 unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT);
1228 Node::iterator NI = N2->find(I->To, Top);
1230 LatticeVal NILV = reversePredicate(NI->LV);
1231 unsigned NILV_s = NILV & (SLT_BIT|SGT_BIT);
1232 unsigned NILV_u = NILV & (ULT_BIT|UGT_BIT);
1234 if ((ILV_s != (SLT_BIT|SGT_BIT) && ILV_s == NILV_s) ||
1235 (ILV_u != (ULT_BIT|UGT_BIT) && ILV_u == NILV_u))
1236 Remove.insert(I->To);
1240 // See if one of the nodes about to be removed is actually a better
1241 // canonical choice than n1.
1242 unsigned orig_n1 = n1;
1243 SetVector<unsigned>::iterator DontRemove = Remove.end();
1244 for (SetVector<unsigned>::iterator I = Remove.begin()+1 /* skip n2 */,
1245 E = Remove.end(); I != E; ++I) {
1247 Value *V = IG.node(n)->getValue();
1248 if (compare(V, V1)) {
1254 if (DontRemove != Remove.end()) {
1255 unsigned n = *DontRemove;
1257 Remove.insert(orig_n1);
1261 // We'd like to allow makeEqual on two values to perform a simple
1262 // substitution without every creating nodes in the IG whenever possible.
1264 // The first iteration through this loop operates on V2 before going
1265 // through the Remove list and operating on those too. If all of the
1266 // iterations performed simple replacements then we exit early.
1267 bool mergeIGNode = false;
1269 for (Value *R = V2; i == 0 || i < Remove.size(); ++i) {
1270 if (i) R = IG.node(Remove[i])->getValue(); // skip n2.
1272 // Try to replace the whole instruction. If we can, we're done.
1273 Instruction *I2 = dyn_cast<Instruction>(R);
1274 if (I2 && below(I2)) {
1275 std::vector<Instruction *> ToNotify;
1276 for (Value::use_iterator UI = R->use_begin(), UE = R->use_end();
1278 Use &TheUse = UI.getUse();
1280 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser()))
1281 ToNotify.push_back(I);
1284 DOUT << "Simply removing " << *I2
1285 << ", replacing with " << *V1 << "\n";
1286 I2->replaceAllUsesWith(V1);
1287 // leave it dead; it'll get erased later.
1291 for (std::vector<Instruction *>::iterator II = ToNotify.begin(),
1292 IE = ToNotify.end(); II != IE; ++II) {
1299 // Otherwise, replace all dominated uses.
1300 for (Value::use_iterator UI = R->use_begin(), UE = R->use_end();
1302 Use &TheUse = UI.getUse();
1304 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1314 // If that killed the instruction, stop here.
1315 if (I2 && isInstructionTriviallyDead(I2)) {
1316 DOUT << "Killed all uses of " << *I2
1317 << ", replacing with " << *V1 << "\n";
1321 // If we make it to here, then we will need to create a node for N1.
1322 // Otherwise, we can skip out early!
1326 if (!isa<Constant>(V1)) {
1327 if (Remove.empty()) {
1328 VR.mergeInto(&V2, 1, V1, Top, this);
1330 std::vector<Value*> RemoveVals;
1331 RemoveVals.reserve(Remove.size());
1333 for (SetVector<unsigned>::iterator I = Remove.begin(),
1334 E = Remove.end(); I != E; ++I) {
1335 Value *V = IG.node(*I)->getValue();
1336 if (!V->use_empty())
1337 RemoveVals.push_back(V);
1339 VR.mergeInto(&RemoveVals[0], RemoveVals.size(), V1, Top, this);
1345 if (!n1) n1 = IG.newNode(V1);
1347 // Migrate relationships from removed nodes to N1.
1348 Node *N1 = IG.node(n1);
1349 for (SetVector<unsigned>::iterator I = Remove.begin(), E = Remove.end();
1352 Node *N = IG.node(n);
1353 for (Node::iterator NI = N->begin(), NE = N->end(); NI != NE; ++NI) {
1354 if (NI->Subtree->DominatedBy(Top)) {
1356 assert((NI->LV & EQ_BIT) && "Node inequal to itself.");
1359 if (Remove.count(NI->To))
1362 IG.node(NI->To)->update(n1, reversePredicate(NI->LV), Top);
1363 N1->update(NI->To, NI->LV, Top);
1368 // Point V2 (and all items in Remove) to N1.
1370 IG.addEquality(n1, V2, Top);
1372 for (SetVector<unsigned>::iterator I = Remove.begin(),
1373 E = Remove.end(); I != E; ++I) {
1374 IG.addEquality(n1, IG.node(*I)->getValue(), Top);
1378 // If !Remove.empty() then V2 = Remove[0]->getValue().
1379 // Even when Remove is empty, we still want to process V2.
1381 for (Value *R = V2; i == 0 || i < Remove.size(); ++i) {
1382 if (i) R = IG.node(Remove[i])->getValue(); // skip n2.
1384 if (Instruction *I2 = dyn_cast<Instruction>(R)) {
1386 Top->DominatedBy(Forest->getNodeForBlock(I2->getParent())))
1389 for (Value::use_iterator UI = V2->use_begin(), UE = V2->use_end();
1391 Use &TheUse = UI.getUse();
1393 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1395 Top->DominatedBy(Forest->getNodeForBlock(I->getParent())))
1402 // re-opsToDef all dominated users of V1.
1403 if (Instruction *I = dyn_cast<Instruction>(V1)) {
1404 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1406 Use &TheUse = UI.getUse();
1408 Value *V = TheUse.getUser();
1409 if (!V->use_empty()) {
1410 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
1412 Top->DominatedBy(Forest->getNodeForBlock(Inst->getParent())))
1422 /// cmpInstToLattice - converts an CmpInst::Predicate to lattice value
1423 /// Requires that the lattice value be valid; does not accept ICMP_EQ.
1424 static LatticeVal cmpInstToLattice(ICmpInst::Predicate Pred) {
1426 case ICmpInst::ICMP_EQ:
1427 assert(!"No matching lattice value.");
1428 return static_cast<LatticeVal>(EQ_BIT);
1430 assert(!"Invalid 'icmp' predicate.");
1431 case ICmpInst::ICMP_NE:
1433 case ICmpInst::ICMP_UGT:
1435 case ICmpInst::ICMP_UGE:
1437 case ICmpInst::ICMP_ULT:
1439 case ICmpInst::ICMP_ULE:
1441 case ICmpInst::ICMP_SGT:
1443 case ICmpInst::ICMP_SGE:
1445 case ICmpInst::ICMP_SLT:
1447 case ICmpInst::ICMP_SLE:
1453 VRPSolver(InequalityGraph &IG, UnreachableBlocks &UB, ValueRanges &VR,
1454 ETForest *Forest, bool &modified, BasicBlock *TopBB)
1459 Top(Forest->getNodeForBlock(TopBB)),
1462 modified(modified) {}
1464 VRPSolver(InequalityGraph &IG, UnreachableBlocks &UB, ValueRanges &VR,
1465 ETForest *Forest, bool &modified, Instruction *TopInst)
1473 TopBB = TopInst->getParent();
1474 Top = Forest->getNodeForBlock(TopBB);
1477 bool isRelatedBy(Value *V1, Value *V2, ICmpInst::Predicate Pred) const {
1478 if (Constant *C1 = dyn_cast<Constant>(V1))
1479 if (Constant *C2 = dyn_cast<Constant>(V2))
1480 return ConstantExpr::getCompare(Pred, C1, C2) ==
1481 ConstantInt::getTrue();
1483 if (unsigned n1 = IG.getNode(V1, Top))
1484 if (unsigned n2 = IG.getNode(V2, Top)) {
1485 if (n1 == n2) return Pred == ICmpInst::ICMP_EQ ||
1486 Pred == ICmpInst::ICMP_ULE ||
1487 Pred == ICmpInst::ICMP_UGE ||
1488 Pred == ICmpInst::ICMP_SLE ||
1489 Pred == ICmpInst::ICMP_SGE;
1490 if (Pred == ICmpInst::ICMP_EQ) return false;
1491 if (IG.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true;
1494 if (Pred == ICmpInst::ICMP_EQ) return V1 == V2;
1495 return VR.isRelatedBy(V1, V2, Top, cmpInstToLattice(Pred));
1498 /// add - adds a new property to the work queue
1499 void add(Value *V1, Value *V2, ICmpInst::Predicate Pred,
1500 Instruction *I = NULL) {
1501 DOUT << "adding " << *V1 << " " << Pred << " " << *V2;
1502 if (I) DOUT << " context: " << *I;
1503 else DOUT << " default context";
1506 assert(V1->getType() == V2->getType() &&
1507 "Can't relate two values with different types.");
1509 WorkList.push_back(Operation());
1510 Operation &O = WorkList.back();
1511 O.LHS = V1, O.RHS = V2, O.Op = Pred, O.ContextInst = I;
1512 O.ContextBB = I ? I->getParent() : TopBB;
1515 /// defToOps - Given an instruction definition that we've learned something
1516 /// new about, find any new relationships between its operands.
1517 void defToOps(Instruction *I) {
1518 Instruction *NewContext = below(I) ? I : TopInst;
1519 Value *Canonical = IG.canonicalize(I, Top);
1521 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1522 const Type *Ty = BO->getType();
1523 assert(!Ty->isFPOrFPVector() && "Float in work queue!");
1525 Value *Op0 = IG.canonicalize(BO->getOperand(0), Top);
1526 Value *Op1 = IG.canonicalize(BO->getOperand(1), Top);
1528 // TODO: "and i32 -1, %x" EQ %y then %x EQ %y.
1530 switch (BO->getOpcode()) {
1531 case Instruction::And: {
1532 // "and i32 %a, %b" EQ -1 then %a EQ -1 and %b EQ -1
1533 ConstantInt *CI = ConstantInt::getAllOnesValue(Ty);
1534 if (Canonical == CI) {
1535 add(CI, Op0, ICmpInst::ICMP_EQ, NewContext);
1536 add(CI, Op1, ICmpInst::ICMP_EQ, NewContext);
1539 case Instruction::Or: {
1540 // "or i32 %a, %b" EQ 0 then %a EQ 0 and %b EQ 0
1541 Constant *Zero = Constant::getNullValue(Ty);
1542 if (Canonical == Zero) {
1543 add(Zero, Op0, ICmpInst::ICMP_EQ, NewContext);
1544 add(Zero, Op1, ICmpInst::ICMP_EQ, NewContext);
1547 case Instruction::Xor: {
1548 // "xor i32 %c, %a" EQ %b then %a EQ %c ^ %b
1549 // "xor i32 %c, %a" EQ %c then %a EQ 0
1550 // "xor i32 %c, %a" NE %c then %a NE 0
1551 // Repeat the above, with order of operands reversed.
1554 if (!isa<Constant>(LHS)) std::swap(LHS, RHS);
1556 if (ConstantInt *CI = dyn_cast<ConstantInt>(Canonical)) {
1557 if (ConstantInt *Arg = dyn_cast<ConstantInt>(LHS)) {
1558 add(RHS, ConstantInt::get(CI->getValue() ^ Arg->getValue()),
1559 ICmpInst::ICMP_EQ, NewContext);
1562 if (Canonical == LHS) {
1563 if (isa<ConstantInt>(Canonical))
1564 add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_EQ,
1566 } else if (isRelatedBy(LHS, Canonical, ICmpInst::ICMP_NE)) {
1567 add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_NE,
1574 } else if (ICmpInst *IC = dyn_cast<ICmpInst>(I)) {
1575 // "icmp ult i32 %a, %y" EQ true then %a u< y
1578 if (Canonical == ConstantInt::getTrue()) {
1579 add(IC->getOperand(0), IC->getOperand(1), IC->getPredicate(),
1581 } else if (Canonical == ConstantInt::getFalse()) {
1582 add(IC->getOperand(0), IC->getOperand(1),
1583 ICmpInst::getInversePredicate(IC->getPredicate()), NewContext);
1585 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
1586 if (I->getType()->isFPOrFPVector()) return;
1588 // Given: "%a = select i1 %x, i32 %b, i32 %c"
1589 // %a EQ %b and %b NE %c then %x EQ true
1590 // %a EQ %c and %b NE %c then %x EQ false
1592 Value *True = SI->getTrueValue();
1593 Value *False = SI->getFalseValue();
1594 if (isRelatedBy(True, False, ICmpInst::ICMP_NE)) {
1595 if (Canonical == IG.canonicalize(True, Top) ||
1596 isRelatedBy(Canonical, False, ICmpInst::ICMP_NE))
1597 add(SI->getCondition(), ConstantInt::getTrue(),
1598 ICmpInst::ICMP_EQ, NewContext);
1599 else if (Canonical == IG.canonicalize(False, Top) ||
1600 isRelatedBy(Canonical, True, ICmpInst::ICMP_NE))
1601 add(SI->getCondition(), ConstantInt::getFalse(),
1602 ICmpInst::ICMP_EQ, NewContext);
1604 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1605 for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(),
1606 OE = GEPI->idx_end(); OI != OE; ++OI) {
1607 ConstantInt *Op = dyn_cast<ConstantInt>(IG.canonicalize(*OI, Top));
1608 if (!Op || !Op->isZero()) return;
1610 // TODO: The GEPI indices are all zero. Copy from definition to operand,
1611 // jumping the type plane as needed.
1612 if (isRelatedBy(GEPI, Constant::getNullValue(GEPI->getType()),
1613 ICmpInst::ICMP_NE)) {
1614 Value *Ptr = GEPI->getPointerOperand();
1615 add(Ptr, Constant::getNullValue(Ptr->getType()), ICmpInst::ICMP_NE,
1618 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
1619 const Type *SrcTy = CI->getSrcTy();
1621 Value *TheCI = IG.canonicalize(CI, Top);
1622 uint32_t W = VR.typeToWidth(SrcTy);
1624 ConstantRange CR = VR.rangeFromValue(TheCI, Top, W);
1626 if (CR.isFullSet()) return;
1628 switch (CI->getOpcode()) {
1630 case Instruction::ZExt:
1631 case Instruction::SExt:
1632 VR.applyRange(IG.canonicalize(CI->getOperand(0), Top),
1633 CR.truncate(W), Top, this);
1635 case Instruction::BitCast:
1636 VR.applyRange(IG.canonicalize(CI->getOperand(0), Top),
1643 /// opsToDef - A new relationship was discovered involving one of this
1644 /// instruction's operands. Find any new relationship involving the
1645 /// definition, or another operand.
1646 void opsToDef(Instruction *I) {
1647 Instruction *NewContext = below(I) ? I : TopInst;
1649 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1650 Value *Op0 = IG.canonicalize(BO->getOperand(0), Top);
1651 Value *Op1 = IG.canonicalize(BO->getOperand(1), Top);
1653 if (ConstantInt *CI0 = dyn_cast<ConstantInt>(Op0))
1654 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Op1)) {
1655 add(BO, ConstantExpr::get(BO->getOpcode(), CI0, CI1),
1656 ICmpInst::ICMP_EQ, NewContext);
1660 // "%y = and i1 true, %x" then %x EQ %y
1661 // "%y = or i1 false, %x" then %x EQ %y
1662 // "%x = add i32 %y, 0" then %x EQ %y
1663 // "%x = mul i32 %y, 0" then %x EQ 0
1665 Instruction::BinaryOps Opcode = BO->getOpcode();
1666 const Type *Ty = BO->getType();
1667 assert(!Ty->isFPOrFPVector() && "Float in work queue!");
1669 Constant *Zero = Constant::getNullValue(Ty);
1670 ConstantInt *AllOnes = ConstantInt::getAllOnesValue(Ty);
1674 case Instruction::LShr:
1675 case Instruction::AShr:
1676 case Instruction::Shl:
1677 case Instruction::Sub:
1679 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1683 case Instruction::Or:
1684 if (Op0 == AllOnes || Op1 == AllOnes) {
1685 add(BO, AllOnes, ICmpInst::ICMP_EQ, NewContext);
1688 case Instruction::Xor:
1689 case Instruction::Add:
1691 add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
1693 } else if (Op1 == Zero) {
1694 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1698 case Instruction::And:
1699 if (Op0 == AllOnes) {
1700 add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
1702 } else if (Op1 == AllOnes) {
1703 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1707 case Instruction::Mul:
1708 if (Op0 == Zero || Op1 == Zero) {
1709 add(BO, Zero, ICmpInst::ICMP_EQ, NewContext);
1715 // "%x = add i32 %y, %z" and %x EQ %y then %z EQ 0
1716 // "%x = add i32 %y, %z" and %x EQ %z then %y EQ 0
1717 // "%x = shl i32 %y, %z" and %x EQ %y and %y NE 0 then %z EQ 0
1718 // "%x = udiv i32 %y, %z" and %x EQ %y then %z EQ 1
1720 Value *Known = Op0, *Unknown = Op1,
1721 *TheBO = IG.canonicalize(BO, Top);
1722 if (Known != TheBO) std::swap(Known, Unknown);
1723 if (Known == TheBO) {
1726 case Instruction::LShr:
1727 case Instruction::AShr:
1728 case Instruction::Shl:
1729 if (!isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) break;
1730 // otherwise, fall-through.
1731 case Instruction::Sub:
1732 if (Unknown == Op1) break;
1733 // otherwise, fall-through.
1734 case Instruction::Xor:
1735 case Instruction::Add:
1736 add(Unknown, Zero, ICmpInst::ICMP_EQ, NewContext);
1738 case Instruction::UDiv:
1739 case Instruction::SDiv:
1740 if (Unknown == Op1) break;
1741 if (isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) {
1742 Constant *One = ConstantInt::get(Ty, 1);
1743 add(Unknown, One, ICmpInst::ICMP_EQ, NewContext);
1749 // TODO: "%a = add i32 %b, 1" and %b > %z then %a >= %z.
1751 } else if (ICmpInst *IC = dyn_cast<ICmpInst>(I)) {
1752 // "%a = icmp ult i32 %b, %c" and %b u< %c then %a EQ true
1753 // "%a = icmp ult i32 %b, %c" and %b u>= %c then %a EQ false
1756 Value *Op0 = IG.canonicalize(IC->getOperand(0), Top);
1757 Value *Op1 = IG.canonicalize(IC->getOperand(1), Top);
1759 ICmpInst::Predicate Pred = IC->getPredicate();
1760 if (isRelatedBy(Op0, Op1, Pred)) {
1761 add(IC, ConstantInt::getTrue(), ICmpInst::ICMP_EQ, NewContext);
1762 } else if (isRelatedBy(Op0, Op1, ICmpInst::getInversePredicate(Pred))) {
1763 add(IC, ConstantInt::getFalse(), ICmpInst::ICMP_EQ, NewContext);
1766 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
1767 if (I->getType()->isFPOrFPVector()) return;
1769 // Given: "%a = select i1 %x, i32 %b, i32 %c"
1770 // %x EQ true then %a EQ %b
1771 // %x EQ false then %a EQ %c
1772 // %b EQ %c then %a EQ %b
1774 Value *Canonical = IG.canonicalize(SI->getCondition(), Top);
1775 if (Canonical == ConstantInt::getTrue()) {
1776 add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext);
1777 } else if (Canonical == ConstantInt::getFalse()) {
1778 add(SI, SI->getFalseValue(), ICmpInst::ICMP_EQ, NewContext);
1779 } else if (IG.canonicalize(SI->getTrueValue(), Top) ==
1780 IG.canonicalize(SI->getFalseValue(), Top)) {
1781 add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext);
1783 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
1784 const Type *DestTy = CI->getDestTy();
1785 if (DestTy->isFPOrFPVector()) return;
1787 Value *Op = IG.canonicalize(CI->getOperand(0), Top);
1788 Instruction::CastOps Opcode = CI->getOpcode();
1790 if (Constant *C = dyn_cast<Constant>(Op)) {
1791 add(CI, ConstantExpr::getCast(Opcode, C, DestTy),
1792 ICmpInst::ICMP_EQ, NewContext);
1795 uint32_t W = VR.typeToWidth(DestTy);
1796 Value *TheCI = IG.canonicalize(CI, Top);
1797 ConstantRange CR = VR.rangeFromValue(Op, Top, W);
1799 if (!CR.isFullSet()) {
1802 case Instruction::ZExt:
1803 VR.applyRange(TheCI, CR.zeroExtend(W), Top, this);
1805 case Instruction::SExt:
1806 VR.applyRange(TheCI, CR.signExtend(W), Top, this);
1808 case Instruction::Trunc: {
1809 ConstantRange Result = CR.truncate(W);
1810 if (!Result.isFullSet())
1811 VR.applyRange(TheCI, Result, Top, this);
1813 case Instruction::BitCast:
1814 VR.applyRange(TheCI, CR, Top, this);
1816 // TODO: other casts?
1819 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1820 for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(),
1821 OE = GEPI->idx_end(); OI != OE; ++OI) {
1822 ConstantInt *Op = dyn_cast<ConstantInt>(IG.canonicalize(*OI, Top));
1823 if (!Op || !Op->isZero()) return;
1825 // TODO: The GEPI indices are all zero. Copy from operand to definition,
1826 // jumping the type plane as needed.
1827 Value *Ptr = GEPI->getPointerOperand();
1828 if (isRelatedBy(Ptr, Constant::getNullValue(Ptr->getType()),
1829 ICmpInst::ICMP_NE)) {
1830 add(GEPI, Constant::getNullValue(GEPI->getType()), ICmpInst::ICMP_NE,
1836 /// solve - process the work queue
1838 //DOUT << "WorkList entry, size: " << WorkList.size() << "\n";
1839 while (!WorkList.empty()) {
1840 //DOUT << "WorkList size: " << WorkList.size() << "\n";
1842 Operation &O = WorkList.front();
1843 TopInst = O.ContextInst;
1844 TopBB = O.ContextBB;
1845 Top = Forest->getNodeForBlock(TopBB);
1847 O.LHS = IG.canonicalize(O.LHS, Top);
1848 O.RHS = IG.canonicalize(O.RHS, Top);
1850 assert(O.LHS == IG.canonicalize(O.LHS, Top) && "Canonicalize isn't.");
1851 assert(O.RHS == IG.canonicalize(O.RHS, Top) && "Canonicalize isn't.");
1853 DOUT << "solving " << *O.LHS << " " << O.Op << " " << *O.RHS;
1854 if (O.ContextInst) DOUT << " context inst: " << *O.ContextInst;
1855 else DOUT << " context block: " << O.ContextBB->getName();
1860 // If they're both Constant, skip it. Check for contradiction and mark
1861 // the BB as unreachable if so.
1862 if (Constant *CI_L = dyn_cast<Constant>(O.LHS)) {
1863 if (Constant *CI_R = dyn_cast<Constant>(O.RHS)) {
1864 if (ConstantExpr::getCompare(O.Op, CI_L, CI_R) ==
1865 ConstantInt::getFalse())
1868 WorkList.pop_front();
1873 if (compare(O.LHS, O.RHS)) {
1874 std::swap(O.LHS, O.RHS);
1875 O.Op = ICmpInst::getSwappedPredicate(O.Op);
1878 if (O.Op == ICmpInst::ICMP_EQ) {
1879 if (!makeEqual(O.RHS, O.LHS))
1882 LatticeVal LV = cmpInstToLattice(O.Op);
1884 if ((LV & EQ_BIT) &&
1885 isRelatedBy(O.LHS, O.RHS, ICmpInst::getSwappedPredicate(O.Op))) {
1886 if (!makeEqual(O.RHS, O.LHS))
1889 if (isRelatedBy(O.LHS, O.RHS, ICmpInst::getInversePredicate(O.Op))){
1891 WorkList.pop_front();
1895 unsigned n1 = IG.getNode(O.LHS, Top);
1896 unsigned n2 = IG.getNode(O.RHS, Top);
1898 if (n1 && n1 == n2) {
1899 if (O.Op != ICmpInst::ICMP_UGE && O.Op != ICmpInst::ICMP_ULE &&
1900 O.Op != ICmpInst::ICMP_SGE && O.Op != ICmpInst::ICMP_SLE)
1903 WorkList.pop_front();
1907 if (VR.isRelatedBy(O.LHS, O.RHS, Top, LV) ||
1908 (n1 && n2 && IG.isRelatedBy(n1, n2, Top, LV))) {
1909 WorkList.pop_front();
1913 VR.addInequality(O.LHS, O.RHS, Top, LV, this);
1914 if ((!isa<ConstantInt>(O.RHS) && !isa<ConstantInt>(O.LHS)) ||
1916 if (!n1) n1 = IG.newNode(O.LHS);
1917 if (!n2) n2 = IG.newNode(O.RHS);
1918 IG.addInequality(n1, n2, Top, LV);
1921 if (Instruction *I1 = dyn_cast<Instruction>(O.LHS)) {
1923 Top->DominatedBy(Forest->getNodeForBlock(I1->getParent())))
1926 if (isa<Instruction>(O.LHS) || isa<Argument>(O.LHS)) {
1927 for (Value::use_iterator UI = O.LHS->use_begin(),
1928 UE = O.LHS->use_end(); UI != UE;) {
1929 Use &TheUse = UI.getUse();
1931 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1933 Top->DominatedBy(Forest->getNodeForBlock(I->getParent())))
1938 if (Instruction *I2 = dyn_cast<Instruction>(O.RHS)) {
1940 Top->DominatedBy(Forest->getNodeForBlock(I2->getParent())))
1943 if (isa<Instruction>(O.RHS) || isa<Argument>(O.RHS)) {
1944 for (Value::use_iterator UI = O.RHS->use_begin(),
1945 UE = O.RHS->use_end(); UI != UE;) {
1946 Use &TheUse = UI.getUse();
1948 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1950 Top->DominatedBy(Forest->getNodeForBlock(I->getParent())))
1958 WorkList.pop_front();
1963 void ValueRanges::addToWorklist(Value *V, Constant *C,
1964 ICmpInst::Predicate Pred, VRPSolver *VRP) {
1965 VRP->add(V, C, Pred, VRP->TopInst);
1968 void ValueRanges::markBlock(VRPSolver *VRP) {
1969 VRP->UB.mark(VRP->TopBB);
1973 bool ValueRanges::isCanonical(Value *V, ETNode *Subtree, VRPSolver *VRP) {
1974 return V == VRP->IG.canonicalize(V, Subtree);
1978 /// PredicateSimplifier - This class is a simplifier that replaces
1979 /// one equivalent variable with another. It also tracks what
1980 /// can't be equal and will solve setcc instructions when possible.
1981 /// @brief Root of the predicate simplifier optimization.
1982 class VISIBILITY_HIDDEN PredicateSimplifier : public FunctionPass {
1985 InequalityGraph *IG;
1986 UnreachableBlocks UB;
1989 std::vector<BasicBlock *> WorkList;
1992 bool runOnFunction(Function &F);
1994 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1995 AU.addRequiredID(BreakCriticalEdgesID);
1996 AU.addRequired<ETForest>();
1997 AU.addRequired<TargetData>();
1998 AU.addPreserved<TargetData>();
2002 /// Forwards - Adds new properties into PropertySet and uses them to
2003 /// simplify instructions. Because new properties sometimes apply to
2004 /// a transition from one BasicBlock to another, this will use the
2005 /// PredicateSimplifier::proceedToSuccessor(s) interface to enter the
2006 /// basic block with the new PropertySet.
2007 /// @brief Performs abstract execution of the program.
2008 class VISIBILITY_HIDDEN Forwards : public InstVisitor<Forwards> {
2009 friend class InstVisitor<Forwards>;
2010 PredicateSimplifier *PS;
2014 InequalityGraph &IG;
2015 UnreachableBlocks &UB;
2018 Forwards(PredicateSimplifier *PS, BasicBlock* node)
2019 : PS(PS), Node(node), IG(*PS->IG), UB(PS->UB), VR(*PS->VR) {}
2021 void visitTerminatorInst(TerminatorInst &TI);
2022 void visitBranchInst(BranchInst &BI);
2023 void visitSwitchInst(SwitchInst &SI);
2025 void visitAllocaInst(AllocaInst &AI);
2026 void visitLoadInst(LoadInst &LI);
2027 void visitStoreInst(StoreInst &SI);
2029 void visitSExtInst(SExtInst &SI);
2030 void visitZExtInst(ZExtInst &ZI);
2032 void visitBinaryOperator(BinaryOperator &BO);
2033 void visitICmpInst(ICmpInst &IC);
2036 // Used by terminator instructions to proceed from the current basic
2037 // block to the next. Verifies that "current" dominates "next",
2038 // then calls visitBasicBlock.
2039 void proceedToSuccessors(BasicBlock *Current) {
2040 std::vector<BasicBlock*> Children;
2041 Forest->getChildren(Current, Children);
2042 for (std::vector<BasicBlock*>::iterator I = Children.begin(),
2043 E = Children.end(); I != E; ++I) {
2044 WorkList.push_back(*I);
2048 void proceedToSuccessor(BasicBlock *Next) {
2049 WorkList.push_back(Next);
2052 // Visits each instruction in the basic block.
2053 void visitBasicBlock(BasicBlock *BB) {
2054 ETNode *ET = Forest->getNodeForBlock(BB);
2055 DOUT << "Entering Basic Block: " << BB->getName()
2056 << " (" << ET->getDFSNumIn() << ")\n";
2057 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2058 visitInstruction(I++, BB, ET);
2062 // Tries to simplify each Instruction and add new properties to
2064 void visitInstruction(Instruction *I, BasicBlock *node, ETNode *ET) {
2065 DOUT << "Considering instruction " << *I << "\n";
2068 // Sometimes instructions are killed in earlier analysis.
2069 if (isInstructionTriviallyDead(I)) {
2073 I->eraseFromParent();
2078 // Try to replace the whole instruction.
2079 Value *V = IG->canonicalize(I, ET);
2080 assert(V == I && "Late instruction canonicalization.");
2084 DOUT << "Removing " << *I << ", replacing with " << *V << "\n";
2086 I->replaceAllUsesWith(V);
2087 I->eraseFromParent();
2091 // Try to substitute operands.
2092 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
2093 Value *Oper = I->getOperand(i);
2094 Value *V = IG->canonicalize(Oper, ET);
2095 assert(V == Oper && "Late operand canonicalization.");
2099 DOUT << "Resolving " << *I;
2100 I->setOperand(i, V);
2101 DOUT << " into " << *I;
2106 std::string name = I->getParent()->getName();
2107 DOUT << "push (%" << name << ")\n";
2108 Forwards visit(this, node);
2110 DOUT << "pop (%" << name << ")\n";
2114 bool PredicateSimplifier::runOnFunction(Function &F) {
2115 Forest = &getAnalysis<ETForest>();
2117 TargetData *TD = &getAnalysis<TargetData>();
2119 // XXX: should only act when numbers are out of date
2120 Forest->updateDFSNumbers();
2122 DOUT << "Entering Function: " << F.getName() << "\n";
2125 BasicBlock *RootBlock = &F.getEntryBlock();
2126 IG = new InequalityGraph(Forest->getNodeForBlock(RootBlock));
2127 VR = new ValueRanges(TD);
2128 WorkList.push_back(Forest->getRoot());
2131 BasicBlock *node = WorkList.back();
2132 WorkList.pop_back();
2133 if (!UB.isDead(node)) visitBasicBlock(node);
2134 } while (!WorkList.empty());
2139 modified |= UB.kill();
2144 void PredicateSimplifier::Forwards::visitTerminatorInst(TerminatorInst &TI) {
2145 PS->proceedToSuccessors(Node);
2148 void PredicateSimplifier::Forwards::visitBranchInst(BranchInst &BI) {
2149 if (BI.isUnconditional()) {
2150 PS->proceedToSuccessors(Node);
2154 Value *Condition = BI.getCondition();
2155 BasicBlock *TrueDest = BI.getSuccessor(0);
2156 BasicBlock *FalseDest = BI.getSuccessor(1);
2158 if (isa<Constant>(Condition) || TrueDest == FalseDest) {
2159 PS->proceedToSuccessors(Node);
2163 std::vector<BasicBlock*> Children;
2164 PS->Forest->getChildren(Node, Children);
2165 for (std::vector<BasicBlock*>::iterator I = Children.begin(),
2166 E = Children.end(); I != E; ++I) {
2167 BasicBlock *Dest = *I;
2168 DOUT << "Branch thinking about %" << Dest->getName()
2169 << "(" << PS->Forest->getNodeForBlock(Dest)->getDFSNumIn() << ")\n";
2171 if (Dest == TrueDest) {
2172 DOUT << "(" << Node->getName() << ") true set:\n";
2173 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, Dest);
2174 VRP.add(ConstantInt::getTrue(), Condition, ICmpInst::ICMP_EQ);
2177 } else if (Dest == FalseDest) {
2178 DOUT << "(" << Node->getName() << ") false set:\n";
2179 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, Dest);
2180 VRP.add(ConstantInt::getFalse(), Condition, ICmpInst::ICMP_EQ);
2185 PS->proceedToSuccessor(*I);
2189 void PredicateSimplifier::Forwards::visitSwitchInst(SwitchInst &SI) {
2190 Value *Condition = SI.getCondition();
2192 // Set the EQProperty in each of the cases BBs, and the NEProperties
2193 // in the default BB.
2194 std::vector<BasicBlock*> Children;
2195 PS->Forest->getChildren(Node, Children);
2196 for (std::vector<BasicBlock*>::iterator I = Children.begin(),
2197 E = Children.end(); I != E; ++I) {
2198 BasicBlock *BB = *I;
2199 DOUT << "Switch thinking about BB %" << BB->getName()
2200 << "(" << PS->Forest->getNodeForBlock(BB)->getDFSNumIn() << ")\n";
2202 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, BB);
2203 if (BB == SI.getDefaultDest()) {
2204 for (unsigned i = 1, e = SI.getNumCases(); i < e; ++i)
2205 if (SI.getSuccessor(i) != BB)
2206 VRP.add(Condition, SI.getCaseValue(i), ICmpInst::ICMP_NE);
2208 } else if (ConstantInt *CI = SI.findCaseDest(BB)) {
2209 VRP.add(Condition, CI, ICmpInst::ICMP_EQ);
2212 PS->proceedToSuccessor(*I);
2216 void PredicateSimplifier::Forwards::visitAllocaInst(AllocaInst &AI) {
2217 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &AI);
2218 VRP.add(Constant::getNullValue(AI.getType()), &AI, ICmpInst::ICMP_NE);
2222 void PredicateSimplifier::Forwards::visitLoadInst(LoadInst &LI) {
2223 Value *Ptr = LI.getPointerOperand();
2224 // avoid "load uint* null" -> null NE null.
2225 if (isa<Constant>(Ptr)) return;
2227 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &LI);
2228 VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
2232 void PredicateSimplifier::Forwards::visitStoreInst(StoreInst &SI) {
2233 Value *Ptr = SI.getPointerOperand();
2234 if (isa<Constant>(Ptr)) return;
2236 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &SI);
2237 VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
2241 void PredicateSimplifier::Forwards::visitSExtInst(SExtInst &SI) {
2242 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &SI);
2243 uint32_t SrcBitWidth = cast<IntegerType>(SI.getSrcTy())->getBitWidth();
2244 uint32_t DstBitWidth = cast<IntegerType>(SI.getDestTy())->getBitWidth();
2245 APInt Min(APInt::getHighBitsSet(DstBitWidth, DstBitWidth-SrcBitWidth+1));
2246 APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth-1));
2247 VRP.add(ConstantInt::get(Min), &SI, ICmpInst::ICMP_SLE);
2248 VRP.add(ConstantInt::get(Max), &SI, ICmpInst::ICMP_SGE);
2252 void PredicateSimplifier::Forwards::visitZExtInst(ZExtInst &ZI) {
2253 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &ZI);
2254 uint32_t SrcBitWidth = cast<IntegerType>(ZI.getSrcTy())->getBitWidth();
2255 uint32_t DstBitWidth = cast<IntegerType>(ZI.getDestTy())->getBitWidth();
2256 APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth));
2257 VRP.add(ConstantInt::get(Max), &ZI, ICmpInst::ICMP_UGE);
2261 void PredicateSimplifier::Forwards::visitBinaryOperator(BinaryOperator &BO) {
2262 Instruction::BinaryOps ops = BO.getOpcode();
2266 case Instruction::URem:
2267 case Instruction::SRem:
2268 case Instruction::UDiv:
2269 case Instruction::SDiv: {
2270 Value *Divisor = BO.getOperand(1);
2271 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2272 VRP.add(Constant::getNullValue(Divisor->getType()), Divisor,
2281 case Instruction::Shl: {
2282 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2283 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE);
2286 case Instruction::AShr: {
2287 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2288 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_SLE);
2291 case Instruction::LShr:
2292 case Instruction::UDiv: {
2293 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2294 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE);
2297 case Instruction::URem: {
2298 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2299 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE);
2302 case Instruction::And: {
2303 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2304 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE);
2305 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE);
2308 case Instruction::Or: {
2309 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2310 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE);
2311 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_UGE);
2317 void PredicateSimplifier::Forwards::visitICmpInst(ICmpInst &IC) {
2318 // If possible, squeeze the ICmp predicate into something simpler.
2319 // Eg., if x = [0, 4) and we're being asked icmp uge %x, 3 then change
2320 // the predicate to eq.
2322 // XXX: once we do full PHI handling, modifying the instruction in the
2323 // Forwards visitor will cause missed optimizations.
2325 ICmpInst::Predicate Pred = IC.getPredicate();
2329 case ICmpInst::ICMP_ULE: Pred = ICmpInst::ICMP_ULT; break;
2330 case ICmpInst::ICMP_UGE: Pred = ICmpInst::ICMP_UGT; break;
2331 case ICmpInst::ICMP_SLE: Pred = ICmpInst::ICMP_SLT; break;
2332 case ICmpInst::ICMP_SGE: Pred = ICmpInst::ICMP_SGT; break;
2334 if (Pred != IC.getPredicate()) {
2335 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &IC);
2336 if (VRP.isRelatedBy(IC.getOperand(1), IC.getOperand(0),
2337 ICmpInst::ICMP_NE)) {
2339 PS->modified = true;
2340 IC.setPredicate(Pred);
2344 Pred = IC.getPredicate();
2346 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(IC.getOperand(1))) {
2347 ConstantInt *NextVal = 0;
2350 case ICmpInst::ICMP_SLT:
2351 case ICmpInst::ICMP_ULT:
2352 if (Op1->getValue() != 0)
2353 NextVal = ConstantInt::get(Op1->getValue()-1);
2355 case ICmpInst::ICMP_SGT:
2356 case ICmpInst::ICMP_UGT:
2357 if (!Op1->getValue().isAllOnesValue())
2358 NextVal = ConstantInt::get(Op1->getValue()+1);
2363 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &IC);
2364 if (VRP.isRelatedBy(IC.getOperand(0), NextVal,
2365 ICmpInst::getInversePredicate(Pred))) {
2366 ICmpInst *NewIC = new ICmpInst(ICmpInst::ICMP_EQ, IC.getOperand(0),
2368 NewIC->takeName(&IC);
2369 IC.replaceAllUsesWith(NewIC);
2370 IG.remove(&IC); // XXX: prove this isn't necessary
2371 IC.eraseFromParent();
2373 PS->modified = true;
2379 RegisterPass<PredicateSimplifier> X("predsimplify",
2380 "Predicate Simplifier");
2383 FunctionPass *llvm::createPredicateSimplifierPass() {
2384 return new PredicateSimplifier();