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.
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/Analysis/ET-Forest.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/Transforms/Utils/Local.h"
104 using namespace llvm;
106 STATISTIC(NumVarsReplaced, "Number of argument substitutions");
107 STATISTIC(NumInstruction , "Number of instructions removed");
108 STATISTIC(NumSimple , "Number of simple replacements");
109 STATISTIC(NumBlocks , "Number of blocks marked unreachable");
110 STATISTIC(NumSnuggle , "Number of comparisons snuggled");
113 // SLT SGT ULT UGT EQ
114 // 0 1 0 1 0 -- GT 10
115 // 0 1 0 1 1 -- GE 11
116 // 0 1 1 0 0 -- SGTULT 12
117 // 0 1 1 0 1 -- SGEULE 13
118 // 0 1 1 1 0 -- SGT 14
119 // 0 1 1 1 1 -- SGE 15
120 // 1 0 0 1 0 -- SLTUGT 18
121 // 1 0 0 1 1 -- SLEUGE 19
122 // 1 0 1 0 0 -- LT 20
123 // 1 0 1 0 1 -- LE 21
124 // 1 0 1 1 0 -- SLT 22
125 // 1 0 1 1 1 -- SLE 23
126 // 1 1 0 1 0 -- UGT 26
127 // 1 1 0 1 1 -- UGE 27
128 // 1 1 1 0 0 -- ULT 28
129 // 1 1 1 0 1 -- ULE 29
130 // 1 1 1 1 0 -- NE 30
132 EQ_BIT = 1, UGT_BIT = 2, ULT_BIT = 4, SGT_BIT = 8, SLT_BIT = 16
135 GT = SGT_BIT | UGT_BIT,
137 LT = SLT_BIT | ULT_BIT,
139 NE = SLT_BIT | SGT_BIT | ULT_BIT | UGT_BIT,
140 SGTULT = SGT_BIT | ULT_BIT,
141 SGEULE = SGTULT | EQ_BIT,
142 SLTUGT = SLT_BIT | UGT_BIT,
143 SLEUGE = SLTUGT | EQ_BIT,
144 ULT = SLT_BIT | SGT_BIT | ULT_BIT,
145 UGT = SLT_BIT | SGT_BIT | UGT_BIT,
146 SLT = SLT_BIT | ULT_BIT | UGT_BIT,
147 SGT = SGT_BIT | ULT_BIT | UGT_BIT,
154 static bool validPredicate(LatticeVal LV) {
156 case GT: case GE: case LT: case LE: case NE:
157 case SGTULT: case SGT: case SGEULE:
158 case SLTUGT: case SLT: case SLEUGE:
160 case SLE: case SGE: case ULE: case UGE:
167 /// reversePredicate - reverse the direction of the inequality
168 static LatticeVal reversePredicate(LatticeVal LV) {
169 unsigned reverse = LV ^ (SLT_BIT|SGT_BIT|ULT_BIT|UGT_BIT); //preserve EQ_BIT
171 if ((reverse & (SLT_BIT|SGT_BIT)) == 0)
172 reverse |= (SLT_BIT|SGT_BIT);
174 if ((reverse & (ULT_BIT|UGT_BIT)) == 0)
175 reverse |= (ULT_BIT|UGT_BIT);
177 LatticeVal Rev = static_cast<LatticeVal>(reverse);
178 assert(validPredicate(Rev) && "Failed reversing predicate.");
182 /// This is a StrictWeakOrdering predicate that sorts ETNodes by how many
183 /// descendants they have. With this, you can iterate through a list sorted
184 /// by this operation and the first matching entry is the most specific
185 /// match for your basic block. The order provided is stable; ETNodes with
186 /// the same number of children are sorted by pointer address.
187 struct VISIBILITY_HIDDEN OrderByDominance {
188 bool operator()(const ETNode *LHS, const ETNode *RHS) const {
189 unsigned LHS_spread = LHS->getDFSNumOut() - LHS->getDFSNumIn();
190 unsigned RHS_spread = RHS->getDFSNumOut() - RHS->getDFSNumIn();
191 if (LHS_spread != RHS_spread) return LHS_spread < RHS_spread;
192 else return LHS < RHS;
196 /// The InequalityGraph stores the relationships between values.
197 /// Each Value in the graph is assigned to a Node. Nodes are pointer
198 /// comparable for equality. The caller is expected to maintain the logical
199 /// consistency of the system.
201 /// The InequalityGraph class may invalidate Node*s after any mutator call.
202 /// @brief The InequalityGraph stores the relationships between values.
203 class VISIBILITY_HIDDEN InequalityGraph {
206 InequalityGraph(); // DO NOT IMPLEMENT
207 InequalityGraph(InequalityGraph &); // DO NOT IMPLEMENT
209 explicit InequalityGraph(ETNode *TreeRoot) : TreeRoot(TreeRoot) {}
213 /// An Edge is contained inside a Node making one end of the edge implicit
214 /// and contains a pointer to the other end. The edge contains a lattice
215 /// value specifying the relationship and an ETNode specifying the root
216 /// in the dominator tree to which this edge applies.
217 class VISIBILITY_HIDDEN Edge {
219 Edge(unsigned T, LatticeVal V, ETNode *ST)
220 : To(T), LV(V), Subtree(ST) {}
226 bool operator<(const Edge &edge) const {
227 if (To != edge.To) return To < edge.To;
228 else return OrderByDominance()(Subtree, edge.Subtree);
230 bool operator<(unsigned to) const {
235 /// A single node in the InequalityGraph. This stores the canonical Value
236 /// for the node, as well as the relationships with the neighbours.
238 /// @brief A single node in the InequalityGraph.
239 class VISIBILITY_HIDDEN Node {
240 friend class InequalityGraph;
242 typedef SmallVector<Edge, 4> RelationsType;
243 RelationsType Relations;
247 // TODO: can this idea improve performance?
248 //friend class std::vector<Node>;
249 //Node(Node &N) { RelationsType.swap(N.RelationsType); }
252 typedef RelationsType::iterator iterator;
253 typedef RelationsType::const_iterator const_iterator;
255 Node(Value *V) : Canonical(V) {}
261 virtual void dump() const {
262 dump(*cerr.stream());
265 void dump(std::ostream &os) const {
266 os << *getValue() << ":\n";
267 for (Node::const_iterator NI = begin(), NE = end(); NI != NE; ++NI) {
268 static const std::string names[32] =
269 { "000000", "000001", "000002", "000003", "000004", "000005",
270 "000006", "000007", "000008", "000009", " >", " >=",
271 " s>u<", "s>=u<=", " s>", " s>=", "000016", "000017",
272 " s<u>", "s<=u>=", " <", " <=", " s<", " s<=",
273 "000024", "000025", " u>", " u>=", " u<", " u<=",
275 os << " " << names[NI->LV] << " " << NI->To
276 << " (" << NI->Subtree->getDFSNumIn() << ")\n";
282 iterator begin() { return Relations.begin(); }
283 iterator end() { return Relations.end(); }
284 const_iterator begin() const { return Relations.begin(); }
285 const_iterator end() const { return Relations.end(); }
287 iterator find(unsigned n, ETNode *Subtree) {
289 for (iterator I = std::lower_bound(begin(), E, n);
290 I != E && I->To == n; ++I) {
291 if (Subtree->DominatedBy(I->Subtree))
297 const_iterator find(unsigned n, ETNode *Subtree) const {
298 const_iterator E = end();
299 for (const_iterator I = std::lower_bound(begin(), E, n);
300 I != E && I->To == n; ++I) {
301 if (Subtree->DominatedBy(I->Subtree))
307 Value *getValue() const
312 /// Updates the lattice value for a given node. Create a new entry if
313 /// one doesn't exist, otherwise it merges the values. The new lattice
314 /// value must not be inconsistent with any previously existing value.
315 void update(unsigned n, LatticeVal R, ETNode *Subtree) {
316 assert(validPredicate(R) && "Invalid predicate.");
317 iterator I = find(n, Subtree);
319 Edge edge(n, R, Subtree);
320 iterator Insert = std::lower_bound(begin(), end(), edge);
321 Relations.insert(Insert, edge);
323 LatticeVal LV = static_cast<LatticeVal>(I->LV & R);
324 assert(validPredicate(LV) && "Invalid union of lattice values.");
326 if (Subtree != I->Subtree) {
327 assert(Subtree->DominatedBy(I->Subtree) &&
328 "Find returned subtree that doesn't apply.");
330 Edge edge(n, R, Subtree);
331 iterator Insert = std::lower_bound(begin(), end(), edge);
332 Relations.insert(Insert, edge); // invalidates I
333 I = find(n, Subtree);
336 // Also, we have to tighten any edge that Subtree dominates.
337 for (iterator B = begin(); I->To == n; --I) {
338 if (I->Subtree->DominatedBy(Subtree)) {
339 LatticeVal LV = static_cast<LatticeVal>(I->LV & R);
340 assert(validPredicate(LV) && "Invalid union of lattice values.");
351 struct VISIBILITY_HIDDEN NodeMapEdge {
356 NodeMapEdge(Value *V, unsigned index, ETNode *Subtree)
357 : V(V), index(index), Subtree(Subtree) {}
359 bool operator==(const NodeMapEdge &RHS) const {
361 Subtree == RHS.Subtree;
364 bool operator<(const NodeMapEdge &RHS) const {
365 if (V != RHS.V) return V < RHS.V;
366 return OrderByDominance()(Subtree, RHS.Subtree);
369 bool operator<(Value *RHS) const {
374 typedef std::vector<NodeMapEdge> NodeMapType;
377 std::vector<Node> Nodes;
380 /// node - returns the node object at a given index retrieved from getNode.
381 /// Index zero is reserved and may not be passed in here. The pointer
382 /// returned is valid until the next call to newNode or getOrInsertNode.
383 Node *node(unsigned index) {
384 assert(index != 0 && "Zero index is reserved for not found.");
385 assert(index <= Nodes.size() && "Index out of range.");
386 return &Nodes[index-1];
389 /// Returns the node currently representing Value V, or zero if no such
391 unsigned getNode(Value *V, ETNode *Subtree) {
392 NodeMapType::iterator E = NodeMap.end();
393 NodeMapEdge Edge(V, 0, Subtree);
394 NodeMapType::iterator I = std::lower_bound(NodeMap.begin(), E, Edge);
395 while (I != E && I->V == V) {
396 if (Subtree->DominatedBy(I->Subtree))
403 /// getOrInsertNode - always returns a valid node index, creating a node
404 /// to match the Value if needed.
405 unsigned getOrInsertNode(Value *V, ETNode *Subtree) {
406 if (unsigned n = getNode(V, Subtree))
412 /// newNode - creates a new node for a given Value and returns the index.
413 unsigned newNode(Value *V) {
414 Nodes.push_back(Node(V));
416 NodeMapEdge MapEntry = NodeMapEdge(V, Nodes.size(), TreeRoot);
417 assert(!std::binary_search(NodeMap.begin(), NodeMap.end(), MapEntry) &&
418 "Attempt to create a duplicate Node.");
419 NodeMap.insert(std::lower_bound(NodeMap.begin(), NodeMap.end(),
420 MapEntry), MapEntry);
421 return MapEntry.index;
424 /// If the Value is in the graph, return the canonical form. Otherwise,
425 /// return the original Value.
426 Value *canonicalize(Value *V, ETNode *Subtree) {
427 if (isa<Constant>(V)) return V;
429 if (unsigned n = getNode(V, Subtree))
430 return node(n)->getValue();
435 /// isRelatedBy - true iff n1 op n2
436 bool isRelatedBy(unsigned n1, unsigned n2, ETNode *Subtree, LatticeVal LV) {
437 if (n1 == n2) return LV & EQ_BIT;
440 Node::iterator I = N1->find(n2, Subtree), E = N1->end();
441 if (I != E) return (I->LV & LV) == I->LV;
446 // The add* methods assume that your input is logically valid and may
447 // assertion-fail or infinitely loop if you attempt a contradiction.
449 void addEquality(unsigned n, Value *V, ETNode *Subtree) {
450 assert(canonicalize(node(n)->getValue(), Subtree) == node(n)->getValue()
451 && "Node's 'canonical' choice isn't best within this subtree.");
453 // Suppose that we are given "%x -> node #1 (%y)". The problem is that
454 // we may already have "%z -> node #2 (%x)" somewhere above us in the
455 // graph. We need to find those edges and add "%z -> node #1 (%y)"
456 // to keep the lookups canonical.
458 std::vector<Value *> ToRepoint;
459 ToRepoint.push_back(V);
461 if (unsigned Conflict = getNode(V, Subtree)) {
462 // XXX: NodeMap.size() exceeds 68,000 entries compiling kimwitu++!
463 for (NodeMapType::iterator I = NodeMap.begin(), E = NodeMap.end();
465 if (I->index == Conflict && Subtree->DominatedBy(I->Subtree))
466 ToRepoint.push_back(I->V);
470 for (std::vector<Value *>::iterator VI = ToRepoint.begin(),
471 VE = ToRepoint.end(); VI != VE; ++VI) {
474 // XXX: review this code. This may be doing too many insertions.
475 NodeMapEdge Edge(V, n, Subtree);
476 NodeMapType::iterator E = NodeMap.end();
477 NodeMapType::iterator I = std::lower_bound(NodeMap.begin(), E, Edge);
478 if (I == E || I->V != V || I->Subtree != Subtree) {
480 NodeMap.insert(I, Edge);
481 } else if (I != E && I->V == V && I->Subtree == Subtree) {
482 // Update best choice
488 if (isa<Constant>(V)) {
489 if (isa<Constant>(N->getValue())) {
490 assert(V == N->getValue() && "Constant equals different constant?");
497 /// addInequality - Sets n1 op n2.
498 /// It is also an error to call this on an inequality that is already true.
499 void addInequality(unsigned n1, unsigned n2, ETNode *Subtree,
501 assert(n1 != n2 && "A node can't be inequal to itself.");
504 assert(!isRelatedBy(n1, n2, Subtree, reversePredicate(LV1)) &&
505 "Contradictory inequality.");
510 // Suppose we're adding %n1 < %n2. Find all the %a < %n1 and
511 // add %a < %n2 too. This keeps the graph fully connected.
513 // Someone with a head for this sort of logic, please review this.
514 // Given that %x SLTUGT %y and %a SLE %x, what is the relationship
515 // between %a and %y? I believe the below code is correct, but I don't
516 // think it's the most efficient solution.
518 unsigned LV1_s = LV1 & (SLT_BIT|SGT_BIT);
519 unsigned LV1_u = LV1 & (ULT_BIT|UGT_BIT);
520 for (Node::iterator I = N1->begin(), E = N1->end(); I != E; ++I) {
521 if (I->LV != NE && I->To != n2) {
522 ETNode *Local_Subtree = NULL;
523 if (Subtree->DominatedBy(I->Subtree))
524 Local_Subtree = Subtree;
525 else if (I->Subtree->DominatedBy(Subtree))
526 Local_Subtree = I->Subtree;
529 unsigned new_relationship = 0;
530 LatticeVal ILV = reversePredicate(I->LV);
531 unsigned ILV_s = ILV & (SLT_BIT|SGT_BIT);
532 unsigned ILV_u = ILV & (ULT_BIT|UGT_BIT);
534 if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s)
535 new_relationship |= ILV_s;
537 if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u)
538 new_relationship |= ILV_u;
540 if (new_relationship) {
541 if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0)
542 new_relationship |= (SLT_BIT|SGT_BIT);
543 if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0)
544 new_relationship |= (ULT_BIT|UGT_BIT);
545 if ((LV1 & EQ_BIT) && (ILV & EQ_BIT))
546 new_relationship |= EQ_BIT;
548 LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
550 node(I->To)->update(n2, NewLV, Local_Subtree);
551 N2->update(I->To, reversePredicate(NewLV), Local_Subtree);
557 for (Node::iterator I = N2->begin(), E = N2->end(); I != E; ++I) {
558 if (I->LV != NE && I->To != n1) {
559 ETNode *Local_Subtree = NULL;
560 if (Subtree->DominatedBy(I->Subtree))
561 Local_Subtree = Subtree;
562 else if (I->Subtree->DominatedBy(Subtree))
563 Local_Subtree = I->Subtree;
566 unsigned new_relationship = 0;
567 unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT);
568 unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT);
570 if (LV1_s != (SLT_BIT|SGT_BIT) && ILV_s == LV1_s)
571 new_relationship |= ILV_s;
573 if (LV1_u != (ULT_BIT|UGT_BIT) && ILV_u == LV1_u)
574 new_relationship |= ILV_u;
576 if (new_relationship) {
577 if ((new_relationship & (SLT_BIT|SGT_BIT)) == 0)
578 new_relationship |= (SLT_BIT|SGT_BIT);
579 if ((new_relationship & (ULT_BIT|UGT_BIT)) == 0)
580 new_relationship |= (ULT_BIT|UGT_BIT);
581 if ((LV1 & EQ_BIT) && (I->LV & EQ_BIT))
582 new_relationship |= EQ_BIT;
584 LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
586 N1->update(I->To, NewLV, Local_Subtree);
587 node(I->To)->update(n1, reversePredicate(NewLV), Local_Subtree);
594 N1->update(n2, LV1, Subtree);
595 N2->update(n1, reversePredicate(LV1), Subtree);
598 /// remove - Removes a Value from the graph. If the value is the canonical
599 /// choice for a Node, destroys the Node from the graph deleting all edges
600 /// to and from it. This method does not renumber the nodes.
601 void remove(Value *V) {
602 for (unsigned i = 0; i < NodeMap.size();) {
603 NodeMapType::iterator I = NodeMap.begin()+i;
605 Node *N = node(I->index);
606 if (node(I->index)->getValue() == V) {
607 for (Node::iterator NI = N->begin(), NE = N->end(); NI != NE; ++NI){
608 Node::iterator Iter = node(NI->To)->find(I->index, TreeRoot);
610 node(NI->To)->Relations.erase(Iter);
611 Iter = node(NI->To)->find(I->index, TreeRoot);
612 } while (Iter != node(NI->To)->end());
616 N->Relations.clear();
623 virtual ~InequalityGraph() {}
624 virtual void dump() {
625 dump(*cerr.stream());
628 void dump(std::ostream &os) {
629 std::set<Node *> VisitedNodes;
630 for (NodeMapType::const_iterator I = NodeMap.begin(), E = NodeMap.end();
632 Node *N = node(I->index);
633 os << *I->V << " == " << I->index
634 << "(" << I->Subtree->getDFSNumIn() << ")\n";
635 if (VisitedNodes.insert(N).second) {
636 os << I->index << ". ";
637 if (!N->getValue()) os << "(deleted node)\n";
647 /// ValueRanges tracks the known integer ranges and anti-ranges of the nodes
648 /// in the InequalityGraph.
649 class VISIBILITY_HIDDEN ValueRanges {
651 /// A ScopedRange ties an InequalityGraph node with a ConstantRange under
652 /// the scope of a rooted subtree in the dominator tree.
653 class VISIBILITY_HIDDEN ScopedRange {
655 ScopedRange(Value *V, ConstantRange CR, ETNode *ST)
656 : V(V), CR(CR), Subtree(ST) {}
662 bool operator<(const ScopedRange &range) const {
663 if (V != range.V) return V < range.V;
664 else return OrderByDominance()(Subtree, range.Subtree);
667 bool operator<(const Value *value) const {
672 std::vector<ScopedRange> Ranges;
673 typedef std::vector<ScopedRange>::iterator iterator;
675 // XXX: this is a copy of the code in InequalityGraph::Node. Perhaps a
676 // intrusive domtree-scoped container is in order?
678 iterator begin() { return Ranges.begin(); }
679 iterator end() { return Ranges.end(); }
681 iterator find(Value *V, ETNode *Subtree) {
683 for (iterator I = std::lower_bound(begin(), E, V);
684 I != E && I->V == V; ++I) {
685 if (Subtree->DominatedBy(I->Subtree))
691 void update(Value *V, ConstantRange CR, ETNode *Subtree) {
692 assert(!CR.isEmptySet() && "Empty ConstantRange!");
693 if (CR.isFullSet()) return;
695 iterator I = find(V, Subtree);
697 ScopedRange range(V, CR, Subtree);
698 iterator Insert = std::lower_bound(begin(), end(), range);
699 Ranges.insert(Insert, range);
701 CR = CR.intersectWith(I->CR);
702 assert(!CR.isEmptySet() && "Empty intersection of ConstantRanges!");
705 if (Subtree != I->Subtree) {
706 assert(Subtree->DominatedBy(I->Subtree) &&
707 "Find returned subtree that doesn't apply.");
709 ScopedRange range(V, CR, Subtree);
710 iterator Insert = std::lower_bound(begin(), end(), range);
711 Ranges.insert(Insert, range); // invalidates I
712 I = find(V, Subtree);
715 // Also, we have to tighten any edge that Subtree dominates.
716 for (iterator B = begin(); I->V == V; --I) {
717 if (I->Subtree->DominatedBy(Subtree)) {
718 CR = CR.intersectWith(I->CR);
719 assert(!CR.isEmptySet() &&
720 "Empty intersection of ConstantRanges!");
729 /// range - Creates a ConstantRange representing the set of all values
730 /// that match the ICmpInst::Predicate with any of the values in CR.
731 ConstantRange range(ICmpInst::Predicate ICmpOpcode,
732 const ConstantRange &CR) {
733 uint32_t W = CR.getBitWidth();
734 switch (ICmpOpcode) {
735 default: assert(!"Invalid ICmp opcode to range()");
736 case ICmpInst::ICMP_EQ:
737 return ConstantRange(CR.getLower(), CR.getUpper());
738 case ICmpInst::ICMP_NE:
739 if (CR.isSingleElement())
740 return ConstantRange(CR.getUpper(), CR.getLower());
741 return ConstantRange(W);
742 case ICmpInst::ICMP_ULT:
743 return ConstantRange(APInt::getMinValue(W), CR.getUnsignedMax());
744 case ICmpInst::ICMP_SLT:
745 return ConstantRange(APInt::getSignedMinValue(W), CR.getSignedMax());
746 case ICmpInst::ICMP_ULE: {
747 APInt UMax = CR.getUnsignedMax();
748 if (UMax == APInt::getMaxValue(W))
749 return ConstantRange(W);
750 return ConstantRange(APInt::getMinValue(W), UMax + 1);
752 case ICmpInst::ICMP_SLE: {
753 APInt SMax = CR.getSignedMax();
754 if (SMax == APInt::getSignedMaxValue(W) ||
755 SMax + 1 == APInt::getSignedMaxValue(W))
756 return ConstantRange(W);
757 return ConstantRange(APInt::getSignedMinValue(W), SMax + 1);
759 case ICmpInst::ICMP_UGT:
760 return ConstantRange(CR.getUnsignedMin() + 1,
761 APInt::getMaxValue(W) + 1);
762 case ICmpInst::ICMP_SGT:
763 return ConstantRange(CR.getSignedMin() + 1,
764 APInt::getSignedMaxValue(W) + 1);
765 case ICmpInst::ICMP_UGE: {
766 APInt UMin = CR.getUnsignedMin();
767 if (UMin == APInt::getMinValue(W))
768 return ConstantRange(W);
769 return ConstantRange(UMin, APInt::getMaxValue(W) + 1);
771 case ICmpInst::ICMP_SGE: {
772 APInt SMin = CR.getSignedMin();
773 if (SMin == APInt::getSignedMinValue(W))
774 return ConstantRange(W);
775 return ConstantRange(SMin, APInt::getSignedMaxValue(W) + 1);
780 /// create - Creates a ConstantRange that matches the given LatticeVal
781 /// relation with a given integer.
782 ConstantRange create(LatticeVal LV, const ConstantRange &CR) {
783 assert(!CR.isEmptySet() && "Can't deal with empty set.");
786 return range(ICmpInst::ICMP_NE, CR);
788 unsigned LV_s = LV & (SGT_BIT|SLT_BIT);
789 unsigned LV_u = LV & (UGT_BIT|ULT_BIT);
790 bool hasEQ = LV & EQ_BIT;
792 ConstantRange Range(CR.getBitWidth());
794 if (LV_s == SGT_BIT) {
795 Range = Range.intersectWith(range(
796 hasEQ ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SGT, CR));
797 } else if (LV_s == SLT_BIT) {
798 Range = Range.intersectWith(range(
799 hasEQ ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_SLT, CR));
802 if (LV_u == UGT_BIT) {
803 Range = Range.intersectWith(range(
804 hasEQ ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_UGT, CR));
805 } else if (LV_u == ULT_BIT) {
806 Range = Range.intersectWith(range(
807 hasEQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT, CR));
813 ConstantRange rangeFromValue(Value *V, ETNode *Subtree, uint32_t W) {
814 ConstantInt *C = dyn_cast<ConstantInt>(V);
816 return ConstantRange(C->getValue());
818 iterator I = find(V, Subtree);
822 return ConstantRange(W);
825 static uint32_t widthOfValue(Value *V) {
826 const Type *Ty = V->getType();
827 if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty))
828 return ITy->getBitWidth();
830 // XXX: I'd like to transform T* into the appropriate integer by
831 // bit length, however that data may not be available.
837 bool isCanonical(Value *V, ETNode *Subtree, VRPSolver *VRP);
842 bool isRelatedBy(Value *V1, Value *V2, ETNode *Subtree, LatticeVal LV) {
843 uint32_t W = widthOfValue(V1);
844 if (!W) return false;
846 ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
847 ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
849 // True iff all values in CR1 are LV to all values in CR2.
851 default: assert(!"Impossible lattice value!");
853 return CR1.intersectWith(CR2).isEmptySet();
855 return CR1.getUnsignedMax().ult(CR2.getUnsignedMin());
857 return CR1.getUnsignedMax().ule(CR2.getUnsignedMin());
859 return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax());
861 return CR1.getUnsignedMin().uge(CR2.getUnsignedMax());
863 return CR1.getSignedMax().slt(CR2.getSignedMin());
865 return CR1.getSignedMax().sle(CR2.getSignedMin());
867 return CR1.getSignedMin().sgt(CR2.getSignedMax());
869 return CR1.getSignedMin().sge(CR2.getSignedMax());
871 return CR1.getUnsignedMax().ult(CR2.getUnsignedMin()) &&
872 CR1.getSignedMax().slt(CR2.getUnsignedMin());
874 return CR1.getUnsignedMax().ule(CR2.getUnsignedMin()) &&
875 CR1.getSignedMax().sle(CR2.getUnsignedMin());
877 return CR1.getUnsignedMin().ugt(CR2.getUnsignedMax()) &&
878 CR1.getSignedMin().sgt(CR2.getSignedMax());
880 return CR1.getUnsignedMin().uge(CR2.getUnsignedMax()) &&
881 CR1.getSignedMin().sge(CR2.getSignedMax());
883 return CR1.getSignedMax().slt(CR2.getSignedMin()) &&
884 CR1.getUnsignedMin().ugt(CR2.getUnsignedMax());
886 return CR1.getSignedMax().sle(CR2.getSignedMin()) &&
887 CR1.getUnsignedMin().uge(CR2.getUnsignedMax());
889 return CR1.getSignedMin().sgt(CR2.getSignedMax()) &&
890 CR1.getUnsignedMax().ult(CR2.getUnsignedMin());
892 return CR1.getSignedMin().sge(CR2.getSignedMax()) &&
893 CR1.getUnsignedMax().ule(CR2.getUnsignedMin());
897 void addToWorklist(Value *V, const APInt *I, ICmpInst::Predicate Pred,
900 void mergeInto(Value **I, unsigned n, Value *New, ETNode *Subtree,
902 assert(isCanonical(New, Subtree, VRP) && "Best choice not canonical?");
904 uint32_t W = widthOfValue(New);
907 ConstantRange CR_New = rangeFromValue(New, Subtree, W);
908 ConstantRange Merged = CR_New;
910 for (; n != 0; ++I, --n) {
911 ConstantRange CR_Kill = rangeFromValue(*I, Subtree, W);
912 if (CR_Kill.isFullSet()) continue;
913 Merged = Merged.intersectWith(CR_Kill);
916 if (Merged.isFullSet() || Merged == CR_New) return;
918 if (Merged.isSingleElement())
919 addToWorklist(New, Merged.getSingleElement(),
920 ICmpInst::ICMP_EQ, VRP);
922 update(New, Merged, Subtree);
925 void addInequality(Value *V1, Value *V2, ETNode *Subtree, LatticeVal LV,
927 assert(!isRelatedBy(V1, V2, Subtree, LV) && "Asked to do useless work.");
929 assert(isCanonical(V1, Subtree, VRP) && "Value not canonical.");
930 assert(isCanonical(V2, Subtree, VRP) && "Value not canonical.");
932 if (LV == NE) return; // we can't represent those.
933 // XXX: except in the case where isSingleElement and equal to either
934 // Lower or Upper. That's probably not profitable. (Type::Int1Ty?)
936 uint32_t W = widthOfValue(V1);
939 ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
940 ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
942 if (!CR1.isSingleElement()) {
943 ConstantRange NewCR1 = CR1.intersectWith(create(LV, CR2));
945 if (NewCR1.isSingleElement())
946 addToWorklist(V1, NewCR1.getSingleElement(),
947 ICmpInst::ICMP_EQ, VRP);
949 update(V1, NewCR1, Subtree);
953 if (!CR2.isSingleElement()) {
954 ConstantRange NewCR2 = CR2.intersectWith(create(reversePredicate(LV),
957 if (NewCR2.isSingleElement())
958 addToWorklist(V2, NewCR2.getSingleElement(),
959 ICmpInst::ICMP_EQ, VRP);
961 update(V2, NewCR2, Subtree);
967 /// UnreachableBlocks keeps tracks of blocks that are for one reason or
968 /// another discovered to be unreachable. This is used to cull the graph when
969 /// analyzing instructions, and to mark blocks with the "unreachable"
970 /// terminator instruction after the function has executed.
971 class VISIBILITY_HIDDEN UnreachableBlocks {
973 std::vector<BasicBlock *> DeadBlocks;
976 /// mark - mark a block as dead
977 void mark(BasicBlock *BB) {
978 std::vector<BasicBlock *>::iterator E = DeadBlocks.end();
979 std::vector<BasicBlock *>::iterator I =
980 std::lower_bound(DeadBlocks.begin(), E, BB);
982 if (I == E || *I != BB) DeadBlocks.insert(I, BB);
985 /// isDead - returns whether a block is known to be dead already
986 bool isDead(BasicBlock *BB) {
987 std::vector<BasicBlock *>::iterator E = DeadBlocks.end();
988 std::vector<BasicBlock *>::iterator I =
989 std::lower_bound(DeadBlocks.begin(), E, BB);
991 return I != E && *I == BB;
994 /// kill - replace the dead blocks' terminator with an UnreachableInst.
996 bool modified = false;
997 for (std::vector<BasicBlock *>::iterator I = DeadBlocks.begin(),
998 E = DeadBlocks.end(); I != E; ++I) {
1001 DOUT << "unreachable block: " << BB->getName() << "\n";
1003 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
1005 BasicBlock *Succ = *SI;
1006 Succ->removePredecessor(BB);
1009 TerminatorInst *TI = BB->getTerminator();
1010 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
1011 TI->eraseFromParent();
1012 new UnreachableInst(BB);
1021 /// VRPSolver keeps track of how changes to one variable affect other
1022 /// variables, and forwards changes along to the InequalityGraph. It
1023 /// also maintains the correct choice for "canonical" in the IG.
1024 /// @brief VRPSolver calculates inferences from a new relationship.
1025 class VISIBILITY_HIDDEN VRPSolver {
1027 friend class ValueRanges;
1031 ICmpInst::Predicate Op;
1033 BasicBlock *ContextBB;
1034 Instruction *ContextInst;
1036 std::deque<Operation> WorkList;
1038 InequalityGraph &IG;
1039 UnreachableBlocks &UB;
1045 Instruction *TopInst;
1048 typedef InequalityGraph::Node Node;
1050 /// IdomI - Determines whether one Instruction dominates another.
1051 bool IdomI(Instruction *I1, Instruction *I2) const {
1052 BasicBlock *BB1 = I1->getParent(),
1053 *BB2 = I2->getParent();
1055 if (isa<TerminatorInst>(I1)) return false;
1056 if (isa<TerminatorInst>(I2)) return true;
1057 if (isa<PHINode>(I1) && !isa<PHINode>(I2)) return true;
1058 if (!isa<PHINode>(I1) && isa<PHINode>(I2)) return false;
1060 for (BasicBlock::const_iterator I = BB1->begin(), E = BB1->end();
1062 if (&*I == I1) return true;
1063 if (&*I == I2) return false;
1065 assert(!"Instructions not found in parent BasicBlock?");
1067 return Forest->properlyDominates(BB1, BB2);
1072 /// Returns true if V1 is a better canonical value than V2.
1073 bool compare(Value *V1, Value *V2) const {
1074 if (isa<Constant>(V1))
1075 return !isa<Constant>(V2);
1076 else if (isa<Constant>(V2))
1078 else if (isa<Argument>(V1))
1079 return !isa<Argument>(V2);
1080 else if (isa<Argument>(V2))
1083 Instruction *I1 = dyn_cast<Instruction>(V1);
1084 Instruction *I2 = dyn_cast<Instruction>(V2);
1087 return V1->getNumUses() < V2->getNumUses();
1089 return IdomI(I1, I2);
1092 // below - true if the Instruction is dominated by the current context
1093 // block or instruction
1094 bool below(Instruction *I) {
1096 return IdomI(TopInst, I);
1098 ETNode *Node = Forest->getNodeForBlock(I->getParent());
1099 return Node->DominatedBy(Top);
1103 bool makeEqual(Value *V1, Value *V2) {
1104 DOUT << "makeEqual(" << *V1 << ", " << *V2 << ")\n";
1106 if (V1 == V2) return true;
1108 if (isa<Constant>(V1) && isa<Constant>(V2))
1111 unsigned n1 = IG.getNode(V1, Top), n2 = IG.getNode(V2, Top);
1114 if (n1 == n2) return true;
1115 if (IG.isRelatedBy(n1, n2, Top, NE)) return false;
1118 if (n1) assert(V1 == IG.node(n1)->getValue() && "Value isn't canonical.");
1119 if (n2) assert(V2 == IG.node(n2)->getValue() && "Value isn't canonical.");
1121 assert(!compare(V2, V1) && "Please order parameters to makeEqual.");
1123 assert(!isa<Constant>(V2) && "Tried to remove a constant.");
1125 SetVector<unsigned> Remove;
1126 if (n2) Remove.insert(n2);
1129 // Suppose we're being told that %x == %y, and %x <= %z and %y >= %z.
1130 // We can't just merge %x and %y because the relationship with %z would
1131 // be EQ and that's invalid. What we're doing is looking for any nodes
1132 // %z such that %x <= %z and %y >= %z, and vice versa.
1134 Node *N1 = IG.node(n1);
1135 Node *N2 = IG.node(n2);
1136 Node::iterator end = N2->end();
1138 // Find the intersection between N1 and N2 which is dominated by
1139 // Top. If we find %x where N1 <= %x <= N2 (or >=) then add %x to
1141 for (Node::iterator I = N1->begin(), E = N1->end(); I != E; ++I) {
1142 if (!(I->LV & EQ_BIT) || !Top->DominatedBy(I->Subtree)) continue;
1144 unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT);
1145 unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT);
1146 Node::iterator NI = N2->find(I->To, Top);
1148 LatticeVal NILV = reversePredicate(NI->LV);
1149 unsigned NILV_s = NILV & (SLT_BIT|SGT_BIT);
1150 unsigned NILV_u = NILV & (ULT_BIT|UGT_BIT);
1152 if ((ILV_s != (SLT_BIT|SGT_BIT) && ILV_s == NILV_s) ||
1153 (ILV_u != (ULT_BIT|UGT_BIT) && ILV_u == NILV_u))
1154 Remove.insert(I->To);
1158 // See if one of the nodes about to be removed is actually a better
1159 // canonical choice than n1.
1160 unsigned orig_n1 = n1;
1161 SetVector<unsigned>::iterator DontRemove = Remove.end();
1162 for (SetVector<unsigned>::iterator I = Remove.begin()+1 /* skip n2 */,
1163 E = Remove.end(); I != E; ++I) {
1165 Value *V = IG.node(n)->getValue();
1166 if (compare(V, V1)) {
1172 if (DontRemove != Remove.end()) {
1173 unsigned n = *DontRemove;
1175 Remove.insert(orig_n1);
1179 // We'd like to allow makeEqual on two values to perform a simple
1180 // substitution without every creating nodes in the IG whenever possible.
1182 // The first iteration through this loop operates on V2 before going
1183 // through the Remove list and operating on those too. If all of the
1184 // iterations performed simple replacements then we exit early.
1185 bool mergeIGNode = false;
1187 for (Value *R = V2; i == 0 || i < Remove.size(); ++i) {
1188 if (i) R = IG.node(Remove[i])->getValue(); // skip n2.
1190 // Try to replace the whole instruction. If we can, we're done.
1191 Instruction *I2 = dyn_cast<Instruction>(R);
1192 if (I2 && below(I2)) {
1193 std::vector<Instruction *> ToNotify;
1194 for (Value::use_iterator UI = R->use_begin(), UE = R->use_end();
1196 Use &TheUse = UI.getUse();
1198 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser()))
1199 ToNotify.push_back(I);
1202 DOUT << "Simply removing " << *I2
1203 << ", replacing with " << *V1 << "\n";
1204 I2->replaceAllUsesWith(V1);
1205 // leave it dead; it'll get erased later.
1209 for (std::vector<Instruction *>::iterator II = ToNotify.begin(),
1210 IE = ToNotify.end(); II != IE; ++II) {
1217 // Otherwise, replace all dominated uses.
1218 for (Value::use_iterator UI = R->use_begin(), UE = R->use_end();
1220 Use &TheUse = UI.getUse();
1222 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1232 // If that killed the instruction, stop here.
1233 if (I2 && isInstructionTriviallyDead(I2)) {
1234 DOUT << "Killed all uses of " << *I2
1235 << ", replacing with " << *V1 << "\n";
1239 // If we make it to here, then we will need to create a node for N1.
1240 // Otherwise, we can skip out early!
1244 if (!isa<Constant>(V1)) {
1245 if (Remove.empty()) {
1246 VR.mergeInto(&V2, 1, V1, Top, this);
1248 std::vector<Value*> RemoveVals;
1249 RemoveVals.reserve(Remove.size());
1251 for (SetVector<unsigned>::iterator I = Remove.begin(),
1252 E = Remove.end(); I != E; ++I) {
1253 Value *V = IG.node(*I)->getValue();
1254 if (!V->use_empty())
1255 RemoveVals.push_back(V);
1257 VR.mergeInto(&RemoveVals[0], RemoveVals.size(), V1, Top, this);
1263 if (!n1) n1 = IG.newNode(V1);
1265 // Migrate relationships from removed nodes to N1.
1266 Node *N1 = IG.node(n1);
1267 for (SetVector<unsigned>::iterator I = Remove.begin(), E = Remove.end();
1270 Node *N = IG.node(n);
1271 for (Node::iterator NI = N->begin(), NE = N->end(); NI != NE; ++NI) {
1272 if (NI->Subtree->DominatedBy(Top)) {
1274 assert((NI->LV & EQ_BIT) && "Node inequal to itself.");
1277 if (Remove.count(NI->To))
1280 IG.node(NI->To)->update(n1, reversePredicate(NI->LV), Top);
1281 N1->update(NI->To, NI->LV, Top);
1286 // Point V2 (and all items in Remove) to N1.
1288 IG.addEquality(n1, V2, Top);
1290 for (SetVector<unsigned>::iterator I = Remove.begin(),
1291 E = Remove.end(); I != E; ++I) {
1292 IG.addEquality(n1, IG.node(*I)->getValue(), Top);
1296 // If !Remove.empty() then V2 = Remove[0]->getValue().
1297 // Even when Remove is empty, we still want to process V2.
1299 for (Value *R = V2; i == 0 || i < Remove.size(); ++i) {
1300 if (i) R = IG.node(Remove[i])->getValue(); // skip n2.
1302 if (Instruction *I2 = dyn_cast<Instruction>(R)) {
1304 Top->DominatedBy(Forest->getNodeForBlock(I2->getParent())))
1307 for (Value::use_iterator UI = V2->use_begin(), UE = V2->use_end();
1309 Use &TheUse = UI.getUse();
1311 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1313 Top->DominatedBy(Forest->getNodeForBlock(I->getParent())))
1320 // re-opsToDef all dominated users of V1.
1321 if (Instruction *I = dyn_cast<Instruction>(V1)) {
1322 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1324 Use &TheUse = UI.getUse();
1326 Value *V = TheUse.getUser();
1327 if (!V->use_empty()) {
1328 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
1330 Top->DominatedBy(Forest->getNodeForBlock(Inst->getParent())))
1340 /// cmpInstToLattice - converts an CmpInst::Predicate to lattice value
1341 /// Requires that the lattice value be valid; does not accept ICMP_EQ.
1342 static LatticeVal cmpInstToLattice(ICmpInst::Predicate Pred) {
1344 case ICmpInst::ICMP_EQ:
1345 assert(!"No matching lattice value.");
1346 return static_cast<LatticeVal>(EQ_BIT);
1348 assert(!"Invalid 'icmp' predicate.");
1349 case ICmpInst::ICMP_NE:
1351 case ICmpInst::ICMP_UGT:
1353 case ICmpInst::ICMP_UGE:
1355 case ICmpInst::ICMP_ULT:
1357 case ICmpInst::ICMP_ULE:
1359 case ICmpInst::ICMP_SGT:
1361 case ICmpInst::ICMP_SGE:
1363 case ICmpInst::ICMP_SLT:
1365 case ICmpInst::ICMP_SLE:
1371 VRPSolver(InequalityGraph &IG, UnreachableBlocks &UB, ValueRanges &VR,
1372 ETForest *Forest, bool &modified, BasicBlock *TopBB)
1377 Top(Forest->getNodeForBlock(TopBB)),
1380 modified(modified) {}
1382 VRPSolver(InequalityGraph &IG, UnreachableBlocks &UB, ValueRanges &VR,
1383 ETForest *Forest, bool &modified, Instruction *TopInst)
1391 TopBB = TopInst->getParent();
1392 Top = Forest->getNodeForBlock(TopBB);
1395 bool isRelatedBy(Value *V1, Value *V2, ICmpInst::Predicate Pred) const {
1396 if (Constant *C1 = dyn_cast<Constant>(V1))
1397 if (Constant *C2 = dyn_cast<Constant>(V2))
1398 return ConstantExpr::getCompare(Pred, C1, C2) ==
1399 ConstantInt::getTrue();
1401 if (unsigned n1 = IG.getNode(V1, Top))
1402 if (unsigned n2 = IG.getNode(V2, Top)) {
1403 if (n1 == n2) return Pred == ICmpInst::ICMP_EQ ||
1404 Pred == ICmpInst::ICMP_ULE ||
1405 Pred == ICmpInst::ICMP_UGE ||
1406 Pred == ICmpInst::ICMP_SLE ||
1407 Pred == ICmpInst::ICMP_SGE;
1408 if (Pred == ICmpInst::ICMP_EQ) return false;
1409 if (IG.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true;
1412 if (Pred == ICmpInst::ICMP_EQ) return V1 == V2;
1413 return VR.isRelatedBy(V1, V2, Top, cmpInstToLattice(Pred));
1416 /// add - adds a new property to the work queue
1417 void add(Value *V1, Value *V2, ICmpInst::Predicate Pred,
1418 Instruction *I = NULL) {
1419 DOUT << "adding " << *V1 << " " << Pred << " " << *V2;
1420 if (I) DOUT << " context: " << *I;
1421 else DOUT << " default context";
1424 WorkList.push_back(Operation());
1425 Operation &O = WorkList.back();
1426 O.LHS = V1, O.RHS = V2, O.Op = Pred, O.ContextInst = I;
1427 O.ContextBB = I ? I->getParent() : TopBB;
1430 /// defToOps - Given an instruction definition that we've learned something
1431 /// new about, find any new relationships between its operands.
1432 void defToOps(Instruction *I) {
1433 Instruction *NewContext = below(I) ? I : TopInst;
1434 Value *Canonical = IG.canonicalize(I, Top);
1436 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1437 const Type *Ty = BO->getType();
1438 assert(!Ty->isFPOrFPVector() && "Float in work queue!");
1440 Value *Op0 = IG.canonicalize(BO->getOperand(0), Top);
1441 Value *Op1 = IG.canonicalize(BO->getOperand(1), Top);
1443 // TODO: "and i32 -1, %x" EQ %y then %x EQ %y.
1445 switch (BO->getOpcode()) {
1446 case Instruction::And: {
1447 // "and i32 %a, %b" EQ -1 then %a EQ -1 and %b EQ -1
1448 ConstantInt *CI = ConstantInt::getAllOnesValue(Ty);
1449 if (Canonical == CI) {
1450 add(CI, Op0, ICmpInst::ICMP_EQ, NewContext);
1451 add(CI, Op1, ICmpInst::ICMP_EQ, NewContext);
1454 case Instruction::Or: {
1455 // "or i32 %a, %b" EQ 0 then %a EQ 0 and %b EQ 0
1456 Constant *Zero = Constant::getNullValue(Ty);
1457 if (Canonical == Zero) {
1458 add(Zero, Op0, ICmpInst::ICMP_EQ, NewContext);
1459 add(Zero, Op1, ICmpInst::ICMP_EQ, NewContext);
1462 case Instruction::Xor: {
1463 // "xor i32 %c, %a" EQ %b then %a EQ %c ^ %b
1464 // "xor i32 %c, %a" EQ %c then %a EQ 0
1465 // "xor i32 %c, %a" NE %c then %a NE 0
1466 // Repeat the above, with order of operands reversed.
1469 if (!isa<Constant>(LHS)) std::swap(LHS, RHS);
1471 if (ConstantInt *CI = dyn_cast<ConstantInt>(Canonical)) {
1472 if (ConstantInt *Arg = dyn_cast<ConstantInt>(LHS)) {
1473 add(RHS, ConstantInt::get(CI->getValue() ^ Arg->getValue()),
1474 ICmpInst::ICMP_EQ, NewContext);
1477 if (Canonical == LHS) {
1478 if (isa<ConstantInt>(Canonical))
1479 add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_EQ,
1481 } else if (isRelatedBy(LHS, Canonical, ICmpInst::ICMP_NE)) {
1482 add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_NE,
1489 } else if (ICmpInst *IC = dyn_cast<ICmpInst>(I)) {
1490 // "icmp ult i32 %a, %y" EQ true then %a u< y
1493 if (Canonical == ConstantInt::getTrue()) {
1494 add(IC->getOperand(0), IC->getOperand(1), IC->getPredicate(),
1496 } else if (Canonical == ConstantInt::getFalse()) {
1497 add(IC->getOperand(0), IC->getOperand(1),
1498 ICmpInst::getInversePredicate(IC->getPredicate()), NewContext);
1500 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
1501 if (I->getType()->isFPOrFPVector()) return;
1503 // Given: "%a = select i1 %x, i32 %b, i32 %c"
1504 // %a EQ %b and %b NE %c then %x EQ true
1505 // %a EQ %c and %b NE %c then %x EQ false
1507 Value *True = SI->getTrueValue();
1508 Value *False = SI->getFalseValue();
1509 if (isRelatedBy(True, False, ICmpInst::ICMP_NE)) {
1510 if (Canonical == IG.canonicalize(True, Top) ||
1511 isRelatedBy(Canonical, False, ICmpInst::ICMP_NE))
1512 add(SI->getCondition(), ConstantInt::getTrue(),
1513 ICmpInst::ICMP_EQ, NewContext);
1514 else if (Canonical == IG.canonicalize(False, Top) ||
1515 isRelatedBy(Canonical, True, ICmpInst::ICMP_NE))
1516 add(SI->getCondition(), ConstantInt::getFalse(),
1517 ICmpInst::ICMP_EQ, NewContext);
1519 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1520 for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(),
1521 OE = GEPI->idx_end(); OI != OE; ++OI) {
1522 ConstantInt *Op = dyn_cast<ConstantInt>(IG.canonicalize(*OI, Top));
1523 if (!Op || !Op->isZero()) return;
1525 // TODO: The GEPI indices are all zero. Copy from definition to operand,
1526 // jumping the type plane as needed.
1527 if (isRelatedBy(GEPI, Constant::getNullValue(GEPI->getType()),
1528 ICmpInst::ICMP_NE)) {
1529 Value *Ptr = GEPI->getPointerOperand();
1530 add(Ptr, Constant::getNullValue(Ptr->getType()), ICmpInst::ICMP_NE,
1534 // TODO: CastInst "%a = cast ... %b" where %a is EQ or NE a constant.
1537 /// opsToDef - A new relationship was discovered involving one of this
1538 /// instruction's operands. Find any new relationship involving the
1539 /// definition, or another operand.
1540 void opsToDef(Instruction *I) {
1541 Instruction *NewContext = below(I) ? I : TopInst;
1543 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1544 Value *Op0 = IG.canonicalize(BO->getOperand(0), Top);
1545 Value *Op1 = IG.canonicalize(BO->getOperand(1), Top);
1547 if (ConstantInt *CI0 = dyn_cast<ConstantInt>(Op0))
1548 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Op1)) {
1549 add(BO, ConstantExpr::get(BO->getOpcode(), CI0, CI1),
1550 ICmpInst::ICMP_EQ, NewContext);
1554 // "%y = and i1 true, %x" then %x EQ %y
1555 // "%y = or i1 false, %x" then %x EQ %y
1556 // "%x = add i32 %y, 0" then %x EQ %y
1557 // "%x = mul i32 %y, 0" then %x EQ 0
1559 Instruction::BinaryOps Opcode = BO->getOpcode();
1560 const Type *Ty = BO->getType();
1561 assert(!Ty->isFPOrFPVector() && "Float in work queue!");
1563 Constant *Zero = Constant::getNullValue(Ty);
1564 Constant *AllOnes = ConstantInt::getAllOnesValue(Ty);
1568 case Instruction::LShr:
1569 case Instruction::AShr:
1570 case Instruction::Shl:
1571 case Instruction::Sub:
1573 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1577 case Instruction::Or:
1578 if (Op0 == AllOnes || Op1 == AllOnes) {
1579 add(BO, AllOnes, ICmpInst::ICMP_EQ, NewContext);
1582 case Instruction::Xor:
1583 case Instruction::Add:
1585 add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
1587 } else if (Op1 == Zero) {
1588 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1592 case Instruction::And:
1593 if (Op0 == AllOnes) {
1594 add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
1596 } else if (Op1 == AllOnes) {
1597 add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
1601 case Instruction::Mul:
1602 if (Op0 == Zero || Op1 == Zero) {
1603 add(BO, Zero, ICmpInst::ICMP_EQ, NewContext);
1609 // "%x = add i32 %y, %z" and %x EQ %y then %z EQ 0
1610 // "%x = add i32 %y, %z" and %x EQ %z then %y EQ 0
1611 // "%x = shl i32 %y, %z" and %x EQ %y and %y NE 0 then %z EQ 0
1612 // "%x = udiv i32 %y, %z" and %x EQ %y then %z EQ 1
1614 Value *Known = Op0, *Unknown = Op1,
1615 *TheBO = IG.canonicalize(BO, Top);
1616 if (Known != TheBO) std::swap(Known, Unknown);
1617 if (Known == TheBO) {
1620 case Instruction::LShr:
1621 case Instruction::AShr:
1622 case Instruction::Shl:
1623 if (!isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) break;
1624 // otherwise, fall-through.
1625 case Instruction::Sub:
1626 if (Unknown == Op1) break;
1627 // otherwise, fall-through.
1628 case Instruction::Xor:
1629 case Instruction::Add:
1630 add(Unknown, Zero, ICmpInst::ICMP_EQ, NewContext);
1632 case Instruction::UDiv:
1633 case Instruction::SDiv:
1634 if (Unknown == Op1) break;
1635 if (isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) {
1636 Constant *One = ConstantInt::get(Ty, 1);
1637 add(Unknown, One, ICmpInst::ICMP_EQ, NewContext);
1643 // TODO: "%a = add i32 %b, 1" and %b > %z then %a >= %z.
1645 } else if (ICmpInst *IC = dyn_cast<ICmpInst>(I)) {
1646 // "%a = icmp ult i32 %b, %c" and %b u< %c then %a EQ true
1647 // "%a = icmp ult i32 %b, %c" and %b u>= %c then %a EQ false
1650 Value *Op0 = IG.canonicalize(IC->getOperand(0), Top);
1651 Value *Op1 = IG.canonicalize(IC->getOperand(1), Top);
1653 ICmpInst::Predicate Pred = IC->getPredicate();
1654 if (isRelatedBy(Op0, Op1, Pred)) {
1655 add(IC, ConstantInt::getTrue(), ICmpInst::ICMP_EQ, NewContext);
1656 } else if (isRelatedBy(Op0, Op1, ICmpInst::getInversePredicate(Pred))) {
1657 add(IC, ConstantInt::getFalse(), ICmpInst::ICMP_EQ, NewContext);
1660 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
1661 if (I->getType()->isFPOrFPVector()) return;
1663 // Given: "%a = select i1 %x, i32 %b, i32 %c"
1664 // %x EQ true then %a EQ %b
1665 // %x EQ false then %a EQ %c
1666 // %b EQ %c then %a EQ %b
1668 Value *Canonical = IG.canonicalize(SI->getCondition(), Top);
1669 if (Canonical == ConstantInt::getTrue()) {
1670 add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext);
1671 } else if (Canonical == ConstantInt::getFalse()) {
1672 add(SI, SI->getFalseValue(), ICmpInst::ICMP_EQ, NewContext);
1673 } else if (IG.canonicalize(SI->getTrueValue(), Top) ==
1674 IG.canonicalize(SI->getFalseValue(), Top)) {
1675 add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext);
1677 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
1678 const Type *Ty = CI->getDestTy();
1679 if (Ty->isFPOrFPVector()) return;
1681 if (Constant *C = dyn_cast<Constant>(
1682 IG.canonicalize(CI->getOperand(0), Top))) {
1683 add(CI, ConstantExpr::getCast(CI->getOpcode(), C, Ty),
1684 ICmpInst::ICMP_EQ, NewContext);
1687 // TODO: "%a = cast ... %b" where %b is NE/LT/GT a constant.
1688 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1689 for (GetElementPtrInst::op_iterator OI = GEPI->idx_begin(),
1690 OE = GEPI->idx_end(); OI != OE; ++OI) {
1691 ConstantInt *Op = dyn_cast<ConstantInt>(IG.canonicalize(*OI, Top));
1692 if (!Op || !Op->isZero()) return;
1694 // TODO: The GEPI indices are all zero. Copy from operand to definition,
1695 // jumping the type plane as needed.
1696 Value *Ptr = GEPI->getPointerOperand();
1697 if (isRelatedBy(Ptr, Constant::getNullValue(Ptr->getType()),
1698 ICmpInst::ICMP_NE)) {
1699 add(GEPI, Constant::getNullValue(GEPI->getType()), ICmpInst::ICMP_NE,
1705 /// solve - process the work queue
1706 /// Return false if a logical contradiction occurs.
1708 //DOUT << "WorkList entry, size: " << WorkList.size() << "\n";
1709 while (!WorkList.empty()) {
1710 //DOUT << "WorkList size: " << WorkList.size() << "\n";
1712 Operation &O = WorkList.front();
1713 TopInst = O.ContextInst;
1714 TopBB = O.ContextBB;
1715 Top = Forest->getNodeForBlock(TopBB);
1717 O.LHS = IG.canonicalize(O.LHS, Top);
1718 O.RHS = IG.canonicalize(O.RHS, Top);
1720 assert(O.LHS == IG.canonicalize(O.LHS, Top) && "Canonicalize isn't.");
1721 assert(O.RHS == IG.canonicalize(O.RHS, Top) && "Canonicalize isn't.");
1723 DOUT << "solving " << *O.LHS << " " << O.Op << " " << *O.RHS;
1724 if (O.ContextInst) DOUT << " context inst: " << *O.ContextInst;
1725 else DOUT << " context block: " << O.ContextBB->getName();
1730 // If they're both Constant, skip it. Check for contradiction and mark
1731 // the BB as unreachable if so.
1732 if (Constant *CI_L = dyn_cast<Constant>(O.LHS)) {
1733 if (Constant *CI_R = dyn_cast<Constant>(O.RHS)) {
1734 if (ConstantExpr::getCompare(O.Op, CI_L, CI_R) ==
1735 ConstantInt::getFalse())
1738 WorkList.pop_front();
1743 if (compare(O.LHS, O.RHS)) {
1744 std::swap(O.LHS, O.RHS);
1745 O.Op = ICmpInst::getSwappedPredicate(O.Op);
1748 if (O.Op == ICmpInst::ICMP_EQ) {
1749 if (!makeEqual(O.RHS, O.LHS))
1752 LatticeVal LV = cmpInstToLattice(O.Op);
1754 if ((LV & EQ_BIT) &&
1755 isRelatedBy(O.LHS, O.RHS, ICmpInst::getSwappedPredicate(O.Op))) {
1756 if (!makeEqual(O.RHS, O.LHS))
1759 if (isRelatedBy(O.LHS, O.RHS, ICmpInst::getInversePredicate(O.Op))){
1761 WorkList.pop_front();
1765 unsigned n1 = IG.getNode(O.LHS, Top);
1766 unsigned n2 = IG.getNode(O.RHS, Top);
1768 if (n1 && n1 == n2) {
1769 if (O.Op != ICmpInst::ICMP_UGE && O.Op != ICmpInst::ICMP_ULE &&
1770 O.Op != ICmpInst::ICMP_SGE && O.Op != ICmpInst::ICMP_SLE)
1773 WorkList.pop_front();
1777 if (VR.isRelatedBy(O.LHS, O.RHS, Top, LV) ||
1778 (n1 && n2 && IG.isRelatedBy(n1, n2, Top, LV))) {
1779 WorkList.pop_front();
1783 VR.addInequality(O.LHS, O.RHS, Top, LV, this);
1784 if ((!isa<ConstantInt>(O.RHS) && !isa<ConstantInt>(O.LHS)) ||
1786 if (!n1) n1 = IG.newNode(O.LHS);
1787 if (!n2) n2 = IG.newNode(O.RHS);
1788 IG.addInequality(n1, n2, Top, LV);
1791 if (Instruction *I1 = dyn_cast<Instruction>(O.LHS)) {
1793 Top->DominatedBy(Forest->getNodeForBlock(I1->getParent())))
1796 if (isa<Instruction>(O.LHS) || isa<Argument>(O.LHS)) {
1797 for (Value::use_iterator UI = O.LHS->use_begin(),
1798 UE = O.LHS->use_end(); UI != UE;) {
1799 Use &TheUse = UI.getUse();
1801 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1803 Top->DominatedBy(Forest->getNodeForBlock(I->getParent())))
1808 if (Instruction *I2 = dyn_cast<Instruction>(O.RHS)) {
1810 Top->DominatedBy(Forest->getNodeForBlock(I2->getParent())))
1813 if (isa<Instruction>(O.RHS) || isa<Argument>(O.RHS)) {
1814 for (Value::use_iterator UI = O.RHS->use_begin(),
1815 UE = O.RHS->use_end(); UI != UE;) {
1816 Use &TheUse = UI.getUse();
1818 if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
1820 Top->DominatedBy(Forest->getNodeForBlock(I->getParent())))
1828 WorkList.pop_front();
1833 void ValueRanges::addToWorklist(Value *V, const APInt *I,
1834 ICmpInst::Predicate Pred, VRPSolver *VRP) {
1835 VRP->add(V, ConstantInt::get(*I), Pred, VRP->TopInst);
1839 bool ValueRanges::isCanonical(Value *V, ETNode *Subtree, VRPSolver *VRP) {
1840 return V == VRP->IG.canonicalize(V, Subtree);
1844 /// PredicateSimplifier - This class is a simplifier that replaces
1845 /// one equivalent variable with another. It also tracks what
1846 /// can't be equal and will solve setcc instructions when possible.
1847 /// @brief Root of the predicate simplifier optimization.
1848 class VISIBILITY_HIDDEN PredicateSimplifier : public FunctionPass {
1852 InequalityGraph *IG;
1853 UnreachableBlocks UB;
1856 std::vector<DominatorTree::Node *> WorkList;
1859 bool runOnFunction(Function &F);
1861 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
1862 AU.addRequiredID(BreakCriticalEdgesID);
1863 AU.addRequired<DominatorTree>();
1864 AU.addRequired<ETForest>();
1868 /// Forwards - Adds new properties into PropertySet and uses them to
1869 /// simplify instructions. Because new properties sometimes apply to
1870 /// a transition from one BasicBlock to another, this will use the
1871 /// PredicateSimplifier::proceedToSuccessor(s) interface to enter the
1872 /// basic block with the new PropertySet.
1873 /// @brief Performs abstract execution of the program.
1874 class VISIBILITY_HIDDEN Forwards : public InstVisitor<Forwards> {
1875 friend class InstVisitor<Forwards>;
1876 PredicateSimplifier *PS;
1877 DominatorTree::Node *DTNode;
1880 InequalityGraph &IG;
1881 UnreachableBlocks &UB;
1884 Forwards(PredicateSimplifier *PS, DominatorTree::Node *DTNode)
1885 : PS(PS), DTNode(DTNode), IG(*PS->IG), UB(PS->UB), VR(*PS->VR) {}
1887 void visitTerminatorInst(TerminatorInst &TI);
1888 void visitBranchInst(BranchInst &BI);
1889 void visitSwitchInst(SwitchInst &SI);
1891 void visitAllocaInst(AllocaInst &AI);
1892 void visitLoadInst(LoadInst &LI);
1893 void visitStoreInst(StoreInst &SI);
1895 void visitSExtInst(SExtInst &SI);
1896 void visitZExtInst(ZExtInst &ZI);
1898 void visitBinaryOperator(BinaryOperator &BO);
1899 void visitICmpInst(ICmpInst &IC);
1902 // Used by terminator instructions to proceed from the current basic
1903 // block to the next. Verifies that "current" dominates "next",
1904 // then calls visitBasicBlock.
1905 void proceedToSuccessors(DominatorTree::Node *Current) {
1906 for (DominatorTree::Node::iterator I = Current->begin(),
1907 E = Current->end(); I != E; ++I) {
1908 WorkList.push_back(*I);
1912 void proceedToSuccessor(DominatorTree::Node *Next) {
1913 WorkList.push_back(Next);
1916 // Visits each instruction in the basic block.
1917 void visitBasicBlock(DominatorTree::Node *Node) {
1918 BasicBlock *BB = Node->getBlock();
1919 ETNode *ET = Forest->getNodeForBlock(BB);
1920 DOUT << "Entering Basic Block: " << BB->getName()
1921 << " (" << ET->getDFSNumIn() << ")\n";
1922 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
1923 visitInstruction(I++, Node, ET);
1927 // Tries to simplify each Instruction and add new properties to
1929 void visitInstruction(Instruction *I, DominatorTree::Node *DT, ETNode *ET) {
1930 DOUT << "Considering instruction " << *I << "\n";
1933 // Sometimes instructions are killed in earlier analysis.
1934 if (isInstructionTriviallyDead(I)) {
1938 I->eraseFromParent();
1943 // Try to replace the whole instruction.
1944 Value *V = IG->canonicalize(I, ET);
1945 assert(V == I && "Late instruction canonicalization.");
1949 DOUT << "Removing " << *I << ", replacing with " << *V << "\n";
1951 I->replaceAllUsesWith(V);
1952 I->eraseFromParent();
1956 // Try to substitute operands.
1957 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
1958 Value *Oper = I->getOperand(i);
1959 Value *V = IG->canonicalize(Oper, ET);
1960 assert(V == Oper && "Late operand canonicalization.");
1964 DOUT << "Resolving " << *I;
1965 I->setOperand(i, V);
1966 DOUT << " into " << *I;
1971 std::string name = I->getParent()->getName();
1972 DOUT << "push (%" << name << ")\n";
1973 Forwards visit(this, DT);
1975 DOUT << "pop (%" << name << ")\n";
1979 bool PredicateSimplifier::runOnFunction(Function &F) {
1980 DT = &getAnalysis<DominatorTree>();
1981 Forest = &getAnalysis<ETForest>();
1983 Forest->updateDFSNumbers(); // XXX: should only act when numbers are out of date
1985 DOUT << "Entering Function: " << F.getName() << "\n";
1988 BasicBlock *RootBlock = &F.getEntryBlock();
1989 IG = new InequalityGraph(Forest->getNodeForBlock(RootBlock));
1990 VR = new ValueRanges();
1991 WorkList.push_back(DT->getRootNode());
1994 DominatorTree::Node *DTNode = WorkList.back();
1995 WorkList.pop_back();
1996 if (!UB.isDead(DTNode->getBlock())) visitBasicBlock(DTNode);
1997 } while (!WorkList.empty());
2002 modified |= UB.kill();
2007 void PredicateSimplifier::Forwards::visitTerminatorInst(TerminatorInst &TI) {
2008 PS->proceedToSuccessors(DTNode);
2011 void PredicateSimplifier::Forwards::visitBranchInst(BranchInst &BI) {
2012 if (BI.isUnconditional()) {
2013 PS->proceedToSuccessors(DTNode);
2017 Value *Condition = BI.getCondition();
2018 BasicBlock *TrueDest = BI.getSuccessor(0);
2019 BasicBlock *FalseDest = BI.getSuccessor(1);
2021 if (isa<Constant>(Condition) || TrueDest == FalseDest) {
2022 PS->proceedToSuccessors(DTNode);
2026 for (DominatorTree::Node::iterator I = DTNode->begin(), E = DTNode->end();
2028 BasicBlock *Dest = (*I)->getBlock();
2029 DOUT << "Branch thinking about %" << Dest->getName()
2030 << "(" << PS->Forest->getNodeForBlock(Dest)->getDFSNumIn() << ")\n";
2032 if (Dest == TrueDest) {
2033 DOUT << "(" << DTNode->getBlock()->getName() << ") true set:\n";
2034 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, Dest);
2035 VRP.add(ConstantInt::getTrue(), Condition, ICmpInst::ICMP_EQ);
2038 } else if (Dest == FalseDest) {
2039 DOUT << "(" << DTNode->getBlock()->getName() << ") false set:\n";
2040 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, Dest);
2041 VRP.add(ConstantInt::getFalse(), Condition, ICmpInst::ICMP_EQ);
2046 PS->proceedToSuccessor(*I);
2050 void PredicateSimplifier::Forwards::visitSwitchInst(SwitchInst &SI) {
2051 Value *Condition = SI.getCondition();
2053 // Set the EQProperty in each of the cases BBs, and the NEProperties
2054 // in the default BB.
2056 for (DominatorTree::Node::iterator I = DTNode->begin(), E = DTNode->end();
2058 BasicBlock *BB = (*I)->getBlock();
2059 DOUT << "Switch thinking about BB %" << BB->getName()
2060 << "(" << PS->Forest->getNodeForBlock(BB)->getDFSNumIn() << ")\n";
2062 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, BB);
2063 if (BB == SI.getDefaultDest()) {
2064 for (unsigned i = 1, e = SI.getNumCases(); i < e; ++i)
2065 if (SI.getSuccessor(i) != BB)
2066 VRP.add(Condition, SI.getCaseValue(i), ICmpInst::ICMP_NE);
2068 } else if (ConstantInt *CI = SI.findCaseDest(BB)) {
2069 VRP.add(Condition, CI, ICmpInst::ICMP_EQ);
2072 PS->proceedToSuccessor(*I);
2076 void PredicateSimplifier::Forwards::visitAllocaInst(AllocaInst &AI) {
2077 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &AI);
2078 VRP.add(Constant::getNullValue(AI.getType()), &AI, ICmpInst::ICMP_NE);
2082 void PredicateSimplifier::Forwards::visitLoadInst(LoadInst &LI) {
2083 Value *Ptr = LI.getPointerOperand();
2084 // avoid "load uint* null" -> null NE null.
2085 if (isa<Constant>(Ptr)) return;
2087 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &LI);
2088 VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
2092 void PredicateSimplifier::Forwards::visitStoreInst(StoreInst &SI) {
2093 Value *Ptr = SI.getPointerOperand();
2094 if (isa<Constant>(Ptr)) return;
2096 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &SI);
2097 VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
2101 void PredicateSimplifier::Forwards::visitSExtInst(SExtInst &SI) {
2102 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &SI);
2103 uint32_t SrcBitWidth = cast<IntegerType>(SI.getSrcTy())->getBitWidth();
2104 uint32_t DstBitWidth = cast<IntegerType>(SI.getDestTy())->getBitWidth();
2105 APInt Min(APInt::getSignedMinValue(SrcBitWidth));
2106 APInt Max(APInt::getSignedMaxValue(SrcBitWidth));
2107 Min.sext(DstBitWidth);
2108 Max.sext(DstBitWidth);
2109 VRP.add(ConstantInt::get(Min), &SI, ICmpInst::ICMP_SLE);
2110 VRP.add(ConstantInt::get(Max), &SI, ICmpInst::ICMP_SGE);
2114 void PredicateSimplifier::Forwards::visitZExtInst(ZExtInst &ZI) {
2115 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &ZI);
2116 uint32_t SrcBitWidth = cast<IntegerType>(ZI.getSrcTy())->getBitWidth();
2117 uint32_t DstBitWidth = cast<IntegerType>(ZI.getDestTy())->getBitWidth();
2118 APInt Max(APInt::getMaxValue(SrcBitWidth));
2119 Max.zext(DstBitWidth);
2120 VRP.add(ConstantInt::get(Max), &ZI, ICmpInst::ICMP_UGE);
2124 void PredicateSimplifier::Forwards::visitBinaryOperator(BinaryOperator &BO) {
2125 Instruction::BinaryOps ops = BO.getOpcode();
2129 case Instruction::URem:
2130 case Instruction::SRem:
2131 case Instruction::UDiv:
2132 case Instruction::SDiv: {
2133 Value *Divisor = BO.getOperand(1);
2134 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2135 VRP.add(Constant::getNullValue(Divisor->getType()), Divisor,
2144 case Instruction::Shl: {
2145 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2146 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE);
2149 case Instruction::AShr: {
2150 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2151 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_SLE);
2154 case Instruction::LShr:
2155 case Instruction::UDiv: {
2156 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2157 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE);
2160 case Instruction::URem: {
2161 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2162 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE);
2165 case Instruction::And: {
2166 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2167 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_ULE);
2168 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_ULE);
2171 case Instruction::Or: {
2172 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &BO);
2173 VRP.add(&BO, BO.getOperand(0), ICmpInst::ICMP_UGE);
2174 VRP.add(&BO, BO.getOperand(1), ICmpInst::ICMP_UGE);
2180 void PredicateSimplifier::Forwards::visitICmpInst(ICmpInst &IC) {
2181 // If possible, squeeze the ICmp predicate into something simpler.
2182 // Eg., if x = [0, 4) and we're being asked icmp uge %x, 3 then change
2183 // the predicate to eq.
2185 ICmpInst::Predicate Pred = IC.getPredicate();
2187 if (ConstantInt *Op1 = dyn_cast<ConstantInt>(IC.getOperand(1))) {
2188 ConstantInt *NextVal = 0;
2191 case ICmpInst::ICMP_SLT:
2192 case ICmpInst::ICMP_ULT:
2193 if (Op1->getValue() != 0)
2194 NextVal = cast<ConstantInt>(ConstantExpr::getSub(
2195 Op1, ConstantInt::get(Op1->getType(), 1)));
2197 case ICmpInst::ICMP_SGT:
2198 case ICmpInst::ICMP_UGT:
2199 if (!Op1->getValue().isAllOnesValue())
2200 NextVal = cast<ConstantInt>(ConstantExpr::getAdd(
2201 Op1, ConstantInt::get(Op1->getType(), 1)));
2206 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &IC);
2207 if (VRP.isRelatedBy(IC.getOperand(0), NextVal,
2208 ICmpInst::getInversePredicate(Pred))) {
2209 ICmpInst *NewIC = new ICmpInst(ICmpInst::ICMP_EQ, IC.getOperand(0),
2211 NewIC->takeName(&IC);
2212 IC.replaceAllUsesWith(NewIC);
2213 IG.remove(&IC); // XXX: prove this isn't necessary
2214 IC.eraseFromParent();
2216 PS->modified = true;
2224 case ICmpInst::ICMP_ULE: Pred = ICmpInst::ICMP_ULT; break;
2225 case ICmpInst::ICMP_UGE: Pred = ICmpInst::ICMP_UGT; break;
2226 case ICmpInst::ICMP_SLE: Pred = ICmpInst::ICMP_SLT; break;
2227 case ICmpInst::ICMP_SGE: Pred = ICmpInst::ICMP_SGT; break;
2229 VRPSolver VRP(IG, UB, VR, PS->Forest, PS->modified, &IC);
2230 if (VRP.isRelatedBy(IC.getOperand(1), IC.getOperand(0), Pred)) {
2232 PS->modified = true;
2233 IC.setPredicate(Pred);
2237 RegisterPass<PredicateSimplifier> X("predsimplify",
2238 "Predicate Simplifier");
2241 FunctionPass *llvm::createPredicateSimplifierPass() {
2242 return new PredicateSimplifier();