//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Nick Lewycky and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
//
// The ValueRanges class stores the known integer bounds of a Value. When we
// encounter i8 %a u< %b, the ValueRanges stores that %a = [1, 255] and
-// %b = [0, 254]. Because we store these by Value*, you should always
-// canonicalize through the InequalityGraph first.
+// %b = [0, 254].
//
// It never stores an empty range, because that means that the code is
// unreachable. It never stores a single-element range since that's an equality
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Analysis/Dominators.h"
+#include "llvm/Assembly/Writer.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Transforms/Utils/Local.h"
#include <algorithm>
#include <deque>
-#include <sstream>
#include <stack>
using namespace llvm;
STATISTIC(NumBlocks , "Number of blocks marked unreachable");
STATISTIC(NumSnuggle , "Number of comparisons snuggled");
+static const ConstantRange empty(1, false);
+
namespace {
class DomTreeDFS {
public:
}
}
+ /// getRootNode - This returns the entry node for the CFG of the function.
Node *getRootNode() const { return Entry; }
+ /// getNodeForBlock - return the node for the specified basic block.
Node *getNodeForBlock(BasicBlock *BB) const {
if (!NodeMap.count(BB)) return 0;
return const_cast<DomTreeDFS*>(this)->NodeMap[BB];
}
+ /// dominates - returns true if the basic block for I1 dominates that of
+ /// the basic block for I2. If the instructions belong to the same basic
+ /// block, the instruction first instruction sequentially in the block is
+ /// considered dominating.
bool dominates(Instruction *I1, Instruction *I2) {
BasicBlock *BB1 = I1->getParent(),
*BB2 = I2->getParent();
*Node2 = getNodeForBlock(BB2);
return Node1 && Node2 && Node1->dominates(Node2);
}
+ return false; // Not reached
}
+
private:
+ /// renumber - calculates the depth first search numberings and applies
+ /// them onto the nodes.
void renumber() {
std::stack<std::pair<Node *, Node::iterator> > S;
unsigned n = 0;
UGE = UGT | EQ_BIT
};
+#ifndef NDEBUG
+ /// validPredicate - determines whether a given value is actually a lattice
+ /// value. Only used in assertions or debugging.
static bool validPredicate(LatticeVal LV) {
switch (LV) {
case GT: case GE: case LT: case LE: case NE:
return false;
}
}
+#endif
/// reversePredicate - reverse the direction of the inequality
static LatticeVal reversePredicate(LatticeVal LV) {
/// ValueNumbering stores the scope-specific value numbers for a given Value.
class VISIBILITY_HIDDEN ValueNumbering {
+
+ /// VNPair is a tuple of {Value, index number, DomTreeDFS::Node}. It
+ /// includes the comparison operators necessary to allow you to store it
+ /// in a sorted vector.
class VISIBILITY_HIDDEN VNPair {
public:
Value *V;
bool operator<(Value *RHS) const {
return V < RHS;
}
+
+ bool operator>(Value *RHS) const {
+ return V > RHS;
+ }
+
+ friend bool operator<(Value *RHS, const VNPair &pair) {
+ return pair.operator>(RHS);
+ }
};
typedef std::vector<VNPair> VNMapType;
VNMapType VNMap;
+ /// The canonical choice for value number at index.
std::vector<Value *> Values;
DomTreeDFS *DTDFS;
public:
+#ifndef NDEBUG
+ virtual ~ValueNumbering() {}
+ virtual void dump() {
+ dump(*cerr.stream());
+ }
+
+ void dump(std::ostream &os) {
+ for (unsigned i = 1; i <= Values.size(); ++i) {
+ os << i << " = ";
+ WriteAsOperand(os, Values[i-1]);
+ os << " {";
+ for (unsigned j = 0; j < VNMap.size(); ++j) {
+ if (VNMap[j].index == i) {
+ WriteAsOperand(os, VNMap[j].V);
+ os << " (" << VNMap[j].Subtree->getDFSNumIn() << ") ";
+ }
+ }
+ os << "}\n";
+ }
+ }
+#endif
+
/// compare - returns true if V1 is a better canonical value than V2.
bool compare(Value *V1, Value *V2) const {
if (isa<Constant>(V1))
/// valueNumber - finds the value number for V under the Subtree. If
/// there is no value number, returns zero.
unsigned valueNumber(Value *V, DomTreeDFS::Node *Subtree) {
+ if (!(isa<Constant>(V) || isa<Argument>(V) || isa<Instruction>(V))
+ || V->getType() == Type::VoidTy) return 0;
+
VNMapType::iterator E = VNMap.end();
VNPair pair(V, 0, Subtree);
VNMapType::iterator I = std::lower_bound(VNMap.begin(), E, pair);
return 0;
}
+ /// getOrInsertVN - always returns a value number, creating it if necessary.
+ unsigned getOrInsertVN(Value *V, DomTreeDFS::Node *Subtree) {
+ if (unsigned n = valueNumber(V, Subtree))
+ return n;
+ else
+ return newVN(V);
+ }
+
/// newVN - creates a new value number. Value V must not already have a
/// value number assigned.
unsigned newVN(Value *V) {
+ assert((isa<Constant>(V) || isa<Argument>(V) || isa<Instruction>(V)) &&
+ "Bad Value for value numbering.");
+ assert(V->getType() != Type::VoidTy && "Won't value number a void value");
+
Values.push_back(V);
VNPair pair = VNPair(V, Values.size(), DTDFS->getRootNode());
- assert(!std::binary_search(VNMap.begin(), VNMap.end(), pair) &&
+ VNMapType::iterator I = std::lower_bound(VNMap.begin(), VNMap.end(), pair);
+ assert((I == VNMap.end() || value(I->index) != V) &&
"Attempt to create a duplicate value number.");
- VNMap.insert(std::lower_bound(VNMap.begin(), VNMap.end(), pair), pair);
+ VNMap.insert(I, pair);
return Values.size();
}
VNMapType::iterator I = std::lower_bound(B, E, pair);
if (I != E && I->V == V && I->Subtree == Subtree)
I->index = n; // Update best choice
- else
+ else
VNMap.insert(I, pair); // New Value
// XXX: we currently don't have to worry about updating values with
/// remove - removes all references to value V.
void remove(Value *V) {
- VNMapType::iterator B = VNMap.begin();
+ VNMapType::iterator B = VNMap.begin(), E = VNMap.end();
VNPair pair(V, 0, DTDFS->getRootNode());
- VNMapType::iterator J = std::upper_bound(B, VNMap.end(), pair);
+ VNMapType::iterator J = std::upper_bound(B, E, pair);
VNMapType::iterator I = J;
- while (I != B && I->V == V) --I;
+ while (I != B && (I == E || I->V == V)) --I;
VNMap.erase(I, J);
}
" !=", "000031" };
for (Node::const_iterator NI = begin(), NE = end(); NI != NE; ++NI) {
os << names[NI->LV] << " " << NI->To
- << " (" << NI->Subtree->getDFSNumIn() << ")";
- if (NI != NE) os << ", ";
+ << " (" << NI->Subtree->getDFSNumIn() << "), ";
}
}
public:
return E;
}
- /// Updates the lattice value for a given node. Create a new entry if
- /// one doesn't exist, otherwise it merges the values. The new lattice
- /// value must not be inconsistent with any previously existing value.
+ /// update - updates the lattice value for a given node, creating a new
+ /// entry if one doesn't exist. The new lattice value must not be
+ /// inconsistent with any previously existing value.
void update(unsigned n, LatticeVal R, DomTreeDFS::Node *Subtree) {
assert(validPredicate(R) && "Invalid predicate.");
- iterator I = find(n, Subtree);
- if (I == end()) {
- Edge edge(n, R, Subtree);
- iterator Insert = std::lower_bound(begin(), end(), edge);
- Relations.insert(Insert, edge);
- } else {
- LatticeVal LV = static_cast<LatticeVal>(I->LV & R);
- assert(validPredicate(LV) && "Invalid union of lattice values.");
- if (LV != I->LV) {
- if (Subtree != I->Subtree) {
- assert(Subtree->DominatedBy(I->Subtree) &&
- "Find returned subtree that doesn't apply.");
-
- Edge edge(n, R, Subtree);
- iterator Insert = std::lower_bound(begin(), end(), edge);
- Relations.insert(Insert, edge); // invalidates I
- I = find(n, Subtree);
- }
- // Also, we have to tighten any edge that Subtree dominates.
- for (iterator B = begin(); I->To == n; --I) {
- if (I->Subtree->DominatedBy(Subtree)) {
- LatticeVal LV = static_cast<LatticeVal>(I->LV & R);
- assert(validPredicate(LV) && "Invalid union of lattice values");
- I->LV = LV;
- }
- if (I == B) break;
+ Edge edge(n, R, Subtree);
+ iterator B = begin(), E = end();
+ iterator I = std::lower_bound(B, E, edge);
+
+ iterator J = I;
+ while (J != E && J->To == n) {
+ if (Subtree->DominatedBy(J->Subtree))
+ break;
+ ++J;
+ }
+
+ if (J != E && J->To == n) {
+ edge.LV = static_cast<LatticeVal>(J->LV & R);
+ assert(validPredicate(edge.LV) && "Invalid union of lattice values.");
+
+ if (edge.LV == J->LV)
+ return; // This update adds nothing new.
+ }
+
+ if (I != B) {
+ // We also have to tighten any edge beneath our update.
+ for (iterator K = I - 1; K->To == n; --K) {
+ if (K->Subtree->DominatedBy(Subtree)) {
+ LatticeVal LV = static_cast<LatticeVal>(K->LV & edge.LV);
+ assert(validPredicate(LV) && "Invalid union of lattice values");
+ K->LV = LV;
}
+ if (K == B) break;
}
}
+
+ // Insert new edge at Subtree if it isn't already there.
+ if (I == E || I->To != n || Subtree != I->Subtree)
+ Relations.insert(I, edge);
}
};
std::vector<Node> Nodes;
public:
- /// node - returns the node object at a given index retrieved from getNode.
- /// Index zero is reserved and may not be passed in here. The pointer
- /// returned is valid until the next call to newNode or getOrInsertNode.
+ /// node - returns the node object at a given value number. The pointer
+ /// returned may be invalidated on the next call to node().
Node *node(unsigned index) {
- assert(index != 0 && "Zero index is reserved for not found.");
- assert(index <= Nodes.size() && "Index out of range.");
+ assert(VN.value(index)); // This triggers the necessary checks.
+ if (Nodes.size() < index) Nodes.resize(index);
return &Nodes[index-1];
}
- /// newNode - creates a new node for a given Value and returns the index.
- unsigned newNode(Value *V) {
- assert((isa<Constant>(V) || isa<Argument>(V) || isa<Instruction>(V)) &&
- "Bad Value for node.");
- assert(V->getType() != Type::VoidTy && "Void node?");
-
- unsigned n = VN.newVN(V);
- if (Nodes.size() < n) Nodes.resize(n);
- return n;
- }
-
/// isRelatedBy - true iff n1 op n2
bool isRelatedBy(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
LatticeVal LV) {
assert(!isRelatedBy(n1, n2, Subtree, reversePredicate(LV1)) &&
"Contradictory inequality.");
- Node *N1 = node(n1);
- Node *N2 = node(n2);
-
// Suppose we're adding %n1 < %n2. Find all the %a < %n1 and
// add %a < %n2 too. This keeps the graph fully connected.
if (LV1 != NE) {
unsigned LV1_s = LV1 & (SLT_BIT|SGT_BIT);
unsigned LV1_u = LV1 & (ULT_BIT|UGT_BIT);
- for (Node::iterator I = N1->begin(), E = N1->end(); I != E; ++I) {
+ for (Node::iterator I = node(n1)->begin(), E = node(n1)->end(); I != E; ++I) {
if (I->LV != NE && I->To != n2) {
DomTreeDFS::Node *Local_Subtree = NULL;
LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
node(I->To)->update(n2, NewLV, Local_Subtree);
- N2->update(I->To, reversePredicate(NewLV), Local_Subtree);
+ node(n2)->update(I->To, reversePredicate(NewLV), Local_Subtree);
}
}
}
}
- for (Node::iterator I = N2->begin(), E = N2->end(); I != E; ++I) {
+ for (Node::iterator I = node(n2)->begin(), E = node(n2)->end(); I != E; ++I) {
if (I->LV != NE && I->To != n1) {
DomTreeDFS::Node *Local_Subtree = NULL;
if (Subtree->DominatedBy(I->Subtree))
LatticeVal NewLV = static_cast<LatticeVal>(new_relationship);
- N1->update(I->To, NewLV, Local_Subtree);
+ node(n1)->update(I->To, NewLV, Local_Subtree);
node(I->To)->update(n1, reversePredicate(NewLV), Local_Subtree);
}
}
}
}
- N1->update(n2, LV1, Subtree);
- N2->update(n1, reversePredicate(LV1), Subtree);
+ node(n1)->update(n2, LV1, Subtree);
+ node(n2)->update(n1, reversePredicate(LV1), Subtree);
}
/// remove - removes a node from the graph by removing all references to
/// ValueRanges tracks the known integer ranges and anti-ranges of the nodes
/// in the InequalityGraph.
class VISIBILITY_HIDDEN ValueRanges {
+ ValueNumbering &VN;
+ TargetData *TD;
+ LLVMContext *Context;
- /// A ScopedRange ties an InequalityGraph node with a ConstantRange under
- /// the scope of a rooted subtree in the dominator tree.
class VISIBILITY_HIDDEN ScopedRange {
- public:
- ScopedRange(Value *V, ConstantRange CR, DomTreeDFS::Node *ST)
- : V(V), CR(CR), Subtree(ST) {}
+ typedef std::vector<std::pair<DomTreeDFS::Node *, ConstantRange> >
+ RangeListType;
+ RangeListType RangeList;
- Value *V;
- ConstantRange CR;
- DomTreeDFS::Node *Subtree;
-
- bool operator<(const ScopedRange &range) const {
- if (V != range.V) return V < range.V;
- return *Subtree < *range.Subtree;
+ static bool swo(const std::pair<DomTreeDFS::Node *, ConstantRange> &LHS,
+ const std::pair<DomTreeDFS::Node *, ConstantRange> &RHS) {
+ return *LHS.first < *RHS.first;
}
- bool operator<(const Value *value) const {
- return V < value;
+ public:
+#ifndef NDEBUG
+ virtual ~ScopedRange() {}
+ virtual void dump() const {
+ dump(*cerr.stream());
}
- bool operator>(const Value *value) const {
- return V > value;
+ void dump(std::ostream &os) const {
+ os << "{";
+ for (const_iterator I = begin(), E = end(); I != E; ++I) {
+ os << &I->second << " (" << I->first->getDFSNumIn() << "), ";
+ }
+ os << "}";
}
+#endif
- friend bool operator<(const Value *value, const ScopedRange &range) {
- return range.operator>(value);
- }
- };
+ typedef RangeListType::iterator iterator;
+ typedef RangeListType::const_iterator const_iterator;
- TargetData *TD;
+ iterator begin() { return RangeList.begin(); }
+ iterator end() { return RangeList.end(); }
+ const_iterator begin() const { return RangeList.begin(); }
+ const_iterator end() const { return RangeList.end(); }
- std::vector<ScopedRange> Ranges;
- typedef std::vector<ScopedRange>::iterator iterator;
+ iterator find(DomTreeDFS::Node *Subtree) {
+ iterator E = end();
+ iterator I = std::lower_bound(begin(), E,
+ std::make_pair(Subtree, empty), swo);
- // XXX: this is a copy of the code in InequalityGraph::Node. Perhaps a
- // intrusive domtree-scoped container is in order?
+ while (I != E && !I->first->dominates(Subtree)) ++I;
+ return I;
+ }
- iterator begin() { return Ranges.begin(); }
- iterator end() { return Ranges.end(); }
+ const_iterator find(DomTreeDFS::Node *Subtree) const {
+ const_iterator E = end();
+ const_iterator I = std::lower_bound(begin(), E,
+ std::make_pair(Subtree, empty), swo);
- iterator find(Value *V, DomTreeDFS::Node *Subtree) {
- iterator E = end();
- for (iterator I = std::lower_bound(begin(), E, V);
- I != E && I->V == V; ++I) {
- if (Subtree->DominatedBy(I->Subtree))
- return I;
+ while (I != E && !I->first->dominates(Subtree)) ++I;
+ return I;
}
- return E;
- }
- void update(Value *V, ConstantRange CR, DomTreeDFS::Node *Subtree) {
- assert(!CR.isEmptySet() && "Empty ConstantRange!");
- if (CR.isFullSet()) return;
+ void update(const ConstantRange &CR, DomTreeDFS::Node *Subtree) {
+ assert(!CR.isEmptySet() && "Empty ConstantRange.");
+ assert(!CR.isSingleElement() && "Refusing to store single element.");
- iterator I = find(V, Subtree);
- if (I == end()) {
- ScopedRange range(V, CR, Subtree);
- iterator Insert = std::lower_bound(begin(), end(), range);
- Ranges.insert(Insert, range);
- } else {
- CR = CR.intersectWith(I->CR);
- assert(!CR.isEmptySet() && "Empty intersection of ConstantRanges!");
-
- if (CR != I->CR) {
- if (Subtree != I->Subtree) {
- assert(Subtree->DominatedBy(I->Subtree) &&
- "Find returned subtree that doesn't apply.");
-
- ScopedRange range(V, CR, Subtree);
- iterator Insert = std::lower_bound(begin(), end(), range);
- Ranges.insert(Insert, range); // invalidates I
- I = find(V, Subtree);
- }
+ iterator E = end();
+ iterator I =
+ std::lower_bound(begin(), E, std::make_pair(Subtree, empty), swo);
- // Also, we have to tighten any edge that Subtree dominates.
- for (iterator B = begin(); I->V == V; --I) {
- if (I->Subtree->DominatedBy(Subtree)) {
- I->CR = CR.intersectWith(I->CR);
- assert(!I->CR.isEmptySet() &&
- "Empty intersection of ConstantRanges!");
- }
- if (I == B) break;
- }
- }
+ if (I != end() && I->first == Subtree) {
+ ConstantRange CR2 = I->second.intersectWith(CR);
+ assert(!CR2.isEmptySet() && !CR2.isSingleElement() &&
+ "Invalid union of ranges.");
+ I->second = CR2;
+ } else
+ RangeList.insert(I, std::make_pair(Subtree, CR));
}
- }
+ };
- /// range - Creates a ConstantRange representing the set of all values
- /// that match the ICmpInst::Predicate with any of the values in CR.
- ConstantRange range(ICmpInst::Predicate ICmpOpcode,
- const ConstantRange &CR) {
- uint32_t W = CR.getBitWidth();
- switch (ICmpOpcode) {
- default: assert(!"Invalid ICmp opcode to range()");
- case ICmpInst::ICMP_EQ:
- return ConstantRange(CR.getLower(), CR.getUpper());
- case ICmpInst::ICMP_NE:
- if (CR.isSingleElement())
- return ConstantRange(CR.getUpper(), CR.getLower());
- return ConstantRange(W);
- case ICmpInst::ICMP_ULT:
- return ConstantRange(APInt::getMinValue(W), CR.getUnsignedMax());
- case ICmpInst::ICMP_SLT:
- return ConstantRange(APInt::getSignedMinValue(W), CR.getSignedMax());
- case ICmpInst::ICMP_ULE: {
- APInt UMax(CR.getUnsignedMax());
- if (UMax.isMaxValue())
- return ConstantRange(W);
- return ConstantRange(APInt::getMinValue(W), UMax + 1);
- }
- case ICmpInst::ICMP_SLE: {
- APInt SMax(CR.getSignedMax());
- if (SMax.isMaxSignedValue() || (SMax+1).isMaxSignedValue())
- return ConstantRange(W);
- return ConstantRange(APInt::getSignedMinValue(W), SMax + 1);
- }
- case ICmpInst::ICMP_UGT:
- return ConstantRange(CR.getUnsignedMin() + 1, APInt::getNullValue(W));
- case ICmpInst::ICMP_SGT:
- return ConstantRange(CR.getSignedMin() + 1,
- APInt::getSignedMinValue(W));
- case ICmpInst::ICMP_UGE: {
- APInt UMin(CR.getUnsignedMin());
- if (UMin.isMinValue())
- return ConstantRange(W);
- return ConstantRange(UMin, APInt::getNullValue(W));
- }
- case ICmpInst::ICMP_SGE: {
- APInt SMin(CR.getSignedMin());
- if (SMin.isMinSignedValue())
- return ConstantRange(W);
- return ConstantRange(SMin, APInt::getSignedMinValue(W));
- }
- }
+ std::vector<ScopedRange> Ranges;
+
+ void update(unsigned n, const ConstantRange &CR, DomTreeDFS::Node *Subtree){
+ if (CR.isFullSet()) return;
+ if (Ranges.size() < n) Ranges.resize(n);
+ Ranges[n-1].update(CR, Subtree);
}
/// create - Creates a ConstantRange that matches the given LatticeVal
assert(!CR.isEmptySet() && "Can't deal with empty set.");
if (LV == NE)
- return range(ICmpInst::ICMP_NE, CR);
+ return ConstantRange::makeICmpRegion(ICmpInst::ICMP_NE, CR);
unsigned LV_s = LV & (SGT_BIT|SLT_BIT);
unsigned LV_u = LV & (UGT_BIT|ULT_BIT);
ConstantRange Range(CR.getBitWidth());
if (LV_s == SGT_BIT) {
- Range = Range.intersectWith(range(
+ Range = Range.intersectWith(ConstantRange::makeICmpRegion(
hasEQ ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SGT, CR));
} else if (LV_s == SLT_BIT) {
- Range = Range.intersectWith(range(
+ Range = Range.intersectWith(ConstantRange::makeICmpRegion(
hasEQ ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_SLT, CR));
}
if (LV_u == UGT_BIT) {
- Range = Range.intersectWith(range(
+ Range = Range.intersectWith(ConstantRange::makeICmpRegion(
hasEQ ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_UGT, CR));
} else if (LV_u == ULT_BIT) {
- Range = Range.intersectWith(range(
+ Range = Range.intersectWith(ConstantRange::makeICmpRegion(
hasEQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT, CR));
}
}
#ifndef NDEBUG
- bool isCanonical(Value *V, DomTreeDFS::Node *Subtree, VRPSolver *VRP);
+ bool isCanonical(Value *V, DomTreeDFS::Node *Subtree) {
+ return V == VN.canonicalize(V, Subtree);
+ }
#endif
public:
- explicit ValueRanges(TargetData *TD) : TD(TD) {}
+ ValueRanges(ValueNumbering &VN, TargetData *TD, LLVMContext *C) :
+ VN(VN), TD(TD), Context(C) {}
- // rangeFromValue - converts a Value into a range. If the value is a
- // constant it constructs the single element range, otherwise it performs
- // a lookup. The width W must be retrieved from typeToWidth and may not
- // be zero.
- ConstantRange rangeFromValue(Value *V, DomTreeDFS::Node *Subtree,
- uint32_t W) {
- if (ConstantInt *C = dyn_cast<ConstantInt>(V)) {
- return ConstantRange(C->getValue());
- } else if (isa<ConstantPointerNull>(V)) {
- return ConstantRange(APInt::getNullValue(W));
- } else {
- iterator I = find(V, Subtree);
- if (I != end())
- return I->CR;
+#ifndef NDEBUG
+ virtual ~ValueRanges() {}
+
+ virtual void dump() const {
+ dump(*cerr.stream());
+ }
+
+ void dump(std::ostream &os) const {
+ for (unsigned i = 0, e = Ranges.size(); i != e; ++i) {
+ os << (i+1) << " = ";
+ Ranges[i].dump(os);
+ os << "\n";
}
- return ConstantRange(W);
+ }
+#endif
+
+ /// range - looks up the ConstantRange associated with a value number.
+ ConstantRange range(unsigned n, DomTreeDFS::Node *Subtree) {
+ assert(VN.value(n)); // performs range checks
+
+ if (n <= Ranges.size()) {
+ ScopedRange::iterator I = Ranges[n-1].find(Subtree);
+ if (I != Ranges[n-1].end()) return I->second;
+ }
+
+ Value *V = VN.value(n);
+ ConstantRange CR = range(V);
+ return CR;
+ }
+
+ /// range - determine a range from a Value without performing any lookups.
+ ConstantRange range(Value *V) const {
+ if (ConstantInt *C = dyn_cast<ConstantInt>(V))
+ return ConstantRange(C->getValue());
+ else if (isa<ConstantPointerNull>(V))
+ return ConstantRange(APInt::getNullValue(typeToWidth(V->getType())));
+ else
+ return ConstantRange(typeToWidth(V->getType()));
}
// typeToWidth - returns the number of bits necessary to store a value of
uint32_t typeToWidth(const Type *Ty) const {
if (TD)
return TD->getTypeSizeInBits(Ty);
-
- if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty))
- return ITy->getBitWidth();
-
- return 0;
+ else
+ return Ty->getPrimitiveSizeInBits();
}
- bool isRelatedBy(Value *V1, Value *V2, DomTreeDFS::Node *Subtree,
- LatticeVal LV) {
- uint32_t W = typeToWidth(V1->getType());
- if (!W) return false;
-
- ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
- ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
-
- // True iff all values in CR1 are LV to all values in CR2.
+ static bool isRelatedBy(const ConstantRange &CR1, const ConstantRange &CR2,
+ LatticeVal LV) {
switch (LV) {
default: assert(!"Impossible lattice value!");
case NE:
}
}
+ bool isRelatedBy(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
+ LatticeVal LV) {
+ ConstantRange CR1 = range(n1, Subtree);
+ ConstantRange CR2 = range(n2, Subtree);
+
+ // True iff all values in CR1 are LV to all values in CR2.
+ return isRelatedBy(CR1, CR2, LV);
+ }
+
void addToWorklist(Value *V, Constant *C, ICmpInst::Predicate Pred,
VRPSolver *VRP);
void markBlock(VRPSolver *VRP);
- void mergeInto(Value **I, unsigned n, Value *New,
+ void mergeInto(Value **I, unsigned n, unsigned New,
DomTreeDFS::Node *Subtree, VRPSolver *VRP) {
- assert(isCanonical(New, Subtree, VRP) && "Best choice not canonical?");
-
- uint32_t W = typeToWidth(New->getType());
- if (!W) return;
-
- ConstantRange CR_New = rangeFromValue(New, Subtree, W);
+ ConstantRange CR_New = range(New, Subtree);
ConstantRange Merged = CR_New;
for (; n != 0; ++I, --n) {
- ConstantRange CR_Kill = rangeFromValue(*I, Subtree, W);
+ unsigned i = VN.valueNumber(*I, Subtree);
+ ConstantRange CR_Kill = i ? range(i, Subtree) : range(*I);
if (CR_Kill.isFullSet()) continue;
Merged = Merged.intersectWith(CR_Kill);
}
applyRange(New, Merged, Subtree, VRP);
}
- void applyRange(Value *V, const ConstantRange &CR,
+ void applyRange(unsigned n, const ConstantRange &CR,
DomTreeDFS::Node *Subtree, VRPSolver *VRP) {
- assert(isCanonical(V, Subtree, VRP) && "Value not canonical.");
+ ConstantRange Merged = CR.intersectWith(range(n, Subtree));
+ if (Merged.isEmptySet()) {
+ markBlock(VRP);
+ return;
+ }
- if (const APInt *I = CR.getSingleElement()) {
+ if (const APInt *I = Merged.getSingleElement()) {
+ Value *V = VN.value(n); // XXX: redesign worklist.
const Type *Ty = V->getType();
if (Ty->isInteger()) {
- addToWorklist(V, ConstantInt::get(*I), ICmpInst::ICMP_EQ, VRP);
+ addToWorklist(V, Context->getConstantInt(*I), ICmpInst::ICMP_EQ, VRP);
return;
} else if (const PointerType *PTy = dyn_cast<PointerType>(Ty)) {
assert(*I == 0 && "Pointer is null but not zero?");
addToWorklist(V, ConstantPointerNull::get(PTy),
- ICmpInst::ICMP_EQ, VRP);
+ ICmpInst::ICMP_EQ, VRP);
return;
}
}
- ConstantRange Merged = CR.intersectWith(
- rangeFromValue(V, Subtree, CR.getBitWidth()));
- if (Merged.isEmptySet()) {
- markBlock(VRP);
- return;
- }
-
- update(V, Merged, Subtree);
+ update(n, Merged, Subtree);
}
- void addNotEquals(Value *V1, Value *V2, DomTreeDFS::Node *Subtree,
+ void addNotEquals(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
VRPSolver *VRP) {
- uint32_t W = typeToWidth(V1->getType());
- if (!W) return;
+ ConstantRange CR1 = range(n1, Subtree);
+ ConstantRange CR2 = range(n2, Subtree);
- ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
- ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
+ uint32_t W = CR1.getBitWidth();
if (const APInt *I = CR1.getSingleElement()) {
if (CR2.isFullSet()) {
ConstantRange NewCR2(CR1.getUpper(), CR1.getLower());
- applyRange(V2, NewCR2, Subtree, VRP);
+ applyRange(n2, NewCR2, Subtree, VRP);
} else if (*I == CR2.getLower()) {
APInt NewLower(CR2.getLower() + 1),
NewUpper(CR2.getUpper());
NewLower = NewUpper = APInt::getMinValue(W);
ConstantRange NewCR2(NewLower, NewUpper);
- applyRange(V2, NewCR2, Subtree, VRP);
+ applyRange(n2, NewCR2, Subtree, VRP);
} else if (*I == CR2.getUpper() - 1) {
APInt NewLower(CR2.getLower()),
NewUpper(CR2.getUpper() - 1);
NewLower = NewUpper = APInt::getMinValue(W);
ConstantRange NewCR2(NewLower, NewUpper);
- applyRange(V2, NewCR2, Subtree, VRP);
+ applyRange(n2, NewCR2, Subtree, VRP);
}
}
if (const APInt *I = CR2.getSingleElement()) {
if (CR1.isFullSet()) {
ConstantRange NewCR1(CR2.getUpper(), CR2.getLower());
- applyRange(V1, NewCR1, Subtree, VRP);
+ applyRange(n1, NewCR1, Subtree, VRP);
} else if (*I == CR1.getLower()) {
APInt NewLower(CR1.getLower() + 1),
NewUpper(CR1.getUpper());
NewLower = NewUpper = APInt::getMinValue(W);
ConstantRange NewCR1(NewLower, NewUpper);
- applyRange(V1, NewCR1, Subtree, VRP);
+ applyRange(n1, NewCR1, Subtree, VRP);
} else if (*I == CR1.getUpper() - 1) {
APInt NewLower(CR1.getLower()),
NewUpper(CR1.getUpper() - 1);
NewLower = NewUpper = APInt::getMinValue(W);
ConstantRange NewCR1(NewLower, NewUpper);
- applyRange(V1, NewCR1, Subtree, VRP);
+ applyRange(n1, NewCR1, Subtree, VRP);
}
}
}
- void addInequality(Value *V1, Value *V2, DomTreeDFS::Node *Subtree,
+ void addInequality(unsigned n1, unsigned n2, DomTreeDFS::Node *Subtree,
LatticeVal LV, VRPSolver *VRP) {
- assert(!isRelatedBy(V1, V2, Subtree, LV) && "Asked to do useless work.");
-
- assert(isCanonical(V1, Subtree, VRP) && "Value not canonical.");
- assert(isCanonical(V2, Subtree, VRP) && "Value not canonical.");
+ assert(!isRelatedBy(n1, n2, Subtree, LV) && "Asked to do useless work.");
if (LV == NE) {
- addNotEquals(V1, V2, Subtree, VRP);
+ addNotEquals(n1, n2, Subtree, VRP);
return;
}
- uint32_t W = typeToWidth(V1->getType());
- if (!W) return;
-
- ConstantRange CR1 = rangeFromValue(V1, Subtree, W);
- ConstantRange CR2 = rangeFromValue(V2, Subtree, W);
+ ConstantRange CR1 = range(n1, Subtree);
+ ConstantRange CR2 = range(n2, Subtree);
if (!CR1.isSingleElement()) {
ConstantRange NewCR1 = CR1.intersectWith(create(LV, CR2));
if (NewCR1 != CR1)
- applyRange(V1, NewCR1, Subtree, VRP);
+ applyRange(n1, NewCR1, Subtree, VRP);
}
if (!CR2.isSingleElement()) {
- ConstantRange NewCR2 = CR2.intersectWith(create(reversePredicate(LV),
- CR1));
+ ConstantRange NewCR2 = CR2.intersectWith(
+ create(reversePredicate(LV), CR1));
if (NewCR2 != CR2)
- applyRange(V2, NewCR2, Subtree, VRP);
+ applyRange(n2, NewCR2, Subtree, VRP);
}
}
};
BasicBlock *TopBB;
Instruction *TopInst;
bool &modified;
+ LLVMContext *Context;
typedef InequalityGraph::Node Node;
if (!Node) return false;
return Top->dominates(Node);
}
+ return false; // Not reached
}
// aboveOrBelow - true if the Instruction either dominates or is dominated
// be EQ and that's invalid. What we're doing is looking for any nodes
// %z such that %x <= %z and %y >= %z, and vice versa.
- Node *N1 = IG.node(n1);
- Node *N2 = IG.node(n2);
- Node::iterator end = N2->end();
+ Node::iterator end = IG.node(n2)->end();
// Find the intersection between N1 and N2 which is dominated by
// Top. If we find %x where N1 <= %x <= N2 (or >=) then add %x to
// Remove.
- for (Node::iterator I = N1->begin(), E = N1->end(); I != E; ++I) {
+ for (Node::iterator I = IG.node(n1)->begin(), E = IG.node(n1)->end();
+ I != E; ++I) {
if (!(I->LV & EQ_BIT) || !Top->DominatedBy(I->Subtree)) continue;
unsigned ILV_s = I->LV & (SLT_BIT|SGT_BIT);
unsigned ILV_u = I->LV & (ULT_BIT|UGT_BIT);
- Node::iterator NI = N2->find(I->To, Top);
+ Node::iterator NI = IG.node(n2)->find(I->To, Top);
if (NI != end) {
LatticeVal NILV = reversePredicate(NI->LV);
unsigned NILV_s = NILV & (SLT_BIT|SGT_BIT);
}
// We'd like to allow makeEqual on two values to perform a simple
- // substitution without every creating nodes in the IG whenever possible.
+ // substitution without creating nodes in the IG whenever possible.
//
// The first iteration through this loop operates on V2 before going
// through the Remove list and operating on those too. If all of the
Instruction *I2 = dyn_cast<Instruction>(R);
if (I2 && below(I2)) {
std::vector<Instruction *> ToNotify;
- for (Value::use_iterator UI = R->use_begin(), UE = R->use_end();
+ for (Value::use_iterator UI = I2->use_begin(), UE = I2->use_end();
UI != UE;) {
Use &TheUse = UI.getUse();
++UI;
- if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser()))
- ToNotify.push_back(I);
+ Instruction *I = cast<Instruction>(TheUse.getUser());
+ ToNotify.push_back(I);
}
DOUT << "Simply removing " << *I2
if (!isa<Constant>(V1)) {
if (Remove.empty()) {
- VR.mergeInto(&V2, 1, V1, Top, this);
+ VR.mergeInto(&V2, 1, VN.getOrInsertVN(V1, Top), Top, this);
} else {
std::vector<Value*> RemoveVals;
RemoveVals.reserve(Remove.size());
if (!V->use_empty())
RemoveVals.push_back(V);
}
- VR.mergeInto(&RemoveVals[0], RemoveVals.size(), V1, Top, this);
+ VR.mergeInto(&RemoveVals[0], RemoveVals.size(),
+ VN.getOrInsertVN(V1, Top), Top, this);
}
}
if (mergeIGNode) {
// Create N1.
- if (!n1) n1 = IG.newNode(V1);
+ if (!n1) n1 = VN.getOrInsertVN(V1, Top);
+ IG.node(n1); // Ensure that IG.Nodes won't get resized
// Migrate relationships from removed nodes to N1.
- Node *N1 = IG.node(n1);
for (SetVector<unsigned>::iterator I = Remove.begin(), E = Remove.end();
I != E; ++I) {
unsigned n = *I;
- Node *N = IG.node(n);
- for (Node::iterator NI = N->begin(), NE = N->end(); NI != NE; ++NI) {
+ for (Node::iterator NI = IG.node(n)->begin(), NE = IG.node(n)->end();
+ NI != NE; ++NI) {
if (NI->Subtree->DominatedBy(Top)) {
if (NI->To == n1) {
assert((NI->LV & EQ_BIT) && "Node inequal to itself.");
continue;
IG.node(NI->To)->update(n1, reversePredicate(NI->LV), Top);
- N1->update(NI->To, NI->LV, Top);
+ IG.node(n1)->update(NI->To, NI->LV, Top);
}
}
}
++UI;
Value *V = TheUse.getUser();
if (!V->use_empty()) {
- if (Instruction *Inst = dyn_cast<Instruction>(V)) {
- if (aboveOrBelow(Inst))
- opsToDef(Inst);
- }
+ Instruction *Inst = cast<Instruction>(V);
+ if (aboveOrBelow(Inst))
+ opsToDef(Inst);
}
}
}
Top(DTDFS->getNodeForBlock(TopBB)),
TopBB(TopBB),
TopInst(NULL),
- modified(modified)
+ modified(modified),
+ Context(&TopBB->getContext())
{
assert(Top && "VRPSolver created for unreachable basic block.");
}
Top(DTDFS->getNodeForBlock(TopInst->getParent())),
TopBB(TopInst->getParent()),
TopInst(TopInst),
- modified(modified)
+ modified(modified),
+ Context(&TopInst->getContext())
{
assert(Top && "VRPSolver created for unreachable basic block.");
assert(Top->getBlock() == TopInst->getParent() && "Context mismatch.");
bool isRelatedBy(Value *V1, Value *V2, ICmpInst::Predicate Pred) const {
if (Constant *C1 = dyn_cast<Constant>(V1))
if (Constant *C2 = dyn_cast<Constant>(V2))
- return ConstantExpr::getCompare(Pred, C1, C2) ==
- ConstantInt::getTrue();
-
- if (unsigned n1 = VN.valueNumber(V1, Top))
- if (unsigned n2 = VN.valueNumber(V2, Top)) {
- if (n1 == n2) return Pred == ICmpInst::ICMP_EQ ||
- Pred == ICmpInst::ICMP_ULE ||
- Pred == ICmpInst::ICMP_UGE ||
- Pred == ICmpInst::ICMP_SLE ||
- Pred == ICmpInst::ICMP_SGE;
- if (Pred == ICmpInst::ICMP_EQ) return false;
- if (IG.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true;
- }
+ return Context->getConstantExprCompare(Pred, C1, C2) ==
+ Context->getTrue();
+
+ unsigned n1 = VN.valueNumber(V1, Top);
+ unsigned n2 = VN.valueNumber(V2, Top);
+
+ if (n1 && n2) {
+ if (n1 == n2) return Pred == ICmpInst::ICMP_EQ ||
+ Pred == ICmpInst::ICMP_ULE ||
+ Pred == ICmpInst::ICMP_UGE ||
+ Pred == ICmpInst::ICMP_SLE ||
+ Pred == ICmpInst::ICMP_SGE;
+ if (Pred == ICmpInst::ICMP_EQ) return false;
+ if (IG.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true;
+ if (VR.isRelatedBy(n1, n2, Top, cmpInstToLattice(Pred))) return true;
+ }
+
+ if ((n1 && !n2 && isa<Constant>(V2)) ||
+ (n2 && !n1 && isa<Constant>(V1))) {
+ ConstantRange CR1 = n1 ? VR.range(n1, Top) : VR.range(V1);
+ ConstantRange CR2 = n2 ? VR.range(n2, Top) : VR.range(V2);
+
+ if (Pred == ICmpInst::ICMP_EQ)
+ return CR1.isSingleElement() &&
+ CR1.getSingleElement() == CR2.getSingleElement();
+ return VR.isRelatedBy(CR1, CR2, cmpInstToLattice(Pred));
+ }
if (Pred == ICmpInst::ICMP_EQ) return V1 == V2;
- return VR.isRelatedBy(V1, V2, Top, cmpInstToLattice(Pred));
+ return false;
}
/// add - adds a new property to the work queue
switch (BO->getOpcode()) {
case Instruction::And: {
// "and i32 %a, %b" EQ -1 then %a EQ -1 and %b EQ -1
- ConstantInt *CI = ConstantInt::getAllOnesValue(Ty);
+ ConstantInt *CI = cast<ConstantInt>(Context->getAllOnesValue(Ty));
if (Canonical == CI) {
add(CI, Op0, ICmpInst::ICMP_EQ, NewContext);
add(CI, Op1, ICmpInst::ICMP_EQ, NewContext);
} break;
case Instruction::Or: {
// "or i32 %a, %b" EQ 0 then %a EQ 0 and %b EQ 0
- Constant *Zero = Constant::getNullValue(Ty);
+ Constant *Zero = Context->getNullValue(Ty);
if (Canonical == Zero) {
add(Zero, Op0, ICmpInst::ICMP_EQ, NewContext);
add(Zero, Op1, ICmpInst::ICMP_EQ, NewContext);
if (ConstantInt *CI = dyn_cast<ConstantInt>(Canonical)) {
if (ConstantInt *Arg = dyn_cast<ConstantInt>(LHS)) {
- add(RHS, ConstantInt::get(CI->getValue() ^ Arg->getValue()),
+ add(RHS,
+ Context->getConstantInt(CI->getValue() ^ Arg->getValue()),
ICmpInst::ICMP_EQ, NewContext);
}
}
if (Canonical == LHS) {
if (isa<ConstantInt>(Canonical))
- add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_EQ,
+ add(RHS, Context->getNullValue(Ty), ICmpInst::ICMP_EQ,
NewContext);
} else if (isRelatedBy(LHS, Canonical, ICmpInst::ICMP_NE)) {
- add(RHS, Constant::getNullValue(Ty), ICmpInst::ICMP_NE,
+ add(RHS, Context->getNullValue(Ty), ICmpInst::ICMP_NE,
NewContext);
}
} break;
// "icmp ult i32 %a, %y" EQ true then %a u< y
// etc.
- if (Canonical == ConstantInt::getTrue()) {
+ if (Canonical == Context->getTrue()) {
add(IC->getOperand(0), IC->getOperand(1), IC->getPredicate(),
NewContext);
- } else if (Canonical == ConstantInt::getFalse()) {
+ } else if (Canonical == Context->getFalse()) {
add(IC->getOperand(0), IC->getOperand(1),
ICmpInst::getInversePredicate(IC->getPredicate()), NewContext);
}
if (isRelatedBy(True, False, ICmpInst::ICMP_NE)) {
if (Canonical == VN.canonicalize(True, Top) ||
isRelatedBy(Canonical, False, ICmpInst::ICMP_NE))
- add(SI->getCondition(), ConstantInt::getTrue(),
+ add(SI->getCondition(), Context->getTrue(),
ICmpInst::ICMP_EQ, NewContext);
else if (Canonical == VN.canonicalize(False, Top) ||
isRelatedBy(Canonical, True, ICmpInst::ICMP_NE))
- add(SI->getCondition(), ConstantInt::getFalse(),
+ add(SI->getCondition(), Context->getFalse(),
ICmpInst::ICMP_EQ, NewContext);
}
} else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
}
// TODO: The GEPI indices are all zero. Copy from definition to operand,
// jumping the type plane as needed.
- if (isRelatedBy(GEPI, Constant::getNullValue(GEPI->getType()),
+ if (isRelatedBy(GEPI, Context->getNullValue(GEPI->getType()),
ICmpInst::ICMP_NE)) {
Value *Ptr = GEPI->getPointerOperand();
- add(Ptr, Constant::getNullValue(Ptr->getType()), ICmpInst::ICMP_NE,
+ add(Ptr, Context->getNullValue(Ptr->getType()), ICmpInst::ICMP_NE,
NewContext);
}
} else if (CastInst *CI = dyn_cast<CastInst>(I)) {
const Type *SrcTy = CI->getSrcTy();
- Value *TheCI = VN.canonicalize(CI, Top);
+ unsigned ci = VN.getOrInsertVN(CI, Top);
uint32_t W = VR.typeToWidth(SrcTy);
if (!W) return;
- ConstantRange CR = VR.rangeFromValue(TheCI, Top, W);
+ ConstantRange CR = VR.range(ci, Top);
if (CR.isFullSet()) return;
default: break;
case Instruction::ZExt:
case Instruction::SExt:
- VR.applyRange(VN.canonicalize(CI->getOperand(0), Top),
+ VR.applyRange(VN.getOrInsertVN(CI->getOperand(0), Top),
CR.truncate(W), Top, this);
break;
case Instruction::BitCast:
- VR.applyRange(VN.canonicalize(CI->getOperand(0), Top),
+ VR.applyRange(VN.getOrInsertVN(CI->getOperand(0), Top),
CR, Top, this);
break;
}
const Type *Ty = BO->getType();
assert(!Ty->isFPOrFPVector() && "Float in work queue!");
- Constant *Zero = Constant::getNullValue(Ty);
- ConstantInt *AllOnes = ConstantInt::getAllOnesValue(Ty);
+ Constant *Zero = Context->getNullValue(Ty);
+ Constant *One = Context->getConstantInt(Ty, 1);
+ ConstantInt *AllOnes = cast<ConstantInt>(Context->getAllOnesValue(Ty));
switch (Opcode) {
default: break;
case Instruction::LShr:
case Instruction::AShr:
case Instruction::Shl:
+ if (Op1 == Zero) {
+ add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
+ return;
+ }
+ break;
case Instruction::Sub:
if (Op1 == Zero) {
add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
return;
}
+ if (ConstantInt *CI0 = dyn_cast<ConstantInt>(Op0)) {
+ unsigned n_ci0 = VN.getOrInsertVN(Op1, Top);
+ ConstantRange CR = VR.range(n_ci0, Top);
+ if (!CR.isFullSet()) {
+ CR.subtract(CI0->getValue());
+ unsigned n_bo = VN.getOrInsertVN(BO, Top);
+ VR.applyRange(n_bo, CR, Top, this);
+ return;
+ }
+ }
+ if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Op1)) {
+ unsigned n_ci1 = VN.getOrInsertVN(Op0, Top);
+ ConstantRange CR = VR.range(n_ci1, Top);
+ if (!CR.isFullSet()) {
+ CR.subtract(CI1->getValue());
+ unsigned n_bo = VN.getOrInsertVN(BO, Top);
+ VR.applyRange(n_bo, CR, Top, this);
+ return;
+ }
+ }
break;
case Instruction::Or:
if (Op0 == AllOnes || Op1 == AllOnes) {
add(BO, AllOnes, ICmpInst::ICMP_EQ, NewContext);
return;
- } // fall-through
- case Instruction::Xor:
+ }
+ if (Op0 == Zero) {
+ add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
+ return;
+ } else if (Op1 == Zero) {
+ add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
+ return;
+ }
+ break;
case Instruction::Add:
+ if (ConstantInt *CI0 = dyn_cast<ConstantInt>(Op0)) {
+ unsigned n_ci0 = VN.getOrInsertVN(Op1, Top);
+ ConstantRange CR = VR.range(n_ci0, Top);
+ if (!CR.isFullSet()) {
+ CR.subtract(-CI0->getValue());
+ unsigned n_bo = VN.getOrInsertVN(BO, Top);
+ VR.applyRange(n_bo, CR, Top, this);
+ return;
+ }
+ }
+ if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Op1)) {
+ unsigned n_ci1 = VN.getOrInsertVN(Op0, Top);
+ ConstantRange CR = VR.range(n_ci1, Top);
+ if (!CR.isFullSet()) {
+ CR.subtract(-CI1->getValue());
+ unsigned n_bo = VN.getOrInsertVN(BO, Top);
+ VR.applyRange(n_bo, CR, Top, this);
+ return;
+ }
+ }
+ // fall-through
+ case Instruction::Xor:
if (Op0 == Zero) {
add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
return;
add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
return;
}
- // fall-through
+ if (Op0 == Zero || Op1 == Zero) {
+ add(BO, Zero, ICmpInst::ICMP_EQ, NewContext);
+ return;
+ }
+ break;
case Instruction::Mul:
if (Op0 == Zero || Op1 == Zero) {
add(BO, Zero, ICmpInst::ICMP_EQ, NewContext);
return;
}
+ if (Op0 == One) {
+ add(BO, Op1, ICmpInst::ICMP_EQ, NewContext);
+ return;
+ } else if (Op1 == One) {
+ add(BO, Op0, ICmpInst::ICMP_EQ, NewContext);
+ return;
+ }
break;
}
// "%x = add i32 %y, %z" and %x EQ %y then %z EQ 0
// "%x = add i32 %y, %z" and %x EQ %z then %y EQ 0
// "%x = shl i32 %y, %z" and %x EQ %y and %y NE 0 then %z EQ 0
- // "%x = udiv i32 %y, %z" and %x EQ %y then %z EQ 1
+ // "%x = udiv i32 %y, %z" and %x EQ %y and %y NE 0 then %z EQ 1
Value *Known = Op0, *Unknown = Op1,
*TheBO = VN.canonicalize(BO, Top);
if (!isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) break;
// otherwise, fall-through.
case Instruction::Sub:
- if (Unknown == Op1) break;
+ if (Unknown == Op0) break;
// otherwise, fall-through.
case Instruction::Xor:
case Instruction::Add:
case Instruction::UDiv:
case Instruction::SDiv:
if (Unknown == Op1) break;
- if (isRelatedBy(Known, Zero, ICmpInst::ICMP_NE)) {
- Constant *One = ConstantInt::get(Ty, 1);
+ if (isRelatedBy(Known, Zero, ICmpInst::ICMP_NE))
add(Unknown, One, ICmpInst::ICMP_EQ, NewContext);
- }
break;
}
}
Value *Op1 = VN.canonicalize(IC->getOperand(1), Top);
ICmpInst::Predicate Pred = IC->getPredicate();
- if (isRelatedBy(Op0, Op1, Pred)) {
- add(IC, ConstantInt::getTrue(), ICmpInst::ICMP_EQ, NewContext);
- } else if (isRelatedBy(Op0, Op1, ICmpInst::getInversePredicate(Pred))) {
- add(IC, ConstantInt::getFalse(), ICmpInst::ICMP_EQ, NewContext);
- }
+ if (isRelatedBy(Op0, Op1, Pred))
+ add(IC, Context->getTrue(), ICmpInst::ICMP_EQ, NewContext);
+ else if (isRelatedBy(Op0, Op1, ICmpInst::getInversePredicate(Pred)))
+ add(IC, Context->getFalse(),
+ ICmpInst::ICMP_EQ, NewContext);
} else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
if (I->getType()->isFPOrFPVector()) return;
// %b EQ %c then %a EQ %b
Value *Canonical = VN.canonicalize(SI->getCondition(), Top);
- if (Canonical == ConstantInt::getTrue()) {
+ if (Canonical == Context->getTrue()) {
add(SI, SI->getTrueValue(), ICmpInst::ICMP_EQ, NewContext);
- } else if (Canonical == ConstantInt::getFalse()) {
+ } else if (Canonical == Context->getFalse()) {
add(SI, SI->getFalseValue(), ICmpInst::ICMP_EQ, NewContext);
} else if (VN.canonicalize(SI->getTrueValue(), Top) ==
VN.canonicalize(SI->getFalseValue(), Top)) {
}
uint32_t W = VR.typeToWidth(DestTy);
- Value *TheCI = VN.canonicalize(CI, Top);
- ConstantRange CR = VR.rangeFromValue(Op, Top, W);
+ unsigned ci = VN.getOrInsertVN(CI, Top);
+ ConstantRange CR = VR.range(VN.getOrInsertVN(Op, Top), Top);
if (!CR.isFullSet()) {
switch (Opcode) {
default: break;
case Instruction::ZExt:
- VR.applyRange(TheCI, CR.zeroExtend(W), Top, this);
+ VR.applyRange(ci, CR.zeroExtend(W), Top, this);
break;
case Instruction::SExt:
- VR.applyRange(TheCI, CR.signExtend(W), Top, this);
+ VR.applyRange(ci, CR.signExtend(W), Top, this);
break;
case Instruction::Trunc: {
ConstantRange Result = CR.truncate(W);
if (!Result.isFullSet())
- VR.applyRange(TheCI, Result, Top, this);
+ VR.applyRange(ci, Result, Top, this);
} break;
case Instruction::BitCast:
- VR.applyRange(TheCI, CR, Top, this);
+ VR.applyRange(ci, CR, Top, this);
break;
// TODO: other casts?
}
// TODO: The GEPI indices are all zero. Copy from operand to definition,
// jumping the type plane as needed.
Value *Ptr = GEPI->getPointerOperand();
- if (isRelatedBy(Ptr, Constant::getNullValue(Ptr->getType()),
+ if (isRelatedBy(Ptr, Context->getNullValue(Ptr->getType()),
ICmpInst::ICMP_NE)) {
- add(GEPI, Constant::getNullValue(GEPI->getType()), ICmpInst::ICMP_NE,
+ add(GEPI, Context->getNullValue(GEPI->getType()), ICmpInst::ICMP_NE,
NewContext);
}
}
else DOUT << " context block: " << O.ContextBB->getName();
DOUT << "\n";
+ DEBUG(VN.dump());
DEBUG(IG.dump());
+ DEBUG(VR.dump());
// If they're both Constant, skip it. Check for contradiction and mark
// the BB as unreachable if so.
if (Constant *CI_L = dyn_cast<Constant>(O.LHS)) {
if (Constant *CI_R = dyn_cast<Constant>(O.RHS)) {
- if (ConstantExpr::getCompare(O.Op, CI_L, CI_R) ==
- ConstantInt::getFalse())
+ if (Context->getConstantExprCompare(O.Op, CI_L, CI_R) ==
+ Context->getFalse())
UB.mark(TopBB);
WorkList.pop_front();
continue;
}
- unsigned n1 = VN.valueNumber(O.LHS, Top);
- unsigned n2 = VN.valueNumber(O.RHS, Top);
+ unsigned n1 = VN.getOrInsertVN(O.LHS, Top);
+ unsigned n2 = VN.getOrInsertVN(O.RHS, Top);
- if (n1 && n1 == n2) {
+ if (n1 == n2) {
if (O.Op != ICmpInst::ICMP_UGE && O.Op != ICmpInst::ICMP_ULE &&
O.Op != ICmpInst::ICMP_SGE && O.Op != ICmpInst::ICMP_SLE)
UB.mark(TopBB);
continue;
}
- if (VR.isRelatedBy(O.LHS, O.RHS, Top, LV) ||
- (n1 && n2 && IG.isRelatedBy(n1, n2, Top, LV))) {
+ if (VR.isRelatedBy(n1, n2, Top, LV) ||
+ IG.isRelatedBy(n1, n2, Top, LV)) {
WorkList.pop_front();
continue;
}
- VR.addInequality(O.LHS, O.RHS, Top, LV, this);
+ VR.addInequality(n1, n2, Top, LV, this);
if ((!isa<ConstantInt>(O.RHS) && !isa<ConstantInt>(O.LHS)) ||
- LV == NE) {
- if (!n1) n1 = IG.newNode(O.LHS);
- if (!n2) n2 = IG.newNode(O.RHS);
+ LV == NE)
IG.addInequality(n1, n2, Top, LV);
- }
if (Instruction *I1 = dyn_cast<Instruction>(O.LHS)) {
if (aboveOrBelow(I1))
UE = O.LHS->use_end(); UI != UE;) {
Use &TheUse = UI.getUse();
++UI;
- if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
- if (aboveOrBelow(I))
- opsToDef(I);
- }
+ Instruction *I = cast<Instruction>(TheUse.getUser());
+ if (aboveOrBelow(I))
+ opsToDef(I);
}
}
if (Instruction *I2 = dyn_cast<Instruction>(O.RHS)) {
UE = O.RHS->use_end(); UI != UE;) {
Use &TheUse = UI.getUse();
++UI;
- if (Instruction *I = dyn_cast<Instruction>(TheUse.getUser())) {
- if (aboveOrBelow(I))
- opsToDef(I);
- }
+ Instruction *I = cast<Instruction>(TheUse.getUser());
+ if (aboveOrBelow(I))
+ opsToDef(I);
}
}
}
VRP->UB.mark(VRP->TopBB);
}
-#ifndef NDEBUG
- bool ValueRanges::isCanonical(Value *V, DomTreeDFS::Node *Subtree,
- VRPSolver *VRP) {
- return V == VRP->VN.canonicalize(V, Subtree);
- }
-#endif
-
/// PredicateSimplifier - This class is a simplifier that replaces
/// one equivalent variable with another. It also tracks what
/// can't be equal and will solve setcc instructions when possible.
std::vector<DomTreeDFS::Node *> WorkList;
+ LLVMContext *Context;
public:
static char ID; // Pass identification, replacement for typeid
- PredicateSimplifier() : FunctionPass((intptr_t)&ID) {}
+ PredicateSimplifier() : FunctionPass(&ID) {}
bool runOnFunction(Function &F);
}
private:
- /// Forwards - Adds new properties into PropertySet and uses them to
+ /// Forwards - Adds new properties to VRPSolver and uses them to
/// simplify instructions. Because new properties sometimes apply to
/// a transition from one BasicBlock to another, this will use the
/// PredicateSimplifier::proceedToSuccessor(s) interface to enter the
- /// basic block with the new PropertySet.
+ /// basic block.
/// @brief Performs abstract execution of the program.
class VISIBILITY_HIDDEN Forwards : public InstVisitor<Forwards> {
friend class InstVisitor<Forwards>;
}
}
- // Tries to simplify each Instruction and add new properties to
- // the PropertySet.
+ // Tries to simplify each Instruction and add new properties.
void visitInstruction(Instruction *I, DomTreeDFS::Node *DT) {
DOUT << "Considering instruction " << *I << "\n";
+ DEBUG(VN->dump());
DEBUG(IG->dump());
+ DEBUG(VR->dump());
// Sometimes instructions are killed in earlier analysis.
if (isInstructionTriviallyDead(I)) {
DominatorTree *DT = &getAnalysis<DominatorTree>();
DTDFS = new DomTreeDFS(DT);
TargetData *TD = &getAnalysis<TargetData>();
+ Context = &F.getContext();
DOUT << "Entering Function: " << F.getName() << "\n";
DomTreeDFS::Node *Root = DTDFS->getRootNode();
VN = new ValueNumbering(DTDFS);
IG = new InequalityGraph(*VN, Root);
- VR = new ValueRanges(TD);
+ VR = new ValueRanges(*VN, TD, Context);
WorkList.push_back(Root);
do {
delete DTDFS;
delete VR;
delete IG;
+ delete VN;
modified |= UB.kill();
return;
}
+ LLVMContext *Context = &BI.getContext();
+
for (DomTreeDFS::Node::iterator I = DTNode->begin(), E = DTNode->end();
I != E; ++I) {
BasicBlock *Dest = (*I)->getBlock();
if (Dest == TrueDest) {
DOUT << "(" << DTNode->getBlock()->getName() << ") true set:\n";
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, Dest);
- VRP.add(ConstantInt::getTrue(), Condition, ICmpInst::ICMP_EQ);
+ VRP.add(Context->getTrue(), Condition, ICmpInst::ICMP_EQ);
VRP.solve();
+ DEBUG(VN.dump());
DEBUG(IG.dump());
+ DEBUG(VR.dump());
} else if (Dest == FalseDest) {
DOUT << "(" << DTNode->getBlock()->getName() << ") false set:\n";
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, Dest);
- VRP.add(ConstantInt::getFalse(), Condition, ICmpInst::ICMP_EQ);
+ VRP.add(Context->getFalse(), Condition, ICmpInst::ICMP_EQ);
VRP.solve();
+ DEBUG(VN.dump());
DEBUG(IG.dump());
+ DEBUG(VR.dump());
}
PS->proceedToSuccessor(*I);
void PredicateSimplifier::Forwards::visitAllocaInst(AllocaInst &AI) {
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &AI);
- VRP.add(Constant::getNullValue(AI.getType()), &AI, ICmpInst::ICMP_NE);
+ VRP.add(AI.getContext().getNullValue(AI.getType()),
+ &AI, ICmpInst::ICMP_NE);
VRP.solve();
}
void PredicateSimplifier::Forwards::visitLoadInst(LoadInst &LI) {
Value *Ptr = LI.getPointerOperand();
- // avoid "load uint* null" -> null NE null.
+ // avoid "load i8* null" -> null NE null.
if (isa<Constant>(Ptr)) return;
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &LI);
- VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
+ VRP.add(LI.getContext().getNullValue(Ptr->getType()),
+ Ptr, ICmpInst::ICMP_NE);
VRP.solve();
}
if (isa<Constant>(Ptr)) return;
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &SI);
- VRP.add(Constant::getNullValue(Ptr->getType()), Ptr, ICmpInst::ICMP_NE);
+ VRP.add(SI.getContext().getNullValue(Ptr->getType()),
+ Ptr, ICmpInst::ICMP_NE);
VRP.solve();
}
void PredicateSimplifier::Forwards::visitSExtInst(SExtInst &SI) {
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &SI);
+ LLVMContext *Context = &SI.getContext();
uint32_t SrcBitWidth = cast<IntegerType>(SI.getSrcTy())->getBitWidth();
uint32_t DstBitWidth = cast<IntegerType>(SI.getDestTy())->getBitWidth();
APInt Min(APInt::getHighBitsSet(DstBitWidth, DstBitWidth-SrcBitWidth+1));
APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth-1));
- VRP.add(ConstantInt::get(Min), &SI, ICmpInst::ICMP_SLE);
- VRP.add(ConstantInt::get(Max), &SI, ICmpInst::ICMP_SGE);
+ VRP.add(Context->getConstantInt(Min), &SI, ICmpInst::ICMP_SLE);
+ VRP.add(Context->getConstantInt(Max), &SI, ICmpInst::ICMP_SGE);
VRP.solve();
}
void PredicateSimplifier::Forwards::visitZExtInst(ZExtInst &ZI) {
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &ZI);
+ LLVMContext *Context = &ZI.getContext();
uint32_t SrcBitWidth = cast<IntegerType>(ZI.getSrcTy())->getBitWidth();
uint32_t DstBitWidth = cast<IntegerType>(ZI.getDestTy())->getBitWidth();
APInt Max(APInt::getLowBitsSet(DstBitWidth, SrcBitWidth));
- VRP.add(ConstantInt::get(Max), &ZI, ICmpInst::ICMP_UGE);
+ VRP.add(Context->getConstantInt(Max), &ZI, ICmpInst::ICMP_UGE);
VRP.solve();
}
case Instruction::SDiv: {
Value *Divisor = BO.getOperand(1);
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &BO);
- VRP.add(Constant::getNullValue(Divisor->getType()), Divisor,
- ICmpInst::ICMP_NE);
+ VRP.add(BO.getContext().getNullValue(Divisor->getType()),
+ Divisor, ICmpInst::ICMP_NE);
VRP.solve();
break;
}
Pred = IC.getPredicate();
+ LLVMContext *Context = &IC.getContext();
+
if (ConstantInt *Op1 = dyn_cast<ConstantInt>(IC.getOperand(1))) {
ConstantInt *NextVal = 0;
switch (Pred) {
case ICmpInst::ICMP_SLT:
case ICmpInst::ICMP_ULT:
if (Op1->getValue() != 0)
- NextVal = ConstantInt::get(Op1->getValue()-1);
+ NextVal = Context->getConstantInt(Op1->getValue()-1);
break;
case ICmpInst::ICMP_SGT:
case ICmpInst::ICMP_UGT:
if (!Op1->getValue().isAllOnesValue())
- NextVal = ConstantInt::get(Op1->getValue()+1);
+ NextVal = Context->getConstantInt(Op1->getValue()+1);
break;
-
}
+
if (NextVal) {
VRPSolver VRP(VN, IG, UB, VR, PS->DTDFS, PS->modified, &IC);
if (VRP.isRelatedBy(IC.getOperand(0), NextVal,
ICmpInst::getInversePredicate(Pred))) {
- ICmpInst *NewIC = new ICmpInst(ICmpInst::ICMP_EQ, IC.getOperand(0),
- NextVal, "", &IC);
+ ICmpInst *NewIC = new ICmpInst(&IC, ICmpInst::ICMP_EQ,
+ IC.getOperand(0), NextVal, "");
NewIC->takeName(&IC);
IC.replaceAllUsesWith(NewIC);
}
}
}
-
- char PredicateSimplifier::ID = 0;
- RegisterPass<PredicateSimplifier> X("predsimplify",
- "Predicate Simplifier");
}
+char PredicateSimplifier::ID = 0;
+static RegisterPass<PredicateSimplifier>
+X("predsimplify", "Predicate Simplifier");
+
FunctionPass *llvm::createPredicateSimplifierPass() {
return new PredicateSimplifier();
}