axmol/scripting/javascript/spidermonkey-win32/include/js/HashTable.h

1371 lines
43 KiB
C
Raw Normal View History

/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*-
* vim: set ts=8 sw=4 et tw=99 ft=cpp:
*
* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
2013-02-27 15:45:23 +08:00
#ifndef js_HashTable_h__
#define js_HashTable_h__
#include "mozilla/Attributes.h"
#include "mozilla/DebugOnly.h"
#include "mozilla/Util.h"
2013-02-27 15:45:23 +08:00
#include "js/TemplateLib.h"
#include "js/Utility.h"
namespace js {
class TempAllocPolicy;
2013-02-27 15:45:23 +08:00
template <class> struct DefaultHasher;
template <class, class> class HashMapEntry;
namespace detail {
template <class T> class HashTableEntry;
template <class T, class HashPolicy, class AllocPolicy> class HashTable;
}
/*****************************************************************************/
2013-02-27 15:45:23 +08:00
// A JS-friendly, STL-like container providing a hash-based map from keys to
// values. In particular, HashMap calls constructors and destructors of all
// objects added so non-PODs may be used safely.
//
// Key/Value requirements:
// - movable, destructible, assignable
// HashPolicy requirements:
// - see Hash Policy section below
// AllocPolicy:
// - see jsalloc.h
//
// Note:
// - HashMap is not reentrant: Key/Value/HashPolicy/AllocPolicy members
// called by HashMap must not call back into the same HashMap object.
// - Due to the lack of exception handling, the user must call |init()|.
template <class Key,
class Value,
class HashPolicy = DefaultHasher<Key>,
class AllocPolicy = TempAllocPolicy>
class HashMap
{
typedef HashMapEntry<Key, Value> TableEntry;
2013-02-27 15:45:23 +08:00
struct MapHashPolicy : HashPolicy
{
typedef Key KeyType;
static const Key &getKey(TableEntry &e) { return e.key; }
static void setKey(TableEntry &e, Key &k) { const_cast<Key &>(e.key) = k; }
};
2013-02-27 15:45:23 +08:00
typedef detail::HashTable<TableEntry, MapHashPolicy, AllocPolicy> Impl;
Impl impl;
2013-02-27 15:45:23 +08:00
public:
typedef typename HashPolicy::Lookup Lookup;
typedef TableEntry Entry;
2013-02-27 15:45:23 +08:00
// HashMap construction is fallible (due to OOM); thus the user must call
// init after constructing a HashMap and check the return value.
HashMap(AllocPolicy a = AllocPolicy()) : impl(a) {}
bool init(uint32_t len = 16) { return impl.init(len); }
bool initialized() const { return impl.initialized(); }
2013-02-27 15:45:23 +08:00
// Return whether the given lookup value is present in the map. E.g.:
//
// typedef HashMap<int,char> HM;
// HM h;
// if (HM::Ptr p = h.lookup(3)) {
// const HM::Entry &e = *p; // p acts like a pointer to Entry
// assert(p->key == 3); // Entry contains the key
// char val = p->value; // and value
// }
//
// Also see the definition of Ptr in HashTable above (with T = Entry).
typedef typename Impl::Ptr Ptr;
Ptr lookup(const Lookup &l) const { return impl.lookup(l); }
2013-02-27 15:45:23 +08:00
// Assuming |p.found()|, remove |*p|.
void remove(Ptr p) { impl.remove(p); }
// Like |lookup(l)|, but on miss, |p = lookupForAdd(l)| allows efficient
// insertion of Key |k| (where |HashPolicy::match(k,l) == true|) using
// |add(p,k,v)|. After |add(p,k,v)|, |p| points to the new Entry. E.g.:
//
// typedef HashMap<int,char> HM;
// HM h;
// HM::AddPtr p = h.lookupForAdd(3);
// if (!p) {
// if (!h.add(p, 3, 'a'))
// return false;
// }
// const HM::Entry &e = *p; // p acts like a pointer to Entry
// assert(p->key == 3); // Entry contains the key
// char val = p->value; // and value
//
// Also see the definition of AddPtr in HashTable above (with T = Entry).
//
// N.B. The caller must ensure that no mutating hash table operations
// occur between a pair of |lookupForAdd| and |add| calls. To avoid
// looking up the key a second time, the caller may use the more efficient
// relookupOrAdd method. This method reuses part of the hashing computation
// to more efficiently insert the key if it has not been added. For
// example, a mutation-handling version of the previous example:
//
// HM::AddPtr p = h.lookupForAdd(3);
// if (!p) {
// call_that_may_mutate_h();
// if (!h.relookupOrAdd(p, 3, 'a'))
// return false;
// }
// const HM::Entry &e = *p;
// assert(p->key == 3);
// char val = p->value;
typedef typename Impl::AddPtr AddPtr;
AddPtr lookupForAdd(const Lookup &l) const {
return impl.lookupForAdd(l);
}
2013-02-27 15:45:23 +08:00
template<typename KeyInput, typename ValueInput>
bool add(AddPtr &p, const KeyInput &k, const ValueInput &v) {
Entry e(k, v);
return impl.add(p, Move(e));
}
2013-02-27 15:45:23 +08:00
bool add(AddPtr &p, const Key &k) {
Entry e(k, Value());
return impl.add(p, Move(e));
}
2013-02-27 15:45:23 +08:00
template<typename KeyInput, typename ValueInput>
bool relookupOrAdd(AddPtr &p, const KeyInput &k, const ValueInput &v) {
Entry e(k, v);
return impl.relookupOrAdd(p, k, Move(e));
}
// |all()| returns a Range containing |count()| elements. E.g.:
//
// typedef HashMap<int,char> HM;
// HM h;
// for (HM::Range r = h.all(); !r.empty(); r.popFront())
// char c = r.front().value;
//
// Also see the definition of Range in HashTable above (with T = Entry).
typedef typename Impl::Range Range;
Range all() const { return impl.all(); }
2013-02-27 15:45:23 +08:00
// Typedef for the enumeration class. An Enum may be used to examine and
// remove table entries:
//
// typedef HashMap<int,char> HM;
// HM s;
// for (HM::Enum e(s); !e.empty(); e.popFront())
// if (e.front().value == 'l')
// e.removeFront();
//
// Table resize may occur in Enum's destructor. Also see the definition of
// Enum in HashTable above (with T = Entry).
typedef typename Impl::Enum Enum;
2013-02-27 15:45:23 +08:00
// Remove all entries. This does not shrink the table. For that consider
// using the finish() method.
void clear() { impl.clear(); }
2013-02-27 15:45:23 +08:00
// Remove all the entries and release all internal buffers. The map must
// be initialized again before any use.
void finish() { impl.finish(); }
2013-02-27 15:45:23 +08:00
// Does the table contain any entries?
bool empty() const { return impl.empty(); }
2013-02-27 15:45:23 +08:00
// Number of live elements in the map.
uint32_t count() const { return impl.count(); }
2013-02-27 15:45:23 +08:00
// Total number of allocation in the dynamic table. Note: resize will
// happen well before count() == capacity().
size_t capacity() const { return impl.capacity(); }
2013-02-27 15:45:23 +08:00
// Don't just call |impl.sizeOfExcludingThis()| because there's no
// guarantee that |impl| is the first field in HashMap.
size_t sizeOfExcludingThis(JSMallocSizeOfFun mallocSizeOf) const {
return impl.sizeOfExcludingThis(mallocSizeOf);
}
size_t sizeOfIncludingThis(JSMallocSizeOfFun mallocSizeOf) const {
return mallocSizeOf(this) + impl.sizeOfExcludingThis(mallocSizeOf);
}
2013-02-27 15:45:23 +08:00
// If |generation()| is the same before and after a HashMap operation,
// pointers into the table remain valid.
unsigned generation() const { return impl.generation(); }
2013-02-27 15:45:23 +08:00
/************************************************** Shorthand operations */
2013-02-27 15:45:23 +08:00
bool has(const Lookup &l) const {
return impl.lookup(l) != NULL;
}
2013-02-27 15:45:23 +08:00
// Overwrite existing value with v. Return false on oom.
template<typename KeyInput, typename ValueInput>
bool put(const KeyInput &k, const ValueInput &v) {
AddPtr p = lookupForAdd(k);
if (p) {
p->value = v;
return true;
}
2013-02-27 15:45:23 +08:00
return add(p, k, v);
}
2013-02-27 15:45:23 +08:00
// Like put, but assert that the given key is not already present.
template<typename KeyInput, typename ValueInput>
bool putNew(const KeyInput &k, const ValueInput &v) {
Entry e(k, v);
return impl.putNew(k, Move(e));
}
2013-02-27 15:45:23 +08:00
// Add (k,defaultValue) if |k| is not found. Return a false-y Ptr on oom.
Ptr lookupWithDefault(const Key &k, const Value &defaultValue) {
AddPtr p = lookupForAdd(k);
if (p)
return p;
(void)add(p, k, defaultValue); // p is left false-y on oom.
return p;
}
2013-02-27 15:45:23 +08:00
// Remove if present.
void remove(const Lookup &l) {
if (Ptr p = lookup(l))
remove(p);
}
2013-02-27 15:45:23 +08:00
private:
// Not implicitly copyable (expensive). May add explicit |clone| later.
HashMap(const HashMap &hm) MOZ_DELETE;
HashMap &operator=(const HashMap &hm) MOZ_DELETE;
2013-02-27 15:45:23 +08:00
friend class Impl::Enum;
2013-02-27 15:45:23 +08:00
typedef typename tl::StaticAssert<tl::IsRelocatableHeapType<Key>::result>::result keyAssert;
typedef typename tl::StaticAssert<tl::IsRelocatableHeapType<Value>::result>::result valAssert;
};
2013-02-27 15:45:23 +08:00
/*****************************************************************************/
2013-02-27 15:45:23 +08:00
// A JS-friendly, STL-like container providing a hash-based set of values. In
// particular, HashSet calls constructors and destructors of all objects added
// so non-PODs may be used safely.
//
// T requirements:
// - movable, destructible, assignable
// HashPolicy requirements:
// - see Hash Policy section below
// AllocPolicy:
// - see jsalloc.h
//
// Note:
// - HashSet is not reentrant: T/HashPolicy/AllocPolicy members called by
// HashSet must not call back into the same HashSet object.
// - Due to the lack of exception handling, the user must call |init()|.
template <class T,
class HashPolicy = DefaultHasher<T>,
class AllocPolicy = TempAllocPolicy>
class HashSet
{
struct SetOps : HashPolicy
{
typedef T KeyType;
static const KeyType &getKey(const T &t) { return t; }
static void setKey(T &t, KeyType &k) { t = k; }
};
2013-02-27 15:45:23 +08:00
typedef detail::HashTable<const T, SetOps, AllocPolicy> Impl;
Impl impl;
2013-02-27 15:45:23 +08:00
public:
typedef typename HashPolicy::Lookup Lookup;
typedef T Entry;
2013-02-27 15:45:23 +08:00
// HashSet construction is fallible (due to OOM); thus the user must call
// init after constructing a HashSet and check the return value.
HashSet(AllocPolicy a = AllocPolicy()) : impl(a) {}
bool init(uint32_t len = 16) { return impl.init(len); }
bool initialized() const { return impl.initialized(); }
2013-02-27 15:45:23 +08:00
// Return whether the given lookup value is present in the map. E.g.:
//
// typedef HashSet<int> HS;
// HS h;
// if (HS::Ptr p = h.lookup(3)) {
// assert(*p == 3); // p acts like a pointer to int
// }
//
// Also see the definition of Ptr in HashTable above.
typedef typename Impl::Ptr Ptr;
Ptr lookup(const Lookup &l) const { return impl.lookup(l); }
2013-02-27 15:45:23 +08:00
// Assuming |p.found()|, remove |*p|.
void remove(Ptr p) { impl.remove(p); }
2013-02-27 15:45:23 +08:00
// Like |lookup(l)|, but on miss, |p = lookupForAdd(l)| allows efficient
// insertion of T value |t| (where |HashPolicy::match(t,l) == true|) using
// |add(p,t)|. After |add(p,t)|, |p| points to the new element. E.g.:
//
// typedef HashSet<int> HS;
// HS h;
// HS::AddPtr p = h.lookupForAdd(3);
// if (!p) {
// if (!h.add(p, 3))
// return false;
// }
// assert(*p == 3); // p acts like a pointer to int
//
// Also see the definition of AddPtr in HashTable above.
//
// N.B. The caller must ensure that no mutating hash table operations
// occur between a pair of |lookupForAdd| and |add| calls. To avoid
// looking up the key a second time, the caller may use the more efficient
// relookupOrAdd method. This method reuses part of the hashing computation
// to more efficiently insert the key if it has not been added. For
// example, a mutation-handling version of the previous example:
//
// HS::AddPtr p = h.lookupForAdd(3);
// if (!p) {
// call_that_may_mutate_h();
// if (!h.relookupOrAdd(p, 3, 3))
// return false;
// }
// assert(*p == 3);
//
// Note that relookupOrAdd(p,l,t) performs Lookup using |l| and adds the
// entry |t|, where the caller ensures match(l,t).
typedef typename Impl::AddPtr AddPtr;
AddPtr lookupForAdd(const Lookup &l) const { return impl.lookupForAdd(l); }
2013-02-27 15:45:23 +08:00
bool add(AddPtr &p, const T &t) { return impl.add(p, t); }
bool relookupOrAdd(AddPtr &p, const Lookup &l, const T &t) {
return impl.relookupOrAdd(p, l, t);
}
2013-02-27 15:45:23 +08:00
// |all()| returns a Range containing |count()| elements:
//
// typedef HashSet<int> HS;
// HS h;
// for (HS::Range r = h.all(); !r.empty(); r.popFront())
// int i = r.front();
//
// Also see the definition of Range in HashTable above.
typedef typename Impl::Range Range;
Range all() const { return impl.all(); }
2013-02-27 15:45:23 +08:00
// Typedef for the enumeration class. An Enum may be used to examine and
// remove table entries:
//
// typedef HashSet<int> HS;
// HS s;
// for (HS::Enum e(s); !e.empty(); e.popFront())
// if (e.front() == 42)
// e.removeFront();
//
// Table resize may occur in Enum's destructor. Also see the definition of
// Enum in HashTable above.
typedef typename Impl::Enum Enum;
2013-02-27 15:45:23 +08:00
// Remove all entries. This does not shrink the table. For that consider
// using the finish() method.
void clear() { impl.clear(); }
2013-02-27 15:45:23 +08:00
// Remove all the entries and release all internal buffers. The set must
// be initialized again before any use.
void finish() { impl.finish(); }
2013-02-27 15:45:23 +08:00
// Does the table contain any entries?
bool empty() const { return impl.empty(); }
2013-02-27 15:45:23 +08:00
// Number of live elements in the map.
uint32_t count() const { return impl.count(); }
2013-02-27 15:45:23 +08:00
// Total number of allocation in the dynamic table. Note: resize will
// happen well before count() == capacity().
size_t capacity() const { return impl.capacity(); }
2013-02-27 15:45:23 +08:00
// Don't just call |impl.sizeOfExcludingThis()| because there's no
// guarantee that |impl| is the first field in HashSet.
size_t sizeOfExcludingThis(JSMallocSizeOfFun mallocSizeOf) const {
return impl.sizeOfExcludingThis(mallocSizeOf);
}
2013-02-27 15:45:23 +08:00
size_t sizeOfIncludingThis(JSMallocSizeOfFun mallocSizeOf) const {
return mallocSizeOf(this) + impl.sizeOfExcludingThis(mallocSizeOf);
}
2013-02-27 15:45:23 +08:00
// If |generation()| is the same before and after a HashSet operation,
// pointers into the table remain valid.
unsigned generation() const { return impl.generation(); }
2013-02-27 15:45:23 +08:00
/************************************************** Shorthand operations */
2013-02-27 15:45:23 +08:00
bool has(const Lookup &l) const {
return impl.lookup(l) != NULL;
}
2013-02-27 15:45:23 +08:00
// Overwrite existing value with v. Return false on oom.
bool put(const T &t) {
AddPtr p = lookupForAdd(t);
return p ? true : add(p, t);
}
2013-02-27 15:45:23 +08:00
// Like put, but assert that the given key is not already present.
bool putNew(const T &t) {
return impl.putNew(t, t);
}
2013-02-27 15:45:23 +08:00
bool putNew(const Lookup &l, const T &t) {
return impl.putNew(l, t);
}
2013-02-27 15:45:23 +08:00
void remove(const Lookup &l) {
if (Ptr p = lookup(l))
remove(p);
}
private:
2013-02-27 15:45:23 +08:00
// Not implicitly copyable (expensive). May add explicit |clone| later.
HashSet(const HashSet &hs) MOZ_DELETE;
HashSet &operator=(const HashSet &hs) MOZ_DELETE;
2013-02-27 15:45:23 +08:00
friend class Impl::Enum;
2013-02-27 15:45:23 +08:00
typedef typename tl::StaticAssert<tl::IsRelocatableHeapType<T>::result>::result _;
};
2013-02-27 15:45:23 +08:00
/*****************************************************************************/
2013-02-27 15:45:23 +08:00
// Hash Policy
//
// A hash policy P for a hash table with key-type Key must provide:
// - a type |P::Lookup| to use to lookup table entries;
// - a static member function |P::hash| with signature
//
// static js::HashNumber hash(Lookup)
//
// to use to hash the lookup type; and
// - a static member function |P::match| with signature
//
// static bool match(Key, Lookup)
//
// to use to test equality of key and lookup values.
//
// Normally, Lookup = Key. In general, though, different values and types of
// values can be used to lookup and store. If a Lookup value |l| is != to the
// added Key value |k|, the user must ensure that |P::match(k,l)|. E.g.:
//
// js::HashSet<Key, P>::AddPtr p = h.lookup(l);
// if (!p) {
// assert(P::match(k, l)); // must hold
// h.add(p, k);
// }
// Pointer hashing policy that strips the lowest zeroBits when calculating the
// hash to improve key distribution.
template <typename Key, size_t zeroBits>
struct PointerHasher
{
typedef Key Lookup;
static HashNumber hash(const Lookup &l) {
JS_ASSERT(!js::IsPoisonedPtr(l));
2013-02-27 15:45:23 +08:00
size_t word = reinterpret_cast<size_t>(l) >> zeroBits;
JS_STATIC_ASSERT(sizeof(HashNumber) == 4);
#if JS_BYTES_PER_WORD == 4
return HashNumber(word);
#else
JS_STATIC_ASSERT(sizeof word == 8);
return HashNumber((word >> 32) ^ word);
#endif
}
2013-02-27 15:45:23 +08:00
static bool match(const Key &k, const Lookup &l) {
JS_ASSERT(!js::IsPoisonedPtr(k));
JS_ASSERT(!js::IsPoisonedPtr(l));
2013-02-27 15:45:23 +08:00
return k == l;
}
2013-02-27 15:45:23 +08:00
};
2013-02-27 15:45:23 +08:00
// Default hash policy: just use the 'lookup' value. This of course only
// works if the lookup value is integral. HashTable applies ScrambleHashCode to
// the result of the 'hash' which means that it is 'ok' if the lookup value is
// not well distributed over the HashNumber domain.
template <class Key>
struct DefaultHasher
{
typedef Key Lookup;
static HashNumber hash(const Lookup &l) {
// Hash if can implicitly cast to hash number type.
return l;
}
2013-02-27 15:45:23 +08:00
static bool match(const Key &k, const Lookup &l) {
// Use builtin or overloaded operator==.
return k == l;
}
};
2013-02-27 15:45:23 +08:00
// Specialize hashing policy for pointer types. It assumes that the type is
// at least word-aligned. For types with smaller size use PointerHasher.
template <class T>
struct DefaultHasher<T *> : PointerHasher<T *, tl::FloorLog2<sizeof(void *)>::result>
{};
2013-02-27 15:45:23 +08:00
/*****************************************************************************/
2013-02-27 15:45:23 +08:00
// Both HashMap and HashSet are implemented by a single HashTable that is even
// more heavily parameterized than the other two. This leaves HashTable gnarly
// and extremely coupled to HashMap and HashSet; thus code should not use
// HashTable directly.
2013-02-27 15:45:23 +08:00
template <class Key, class Value>
class HashMapEntry
{
template <class, class, class> friend class detail::HashTable;
template <class> friend class detail::HashTableEntry;
2013-02-27 15:45:23 +08:00
HashMapEntry(const HashMapEntry &) MOZ_DELETE;
void operator=(const HashMapEntry &) MOZ_DELETE;
2013-02-27 15:45:23 +08:00
public:
template<typename KeyInput, typename ValueInput>
HashMapEntry(const KeyInput &k, const ValueInput &v) : key(k), value(v) {}
2013-02-27 15:45:23 +08:00
HashMapEntry(MoveRef<HashMapEntry> rhs)
: key(Move(rhs->key)), value(Move(rhs->value)) { }
2013-02-27 15:45:23 +08:00
const Key key;
Value value;
};
2013-02-27 15:45:23 +08:00
namespace tl {
2013-02-27 15:45:23 +08:00
template <class T>
struct IsPodType<detail::HashTableEntry<T> > {
static const bool result = IsPodType<T>::result;
};
2013-02-27 15:45:23 +08:00
template <class K, class V>
struct IsPodType<HashMapEntry<K, V> >
{
static const bool result = IsPodType<K>::result && IsPodType<V>::result;
};
2013-02-27 15:45:23 +08:00
} // namespace tl
2013-02-27 15:45:23 +08:00
namespace detail {
2013-02-27 15:45:23 +08:00
template <class T, class HashPolicy, class AllocPolicy>
class HashTable;
2013-02-27 15:45:23 +08:00
template <class T>
class HashTableEntry
{
template <class, class, class> friend class HashTable;
typedef typename tl::StripConst<T>::result NonConstT;
2013-02-27 15:45:23 +08:00
HashNumber keyHash;
mozilla::AlignedStorage2<NonConstT> mem;
2013-02-27 15:45:23 +08:00
static const HashNumber sFreeKey = 0;
static const HashNumber sRemovedKey = 1;
static const HashNumber sCollisionBit = 1;
2013-02-27 15:45:23 +08:00
// Assumed by calloc in createTable.
JS_STATIC_ASSERT(sFreeKey == 0);
2013-02-27 15:45:23 +08:00
static bool isLiveHash(HashNumber hash)
{
2013-02-27 15:45:23 +08:00
return hash > sRemovedKey;
}
2013-02-27 15:45:23 +08:00
HashTableEntry(const HashTableEntry &) MOZ_DELETE;
void operator=(const HashTableEntry &) MOZ_DELETE;
~HashTableEntry() MOZ_DELETE;
2013-02-27 15:45:23 +08:00
public:
// NB: HashTableEntry is treated as a POD: no constructor or destructor calls.
2013-02-27 15:45:23 +08:00
void destroyIfLive() {
if (isLive())
mem.addr()->~T();
}
2013-02-27 15:45:23 +08:00
void destroy() {
JS_ASSERT(isLive());
mem.addr()->~T();
}
2013-02-27 15:45:23 +08:00
void swap(HashTableEntry *other) {
Swap(keyHash, other->keyHash);
Swap(mem, other->mem);
}
2013-02-27 15:45:23 +08:00
T &get() { JS_ASSERT(isLive()); return *mem.addr(); }
2013-02-27 15:45:23 +08:00
bool isFree() const { return keyHash == sFreeKey; }
void clearLive() { JS_ASSERT(isLive()); keyHash = sFreeKey; mem.addr()->~T(); }
void clear() { if (isLive()) mem.addr()->~T(); keyHash = sFreeKey; }
bool isRemoved() const { return keyHash == sRemovedKey; }
void removeLive() { JS_ASSERT(isLive()); keyHash = sRemovedKey; mem.addr()->~T(); }
bool isLive() const { return isLiveHash(keyHash); }
void setCollision() { JS_ASSERT(isLive()); keyHash |= sCollisionBit; }
void setCollision(HashNumber bit) { JS_ASSERT(isLive()); keyHash |= bit; }
void unsetCollision() { keyHash &= ~sCollisionBit; }
bool hasCollision() const { return keyHash & sCollisionBit; }
bool matchHash(HashNumber hn) { return (keyHash & ~sCollisionBit) == hn; }
HashNumber getKeyHash() const { return keyHash & ~sCollisionBit; }
2013-02-27 15:45:23 +08:00
template <class U>
void setLive(HashNumber hn, const U &u)
{
2013-02-27 15:45:23 +08:00
JS_ASSERT(!isLive());
keyHash = hn;
new(mem.addr()) T(u);
JS_ASSERT(isLive());
}
2013-02-27 15:45:23 +08:00
};
2013-02-27 15:45:23 +08:00
template <class T, class HashPolicy, class AllocPolicy>
class HashTable : private AllocPolicy
{
typedef typename tl::StripConst<T>::result NonConstT;
typedef typename HashPolicy::KeyType Key;
typedef typename HashPolicy::Lookup Lookup;
public:
typedef HashTableEntry<T> Entry;
// A nullable pointer to a hash table element. A Ptr |p| can be tested
// either explicitly |if (p.found()) p->...| or using boolean conversion
// |if (p) p->...|. Ptr objects must not be used after any mutating hash
// table operations unless |generation()| is tested.
class Ptr
{
2013-02-27 15:45:23 +08:00
friend class HashTable;
typedef void (Ptr::* ConvertibleToBool)();
void nonNull() {}
2013-02-27 15:45:23 +08:00
Entry *entry;
protected:
Ptr(Entry &entry) : entry(&entry) {}
public:
// Leaves Ptr uninitialized.
Ptr() {
#ifdef DEBUG
entry = (Entry *)0xbad;
#endif
}
2013-02-27 15:45:23 +08:00
bool found() const { return entry->isLive(); }
operator ConvertibleToBool() const { return found() ? &Ptr::nonNull : 0; }
bool operator==(const Ptr &rhs) const { JS_ASSERT(found() && rhs.found()); return entry == rhs.entry; }
bool operator!=(const Ptr &rhs) const { return !(*this == rhs); }
T &operator*() const { return entry->get(); }
T *operator->() const { return &entry->get(); }
};
// A Ptr that can be used to add a key after a failed lookup.
class AddPtr : public Ptr
{
2013-02-27 15:45:23 +08:00
friend class HashTable;
HashNumber keyHash;
mozilla::DebugOnly<uint64_t> mutationCount;
2013-02-27 15:45:23 +08:00
AddPtr(Entry &entry, HashNumber hn) : Ptr(entry), keyHash(hn) {}
public:
// Leaves AddPtr uninitialized.
AddPtr() {}
};
// A collection of hash table entries. The collection is enumerated by
// calling |front()| followed by |popFront()| as long as |!empty()|. As
// with Ptr/AddPtr, Range objects must not be used after any mutating hash
// table operation unless the |generation()| is tested.
class Range
{
2013-02-27 15:45:23 +08:00
protected:
friend class HashTable;
2013-02-27 15:45:23 +08:00
Range(Entry *c, Entry *e) : cur(c), end(e), validEntry(true) {
while (cur < end && !cur->isLive())
++cur;
}
2013-02-27 15:45:23 +08:00
Entry *cur, *end;
mozilla::DebugOnly<bool> validEntry;
2013-02-27 15:45:23 +08:00
public:
Range() : cur(NULL), end(NULL), validEntry(false) {}
2013-02-27 15:45:23 +08:00
bool empty() const {
return cur == end;
}
2013-02-27 15:45:23 +08:00
T &front() const {
JS_ASSERT(validEntry);
JS_ASSERT(!empty());
return cur->get();
}
2013-02-27 15:45:23 +08:00
void popFront() {
JS_ASSERT(!empty());
while (++cur < end && !cur->isLive())
continue;
validEntry = true;
}
2013-02-27 15:45:23 +08:00
};
2013-02-27 15:45:23 +08:00
// A Range whose lifetime delimits a mutating enumeration of a hash table.
// Since rehashing when elements were removed during enumeration would be
// bad, it is postponed until |endEnumeration()| is called. If
// |endEnumeration()| is not called before an Enum's constructor, it will
// be called automatically. Since |endEnumeration()| touches the hash
// table, the user must ensure that the hash table is still alive when this
// happens.
class Enum : public Range
{
2013-02-27 15:45:23 +08:00
friend class HashTable;
2013-02-27 15:45:23 +08:00
HashTable &table;
bool rekeyed;
bool removed;
2013-02-27 15:45:23 +08:00
/* Not copyable. */
Enum(const Enum &);
void operator=(const Enum &);
2013-02-27 15:45:23 +08:00
public:
template<class Map> explicit
Enum(Map &map) : Range(map.all()), table(map.impl), rekeyed(false), removed(false) {}
2013-02-27 15:45:23 +08:00
// Removes the |front()| element from the table, leaving |front()|
// invalid until the next call to |popFront()|. For example:
//
// HashSet<int> s;
// for (HashSet<int>::Enum e(s); !e.empty(); e.popFront())
// if (e.front() == 42)
// e.removeFront();
void removeFront() {
table.remove(*this->cur);
removed = true;
this->validEntry = false;
}
2013-02-27 15:45:23 +08:00
// Removes the |front()| element and re-inserts it into the table with
// a new key at the new Lookup position. |front()| is invalid after
// this operation until the next call to |popFront()|.
void rekeyFront(const Lookup &l, const Key &k) {
typename HashTableEntry<T>::NonConstT t(Move(this->cur->get()));
HashPolicy::setKey(t, const_cast<Key &>(k));
table.remove(*this->cur);
table.putNewInfallible(l, Move(t));
rekeyed = true;
this->validEntry = false;
}
2013-02-27 15:45:23 +08:00
void rekeyFront(const Key &k) {
rekeyFront(k, k);
}
2013-02-27 15:45:23 +08:00
// Potentially rehashes the table.
~Enum() {
if (rekeyed)
table.checkOverRemoved();
if (removed)
table.checkUnderloaded();
}
};
2013-02-27 15:45:23 +08:00
private:
uint32_t hashShift; // multiplicative hash shift
uint32_t entryCount; // number of entries in table
uint32_t gen; // entry storage generation number
uint32_t removedCount; // removed entry sentinels in table
Entry *table; // entry storage
2013-02-27 15:45:23 +08:00
void setTableSizeLog2(unsigned sizeLog2)
{
hashShift = sHashBits - sizeLog2;
}
2013-02-27 15:45:23 +08:00
#ifdef DEBUG
mutable struct Stats
{
uint32_t searches; // total number of table searches
uint32_t steps; // hash chain links traversed
uint32_t hits; // searches that found key
uint32_t misses; // searches that didn't find key
uint32_t addOverRemoved; // adds that recycled a removed entry
uint32_t removes; // calls to remove
uint32_t removeFrees; // calls to remove that freed the entry
uint32_t grows; // table expansions
uint32_t shrinks; // table contractions
uint32_t compresses; // table compressions
uint32_t rehashes; // tombstone decontaminations
} stats;
# define METER(x) x
#else
# define METER(x)
#endif
2013-02-27 15:45:23 +08:00
friend class js::ReentrancyGuard;
mutable mozilla::DebugOnly<bool> entered;
mozilla::DebugOnly<uint64_t> mutationCount;
// The default initial capacity is 16, but you can ask for as small as 4.
static const unsigned sMinSizeLog2 = 2;
static const unsigned sMinSize = 1 << sMinSizeLog2;
static const unsigned sMaxInit = JS_BIT(23);
static const unsigned sMaxCapacity = JS_BIT(24);
static const unsigned sHashBits = tl::BitSize<HashNumber>::result;
static const uint8_t sMinAlphaFrac = 64; // (0x100 * .25)
static const uint8_t sMaxAlphaFrac = 192; // (0x100 * .75)
static const uint8_t sInvMaxAlpha = 171; // (ceil(0x100 / .75) >> 1)
static const HashNumber sFreeKey = Entry::sFreeKey;
static const HashNumber sRemovedKey = Entry::sRemovedKey;
static const HashNumber sCollisionBit = Entry::sCollisionBit;
static void staticAsserts()
{
// Rely on compiler "constant overflow warnings".
JS_STATIC_ASSERT(((sMaxInit * sInvMaxAlpha) >> 7) < sMaxCapacity);
JS_STATIC_ASSERT((sMaxCapacity * sInvMaxAlpha) <= UINT32_MAX);
JS_STATIC_ASSERT((sMaxCapacity * sizeof(Entry)) <= UINT32_MAX);
}
2013-02-27 15:45:23 +08:00
static bool isLiveHash(HashNumber hash)
{
2013-02-27 15:45:23 +08:00
return Entry::isLiveHash(hash);
}
2013-02-27 15:45:23 +08:00
static HashNumber prepareHash(const Lookup& l)
{
HashNumber keyHash = ScrambleHashCode(HashPolicy::hash(l));
2013-02-27 15:45:23 +08:00
// Avoid reserved hash codes.
if (!isLiveHash(keyHash))
keyHash -= (sRemovedKey + 1);
return keyHash & ~sCollisionBit;
}
2013-02-27 15:45:23 +08:00
static Entry *createTable(AllocPolicy &alloc, uint32_t capacity)
{
2013-02-27 15:45:23 +08:00
// See JS_STATIC_ASSERT(sFreeKey == 0) in HashTableEntry.
return (Entry *)alloc.calloc_(capacity * sizeof(Entry));
}
2013-02-27 15:45:23 +08:00
static void destroyTable(AllocPolicy &alloc, Entry *oldTable, uint32_t capacity)
{
2013-02-27 15:45:23 +08:00
for (Entry *e = oldTable, *end = e + capacity; e < end; ++e)
e->destroyIfLive();
alloc.free_(oldTable);
}
2013-02-27 15:45:23 +08:00
public:
HashTable(AllocPolicy ap)
: AllocPolicy(ap),
hashShift(sHashBits),
entryCount(0),
gen(0),
removedCount(0),
table(NULL),
entered(false),
mutationCount(0)
{}
MOZ_WARN_UNUSED_RESULT bool init(uint32_t length)
{
2013-02-27 15:45:23 +08:00
JS_ASSERT(!initialized());
2013-02-27 15:45:23 +08:00
// Correct for sMaxAlphaFrac such that the table will not resize
// when adding 'length' entries.
if (length > sMaxInit) {
this->reportAllocOverflow();
return false;
}
uint32_t capacity = (length * sInvMaxAlpha) >> 7;
2013-02-27 15:45:23 +08:00
if (capacity < sMinSize)
capacity = sMinSize;
// FIXME: use JS_CEILING_LOG2 when PGO stops crashing (bug 543034).
uint32_t roundUp = sMinSize, roundUpLog2 = sMinSizeLog2;
while (roundUp < capacity) {
roundUp <<= 1;
++roundUpLog2;
}
2013-02-27 15:45:23 +08:00
capacity = roundUp;
JS_ASSERT(capacity <= sMaxCapacity);
2013-02-27 15:45:23 +08:00
table = createTable(*this, capacity);
if (!table)
return false;
2013-02-27 15:45:23 +08:00
setTableSizeLog2(roundUpLog2);
METER(memset(&stats, 0, sizeof(stats)));
return true;
}
2013-02-27 15:45:23 +08:00
bool initialized() const
{
2013-02-27 15:45:23 +08:00
return !!table;
}
2013-02-27 15:45:23 +08:00
~HashTable()
{
2013-02-27 15:45:23 +08:00
if (table)
destroyTable(*this, table, capacity());
}
2013-02-27 15:45:23 +08:00
private:
HashNumber hash1(HashNumber hash0) const
{
return hash0 >> hashShift;
}
2013-02-27 15:45:23 +08:00
struct DoubleHash
{
HashNumber h2;
HashNumber sizeMask;
};
2013-02-27 15:45:23 +08:00
DoubleHash hash2(HashNumber curKeyHash) const
{
unsigned sizeLog2 = sHashBits - hashShift;
DoubleHash dh = {
((curKeyHash << sizeLog2) >> hashShift) | 1,
(HashNumber(1) << sizeLog2) - 1
};
return dh;
}
2013-02-27 15:45:23 +08:00
static HashNumber applyDoubleHash(HashNumber h1, const DoubleHash &dh)
{
return (h1 - dh.h2) & dh.sizeMask;
}
2013-02-27 15:45:23 +08:00
bool overloaded()
{
return entryCount + removedCount >= ((sMaxAlphaFrac * capacity()) >> 8);
}
2013-02-27 15:45:23 +08:00
bool underloaded()
{
uint32_t tableCapacity = capacity();
return tableCapacity > sMinSize &&
entryCount <= ((sMinAlphaFrac * tableCapacity) >> 8);
}
2013-02-27 15:45:23 +08:00
static bool match(Entry &e, const Lookup &l)
{
return HashPolicy::match(HashPolicy::getKey(e.get()), l);
}
2013-02-27 15:45:23 +08:00
Entry &lookup(const Lookup &l, HashNumber keyHash, unsigned collisionBit) const
{
JS_ASSERT(isLiveHash(keyHash));
JS_ASSERT(!(keyHash & sCollisionBit));
JS_ASSERT(collisionBit == 0 || collisionBit == sCollisionBit);
JS_ASSERT(table);
METER(stats.searches++);
2013-02-27 15:45:23 +08:00
// Compute the primary hash address.
HashNumber h1 = hash1(keyHash);
Entry *entry = &table[h1];
2013-02-27 15:45:23 +08:00
// Miss: return space for a new entry.
if (entry->isFree()) {
METER(stats.misses++);
return *entry;
}
2013-02-27 15:45:23 +08:00
// Hit: return entry.
if (entry->matchHash(keyHash) && match(*entry, l)) {
METER(stats.hits++);
return *entry;
}
2013-02-27 15:45:23 +08:00
// Collision: double hash.
DoubleHash dh = hash2(keyHash);
2013-02-27 15:45:23 +08:00
// Save the first removed entry pointer so we can recycle later.
Entry *firstRemoved = NULL;
2013-02-27 15:45:23 +08:00
while(true) {
if (JS_UNLIKELY(entry->isRemoved())) {
if (!firstRemoved)
firstRemoved = entry;
} else {
entry->setCollision(collisionBit);
}
2013-02-27 15:45:23 +08:00
METER(stats.steps++);
h1 = applyDoubleHash(h1, dh);
2013-02-27 15:45:23 +08:00
entry = &table[h1];
if (entry->isFree()) {
METER(stats.misses++);
return firstRemoved ? *firstRemoved : *entry;
}
2013-02-27 15:45:23 +08:00
if (entry->matchHash(keyHash) && match(*entry, l)) {
METER(stats.hits++);
return *entry;
}
}
}
2013-02-27 15:45:23 +08:00
// This is a copy of lookup hardcoded to the assumptions:
// 1. the lookup is a lookupForAdd
// 2. the key, whose |keyHash| has been passed is not in the table,
// 3. no entries have been removed from the table.
// This specialized search avoids the need for recovering lookup values
// from entries, which allows more flexible Lookup/Key types.
Entry &findFreeEntry(HashNumber keyHash)
{
2013-02-27 15:45:23 +08:00
JS_ASSERT(!(keyHash & sCollisionBit));
JS_ASSERT(table);
METER(stats.searches++);
2013-02-27 15:45:23 +08:00
// We assume 'keyHash' has already been distributed.
2013-02-27 15:45:23 +08:00
// Compute the primary hash address.
HashNumber h1 = hash1(keyHash);
Entry *entry = &table[h1];
2013-02-27 15:45:23 +08:00
// Miss: return space for a new entry.
if (!entry->isLive()) {
METER(stats.misses++);
return *entry;
}
2013-02-27 15:45:23 +08:00
// Collision: double hash.
DoubleHash dh = hash2(keyHash);
2013-02-27 15:45:23 +08:00
while(true) {
JS_ASSERT(!entry->isRemoved());
entry->setCollision();
2013-02-27 15:45:23 +08:00
METER(stats.steps++);
h1 = applyDoubleHash(h1, dh);
2013-02-27 15:45:23 +08:00
entry = &table[h1];
if (!entry->isLive()) {
METER(stats.misses++);
return *entry;
}
}
}
2013-02-27 15:45:23 +08:00
enum RebuildStatus { NotOverloaded, Rehashed, RehashFailed };
2013-02-27 15:45:23 +08:00
RebuildStatus changeTableSize(int deltaLog2)
{
// Look, but don't touch, until we succeed in getting new entry store.
Entry *oldTable = table;
uint32_t oldCap = capacity();
uint32_t newLog2 = sHashBits - hashShift + deltaLog2;
uint32_t newCapacity = JS_BIT(newLog2);
if (newCapacity > sMaxCapacity) {
this->reportAllocOverflow();
return RehashFailed;
}
2013-02-27 15:45:23 +08:00
Entry *newTable = createTable(*this, newCapacity);
if (!newTable)
return RehashFailed;
2013-02-27 15:45:23 +08:00
// We can't fail from here on, so update table parameters.
setTableSizeLog2(newLog2);
removedCount = 0;
gen++;
table = newTable;
2013-02-27 15:45:23 +08:00
// Copy only live entries, leaving removed ones behind.
for (Entry *src = oldTable, *end = src + oldCap; src < end; ++src) {
if (src->isLive()) {
HashNumber hn = src->getKeyHash();
findFreeEntry(hn).setLive(hn, Move(src->get()));
src->destroy();
}
}
2013-02-27 15:45:23 +08:00
// All entries have been destroyed, no need to destroyTable.
this->free_(oldTable);
return Rehashed;
}
2013-02-27 15:45:23 +08:00
RebuildStatus checkOverloaded()
{
if (!overloaded())
return NotOverloaded;
2013-02-27 15:45:23 +08:00
// Compress if a quarter or more of all entries are removed.
int deltaLog2;
if (removedCount >= (capacity() >> 2)) {
METER(stats.compresses++);
deltaLog2 = 0;
} else {
METER(stats.grows++);
deltaLog2 = 1;
}
2013-02-27 15:45:23 +08:00
return changeTableSize(deltaLog2);
}
2013-02-27 15:45:23 +08:00
// Infallibly rehash the table if we are overloaded with removals.
void checkOverRemoved()
{
if (overloaded()) {
METER(stats.rehashes++);
rehashTable();
JS_ASSERT(!overloaded());
}
}
2013-02-27 15:45:23 +08:00
void remove(Entry &e)
{
JS_ASSERT(table);
METER(stats.removes++);
2013-02-27 15:45:23 +08:00
if (e.hasCollision()) {
e.removeLive();
removedCount++;
} else {
METER(stats.removeFrees++);
e.clearLive();
}
entryCount--;
mutationCount++;
}
2013-02-27 15:45:23 +08:00
void checkUnderloaded()
{
if (underloaded()) {
METER(stats.shrinks++);
(void) changeTableSize(-1);
}
}
2013-02-27 15:45:23 +08:00
// This is identical to changeTableSize(currentSize), but without requiring
// a second table. We do this by recycling the collision bits to tell us if
// the element is already inserted or still waiting to be inserted. Since
// already-inserted elements win any conflicts, we get the same table as we
// would have gotten through random insertion order.
void rehashTable()
{
removedCount = 0;
for (size_t i = 0; i < capacity(); ++i)
table[i].unsetCollision();
2013-02-27 15:45:23 +08:00
for (size_t i = 0; i < capacity();) {
Entry *src = &table[i];
2013-02-27 15:45:23 +08:00
if (!src->isLive() || src->hasCollision()) {
++i;
continue;
}
2013-02-27 15:45:23 +08:00
HashNumber keyHash = src->getKeyHash();
HashNumber h1 = hash1(keyHash);
DoubleHash dh = hash2(keyHash);
Entry *tgt = &table[h1];
while (true) {
if (!tgt->hasCollision()) {
src->swap(tgt);
tgt->setCollision();
break;
}
2013-02-27 15:45:23 +08:00
h1 = applyDoubleHash(h1, dh);
tgt = &table[h1];
}
}
// TODO: this algorithm leaves collision bits on *all* elements, even if
// they are on no collision path. We have the option of setting the
// collision bits correctly on a subsequent pass or skipping the rehash
// unless we are totally filled with tombstones: benchmark to find out
// which approach is best.
}
public:
2013-02-27 15:45:23 +08:00
void clear()
{
if (tl::IsPodType<Entry>::result) {
memset(table, 0, sizeof(*table) * capacity());
} else {
uint32_t tableCapacity = capacity();
for (Entry *e = table, *end = table + tableCapacity; e < end; ++e)
e->clear();
}
removedCount = 0;
entryCount = 0;
mutationCount++;
}
2013-02-27 15:45:23 +08:00
void finish()
{
JS_ASSERT(!entered);
2013-02-27 15:45:23 +08:00
if (!table)
return;
2013-02-27 15:45:23 +08:00
destroyTable(*this, table, capacity());
table = NULL;
gen++;
entryCount = 0;
removedCount = 0;
mutationCount++;
}
2013-02-27 15:45:23 +08:00
Range all() const
{
JS_ASSERT(table);
return Range(table, table + capacity());
}
2013-02-27 15:45:23 +08:00
bool empty() const
{
JS_ASSERT(table);
return !entryCount;
}
2013-02-27 15:45:23 +08:00
uint32_t count() const
{
JS_ASSERT(table);
return entryCount;
}
2013-02-27 15:45:23 +08:00
uint32_t capacity() const
{
JS_ASSERT(table);
return JS_BIT(sHashBits - hashShift);
}
2013-02-27 15:45:23 +08:00
uint32_t generation() const
{
JS_ASSERT(table);
return gen;
}
2013-02-27 15:45:23 +08:00
size_t sizeOfExcludingThis(JSMallocSizeOfFun mallocSizeOf) const
{
return mallocSizeOf(table);
}
2013-02-27 15:45:23 +08:00
size_t sizeOfIncludingThis(JSMallocSizeOfFun mallocSizeOf) const
{
return mallocSizeOf(this) + sizeOfExcludingThis(mallocSizeOf);
}
2013-02-27 15:45:23 +08:00
Ptr lookup(const Lookup &l) const
{
ReentrancyGuard g(*this);
HashNumber keyHash = prepareHash(l);
return Ptr(lookup(l, keyHash, 0));
}
2013-02-27 15:45:23 +08:00
AddPtr lookupForAdd(const Lookup &l) const
{
ReentrancyGuard g(*this);
HashNumber keyHash = prepareHash(l);
Entry &entry = lookup(l, keyHash, sCollisionBit);
AddPtr p(entry, keyHash);
p.mutationCount = mutationCount;
return p;
}
2013-02-27 15:45:23 +08:00
template <class U>
bool add(AddPtr &p, const U &rhs)
{
ReentrancyGuard g(*this);
JS_ASSERT(mutationCount == p.mutationCount);
JS_ASSERT(table);
JS_ASSERT(!p.found());
JS_ASSERT(!(p.keyHash & sCollisionBit));
2013-02-27 15:45:23 +08:00
// Changing an entry from removed to live does not affect whether we
// are overloaded and can be handled separately.
if (p.entry->isRemoved()) {
METER(stats.addOverRemoved++);
removedCount--;
p.keyHash |= sCollisionBit;
} else {
// Preserve the validity of |p.entry|.
RebuildStatus status = checkOverloaded();
if (status == RehashFailed)
return false;
if (status == Rehashed)
p.entry = &findFreeEntry(p.keyHash);
}
2013-02-27 15:45:23 +08:00
p.entry->setLive(p.keyHash, rhs);
entryCount++;
mutationCount++;
return true;
}
2013-02-27 15:45:23 +08:00
template <class U>
void putNewInfallible(const Lookup &l, const U &u)
{
JS_ASSERT(table);
HashNumber keyHash = prepareHash(l);
Entry *entry = &findFreeEntry(keyHash);
if (entry->isRemoved()) {
METER(stats.addOverRemoved++);
removedCount--;
keyHash |= sCollisionBit;
}
entry->setLive(keyHash, u);
entryCount++;
mutationCount++;
}
2013-02-27 15:45:23 +08:00
template <class U>
bool putNew(const Lookup &l, const U &u)
{
if (checkOverloaded() == RehashFailed)
return false;
putNewInfallible(l, u);
return true;
}
2013-02-27 15:45:23 +08:00
template <class U>
bool relookupOrAdd(AddPtr& p, const Lookup &l, const U &u)
{
p.mutationCount = mutationCount;
{
ReentrancyGuard g(*this);
p.entry = &lookup(l, p.keyHash, sCollisionBit);
}
return p.found() || add(p, u);
}
2013-02-27 15:45:23 +08:00
void remove(Ptr p)
{
JS_ASSERT(table);
ReentrancyGuard g(*this);
JS_ASSERT(p.found());
remove(*p.entry);
checkUnderloaded();
}
2013-02-27 15:45:23 +08:00
#undef METER
};
2013-02-27 15:45:23 +08:00
} // namespace detail
} // namespace js
#endif // js_HashTable_h__