gecko-dev/mfbt/HashTable.h

2262 строки
63 KiB
C
Исходник Обычный вид История

Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim: set ts=8 sts=2 et sw=2 tw=80: */
/* 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/. */
//---------------------------------------------------------------------------
// Overview
//---------------------------------------------------------------------------
//
// This file defines HashMap<Key, Value> and HashSet<T>, hash tables that are
// fast and have a nice API.
//
// Both hash tables have two optional template parameters.
//
// - HashPolicy. This defines the operations for hashing and matching keys. The
// default HashPolicy is appropriate when both of the following two
// conditions are true.
//
// - The key type stored in the table (|Key| for |HashMap<Key, Value>|, |T|
// for |HashSet<T>|) is an integer, pointer, UniquePtr, float, or double.
//
// - The type used for lookups (|Lookup|) is the same as the key type. This
// is usually the case, but not always.
//
// There is also a |CStringHasher| policy for |char*| keys. If your keys
// don't match any of the above cases, you must provide your own hash policy;
// see the "Hash Policy" section below.
//
// - AllocPolicy. This defines how allocations are done by the table.
//
// - |MallocAllocPolicy| is the default and is usually appropriate; note that
// operations (such as insertions) that might cause allocations are
// fallible and must be checked for OOM. These checks are enforced by the
// use of MOZ_MUST_USE.
//
// - |InfallibleAllocPolicy| is another possibility; it allows the
// abovementioned OOM checks to be done with MOZ_ALWAYS_TRUE().
//
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
// Note that entry storage allocation is lazy, and not done until the first
// lookupForAdd(), put(), or putNew() is performed.
//
// See AllocPolicy.h for more details.
//
// Documentation on how to use HashMap and HashSet, including examples, is
// present within those classes. Search for "class HashMap" and "class
// HashSet".
//
// Both HashMap and HashSet are implemented on top of a third class, HashTable.
// You only need to look at HashTable if you want to understand the
// implementation.
//
// How does mozilla::HashTable (this file) compare with PLDHashTable (and its
// subclasses, such as nsTHashtable)?
//
// - mozilla::HashTable is a lot faster, largely because it uses templates
// throughout *and* inlines everything. PLDHashTable inlines operations much
// less aggressively, and also uses "virtual ops" for operations like hashing
// and matching entries that require function calls.
//
// - Correspondingly, mozilla::HashTable use is likely to increase executable
// size much more than PLDHashTable.
//
// - mozilla::HashTable has a nicer API, with a proper HashSet vs. HashMap
// distinction.
//
// - mozilla::HashTable requires more explicit OOM checking. As mentioned
// above, the use of |InfallibleAllocPolicy| can simplify things.
//
// - mozilla::HashTable has a default capacity on creation of 32 and a minimum
// capacity of 4. PLDHashTable has a default capacity on creation of 8 and a
// minimum capacity of 8.
#ifndef mozilla_HashTable_h
#define mozilla_HashTable_h
#include "mozilla/AllocPolicy.h"
#include "mozilla/Assertions.h"
#include "mozilla/Attributes.h"
#include "mozilla/Casting.h"
#include "mozilla/HashFunctions.h"
#include "mozilla/MathAlgorithms.h"
#include "mozilla/MemoryChecking.h"
#include "mozilla/MemoryReporting.h"
#include "mozilla/Move.h"
#include "mozilla/Opaque.h"
#include "mozilla/PodOperations.h"
#include "mozilla/ReentrancyGuard.h"
#include "mozilla/TypeTraits.h"
#include "mozilla/UniquePtr.h"
namespace mozilla {
template<class>
struct DefaultHasher;
template<class, class>
class HashMapEntry;
namespace detail {
template<typename T>
class HashTableEntry;
template<class T, class HashPolicy, class AllocPolicy>
class HashTable;
} // namespace detail
// The "generation" of a hash table is an opaque value indicating the state of
// modification of the hash table through its lifetime. If the generation of
// a hash table compares equal at times T1 and T2, then lookups in the hash
// table, pointers to (or into) hash table entries, etc. at time T1 are valid
// at time T2. If the generation compares unequal, these computations are all
// invalid and must be performed again to be used.
//
// Generations are meaningfully comparable only with respect to a single hash
// table. It's always nonsensical to compare the generation of distinct hash
// tables H1 and H2.
using Generation = Opaque<uint64_t>;
//---------------------------------------------------------------------------
// HashMap
//---------------------------------------------------------------------------
// HashMap is a fast hash-based map from keys to values.
//
// Template parameter requirements:
// - Key/Value: movable, destructible, assignable.
// - HashPolicy: see the "Hash Policy" section below.
// - AllocPolicy: see AllocPolicy.h.
//
// Note:
// - HashMap is not reentrant: Key/Value/HashPolicy/AllocPolicy members
// called by HashMap must not call back into the same HashMap object.
//
template<class Key,
class Value,
class HashPolicy = DefaultHasher<Key>,
class AllocPolicy = MallocAllocPolicy>
class HashMap
{
// -- Implementation details -----------------------------------------------
// HashMap is not copyable or assignable.
HashMap(const HashMap& hm) = delete;
HashMap& operator=(const HashMap& hm) = delete;
using TableEntry = HashMapEntry<Key, Value>;
struct MapHashPolicy : HashPolicy
{
using Base = HashPolicy;
using KeyType = Key;
static const Key& getKey(TableEntry& aEntry) { return aEntry.key(); }
static void setKey(TableEntry& aEntry, Key& aKey)
{
HashPolicy::rekey(aEntry.mutableKey(), aKey);
}
};
using Impl = detail::HashTable<TableEntry, MapHashPolicy, AllocPolicy>;
Impl mImpl;
friend class Impl::Enum;
public:
using Lookup = typename HashPolicy::Lookup;
using Entry = TableEntry;
// -- Initialization -------------------------------------------------------
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
explicit HashMap(AllocPolicy aAllocPolicy = AllocPolicy(),
uint32_t aLen = Impl::sDefaultLen)
: mImpl(aAllocPolicy, aLen)
{
}
explicit HashMap(uint32_t aLen)
: mImpl(AllocPolicy(), aLen)
{
}
// HashMap is movable.
HashMap(HashMap&& aRhs)
: mImpl(std::move(aRhs.mImpl))
{
}
void operator=(HashMap&& aRhs)
{
MOZ_ASSERT(this != &aRhs, "self-move assignment is prohibited");
mImpl = std::move(aRhs.mImpl);
}
// -- Status and sizing ----------------------------------------------------
// The map's current generation.
Generation generation() const { return mImpl.generation(); }
// Is the map empty?
bool empty() const { return mImpl.empty(); }
// Number of keys/values in the map.
uint32_t count() const { return mImpl.count(); }
// Number of key/value slots in the map. Note: resize will happen well before
// count() == capacity().
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
uint32_t capacity() const { return mImpl.capacity(); }
// The size of the map's entry storage, in bytes. If the keys/values contain
// pointers to other heap blocks, you must iterate over the map and measure
// them separately; hence the "shallow" prefix.
size_t shallowSizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const
{
return mImpl.shallowSizeOfExcludingThis(aMallocSizeOf);
}
size_t shallowSizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const
{
return aMallocSizeOf(this) +
mImpl.shallowSizeOfExcludingThis(aMallocSizeOf);
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
// Attempt to minimize the capacity(). If the table is empty, this will free
// the empty storage and upon regrowth it will be given the minimum capacity.
void compact() { mImpl.compact(); }
// Attempt to reserve enough space to fit at least |aLen| elements. Does
// nothing if the map already has sufficient capacity.
MOZ_MUST_USE bool reserve(uint32_t aLen) { return mImpl.reserve(aLen); }
// -- Lookups --------------------------------------------------------------
// Does the map contain a key/value matching |aLookup|?
bool has(const Lookup& aLookup) const
{
return mImpl.lookup(aLookup).found();
}
// Return a Ptr indicating whether a key/value matching |aLookup| is
// present in the map. E.g.:
//
// using HM = HashMap<int,char>;
// HM h;
// if (HM::Ptr p = h.lookup(3)) {
// assert(p->key() == 3);
// char val = p->value();
// }
//
using Ptr = typename Impl::Ptr;
MOZ_ALWAYS_INLINE Ptr lookup(const Lookup& aLookup) const
{
return mImpl.lookup(aLookup);
}
// Like lookup(), but does not assert if two threads call it at the same
// time. Only use this method when none of the threads will modify the map.
MOZ_ALWAYS_INLINE Ptr readonlyThreadsafeLookup(const Lookup& aLookup) const
{
return mImpl.readonlyThreadsafeLookup(aLookup);
}
// -- Insertions -----------------------------------------------------------
// Overwrite existing value with |aValue|, or add it if not present. Returns
// false on OOM.
template<typename KeyInput, typename ValueInput>
MOZ_MUST_USE bool put(KeyInput&& aKey, ValueInput&& aValue)
{
AddPtr p = lookupForAdd(aKey);
if (p) {
p->value() = std::forward<ValueInput>(aValue);
return true;
}
return add(
p, std::forward<KeyInput>(aKey), std::forward<ValueInput>(aValue));
}
// Like put(), but slightly faster. Must only be used when the given key is
// not already present. (In debug builds, assertions check this.)
template<typename KeyInput, typename ValueInput>
MOZ_MUST_USE bool putNew(KeyInput&& aKey, ValueInput&& aValue)
{
return mImpl.putNew(
aKey, std::forward<KeyInput>(aKey), std::forward<ValueInput>(aValue));
}
// Like putNew(), but should be only used when the table is known to be big
// enough for the insertion, and hashing cannot fail. Typically this is used
// to populate an empty map with known-unique keys after reserving space with
// reserve(), e.g.
//
// using HM = HashMap<int,char>;
// HM h;
// if (!h.reserve(3)) {
// MOZ_CRASH("OOM");
// }
// h.putNewInfallible(1, 'a'); // unique key
// h.putNewInfallible(2, 'b'); // unique key
// h.putNewInfallible(3, 'c'); // unique key
//
template<typename KeyInput, typename ValueInput>
void putNewInfallible(KeyInput&& aKey, ValueInput&& aValue)
{
mImpl.putNewInfallible(
aKey, std::forward<KeyInput>(aKey), std::forward<ValueInput>(aValue));
}
// 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 key/value. E.g.:
//
// using HM = HashMap<int,char>;
// HM h;
// HM::AddPtr p = h.lookupForAdd(3);
// if (!p) {
// if (!h.add(p, 3, 'a')) {
// return false;
// }
// }
// assert(p->key() == 3);
// char val = p->value();
//
// 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;
// }
// }
// assert(p->key() == 3);
// char val = p->value();
//
using AddPtr = typename Impl::AddPtr;
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
MOZ_ALWAYS_INLINE AddPtr lookupForAdd(const Lookup& aLookup)
{
return mImpl.lookupForAdd(aLookup);
}
// Add a key/value. Returns false on OOM.
template<typename KeyInput, typename ValueInput>
MOZ_MUST_USE bool add(AddPtr& aPtr, KeyInput&& aKey, ValueInput&& aValue)
{
return mImpl.add(
aPtr, std::forward<KeyInput>(aKey), std::forward<ValueInput>(aValue));
}
// See the comment above lookupForAdd() for details.
template<typename KeyInput, typename ValueInput>
MOZ_MUST_USE bool relookupOrAdd(AddPtr& aPtr,
KeyInput&& aKey,
ValueInput&& aValue)
{
return mImpl.relookupOrAdd(aPtr,
aKey,
std::forward<KeyInput>(aKey),
std::forward<ValueInput>(aValue));
}
// -- Removal --------------------------------------------------------------
// Lookup and remove the key/value matching |aLookup|, if present.
void remove(const Lookup& aLookup)
{
if (Ptr p = lookup(aLookup)) {
remove(p);
}
}
// Remove a previously found key/value (assuming aPtr.found()). The map must
// not have been mutated in the interim.
void remove(Ptr aPtr) { mImpl.remove(aPtr); }
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
// Remove all keys/values without changing the capacity.
void clear() { mImpl.clear(); }
// Like clear() followed by compact().
void clearAndCompact() { mImpl.clearAndCompact(); }
// -- Rekeying -------------------------------------------------------------
// Infallibly rekey one entry, if necessary. Requires that template
// parameters Key and HashPolicy::Lookup are the same type.
void rekeyIfMoved(const Key& aOldKey, const Key& aNewKey)
{
if (aOldKey != aNewKey) {
rekeyAs(aOldKey, aNewKey, aNewKey);
}
}
// Infallibly rekey one entry if present, and return whether that happened.
bool rekeyAs(const Lookup& aOldLookup,
const Lookup& aNewLookup,
const Key& aNewKey)
{
if (Ptr p = lookup(aOldLookup)) {
mImpl.rekeyAndMaybeRehash(p, aNewLookup, aNewKey);
return true;
}
return false;
}
// -- Iteration ------------------------------------------------------------
// |iter()| returns an Iterator:
//
// HashMap<int, char> h;
// for (auto iter = h.iter(); !iter.done(); iter.next()) {
// char c = iter.get().value();
// }
//
using Iterator = typename Impl::Iterator;
Iterator iter() const { return mImpl.iter(); }
// |modIter()| returns a ModIterator:
//
// HashMap<int, char> h;
// for (auto iter = h.modIter(); !iter.done(); iter.next()) {
// if (iter.get().value() == 'l') {
// iter.remove();
// }
// }
//
// Table resize may occur in ModIterator's destructor.
using ModIterator = typename Impl::ModIterator;
ModIterator modIter() { return mImpl.modIter(); }
// These are similar to Iterator/ModIterator/iter(), but use different
// terminology.
using Range = typename Impl::Range;
using Enum = typename Impl::Enum;
Range all() const { return mImpl.all(); }
};
//---------------------------------------------------------------------------
// HashSet
//---------------------------------------------------------------------------
// HashSet is a fast hash-based set of values.
//
// Template parameter requirements:
// - T: movable, destructible, assignable.
// - HashPolicy: see the "Hash Policy" section below.
// - AllocPolicy: see AllocPolicy.h
//
// Note:
// - HashSet is not reentrant: T/HashPolicy/AllocPolicy members called by
// HashSet must not call back into the same HashSet object.
//
template<class T,
class HashPolicy = DefaultHasher<T>,
class AllocPolicy = MallocAllocPolicy>
class HashSet
{
// -- Implementation details -----------------------------------------------
// HashSet is not copyable or assignable.
HashSet(const HashSet& hs) = delete;
HashSet& operator=(const HashSet& hs) = delete;
struct SetHashPolicy : HashPolicy
{
using Base = HashPolicy;
using KeyType = T;
static const KeyType& getKey(const T& aT) { return aT; }
static void setKey(T& aT, KeyType& aKey) { HashPolicy::rekey(aT, aKey); }
};
using Impl = detail::HashTable<const T, SetHashPolicy, AllocPolicy>;
Impl mImpl;
friend class Impl::Enum;
public:
using Lookup = typename HashPolicy::Lookup;
using Entry = T;
// -- Initialization -------------------------------------------------------
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
explicit HashSet(AllocPolicy aAllocPolicy = AllocPolicy(),
uint32_t aLen = Impl::sDefaultLen)
: mImpl(aAllocPolicy, aLen)
{
}
explicit HashSet(uint32_t aLen)
: mImpl(AllocPolicy(), aLen)
{
}
// HashSet is movable.
HashSet(HashSet&& aRhs)
: mImpl(std::move(aRhs.mImpl))
{
}
void operator=(HashSet&& aRhs)
{
MOZ_ASSERT(this != &aRhs, "self-move assignment is prohibited");
mImpl = std::move(aRhs.mImpl);
}
// -- Status and sizing ----------------------------------------------------
// The set's current generation.
Generation generation() const { return mImpl.generation(); }
// Is the set empty?
bool empty() const { return mImpl.empty(); }
// Number of elements in the set.
uint32_t count() const { return mImpl.count(); }
// Number of element slots in the set. Note: resize will happen well before
// count() == capacity().
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
uint32_t capacity() const { return mImpl.capacity(); }
// The size of the set's entry storage, in bytes. If the elements contain
// pointers to other heap blocks, you must iterate over the set and measure
// them separately; hence the "shallow" prefix.
size_t shallowSizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const
{
return mImpl.shallowSizeOfExcludingThis(aMallocSizeOf);
}
size_t shallowSizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const
{
return aMallocSizeOf(this) +
mImpl.shallowSizeOfExcludingThis(aMallocSizeOf);
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
// Attempt to minimize the capacity(). If the table is empty, this will free
// the empty storage and upon regrowth it will be given the minimum capacity.
void compact() { mImpl.compact(); }
// Attempt to reserve enough space to fit at least |aLen| elements. Does
// nothing if the map already has sufficient capacity.
MOZ_MUST_USE bool reserve(uint32_t aLen) { return mImpl.reserve(aLen); }
// -- Lookups --------------------------------------------------------------
// Does the set contain an element matching |aLookup|?
bool has(const Lookup& aLookup) const
{
return mImpl.lookup(aLookup).found();
}
// Return a Ptr indicating whether an element matching |aLookup| is present
// in the set. E.g.:
//
// using HS = HashSet<int>;
// HS h;
// if (HS::Ptr p = h.lookup(3)) {
// assert(*p == 3); // p acts like a pointer to int
// }
//
using Ptr = typename Impl::Ptr;
MOZ_ALWAYS_INLINE Ptr lookup(const Lookup& aLookup) const
{
return mImpl.lookup(aLookup);
}
// Like lookup(), but does not assert if two threads call it at the same
// time. Only use this method when none of the threads will modify the set.
MOZ_ALWAYS_INLINE Ptr readonlyThreadsafeLookup(const Lookup& aLookup) const
{
return mImpl.readonlyThreadsafeLookup(aLookup);
}
// -- Insertions -----------------------------------------------------------
// Add |aU| if it is not present already. Returns false on OOM.
template<typename U>
MOZ_MUST_USE bool put(U&& aU)
{
AddPtr p = lookupForAdd(aU);
return p ? true : add(p, std::forward<U>(aU));
}
// Like put(), but slightly faster. Must only be used when the given element
// is not already present. (In debug builds, assertions check this.)
template<typename U>
MOZ_MUST_USE bool putNew(U&& aU)
{
return mImpl.putNew(aU, std::forward<U>(aU));
}
// Like the other putNew(), but for when |Lookup| is different to |T|.
template<typename U>
MOZ_MUST_USE bool putNew(const Lookup& aLookup, U&& aU)
{
return mImpl.putNew(aLookup, std::forward<U>(aU));
}
// Like putNew(), but should be only used when the table is known to be big
// enough for the insertion, and hashing cannot fail. Typically this is used
// to populate an empty set with known-unique elements after reserving space
// with reserve(), e.g.
//
// using HS = HashMap<int>;
// HS h;
// if (!h.reserve(3)) {
// MOZ_CRASH("OOM");
// }
// h.putNewInfallible(1); // unique element
// h.putNewInfallible(2); // unique element
// h.putNewInfallible(3); // unique element
//
template<typename U>
void putNewInfallible(const Lookup& aLookup, U&& aU)
{
mImpl.putNewInfallible(aLookup, std::forward<U>(aU));
}
// 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.:
//
// using HS = HashSet<int>;
// 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
//
// 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).
using AddPtr = typename Impl::AddPtr;
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
MOZ_ALWAYS_INLINE AddPtr lookupForAdd(const Lookup& aLookup)
{
return mImpl.lookupForAdd(aLookup);
}
// Add an element. Returns false on OOM.
template<typename U>
MOZ_MUST_USE bool add(AddPtr& aPtr, U&& aU)
{
return mImpl.add(aPtr, std::forward<U>(aU));
}
// See the comment above lookupForAdd() for details.
template<typename U>
MOZ_MUST_USE bool relookupOrAdd(AddPtr& aPtr, const Lookup& aLookup, U&& aU)
{
return mImpl.relookupOrAdd(aPtr, aLookup, std::forward<U>(aU));
}
// -- Removal --------------------------------------------------------------
// Lookup and remove the element matching |aLookup|, if present.
void remove(const Lookup& aLookup)
{
if (Ptr p = lookup(aLookup)) {
remove(p);
}
}
// Remove a previously found element (assuming aPtr.found()). The set must
// not have been mutated in the interim.
void remove(Ptr aPtr) { mImpl.remove(aPtr); }
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
// Remove all keys/values without changing the capacity.
void clear() { mImpl.clear(); }
// Like clear() followed by compact().
void clearAndCompact() { mImpl.clearAndCompact(); }
// -- Rekeying -------------------------------------------------------------
// Infallibly rekey one entry, if present. Requires that template parameters
// T and HashPolicy::Lookup are the same type.
void rekeyIfMoved(const Lookup& aOldValue, const T& aNewValue)
{
if (aOldValue != aNewValue) {
rekeyAs(aOldValue, aNewValue, aNewValue);
}
}
// Infallibly rekey one entry if present, and return whether that happened.
bool rekeyAs(const Lookup& aOldLookup,
const Lookup& aNewLookup,
const T& aNewValue)
{
if (Ptr p = lookup(aOldLookup)) {
mImpl.rekeyAndMaybeRehash(p, aNewLookup, aNewValue);
return true;
}
return false;
}
// Infallibly replace the current key at |aPtr| with an equivalent key.
// Specifically, both HashPolicy::hash and HashPolicy::match must return
// identical results for the new and old key when applied against all
// possible matching values.
void replaceKey(Ptr aPtr, const T& aNewValue)
{
MOZ_ASSERT(aPtr.found());
MOZ_ASSERT(*aPtr != aNewValue);
MOZ_ASSERT(HashPolicy::hash(*aPtr) == HashPolicy::hash(aNewValue));
MOZ_ASSERT(HashPolicy::match(*aPtr, aNewValue));
const_cast<T&>(*aPtr) = aNewValue;
}
// -- Iteration ------------------------------------------------------------
// |iter()| returns an Iterator:
//
// HashSet<int> h;
// for (auto iter = h.iter(); !iter.done(); iter.next()) {
// int i = iter.get();
// }
//
using Iterator = typename Impl::Iterator;
Iterator iter() const { return mImpl.iter(); }
// |modIter()| returns a ModIterator:
//
// HashSet<int> h;
// for (auto iter = h.modIter(); !iter.done(); iter.next()) {
// if (iter.get() == 42) {
// iter.remove();
// }
// }
//
// Table resize may occur in ModIterator's destructor.
using ModIterator = typename Impl::ModIterator;
ModIterator modIter() { return mImpl.modIter(); }
// These are similar to Iterator/ModIterator/iter(), but use different
// terminology.
using Range = typename Impl::Range;
using Enum = typename Impl::Enum;
Range all() const { return mImpl.all(); }
};
//---------------------------------------------------------------------------
// Hash Policy
//---------------------------------------------------------------------------
// A hash policy |HP| for a hash table with key-type |Key| must provide:
//
// - a type |HP::Lookup| to use to lookup table entries;
//
// - a static member function |HP::hash| that hashes lookup values:
//
// static mozilla::HashNumber hash(const Lookup&);
//
// - a static member function |HP::match| that tests equality of key and
// lookup values:
//
// static bool match(const Key&, const Lookup&);
//
// 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 not equal
// to the added Key value |k|, the user must ensure that |HP::match(k,l)| is
// true. E.g.:
//
// mozilla::HashSet<Key, HP>::AddPtr p = h.lookup(l);
// if (!p) {
// assert(HP::match(k, l)); // must hold
// h.add(p, k);
// }
// A pointer hashing policy that uses HashGeneric() to create good hashes for
// pointers. Note that we don't shift out the lowest k bits because we don't
// want to assume anything about the alignment of the pointers.
template<typename Key>
struct PointerHasher
{
using Lookup = Key;
static HashNumber hash(const Lookup& aLookup)
{
size_t word = reinterpret_cast<size_t>(aLookup);
return HashGeneric(word);
}
static bool match(const Key& aKey, const Lookup& aLookup)
{
return aKey == aLookup;
}
static void rekey(Key& aKey, const Key& aNewKey) { aKey = aNewKey; }
};
// The default hash policy, which only works with integers.
template<class Key>
struct DefaultHasher
{
using Lookup = Key;
static HashNumber hash(const Lookup& aLookup)
{
// Just convert the integer to a HashNumber and use that as is. (This
// discards the high 32-bits of 64-bit integers!) ScrambleHashCode() is
// subsequently called on the value to improve the distribution.
return aLookup;
}
static bool match(const Key& aKey, const Lookup& aLookup)
{
// Use builtin or overloaded operator==.
return aKey == aLookup;
}
static void rekey(Key& aKey, const Key& aNewKey) { aKey = aNewKey; }
};
// A DefaultHasher specialization for pointers.
template<class T>
struct DefaultHasher<T*> : PointerHasher<T*>
{
};
// A DefaultHasher specialization for mozilla::UniquePtr.
template<class T, class D>
struct DefaultHasher<UniquePtr<T, D>>
{
using Key = UniquePtr<T, D>;
using Lookup = Key;
using PtrHasher = PointerHasher<T*>;
static HashNumber hash(const Lookup& aLookup)
{
return PtrHasher::hash(aLookup.get());
}
static bool match(const Key& aKey, const Lookup& aLookup)
{
return PtrHasher::match(aKey.get(), aLookup.get());
}
static void rekey(UniquePtr<T, D>& aKey, UniquePtr<T, D>&& aNewKey)
{
aKey = std::move(aNewKey);
}
};
// A DefaultHasher specialization for doubles.
template<>
struct DefaultHasher<double>
{
using Key = double;
using Lookup = Key;
static HashNumber hash(const Lookup& aLookup)
{
// Just xor the high bits with the low bits, and then treat the bits of the
// result as a uint32_t.
static_assert(sizeof(HashNumber) == 4,
"subsequent code assumes a four-byte hash");
uint64_t u = BitwiseCast<uint64_t>(aLookup);
return HashNumber(u ^ (u >> 32));
}
static bool match(const Key& aKey, const Lookup& aLookup)
{
return BitwiseCast<uint64_t>(aKey) == BitwiseCast<uint64_t>(aLookup);
}
};
// A DefaultHasher specialization for floats.
template<>
struct DefaultHasher<float>
{
using Key = float;
using Lookup = Key;
static HashNumber hash(const Lookup& aLookup)
{
// Just use the value as if its bits form an integer. ScrambleHashCode() is
// subsequently called on the value to improve the distribution.
static_assert(sizeof(HashNumber) == 4,
"subsequent code assumes a four-byte hash");
return HashNumber(BitwiseCast<uint32_t>(aLookup));
}
static bool match(const Key& aKey, const Lookup& aLookup)
{
return BitwiseCast<uint32_t>(aKey) == BitwiseCast<uint32_t>(aLookup);
}
};
// A hash policy for C strings.
struct CStringHasher
{
using Key = const char*;
using Lookup = const char*;
static HashNumber hash(const Lookup& aLookup) { return HashString(aLookup); }
static bool match(const Key& aKey, const Lookup& aLookup)
{
return strcmp(aKey, aLookup) == 0;
}
};
//---------------------------------------------------------------------------
// Fallible Hashing Interface
//---------------------------------------------------------------------------
// Most of the time generating a hash code is infallible so this class provides
// default methods that always succeed. Specialize this class for your own hash
// policy to provide fallible hashing.
//
// This is used by MovableCellHasher to handle the fact that generating a unique
// ID for cell pointer may fail due to OOM.
template<typename HashPolicy>
struct FallibleHashMethods
{
// Return true if a hashcode is already available for its argument. Once
// this returns true for a specific argument it must continue to do so.
template<typename Lookup>
static bool hasHash(Lookup&& aLookup)
{
return true;
}
// Fallible method to ensure a hashcode exists for its argument and create
// one if not. Returns false on error, e.g. out of memory.
template<typename Lookup>
static bool ensureHash(Lookup&& aLookup)
{
return true;
}
};
template<typename HashPolicy, typename Lookup>
static bool
HasHash(Lookup&& aLookup)
{
return FallibleHashMethods<typename HashPolicy::Base>::hasHash(
std::forward<Lookup>(aLookup));
}
template<typename HashPolicy, typename Lookup>
static bool
EnsureHash(Lookup&& aLookup)
{
return FallibleHashMethods<typename HashPolicy::Base>::ensureHash(
std::forward<Lookup>(aLookup));
}
//---------------------------------------------------------------------------
// Implementation Details (HashMapEntry, HashTableEntry, HashTable)
//---------------------------------------------------------------------------
// 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.
template<class Key, class Value>
class HashMapEntry
{
Key key_;
Value value_;
template<class, class, class>
friend class detail::HashTable;
template<class>
friend class detail::HashTableEntry;
template<class, class, class, class>
friend class HashMap;
public:
template<typename KeyInput, typename ValueInput>
HashMapEntry(KeyInput&& aKey, ValueInput&& aValue)
: key_(std::forward<KeyInput>(aKey))
, value_(std::forward<ValueInput>(aValue))
{
}
HashMapEntry(HashMapEntry&& aRhs)
: key_(std::move(aRhs.key_))
, value_(std::move(aRhs.value_))
{
}
void operator=(HashMapEntry&& aRhs)
{
key_ = std::move(aRhs.key_);
value_ = std::move(aRhs.value_);
}
using KeyType = Key;
using ValueType = Value;
const Key& key() const { return key_; }
// Use this method with caution! If the key is changed such that its hash
// value also changes, the map will be left in an invalid state.
Key& mutableKey() { return key_; }
const Value& value() const { return value_; }
Value& value() { return value_; }
private:
HashMapEntry(const HashMapEntry&) = delete;
void operator=(const HashMapEntry&) = delete;
};
template<typename K, typename V>
struct IsPod<HashMapEntry<K, V>>
: IntegralConstant<bool, IsPod<K>::value && IsPod<V>::value>
{
};
namespace detail {
template<class T, class HashPolicy, class AllocPolicy>
class HashTable;
template<typename T>
class HashTableEntry
{
private:
using NonConstT = typename RemoveConst<T>::Type;
static const HashNumber sFreeKey = 0;
static const HashNumber sRemovedKey = 1;
static const HashNumber sCollisionBit = 1;
HashNumber mKeyHash = sFreeKey;
alignas(NonConstT) unsigned char mValueData[sizeof(NonConstT)];
private:
template<class, class, class>
friend class HashTable;
// Some versions of GCC treat it as a -Wstrict-aliasing violation (ergo a
// -Werror compile error) to reinterpret_cast<> |mValueData| to |T*|, even
// through |void*|. Placing the latter cast in these separate functions
// breaks the chain such that affected GCC versions no longer warn/error.
void* rawValuePtr() { return mValueData; }
static bool isLiveHash(HashNumber hash) { return hash > sRemovedKey; }
HashTableEntry(const HashTableEntry&) = delete;
void operator=(const HashTableEntry&) = delete;
NonConstT* valuePtr() { return reinterpret_cast<NonConstT*>(rawValuePtr()); }
void destroyStoredT()
{
NonConstT* ptr = valuePtr();
ptr->~T();
MOZ_MAKE_MEM_UNDEFINED(ptr, sizeof(*ptr));
}
public:
HashTableEntry() = default;
~HashTableEntry()
{
if (isLive()) {
destroyStoredT();
}
MOZ_MAKE_MEM_UNDEFINED(this, sizeof(*this));
}
void destroy()
{
MOZ_ASSERT(isLive());
destroyStoredT();
}
void swap(HashTableEntry* aOther)
{
if (this == aOther) {
return;
}
MOZ_ASSERT(isLive());
if (aOther->isLive()) {
Swap(*valuePtr(), *aOther->valuePtr());
} else {
*aOther->valuePtr() = std::move(*valuePtr());
destroy();
}
Swap(mKeyHash, aOther->mKeyHash);
}
T& get()
{
MOZ_ASSERT(isLive());
return *valuePtr();
}
NonConstT& getMutable()
{
MOZ_ASSERT(isLive());
return *valuePtr();
}
bool isFree() const { return mKeyHash == sFreeKey; }
void clearLive()
{
MOZ_ASSERT(isLive());
mKeyHash = sFreeKey;
destroyStoredT();
}
void clear()
{
if (isLive()) {
destroyStoredT();
}
MOZ_MAKE_MEM_UNDEFINED(this, sizeof(*this));
mKeyHash = sFreeKey;
}
bool isRemoved() const { return mKeyHash == sRemovedKey; }
void removeLive()
{
MOZ_ASSERT(isLive());
mKeyHash = sRemovedKey;
destroyStoredT();
}
bool isLive() const { return isLiveHash(mKeyHash); }
void setCollision()
{
MOZ_ASSERT(isLive());
mKeyHash |= sCollisionBit;
}
void unsetCollision() { mKeyHash &= ~sCollisionBit; }
bool hasCollision() const { return mKeyHash & sCollisionBit; }
bool matchHash(HashNumber hn) { return (mKeyHash & ~sCollisionBit) == hn; }
HashNumber getKeyHash() const { return mKeyHash & ~sCollisionBit; }
template<typename... Args>
void setLive(HashNumber aHashNumber, Args&&... aArgs)
{
MOZ_ASSERT(!isLive());
mKeyHash = aHashNumber;
new (valuePtr()) T(std::forward<Args>(aArgs)...);
MOZ_ASSERT(isLive());
}
};
template<class T, class HashPolicy, class AllocPolicy>
class HashTable : private AllocPolicy
{
friend class mozilla::ReentrancyGuard;
using NonConstT = typename RemoveConst<T>::Type;
using Key = typename HashPolicy::KeyType;
using Lookup = typename HashPolicy::Lookup;
public:
using Entry = HashTableEntry<T>;
// 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
{
friend class HashTable;
Entry* mEntry;
#ifdef DEBUG
const HashTable* mTable;
Generation mGeneration;
#endif
protected:
Ptr(Entry& aEntry, const HashTable& aTable)
: mEntry(&aEntry)
#ifdef DEBUG
, mTable(&aTable)
, mGeneration(aTable.generation())
#endif
{
}
public:
Ptr()
: mEntry(nullptr)
#ifdef DEBUG
, mTable(nullptr)
, mGeneration(0)
#endif
{
}
bool isValid() const { return !!mEntry; }
bool found() const
{
if (!isValid()) {
return false;
}
#ifdef DEBUG
MOZ_ASSERT(mGeneration == mTable->generation());
#endif
return mEntry->isLive();
}
explicit operator bool() const { return found(); }
bool operator==(const Ptr& aRhs) const
{
MOZ_ASSERT(found() && aRhs.found());
return mEntry == aRhs.mEntry;
}
bool operator!=(const Ptr& aRhs) const
{
#ifdef DEBUG
MOZ_ASSERT(mGeneration == mTable->generation());
#endif
return !(*this == aRhs);
}
T& operator*() const
{
#ifdef DEBUG
MOZ_ASSERT(found());
MOZ_ASSERT(mGeneration == mTable->generation());
#endif
return mEntry->get();
}
T* operator->() const
{
#ifdef DEBUG
MOZ_ASSERT(found());
MOZ_ASSERT(mGeneration == mTable->generation());
#endif
return &mEntry->get();
}
};
// A Ptr that can be used to add a key after a failed lookup.
class AddPtr : public Ptr
{
friend class HashTable;
HashNumber mKeyHash;
#ifdef DEBUG
uint64_t mMutationCount;
#endif
AddPtr(Entry& aEntry, const HashTable& aTable, HashNumber aHashNumber)
: Ptr(aEntry, aTable)
, mKeyHash(aHashNumber)
#ifdef DEBUG
, mMutationCount(aTable.mMutationCount)
#endif
{
}
public:
AddPtr()
: mKeyHash(0)
{
}
};
// A hash table iterator that (mostly) doesn't allow table modifications.
// As with Ptr/AddPtr, Iterator objects must not be used after any mutating
// hash table operation unless the |generation()| is tested.
class Iterator
{
protected:
friend class HashTable;
explicit Iterator(const HashTable& aTable)
: mCur(aTable.mTable)
, mEnd(aTable.mTable + aTable.capacity())
#ifdef DEBUG
, mTable(aTable)
, mMutationCount(aTable.mMutationCount)
, mGeneration(aTable.generation())
, mValidEntry(true)
#endif
{
while (mCur < mEnd && !mCur->isLive()) {
++mCur;
}
}
Entry* mCur;
Entry* mEnd;
#ifdef DEBUG
const HashTable& mTable;
uint64_t mMutationCount;
Generation mGeneration;
bool mValidEntry;
#endif
public:
bool done() const
{
#ifdef DEBUG
MOZ_ASSERT(mGeneration == mTable.generation());
MOZ_ASSERT(mMutationCount == mTable.mMutationCount);
#endif
return mCur == mEnd;
}
T& get() const
{
MOZ_ASSERT(!done());
#ifdef DEBUG
MOZ_ASSERT(mValidEntry);
MOZ_ASSERT(mGeneration == mTable.generation());
MOZ_ASSERT(mMutationCount == mTable.mMutationCount);
#endif
return mCur->get();
}
void next()
{
MOZ_ASSERT(!done());
#ifdef DEBUG
MOZ_ASSERT(mGeneration == mTable.generation());
MOZ_ASSERT(mMutationCount == mTable.mMutationCount);
#endif
while (++mCur < mEnd && !mCur->isLive()) {
continue;
}
#ifdef DEBUG
mValidEntry = true;
#endif
}
};
// A hash table iterator that permits modification, removal and rekeying.
// Since rehashing when elements were removed during enumeration would be
// bad, it is postponed until the ModIterator is destructed. Since the
// ModIterator's destructor touches the hash table, the user must ensure
// that the hash table is still alive when the destructor runs.
class ModIterator : public Iterator
{
friend class HashTable;
HashTable& mTable;
bool mRekeyed;
bool mRemoved;
// ModIterator is movable but not copyable.
ModIterator(const ModIterator&) = delete;
void operator=(const ModIterator&) = delete;
protected:
explicit ModIterator(HashTable& aTable)
: Iterator(aTable)
, mTable(aTable)
, mRekeyed(false)
, mRemoved(false)
{
}
public:
MOZ_IMPLICIT ModIterator(ModIterator&& aOther)
: Iterator(aOther)
, mTable(aOther.mTable)
, mRekeyed(aOther.mRekeyed)
, mRemoved(aOther.mRemoved)
{
aOther.mRekeyed = false;
aOther.mRemoved = false;
}
// Removes the current element from the table, leaving |get()|
// invalid until the next call to |next()|.
void remove()
{
mTable.remove(*this->mCur);
mRemoved = true;
#ifdef DEBUG
this->mValidEntry = false;
this->mMutationCount = mTable.mMutationCount;
#endif
}
NonConstT& getMutable()
{
MOZ_ASSERT(!this->done());
#ifdef DEBUG
MOZ_ASSERT(this->mValidEntry);
MOZ_ASSERT(this->mGeneration == this->Iterator::mTable.generation());
MOZ_ASSERT(this->mMutationCount == this->Iterator::mTable.mMutationCount);
#endif
return this->mCur->getMutable();
}
// Removes the current element and re-inserts it into the table with
// a new key at the new Lookup position. |get()| is invalid after
// this operation until the next call to |next()|.
void rekey(const Lookup& l, const Key& k)
{
MOZ_ASSERT(&k != &HashPolicy::getKey(this->mCur->get()));
Ptr p(*this->mCur, mTable);
mTable.rekeyWithoutRehash(p, l, k);
mRekeyed = true;
#ifdef DEBUG
this->mValidEntry = false;
this->mMutationCount = mTable.mMutationCount;
#endif
}
void rekey(const Key& k) { rekey(k, k); }
// Potentially rehashes the table.
~ModIterator()
{
if (mRekeyed) {
mTable.mGen++;
mTable.infallibleRehashIfOverloaded();
}
if (mRemoved) {
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
mTable.compact();
}
}
};
// Range is similar to Iterator, but uses different terminology.
class Range
{
friend class HashTable;
Iterator mIter;
protected:
explicit Range(const HashTable& table)
: mIter(table)
{
}
public:
bool empty() const { return mIter.done(); }
T& front() const { return mIter.get(); }
void popFront() { return mIter.next(); }
};
// Enum is similar to ModIterator, but uses different terminology.
class Enum
{
ModIterator mIter;
// Enum is movable but not copyable.
Enum(const Enum&) = delete;
void operator=(const Enum&) = delete;
public:
template<class Map>
explicit Enum(Map& map)
: mIter(map.mImpl)
{
}
MOZ_IMPLICIT Enum(Enum&& other)
: mIter(std::move(other.mIter))
{
}
bool empty() const { return mIter.done(); }
T& front() const { return mIter.get(); }
void popFront() { return mIter.next(); }
void removeFront() { mIter.remove(); }
NonConstT& mutableFront() { return mIter.getMutable(); }
void rekeyFront(const Lookup& aLookup, const Key& aKey)
{
mIter.rekey(aLookup, aKey);
}
void rekeyFront(const Key& aKey) { mIter.rekey(aKey); }
};
// HashTable is movable
HashTable(HashTable&& aRhs)
: AllocPolicy(aRhs)
{
PodAssign(this, &aRhs);
aRhs.mTable = nullptr;
}
void operator=(HashTable&& aRhs)
{
MOZ_ASSERT(this != &aRhs, "self-move assignment is prohibited");
if (mTable) {
destroyTable(*this, mTable, capacity());
}
PodAssign(this, &aRhs);
aRhs.mTable = nullptr;
}
private:
// HashTable is not copyable or assignable
HashTable(const HashTable&) = delete;
void operator=(const HashTable&) = delete;
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
static const uint32_t CAP_BITS = 30;
public:
uint64_t mGen : 56; // entry storage generation number
uint64_t mHashShift : 8; // multiplicative hash shift
Entry* mTable; // entry storage
uint32_t mEntryCount; // number of entries in mTable
uint32_t mRemovedCount; // removed entry sentinels in mTable
#ifdef DEBUG
uint64_t mMutationCount;
mutable bool mEntered;
#endif
// The default initial capacity is 32 (enough to hold 16 elements), but it
// can be as low as 4.
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
static const uint32_t sDefaultLen = 16;
static const uint32_t sMinCapacity = 4;
static const uint32_t sMaxInit = 1u << (CAP_BITS - 1);
static const uint32_t sMaxCapacity = 1u << CAP_BITS;
// Hash-table alpha is conceptually a fraction, but to avoid floating-point
// math we implement it as a ratio of integers.
static const uint8_t sAlphaDenominator = 4;
static const uint8_t sMinAlphaNumerator = 1; // min alpha: 1/4
static const uint8_t sMaxAlphaNumerator = 3; // max alpha: 3/4
static const HashNumber sFreeKey = Entry::sFreeKey;
static const HashNumber sRemovedKey = Entry::sRemovedKey;
static const HashNumber sCollisionBit = Entry::sCollisionBit;
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
static uint32_t bestCapacity(uint32_t aLen)
{
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
static_assert((sMaxInit * sAlphaDenominator) / sAlphaDenominator ==
sMaxInit,
"multiplication in numerator below could overflow");
static_assert(sMaxInit * sAlphaDenominator <=
UINT32_MAX - sMaxAlphaNumerator,
"numerator calculation below could potentially overflow");
// Compute the smallest capacity allowing |aLen| elements to be
// inserted without rehashing: ceil(aLen / max-alpha). (Ceiling
// integral division: <http://stackoverflow.com/a/2745086>.)
uint32_t capacity =
(aLen * sAlphaDenominator + sMaxAlphaNumerator - 1) / sMaxAlphaNumerator;
capacity = (capacity < sMinCapacity)
? sMinCapacity
: RoundUpPow2(capacity);
MOZ_ASSERT(capacity >= aLen);
MOZ_ASSERT(capacity <= sMaxCapacity);
return capacity;
}
static uint32_t hashShift(uint32_t aLen)
{
// Reject all lengths whose initial computed capacity would exceed
// sMaxCapacity. Round that maximum aLen down to the nearest power of two
// for speedier code.
if (MOZ_UNLIKELY(aLen > sMaxInit)) {
MOZ_CRASH("initial length is too large");
}
return kHashNumberBits - mozilla::CeilingLog2(bestCapacity(aLen));
}
static bool isLiveHash(HashNumber aHash) { return Entry::isLiveHash(aHash); }
static HashNumber prepareHash(const Lookup& aLookup)
{
HashNumber keyHash = ScrambleHashCode(HashPolicy::hash(aLookup));
// Avoid reserved hash codes.
if (!isLiveHash(keyHash)) {
keyHash -= (sRemovedKey + 1);
}
return keyHash & ~sCollisionBit;
}
enum FailureBehavior
{
DontReportFailure = false,
ReportFailure = true
};
static Entry* createTable(AllocPolicy& aAllocPolicy,
uint32_t aCapacity,
FailureBehavior aReportFailure = ReportFailure)
{
Entry* table = aReportFailure
? aAllocPolicy.template pod_malloc<Entry>(aCapacity)
: aAllocPolicy.template maybe_pod_malloc<Entry>(aCapacity);
if (table) {
for (uint32_t i = 0; i < aCapacity; i++) {
new (&table[i]) Entry();
}
}
return table;
}
static void destroyTable(AllocPolicy& aAllocPolicy,
Entry* aOldTable,
uint32_t aCapacity)
{
Entry* end = aOldTable + aCapacity;
for (Entry* e = aOldTable; e < end; ++e) {
e->~Entry();
}
aAllocPolicy.free_(aOldTable, aCapacity);
}
public:
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
HashTable(AllocPolicy aAllocPolicy, uint32_t aLen)
: AllocPolicy(aAllocPolicy)
, mGen(0)
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
, mHashShift(hashShift(aLen))
, mTable(nullptr)
, mEntryCount(0)
, mRemovedCount(0)
#ifdef DEBUG
, mMutationCount(0)
, mEntered(false)
#endif
{
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
explicit HashTable(AllocPolicy aAllocPolicy)
: HashTable(aAllocPolicy, sDefaultLen)
{
}
~HashTable()
{
if (mTable) {
destroyTable(*this, mTable, capacity());
}
}
private:
HashNumber hash1(HashNumber aHash0) const { return aHash0 >> mHashShift; }
struct DoubleHash
{
HashNumber mHash2;
HashNumber mSizeMask;
};
DoubleHash hash2(HashNumber aCurKeyHash) const
{
uint32_t sizeLog2 = kHashNumberBits - mHashShift;
DoubleHash dh = { ((aCurKeyHash << sizeLog2) >> mHashShift) | 1,
(HashNumber(1) << sizeLog2) - 1 };
return dh;
}
static HashNumber applyDoubleHash(HashNumber aHash1,
const DoubleHash& aDoubleHash)
{
return (aHash1 - aDoubleHash.mHash2) & aDoubleHash.mSizeMask;
}
static MOZ_ALWAYS_INLINE bool match(Entry& aEntry, const Lookup& aLookup)
{
return HashPolicy::match(HashPolicy::getKey(aEntry.get()), aLookup);
}
enum LookupReason
{
ForNonAdd,
ForAdd
};
// Warning: in order for readonlyThreadsafeLookup() to be safe this
// function must not modify the table in any way when Reason==ForNonAdd.
template<LookupReason Reason>
MOZ_ALWAYS_INLINE Entry& lookup(const Lookup& aLookup,
HashNumber aKeyHash) const
{
MOZ_ASSERT(isLiveHash(aKeyHash));
MOZ_ASSERT(!(aKeyHash & sCollisionBit));
MOZ_ASSERT(mTable);
// Compute the primary hash address.
HashNumber h1 = hash1(aKeyHash);
Entry* entry = &mTable[h1];
// Miss: return space for a new entry.
if (entry->isFree()) {
return *entry;
}
// Hit: return entry.
if (entry->matchHash(aKeyHash) && match(*entry, aLookup)) {
return *entry;
}
// Collision: double hash.
DoubleHash dh = hash2(aKeyHash);
// Save the first removed entry pointer so we can recycle later.
Entry* firstRemoved = nullptr;
while (true) {
if (Reason == ForAdd && !firstRemoved) {
if (MOZ_UNLIKELY(entry->isRemoved())) {
firstRemoved = entry;
} else {
entry->setCollision();
}
}
h1 = applyDoubleHash(h1, dh);
entry = &mTable[h1];
if (entry->isFree()) {
return firstRemoved ? *firstRemoved : *entry;
}
if (entry->matchHash(aKeyHash) && match(*entry, aLookup)) {
return *entry;
}
}
}
// This is a copy of lookup() hardcoded to the assumptions:
// 1. the lookup is for an add;
// 2. the key, whose |keyHash| has been passed, is not in the table.
Entry& findNonLiveEntry(HashNumber aKeyHash)
{
MOZ_ASSERT(!(aKeyHash & sCollisionBit));
MOZ_ASSERT(mTable);
// We assume 'aKeyHash' has already been distributed.
// Compute the primary hash address.
HashNumber h1 = hash1(aKeyHash);
Entry* entry = &mTable[h1];
// Miss: return space for a new entry.
if (!entry->isLive()) {
return *entry;
}
// Collision: double hash.
DoubleHash dh = hash2(aKeyHash);
while (true) {
entry->setCollision();
h1 = applyDoubleHash(h1, dh);
entry = &mTable[h1];
if (!entry->isLive()) {
return *entry;
}
}
}
enum RebuildStatus
{
NotOverloaded,
Rehashed,
RehashFailed
};
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
RebuildStatus changeTableSize(uint32_t newCapacity,
FailureBehavior aReportFailure = ReportFailure)
{
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
MOZ_ASSERT(IsPowerOfTwo(newCapacity));
MOZ_ASSERT(!!mTable == !!capacity());
// Look, but don't touch, until we succeed in getting new entry store.
Entry* oldTable = mTable;
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
uint32_t oldCapacity = capacity();
uint32_t newLog2 = mozilla::CeilingLog2(newCapacity);
if (MOZ_UNLIKELY(newCapacity > sMaxCapacity)) {
if (aReportFailure) {
this->reportAllocOverflow();
}
return RehashFailed;
}
Entry* newTable = createTable(*this, newCapacity, aReportFailure);
if (!newTable) {
return RehashFailed;
}
// We can't fail from here on, so update table parameters.
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
mHashShift = kHashNumberBits - newLog2;
mRemovedCount = 0;
mGen++;
mTable = newTable;
// Copy only live entries, leaving removed ones behind.
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
Entry* end = oldTable + oldCapacity;
for (Entry* src = oldTable; src < end; ++src) {
if (src->isLive()) {
HashNumber hn = src->getKeyHash();
findNonLiveEntry(hn).setLive(
hn, std::move(const_cast<typename Entry::NonConstT&>(src->get())));
}
src->~Entry();
}
// All entries have been destroyed, no need to destroyTable.
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
this->free_(oldTable, oldCapacity);
return Rehashed;
}
RebuildStatus rehashIfOverloaded(
FailureBehavior aReportFailure = ReportFailure)
{
static_assert(sMaxCapacity <= UINT32_MAX / sMaxAlphaNumerator,
"multiplication below could overflow");
// Note: if capacity() is zero, this will always succeed, which is
// what we want.
bool overloaded = mEntryCount + mRemovedCount >=
capacity() * sMaxAlphaNumerator / sAlphaDenominator;
if (!overloaded) {
return NotOverloaded;
}
// Succeed if a quarter or more of all entries are removed. Note that this
// always succeeds if capacity() == 0 (i.e. entry storage has not been
// allocated), which is what we want, because it means changeTableSize()
// will allocate the requested capacity rather than doubling it.
bool manyRemoved = mRemovedCount >= (capacity() >> 2);
uint32_t newCapacity = manyRemoved ? rawCapacity() : rawCapacity() * 2;
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
return changeTableSize(newCapacity, aReportFailure);
}
void infallibleRehashIfOverloaded()
{
if (rehashIfOverloaded(DontReportFailure) == RehashFailed) {
rehashTableInPlace();
}
}
void remove(Entry& aEntry)
{
MOZ_ASSERT(mTable);
if (aEntry.hasCollision()) {
aEntry.removeLive();
mRemovedCount++;
} else {
aEntry.clearLive();
}
mEntryCount--;
#ifdef DEBUG
mMutationCount++;
#endif
}
void shrinkIfUnderloaded()
{
static_assert(sMaxCapacity <= UINT32_MAX / sMinAlphaNumerator,
"multiplication below could overflow");
bool underloaded =
capacity() > sMinCapacity &&
mEntryCount <= capacity() * sMinAlphaNumerator / sAlphaDenominator;
if (underloaded) {
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
(void)changeTableSize(capacity() / 2, DontReportFailure);
}
}
// 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 rehashTableInPlace()
{
mRemovedCount = 0;
mGen++;
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
for (uint32_t i = 0; i < capacity(); ++i) {
mTable[i].unsetCollision();
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
for (uint32_t i = 0; i < capacity();) {
Entry* src = &mTable[i];
if (!src->isLive() || src->hasCollision()) {
++i;
continue;
}
HashNumber keyHash = src->getKeyHash();
HashNumber h1 = hash1(keyHash);
DoubleHash dh = hash2(keyHash);
Entry* tgt = &mTable[h1];
while (true) {
if (!tgt->hasCollision()) {
src->swap(tgt);
tgt->setCollision();
break;
}
h1 = applyDoubleHash(h1, dh);
tgt = &mTable[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.
}
// Note: |aLookup| may be a reference to a piece of |u|, so this function
// must take care not to use |aLookup| after moving |u|.
//
// Prefer to use putNewInfallible; this function does not check
// invariants.
template<typename... Args>
void putNewInfallibleInternal(const Lookup& aLookup, Args&&... aArgs)
{
MOZ_ASSERT(mTable);
HashNumber keyHash = prepareHash(aLookup);
Entry* entry = &findNonLiveEntry(keyHash);
MOZ_ASSERT(entry);
if (entry->isRemoved()) {
mRemovedCount--;
keyHash |= sCollisionBit;
}
entry->setLive(keyHash, std::forward<Args>(aArgs)...);
mEntryCount++;
#ifdef DEBUG
mMutationCount++;
#endif
}
public:
void clear()
{
Entry* end = mTable + capacity();
for (Entry* e = mTable; e < end; ++e) {
e->clear();
}
mRemovedCount = 0;
mEntryCount = 0;
#ifdef DEBUG
mMutationCount++;
#endif
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
// Resize the table down to the smallest capacity that doesn't overload the
// table. Since we call shrinkIfUnderloaded() on every remove, you only need
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
// to call this after a bulk removal of items done without calling remove().
void compact()
{
if (empty()) {
// Free the entry storage.
this->free_(mTable, capacity());
mGen++;
mHashShift = hashShift(0); // gives minimum capacity on regrowth
mTable = nullptr;
mRemovedCount = 0;
return;
}
uint32_t bestCapacity = this->bestCapacity(mEntryCount);
MOZ_ASSERT(bestCapacity <= capacity());
if (bestCapacity < capacity()) {
(void)changeTableSize(bestCapacity, DontReportFailure);
}
}
void clearAndCompact()
{
clear();
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
compact();
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
MOZ_MUST_USE bool reserve(uint32_t aLen)
{
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
if (aLen == 0) {
return true;
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
uint32_t bestCapacity = this->bestCapacity(aLen);
if (bestCapacity <= capacity()) {
return true; // Capacity is already sufficient.
}
RebuildStatus status = changeTableSize(bestCapacity, DontReportFailure);
MOZ_ASSERT(status != NotOverloaded);
return status != RehashFailed;
}
Iterator iter() const
{
return Iterator(*this);
}
ModIterator modIter()
{
return ModIterator(*this);
}
Range all() const
{
return Range(*this);
}
bool empty() const
{
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
return mEntryCount == 0;
}
uint32_t count() const
{
return mEntryCount;
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
uint32_t rawCapacity() const
{
return 1u << (kHashNumberBits - mHashShift);
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
uint32_t capacity() const
{
return mTable ? rawCapacity() : 0;
}
Generation generation() const
{
return Generation(mGen);
}
size_t shallowSizeOfExcludingThis(MallocSizeOf aMallocSizeOf) const
{
return aMallocSizeOf(mTable);
}
size_t shallowSizeOfIncludingThis(MallocSizeOf aMallocSizeOf) const
{
return aMallocSizeOf(this) + shallowSizeOfExcludingThis(aMallocSizeOf);
}
MOZ_ALWAYS_INLINE Ptr readonlyThreadsafeLookup(const Lookup& aLookup) const
{
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
if (!mTable || !HasHash<HashPolicy>(aLookup)) {
return Ptr();
}
HashNumber keyHash = prepareHash(aLookup);
return Ptr(lookup<ForNonAdd>(aLookup, keyHash), *this);
}
MOZ_ALWAYS_INLINE Ptr lookup(const Lookup& aLookup) const
{
ReentrancyGuard g(*this);
return readonlyThreadsafeLookup(aLookup);
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
MOZ_ALWAYS_INLINE AddPtr lookupForAdd(const Lookup& aLookup)
{
ReentrancyGuard g(*this);
if (!EnsureHash<HashPolicy>(aLookup)) {
return AddPtr();
}
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
if (!mTable) {
uint32_t newCapacity = rawCapacity();
RebuildStatus status = changeTableSize(newCapacity, ReportFailure);
MOZ_ASSERT(status != NotOverloaded);
if (status == RehashFailed) {
return AddPtr();
}
}
HashNumber keyHash = prepareHash(aLookup);
// Directly call the constructor in the return statement to avoid
// excess copying when building with Visual Studio 2017.
// See bug 1385181.
return AddPtr(lookup<ForAdd>(aLookup, keyHash), *this, keyHash);
}
template<typename... Args>
MOZ_MUST_USE bool add(AddPtr& aPtr, Args&&... aArgs)
{
ReentrancyGuard g(*this);
Bug 1481998 - Make mozilla::Hash{Map,Set}'s entry storage allocation lazy. r=luke,sfink Entry storage allocation now occurs on the first lookupForAdd()/put()/putNew(). This removes the need for init() and initialized(), and matches how PLDHashTable/nsTHashtable work. It also removes the need for init() functions in a lot of types that are built on top of mozilla::Hash{Map,Set}. Pros: - No need for init() calls and subsequent checks. - No memory allocated for empty tables, which are not that uncommon. Cons: - An extra branch in lookup() and lookupForAdd(), but not in put()/putNew(), because the existing checkOverloaded() can handle it. Specifics: - Construction now can take a length parameter. - init() is removed. Explicit length-setting, when necessary, now occurs in the constructors. - initialized() is removed. - capacity() now returns zero when the entry storage is absent. - lookupForAdd() is no longer `const`, because it can instantiate the storage, which requires modifications. - lookupForAdd() can now return an invalid AddPtr in two cases: - old: hashing failure (due to OOM in the hasher) - new: OOM while instantiating entry storage The existing failure handling paths for the old case work for the new case. - clear(), finish(), and clearAndShrink() are replaced by clear(), compact(), and reserve(). The old compactIfUnderloaded() is also removed. - Capacity computation code is now in its own functions, bestCapacity() and hashShift(). setTableSizeLog2() is removed. - uint32_t is used throughout for capacities, instead of size_t, for consistency with other similar values. - changeTableSize() now takes a capacity instead of a deltaLog2, and it can now handle !mTable. Measurements: - Total source code size is reduced by over 900 lines. Also, lots of existing lines got shorter (i.e. two checks were reduced to one). - Executable size barely changed, down by 2 KiB on Linux64. The extra branches are compensated for by the lack of init() calls. - Speed changed negligibly. The instruction count for Bench_Cpp_MozHash increased from 2.84 billion to 2.89 billion but any execution time change was well below noise.
2018-08-10 11:00:29 +03:00
MOZ_ASSERT_IF(aPtr.isValid(), mTable);
MOZ_ASSERT_IF(aPtr.isValid(), aPtr.mTable == this);
MOZ_ASSERT(!aPtr.found());
MOZ_ASSERT(!(aPtr.mKeyHash & sCollisionBit));
// Check for error from ensureHash() here.
if (!aPtr.isValid()) {
return false;
}
MOZ_ASSERT(aPtr.mGeneration == generation());
#ifdef DEBUG
MOZ_ASSERT(aPtr.mMutationCount == mMutationCount);
#endif
// Changing an entry from removed to live does not affect whether we
// are overloaded and can be handled separately.
if (aPtr.mEntry->isRemoved()) {
if (!this->checkSimulatedOOM()) {
return false;
}
mRemovedCount--;
aPtr.mKeyHash |= sCollisionBit;
} else {
// Preserve the validity of |aPtr.mEntry|.
RebuildStatus status = rehashIfOverloaded();
if (status == RehashFailed) {
return false;
}
if (status == NotOverloaded && !this->checkSimulatedOOM()) {
return false;
}
if (status == Rehashed) {
aPtr.mEntry = &findNonLiveEntry(aPtr.mKeyHash);
}
}
aPtr.mEntry->setLive(aPtr.mKeyHash, std::forward<Args>(aArgs)...);
mEntryCount++;
#ifdef DEBUG
mMutationCount++;
aPtr.mGeneration = generation();
aPtr.mMutationCount = mMutationCount;
#endif
return true;
}
// Note: |aLookup| may be a reference to a piece of |u|, so this function
// must take care not to use |aLookup| after moving |u|.
template<typename... Args>
void putNewInfallible(const Lookup& aLookup, Args&&... aArgs)
{
MOZ_ASSERT(!lookup(aLookup).found());
ReentrancyGuard g(*this);
putNewInfallibleInternal(aLookup, std::forward<Args>(aArgs)...);
}
// Note: |aLookup| may be alias arguments in |aArgs|, so this function must
// take care not to use |aLookup| after moving |aArgs|.
template<typename... Args>
MOZ_MUST_USE bool putNew(const Lookup& aLookup, Args&&... aArgs)
{
if (!this->checkSimulatedOOM()) {
return false;
}
if (!EnsureHash<HashPolicy>(aLookup)) {
return false;
}
if (rehashIfOverloaded() == RehashFailed) {
return false;
}
putNewInfallible(aLookup, std::forward<Args>(aArgs)...);
return true;
}
// Note: |aLookup| may be a reference to a piece of |u|, so this function
// must take care not to use |aLookup| after moving |u|.
template<typename... Args>
MOZ_MUST_USE bool relookupOrAdd(AddPtr& aPtr,
const Lookup& aLookup,
Args&&... aArgs)
{
// Check for error from ensureHash() here.
if (!aPtr.isValid()) {
return false;
}
#ifdef DEBUG
aPtr.mGeneration = generation();
aPtr.mMutationCount = mMutationCount;
#endif
{
ReentrancyGuard g(*this);
// Check that aLookup has not been destroyed.
MOZ_ASSERT(prepareHash(aLookup) == aPtr.mKeyHash);
aPtr.mEntry = &lookup<ForAdd>(aLookup, aPtr.mKeyHash);
}
return aPtr.found() || add(aPtr, std::forward<Args>(aArgs)...);
}
void remove(Ptr aPtr)
{
MOZ_ASSERT(mTable);
ReentrancyGuard g(*this);
MOZ_ASSERT(aPtr.found());
MOZ_ASSERT(aPtr.mGeneration == generation());
remove(*aPtr.mEntry);
shrinkIfUnderloaded();
}
void rekeyWithoutRehash(Ptr aPtr, const Lookup& aLookup, const Key& aKey)
{
MOZ_ASSERT(mTable);
ReentrancyGuard g(*this);
MOZ_ASSERT(aPtr.found());
MOZ_ASSERT(aPtr.mGeneration == generation());
typename HashTableEntry<T>::NonConstT t(std::move(*aPtr));
HashPolicy::setKey(t, const_cast<Key&>(aKey));
remove(*aPtr.mEntry);
putNewInfallibleInternal(aLookup, std::move(t));
}
void rekeyAndMaybeRehash(Ptr aPtr, const Lookup& aLookup, const Key& aKey)
{
rekeyWithoutRehash(aPtr, aLookup, aKey);
infallibleRehashIfOverloaded();
}
};
} // namespace detail
} // namespace mozilla
#endif /* mozilla_HashTable_h */