org.cliffc.high_scale_lib.NonBlockingSetInt Maven / Gradle / Ivy
/*
* Written by Cliff Click and released to the public domain, as explained at
* http://creativecommons.org/licenses/publicdomain
*/
package org.cliffc.high_scale_lib;
import java.io.IOException;
import java.io.Serializable;
import java.lang.reflect.*;
import java.util.*;
import java.util.concurrent.atomic.*;
import sun.misc.Unsafe;
/**
* A multi-threaded bit-vector set, implemented as an array of primitive
* {@code longs}. All operations are non-blocking and multi-threaded safe.
* {@link #contains(int)} calls are roughly the same speed as a {load, mask}
* sequence. {@link #add(int)} and {@link #remove(int)} calls are a tad more
* expensive than a {load, mask, store} sequence because they must use a CAS.
* The bit-vector is auto-sizing.
*
* General note of caution: The Set API allows the use of {@link Integer}
* with silent autoboxing - which can be very expensive if many calls are
* being made. Since autoboxing is silent you may not be aware that this is
* going on. The built-in API takes lower-case {@code ints} and is much more
* efficient.
*
*
Space: space is used in proportion to the largest element, as opposed to
* the number of elements (as is the case with hash-table based Set
* implementations). Space is approximately (largest_element/8 + 64) bytes.
*
* The implementation is a simple bit-vector using CAS for update.
*
* @since 1.5
* @author Cliff Click
*/
public class NonBlockingSetInt extends AbstractSet implements Serializable {
private static final long serialVersionUID = 1234123412341234123L;
private static final Unsafe _unsafe = UtilUnsafe.getUnsafe();
// --- Bits to allow atomic update of the NBSI
private static final long _nbsi_offset;
static { //
Field f = null;
try {
f = NonBlockingSetInt.class.getDeclaredField("_nbsi");
} catch( java.lang.NoSuchFieldException e ) {
}
_nbsi_offset = _unsafe.objectFieldOffset(f);
}
private final boolean CAS_nbsi( NBSI old, NBSI nnn ) {
return _unsafe.compareAndSwapObject(this, _nbsi_offset, old, nnn );
}
// The actual Set of Joy, which changes during a resize event. The
// Only Field for this class, so I can atomically change the entire
// set implementation with a single CAS.
private transient NBSI _nbsi;
/** Create a new empty bit-vector */
public NonBlockingSetInt( ) {
_nbsi = new NBSI(63, new Counter(), this); // The initial 1-word set
}
private NonBlockingSetInt(NonBlockingSetInt a, NonBlockingSetInt b) {
_nbsi = new NBSI(a._nbsi,b._nbsi,new Counter(),this);
}
/**
* Add {@code i} to the set. Uppercase {@link Integer} version of add,
* requires auto-unboxing. When possible use the {@code int} version of
* {@link #add(int)} for efficiency.
* @throws IllegalArgumentException if i is negative.
* @return true if i was added to the set.
*/
public boolean add ( final Integer i ) {
return add(i.intValue());
}
/**
* Test if {@code o} is in the set. This is the uppercase {@link Integer}
* version of contains, requires a type-check and auto-unboxing. When
* possible use the {@code int} version of {@link #contains(int)} for
* efficiency.
* @return true if i was in the set.
*/
public boolean contains( final Object o ) {
return o instanceof Integer ? contains(((Integer)o).intValue()) : false;
}
/**
* Remove {@code o} from the set. This is the uppercase {@link Integer}
* version of remove, requires a type-check and auto-unboxing. When
* possible use the {@code int} version of {@link #remove(int)} for
* efficiency.
* @return true if i was removed to the set.
*/
public boolean remove( final Object o ) {
return o instanceof Integer ? remove (((Integer)o).intValue()) : false;
}
/**
* Add {@code i} to the set. This is the lower-case '{@code int}' version
* of {@link #add} - no autoboxing. Negative values throw
* IllegalArgumentException.
* @throws IllegalArgumentException if i is negative.
* @return true if i was added to the set.
*/
public boolean add( final int i ) {
if( i < 0 ) throw new IllegalArgumentException(""+i);
return _nbsi.add(i);
}
/**
* Test if {@code i} is in the set. This is the lower-case '{@code int}'
* version of {@link #contains} - no autoboxing.
* @return true if i was int the set.
*/
public boolean contains( final int i ) { return i<0 ? false : _nbsi.contains(i); }
/**
* Remove {@code i} from the set. This is the fast lower-case '{@code int}'
* version of {@link #remove} - no autoboxing.
* @return true if i was added to the set.
*/
public boolean remove ( final int i ) { return i<0 ? false : _nbsi.remove (i); }
/**
* Current count of elements in the set. Due to concurrent racing updates,
* the size is only ever approximate. Updates due to the calling thread are
* immediately visible to calling thread.
* @return count of elements.
*/
public int size ( ) { return _nbsi.size( ); }
/** Empty the bitvector. */
public void clear ( ) {
NBSI cleared = new NBSI(63, new Counter(), this); // An empty initial NBSI
while( !CAS_nbsi( _nbsi, cleared ) ) // Spin until clear works
;
}
public int sizeInBytes() { return _nbsi.sizeInBytes(); }
/*****************************************************************
*
* bitwise comparisons optimised for NBSI
*
*****************************************************************/
public NonBlockingSetInt intersect(final NonBlockingSetInt op) {
NonBlockingSetInt res = new NonBlockingSetInt(this,op);
res._nbsi.intersect(res._nbsi, this._nbsi, op._nbsi);
return res;
}
public NonBlockingSetInt union(final NonBlockingSetInt op) {
NonBlockingSetInt res = new NonBlockingSetInt(this,op);
res._nbsi.union(res._nbsi, this._nbsi, op._nbsi);
return res;
}
// public NonBlockingSetInt not(final NonBlockingSetInt op) {
//
// }
/** Verbose printout of internal structure for debugging. */
public void print() { _nbsi.print(0); }
/**
* Standard Java {@link Iterator}. Not very efficient because it
* auto-boxes the returned values.
*/
public Iterator iterator( ) { return new iter(); }
public IntIterator intIterator() { return new NBSIIntIterator(); }
private class NBSIIntIterator implements IntIterator {
NBSI nbsi;
int index = -1;
int prev = -1;
NBSIIntIterator() {
nbsi = _nbsi;
advance();
}
private void advance() {
while( true ) {
index++; // Next index
while( (index>>6) >= nbsi._bits.length ) { // Index out of range?
if( nbsi._new == null ) { // New table?
index = -2; // No, so must be all done
return; //
}
nbsi = nbsi._new; // Carry on, in the new table
}
if( nbsi.contains(index) ) return;
}
}
@Override
public int next() {
if( index == -1 ) throw new NoSuchElementException();
prev = index;
advance();
return prev;
}
@Override
public boolean hasNext() {
return index != -2;
}
public void remove() {
if( prev == -1 ) throw new IllegalStateException();
nbsi.remove(prev);
prev = -1;
}
}
private class iter implements Iterator {
NBSIIntIterator intIterator;
iter() { intIterator = new NBSIIntIterator(); }
public boolean hasNext() { return intIterator.hasNext(); }
public Integer next() { return intIterator.next(); }
public void remove() { intIterator.remove(); }
}
// --- writeObject -------------------------------------------------------
// Write a NBSI to a stream
private void writeObject(java.io.ObjectOutputStream s) throws IOException {
s.defaultWriteObject(); // Nothing to write
final NBSI nbsi = _nbsi; // The One Field is transient
final int len = _nbsi._bits.length<<6;
s.writeInt(len); // Write max element
for( int i=0; i= 0 && idx < ary.length;
return _Lbase + idx * _Lscale;
}
private final boolean CAS( int idx, long old, long nnn ) {
return _unsafe.compareAndSwapLong( _bits, rawIndex(_bits, idx), old, nnn );
}
// --- Resize
// The New Table, only set once to non-zero during a resize.
// Must be atomically set.
private NBSI _new;
private static final long _new_offset;
static { //
Field f = null;
try {
f = NBSI.class.getDeclaredField("_new");
} catch( java.lang.NoSuchFieldException e ) {
}
_new_offset = _unsafe.objectFieldOffset(f);
}
private final boolean CAS_new( NBSI nnn ) {
return _unsafe.compareAndSwapObject(this, _new_offset, null, nnn );
}
private transient final AtomicInteger _copyIdx; // Used to count bits started copying
private transient final AtomicInteger _copyDone; // Used to count words copied in a resize operation
private transient final int _sum_bits_length; // Sum of all nested _bits.lengths
private static final long mask( int i ) { return 1L<<(i&63); }
// I need 1 free bit out of 64 to allow for resize. I do this by stealing
// the high order bit - but then I need to do something with adding element
// number 63 (and friends). I could use a mod63 function but it's more
// efficient to handle the mod-64 case as an exception.
//
// Every 64th bit is put in it's own recursive bitvector. If the low 6 bits
// are all set, we shift them off and recursively operate on the _nbsi64 set.
private final NBSI _nbsi64;
private NBSI( int max_elem, Counter ctr, NonBlockingSetInt nonb ) {
super();
_non_blocking_set_int = nonb;
_size = ctr;
_copyIdx = ctr == null ? null : new AtomicInteger();
_copyDone = ctr == null ? null : new AtomicInteger();
// The main array of bits
_bits = new long[(int)(((long)max_elem+63)>>>6)];
// Every 64th bit is moved off to it's own subarray, so that the
// sign-bit is free for other purposes
_nbsi64 = ((max_elem+1)>>>6) == 0 ? null : new NBSI((max_elem+1)>>>6, null, null);
_sum_bits_length = _bits.length + (_nbsi64==null ? 0 : _nbsi64._sum_bits_length);
}
/** built a new NBSI with buffers large enough to hold bitwise operations on the operands **/
private NBSI(NBSI a, NBSI b, Counter ctr, NonBlockingSetInt nonb) {
super();
_non_blocking_set_int = nonb;
_size = ctr;
_copyIdx = ctr == null ? null : new AtomicInteger();
_copyDone = ctr == null ? null : new AtomicInteger();
if(!has_bits(a) && !has_bits(b)) {
_bits = null;
_nbsi64 = null;
_sum_bits_length = 0;
return;
}
// todo - clean this nastiness up
// essentially just safely creates new empty buffers for each of the recursive bitsets
if(!has_bits(a)) {
_bits = new long[b._bits.length];
_nbsi64 = new NBSI(null,b._nbsi64,null,null);
} else if(!has_bits(b)) {
_bits = new long[a._bits.length];
_nbsi64 = new NBSI(null,a._nbsi64,null,null);
} else {
int bit_length = a._bits.length > b._bits.length ? a._bits.length : b._bits.length;
_bits = new long[bit_length];
_nbsi64 = new NBSI(a._nbsi64,b._nbsi64,null,null);
}
_sum_bits_length = _bits.length + _nbsi64._sum_bits_length;
}
private static boolean has_bits(NBSI n) {
return n != null && n._bits != null;
}
// Lower-case 'int' versions - no autoboxing, very fast.
// 'i' is known positive.
public boolean add( final int i ) {
// Check for out-of-range for the current size bit vector.
// If so we need to grow the bit vector.
if( (i>>6) >= _bits.length )
return install_larger_new_bits(i). // Install larger pile-o-bits (duh)
help_copy().add(i); // Finally, add to the new table
// Handle every 64th bit via using a nested array
NBSI nbsi = this; // The bit array being added into
int j = i; // The bit index being added
while( (j&63) == 63 ) { // Bit 64? (low 6 bits are all set)
nbsi = nbsi._nbsi64; // Recurse
j = j>>6; // Strip off low 6 bits (all set)
}
final long mask = mask(j);
long old;
do {
old = nbsi._bits[j>>6]; // Read old bits
if( old < 0 ) // Not mutable?
// Not mutable: finish copy of word, and retry on copied word
return help_copy_impl(i).help_copy().add(i);
if( (old & mask) != 0 ) return false; // Bit is already set?
} while( !nbsi.CAS( j>>6, old, old | mask ) );
_size.add(1);
return true;
}
public boolean remove( final int i ) {
if( (i>>6) >= _bits.length ) // Out of bounds? Not in this array!
return _new==null ? false : help_copy().remove(i);
// Handle every 64th bit via using a nested array
NBSI nbsi = this; // The bit array being added into
int j = i; // The bit index being added
while( (j&63) == 63 ) { // Bit 64? (low 6 bits are all set)
nbsi = nbsi._nbsi64; // Recurse
j = j>>6; // Strip off low 6 bits (all set)
}
final long mask = mask(j);
long old;
do {
old = nbsi._bits[j>>6]; // Read old bits
if( old < 0 ) // Not mutable?
// Not mutable: finish copy of word, and retry on copied word
return help_copy_impl(i).help_copy().remove(i);
if( (old & mask) == 0 ) return false; // Bit is already clear?
} while( !nbsi.CAS( j>>6, old, old & ~mask ) );
_size.add(-1);
return true;
}
public boolean contains( final int i ) {
if( (i>>6) >= _bits.length ) // Out of bounds? Not in this array!
return _new==null ? false : help_copy().contains(i);
// Handle every 64th bit via using a nested array
NBSI nbsi = this; // The bit array being added into
int j = i; // The bit index being added
while( (j&63) == 63 ) { // Bit 64? (low 6 bits are all set)
nbsi = nbsi._nbsi64; // Recurse
j = j>>6; // Strip off low 6 bits (all set)
}
final long mask = mask(j);
long old = nbsi._bits[j>>6]; // Read old bits
if( old < 0 ) // Not mutable?
// Not mutable: finish copy of word, and retry on copied word
return help_copy_impl(i).help_copy().contains(i);
// Yes mutable: test & return bit
return (old & mask) != 0;
}
/**
* Bitwise operations which store the result in this instance.
* Assumes that this instance contains ample buffer space to store the largest
* buffer from each NBSI in the recursive bitmap.
*
* Also assumes that this method is called during the construction process of
* the bitset before the instance could be leaked to multiple threads.
***/
public boolean intersect(NBSI dest, NBSI a, NBSI b) {
// terminate recursion if one bitset is missing data
// since that word should be left as 0L anyway
if(!has_bits(a) || !has_bits(b))
return true;
for(int i = 0; i < dest._bits.length; i++) {
long left = a.safe_read_word(i,0L);
long right = b.safe_read_word(i,0L);
dest._bits[i] = (left & right) & Long.MAX_VALUE; // mask sign bit
}
// todo - recompute size
return intersect(dest._nbsi64, a._nbsi64, b._nbsi64);
}
public boolean union(NBSI dest, NBSI a, NBSI b) {
// terminate recursion if neiter bitset has data
if(!has_bits(a) && !has_bits(b))
return true;
if(has_bits(a) || has_bits(b)) {
for(int i = 0; i < dest._bits.length; i++) {
long left = a.safe_read_word(i,0);
long right = b.safe_read_word(i,0);
dest._bits[i] = (left | right) & Long.MAX_VALUE;
}
}
return union(dest._nbsi64, a._nbsi64, b._nbsi64);
}
/**************************************************************************/
private long safe_read_word(int i, long default_word) {
if(i >= _bits.length) {
// allow reading past the end of the buffer filling in a default word
return default_word;
}
long word = _bits[i];
if(word < 0) {
word = help_copy_impl(i).help_copy()._bits[i];
}
return word;
}
public int sizeInBytes() { return (int)_bits.length; }
public int size() { return (int)_size.get(); }
// Must grow the current array to hold an element of size i
private NBSI install_larger_new_bits( final int i ) {
if( _new == null ) {
// Grow by powers of 2, to avoid minor grow-by-1's.
// Note: must grow by exact powers-of-2 or the by-64-bit trick doesn't work right
int sz = (_bits.length<<6)<<1;
// CAS to install a new larger size. Did it work? Did it fail? We
// don't know and don't care. Only One can be installed, so if
// another thread installed a too-small size, we can't help it - we
// must simply install our new larger size as a nested-resize table.
CAS_new(new NBSI(sz, _size, _non_blocking_set_int));
}
// Return self for 'fluid' programming style
return this;
}
// Help any top-level NBSI to copy until completed.
// Always return the _new version of *this* NBSI, in case we're nested.
private NBSI help_copy() {
// Pick some words to help with - but only help copy the top-level NBSI.
// Nested NBSI waits until the top is done before we start helping.
NBSI top_nbsi = _non_blocking_set_int._nbsi;
final int HELP = 8; // Tuning number: how much copy pain are we willing to inflict?
// We "help" by forcing individual bit indices to copy. However, bits
// come in lumps of 64 per word, so we just advance the bit counter by 64's.
int idx = top_nbsi._copyIdx.getAndAdd(64*HELP);
for( int i=0; i>6; // Strip off low 6 bits (all set)
}
// Transit from state 1: word is not immutable yet
// Immutable is in bit 63, the sign bit.
long bits = old._bits[j>>6];
while( bits >= 0 ) { // Still in state (1)?
long oldbits = bits;
bits |= mask(63); // Target state of bits: sign-bit means immutable
if( old.CAS( j>>6, oldbits, bits ) ) {
if( oldbits == 0 ) _copyDone.addAndGet(1);
break; // Success - old array word is now immutable
}
bits = old._bits[j>>6]; // Retry if CAS failed
}
// Transit from state 2: non-zero in old and zero in new
if( bits != mask(63) ) { // Non-zero in old?
long new_bits = nnn._bits[j>>6];
if( new_bits == 0 ) { // New array is still zero
new_bits = bits & ~mask(63); // Desired new value: a mutable copy of bits
// One-shot CAS attempt, no loop, from 0 to non-zero.
// If it fails, somebody else did the copy for us
if( !nnn.CAS( j>>6, 0, new_bits ) )
new_bits = nnn._bits[j>>6]; // Since it failed, get the new value
assert new_bits != 0;
}
// Transit from state 3: non-zero in old and non-zero in new
// One-shot CAS attempt, no loop, from non-zero to 0 (but immutable)
if( old.CAS( j>>6, bits, mask(63) ) )
_copyDone.addAndGet(1); // One more word finished copying
}
// Now in state 4: zero (and immutable) in old
// Return the self bitvector for 'fluid' programming style
return this;
}
private void print( int d, String msg ) {
for( int i=0; i