NonBlockingHashMap.java

Go to the documentation of this file.

package com.cliffc.aa.util;
 
import sun.misc.Unsafe;
 
import java.io.IOException;
import java.io.Serializable;
import java.lang.reflect.Field;
import java.util.*;
import java.util.concurrent.ConcurrentMap;
import java.util.concurrent.atomic.AtomicLongFieldUpdater;
import java.util.concurrent.atomic.AtomicReferenceFieldUpdater;
 
/*
 * Written by Cliff Click and released to the public domain, as explained at
 * http://creativecommons.org/licenses/publicdomain
 */
 
public class NonBlockingHashMap<TypeK, TypeV>
  extends AbstractMap<TypeK, TypeV>
  implements ConcurrentMap<TypeK, TypeV>, Cloneable, Serializable {
 
  private static final long serialVersionUID = 1234123412341234123L;
 
  private static final int REPROBE_LIMIT=10; // Too many reprobes then force a table-resize
 
  // --- Bits to allow Unsafe access to arrays
  private static final Unsafe _unsafe = UtilUnsafe.getUnsafe();
  private static final int _Obase  = _unsafe.arrayBaseOffset(Object[].class);
  private static final int _Oscale = _unsafe.arrayIndexScale(Object[].class);
  private static final int _Olog   = _Oscale==4?2:(_Oscale==8?3:9999);
  private static long rawIndex(final Object[] ary, final int idx) {
    assert idx >= 0 && idx < ary.length;
    // Note the long-math requirement, to handle arrays of more than 2^31 bytes
    // - or 2^28 - or about 268M - 8-byte pointer elements.
    return _Obase + ((long)idx << _Olog);
  }
 
  // --- Setup to use Unsafe
  private static final long _kvs_offset;
  static {                      // <clinit>
    Field f = null;
    try { f = NonBlockingHashMap.class.getDeclaredField("_kvs"); }
    catch( java.lang.NoSuchFieldException e ) { throw new RuntimeException(e); }
    _kvs_offset = _unsafe.objectFieldOffset(f);
  }
  private boolean CAS_kvs( final Object[] oldkvs, final Object[] newkvs ) {
    return _unsafe.compareAndSwapObject(this, _kvs_offset, oldkvs, newkvs );
  }
 
  // --- Adding a 'prime' bit onto Values via wrapping with a junk wrapper class
  private static final class Prime {
    final Object _V;
    Prime( Object V ) { _V = V; }
    static Object unbox( Object V ) { return V instanceof Prime ? ((Prime)V)._V : V; }
  }
 
  // --- hash ----------------------------------------------------------------
  // Helper function to spread lousy hashCodes
  private static int hash( final Object key) {
    int h = key.hashCode();     // The real hashCode call
    h ^= (h>>>20) ^ (h>>>12);
    h ^= (h>>> 7) ^ (h>>> 4);
    return h;
  }
 
 
 
  // --- The Hash Table --------------------
  // Slot 0 is always used for a 'CHM' entry below to hold the interesting
  // bits of the hash table.  Slot 1 holds full hashes as an array of ints.
  // Slots {2,3}, {4,5}, etc hold {Key,Value} pairs.  The entire hash table
  // can be atomically replaced by CASing the _kvs field.
  //
  // Why is CHM buried inside the _kvs Object array, instead of the other way
  // around?  The CHM info is used during resize events and updates, but not
  // during standard 'get' operations.  I assume 'get' is much more frequent
  // than 'put'.  'get' can skip the extra indirection of skipping through the
  // CHM to reach the _kvs array.
  private transient Object[] _kvs;
  private static final CHM   chm   (Object[] kvs) { return (CHM  )kvs[0]; }
  private static final int[] hashes(Object[] kvs) { return (int[])kvs[1]; }
  // Number of K,V pairs in the table
  private static final int len(Object[] kvs) { return (kvs.length-2)>>1; }
 
  // Time since last resize
  private transient long _last_resize_milli;
 
  // --- Minimum table size ----------------
  // Pick size 8 K/V pairs, which turns into (8*2+2)*4+12 = 84 bytes on a
  // standard 32-bit HotSpot, and (8*2+2)*8+12 = 156 bytes on 64-bit Azul.
  private static final int MIN_SIZE_LOG=3;             //
  private static final int MIN_SIZE=(1<<MIN_SIZE_LOG); // Must be power of 2
 
  // --- Sentinels -------------------------
  // No-Match-Old - putIfMatch does updates only if it matches the old value,
  // and NO_MATCH_OLD basically counts as a wildcard match.
  private static final Object NO_MATCH_OLD = new Object(); // Sentinel
  // Match-Any-not-null - putIfMatch does updates only if it find a real old
  // value.
  private static final Object MATCH_ANY = new Object(); // Sentinel
  // This K/V pair has been deleted (but the Key slot is forever claimed).
  // The same Key can be reinserted with a new value later.
  public static final Object TOMBSTONE = new Object();
  // Prime'd or box'd version of TOMBSTONE.  This K/V pair was deleted, then a
  // table resize started.  The K/V pair has been marked so that no new
  // updates can happen to the old table (and since the K/V pair was deleted
  // nothing was copied to the new table).
  private static final Prime TOMBPRIME = new Prime(TOMBSTONE);
 
  // --- key,val -------------------------------------------------------------
  // Access K,V for a given idx
  //
  // Note that these are static, so that the caller is forced to read the _kvs
  // field only once, and share that read across all key/val calls - lest the
  // _kvs field move out from under us and back-to-back key & val calls refer
  // to different _kvs arrays.
  private static final Object key(Object[] kvs,int idx) { return kvs[(idx<<1)+2]; }
  private static final Object val(Object[] kvs,int idx) { return kvs[(idx<<1)+3]; }
  private static final boolean CAS_key( Object[] kvs, int idx, Object old, Object key ) {
    return _unsafe.compareAndSwapObject( kvs, rawIndex(kvs,(idx<<1)+2), old, key );
  }
  private static final boolean CAS_val( Object[] kvs, int idx, Object old, Object val ) {
    return _unsafe.compareAndSwapObject( kvs, rawIndex(kvs,(idx<<1)+3), old, val );
  }
 
 
  // --- dump ----------------------------------------------------------------
  public final void print() {
    System.out.println("=========");
    print2(_kvs);
    System.out.println("=========");
  }
  // print the entire state of the table
  private final void print( Object[] kvs ) {
    for( int i=0; i<len(kvs); i++ ) {
      Object K = key(kvs,i);
      if( K != null ) {
        String KS = (K == TOMBSTONE) ? "XXX" : K.toString();
        Object V = val(kvs,i);
        Object U = Prime.unbox(V);
        String p = (V==U) ? "" : "prime_";
        String US = (U == TOMBSTONE) ? "tombstone" : U.toString();
        System.out.println(""+i+" ("+KS+","+p+US+")");
      }
    }
    Object[] newkvs = chm(kvs)._newkvs; // New table, if any
    if( newkvs != null ) {
      System.out.println("----");
      print(newkvs);
    }
  }
  // print only the live values, broken down by the table they are in
  private final void print2( Object[] kvs) {
    for( int i=0; i<len(kvs); i++ ) {
      Object key = key(kvs,i);
      Object val = val(kvs,i);
      Object U = Prime.unbox(val);
      if( key != null && key != TOMBSTONE &&  // key is sane
          val != null && U   != TOMBSTONE ) { // val is sane
        String p = (val==U) ? "" : "prime_";
        System.out.println(""+i+" ("+key+","+p+val+")");
      }
    }
    Object[] newkvs = chm(kvs)._newkvs; // New table, if any
    if( newkvs != null ) {
      System.out.println("----");
      print2(newkvs);
    }
  }
 
  // Count of reprobes
  private transient ConcurrentAutoTable _reprobes = new ConcurrentAutoTable();
  public long reprobes() { long r = _reprobes.get(); _reprobes = new ConcurrentAutoTable(); return r; }
 
 
  // --- reprobe_limit -----------------------------------------------------
  // Heuristic to decide if we have reprobed toooo many times.  Running over
  // the reprobe limit on a 'get' call acts as a 'miss'; on a 'put' call it
  // can trigger a table resize.  Several places must have exact agreement on
  // what the reprobe_limit is, so we share it here.
  private static final int reprobe_limit( int len ) {
    return REPROBE_LIMIT + (len>>4);
  }
 
  // --- NonBlockingHashMap --------------------------------------------------
  // Constructors
 
  public NonBlockingHashMap( ) { this(MIN_SIZE); }
 
  public NonBlockingHashMap( final int initial_sz ) { initialize(initial_sz); }
  private final void initialize( int initial_sz ) {
    if( initial_sz < 0 ) throw new IllegalArgumentException();
    int i;                      // Convert to next largest power-of-2
    if( initial_sz > 1024*1024 ) initial_sz = 1024*1024;
    for( i=MIN_SIZE_LOG; (1<<i) < (initial_sz<<2); i++ ) ;
    // Double size for K,V pairs, add 1 for CHM and 1 for hashes
    _kvs = new Object[((1<<i)<<1)+2];
    _kvs[0] = new CHM(new ConcurrentAutoTable()); // CHM in slot 0
    _kvs[1] = new int[1<<i];          // Matching hash entries
    _last_resize_milli = System.currentTimeMillis();
  }
  // Version for subclassed readObject calls, to be called after the defaultReadObject
  protected final void initialize() { initialize(MIN_SIZE); }
 
  // --- wrappers ------------------------------------------------------------
 
  @Override
  public int     size       ( )                       { return chm(_kvs).size(); }
  @Override
  public boolean isEmpty    ( )                       { return size() == 0;      }
 
  @Override
  public boolean containsKey( Object key )            { return get(key) != null; }
 
  public boolean contains   ( Object val )            { return containsValue(val); }
 
  @Override
  public TypeV   put        ( TypeK  key, TypeV val ) { return putIfMatch( key,      val, NO_MATCH_OLD); }
 
  public TypeV   putIfAbsent( TypeK  key, TypeV val ) { return putIfMatch( key,      val, TOMBSTONE   ); }
 
  @Override
  public TypeV   remove     ( Object key )            { return putIfMatch( key,TOMBSTONE, NO_MATCH_OLD); }
 
  public boolean remove     ( Object key,Object val ) { return putIfMatch( key,TOMBSTONE, val ) == val; }
 
  public TypeV   replace    ( TypeK  key, TypeV val ) { return putIfMatch( key,      val,MATCH_ANY   ); }
 
  public boolean replace    ( TypeK  key, TypeV  oldValue, TypeV newValue ) {
    return putIfMatch( key, newValue, oldValue ) == oldValue;
  }
 
 
  // Atomically replace newVal for oldVal, returning the value that existed
  // there before.  If oldVal is the returned value, then newVal was inserted,
  // otherwise not.  A null oldVal means the key does not exist (only insert if
  // missing); a null newVal means to remove the key.
  @SuppressWarnings("unchecked")
  public final TypeV putIfMatchAllowNull( Object key, Object newVal, Object oldVal ) {
    if( oldVal == null ) oldVal = TOMBSTONE;
    if( newVal == null ) newVal = TOMBSTONE;
    final TypeV res = (TypeV)putIfMatch( this, _kvs, key, newVal, oldVal );
    assert !(res instanceof Prime);
    //assert res != null;
    return res == TOMBSTONE ? null : res;
  }
 
  @SuppressWarnings("unchecked")
  private final TypeV putIfMatch( Object key, Object newVal, Object oldVal ) {
    if (oldVal == null || newVal == null) throw new NullPointerException();
    final Object res = putIfMatch( this, _kvs, key, newVal, oldVal );
    assert !(res instanceof Prime);
    assert res != null;
    return res == TOMBSTONE ? null : (TypeV)res;
  }
 
 
  @Override
  public void putAll(Map<? extends TypeK, ? extends TypeV> m) {
    for (Map.Entry<? extends TypeK, ? extends TypeV> e : m.entrySet())
      put(e.getKey(), e.getValue());
  }
 
  @Override
  public void clear() {         // Smack a new empty table down
    Object[] newkvs = new NonBlockingHashMap(MIN_SIZE)._kvs;
    while( !CAS_kvs(_kvs,newkvs) ) // Spin until the clear works
      ;
  }
 
  @Override
  public boolean containsValue( final Object val ) {
    if( val == null ) throw new NullPointerException();
    for( TypeV V : values() )
      if( V == val || V.equals(val) )
        return true;
    return false;
  }
 
  // This function is supposed to do something for Hashtable, and the JCK
  // tests hang until it gets called... by somebody ... for some reason,
  // any reason....
  protected void rehash() {
  }
 
  @SuppressWarnings("unchecked")
  @Override
  public Object clone() {
    try {
      // Must clone, to get the class right; NBHM might have been
      // extended so it would be wrong to just make a new NBHM.
      NonBlockingHashMap<TypeK,TypeV> t = (NonBlockingHashMap<TypeK,TypeV>) super.clone();
      // But I don't have an atomic clone operation - the underlying _kvs
      // structure is undergoing rapid change.  If I just clone the _kvs
      // field, the CHM in _kvs[0] won't be in sync.
      //
      // Wipe out the cloned array (it was shallow anyways).
      t.clear();
      // Now copy sanely
      for( TypeK K : keySet() ) {
        final TypeV V = get(K);  // Do an official 'get'
        t.put(K,V);
      }
      return t;
    } catch (CloneNotSupportedException e) {
      // this shouldn't happen, since we are Cloneable
      throw new InternalError();
    }
  }
 
  @Override
  public String toString() {
    Iterator<Entry<TypeK,TypeV>> i = entrySet().iterator();
    if( !i.hasNext())
      return "{}";
 
    StringBuilder sb = new StringBuilder();
    sb.append('{');
    for (;;) {
      Entry<TypeK,TypeV> e = i.next();
      TypeK key = e.getKey();
      TypeV value = e.getValue();
      sb.append(key   == this ? "(this Map)" : key);
      sb.append('=');
      sb.append(value == this ? "(this Map)" : value);
      if( !i.hasNext())
        return sb.append('}').toString();
      sb.append(", ");
    }
  }
 
  // Get *any* valid Key.  Useful for making worklists out of a hashtable - the
  // hashtable can remove dup work.  Repeated calls probably return the same
  // Key, so only useful if keys are removed after calling.
  @SuppressWarnings("unchecked")
  public TypeK getKey() { return (TypeK)_getKey(_kvs); }
  private static Object _getKey(Object[] kvs) {
    for( int i=0; i<len(kvs); i++ ) {
      Object key = key(kvs,i);
      Object val = val(kvs,i);
      Object U = Prime.unbox(val);
      if( key != null && key != TOMBSTONE &&  // key is sane
          val != null && U   != TOMBSTONE )   // val is sane
        return key;
    }
    Object[] newkvs = chm(kvs)._newkvs; // New table, if any
    return newkvs == null ? null : _getKey(newkvs);
  }
 
  // --- keyeq ---------------------------------------------------------------
  // Check for key equality.  Try direct pointer compare first, then see if
  // the hashes are unequal (fast negative test) and finally do the full-on
  // 'equals' v-call.
  private static boolean keyeq( Object K, Object key, int[] hashes, int hash, int fullhash ) {
    return
      K==key ||                 // Either keys match exactly OR
      // hash exists and matches?  hash can be zero during the install of a
      // new key/value pair.
      ((hashes[hash] == 0 || hashes[hash] == fullhash) &&
       // Do not call the users' "equals()" call with a Tombstone, as this can
       // surprise poorly written "equals()" calls that throw exceptions
       // instead of simply returning false.
       K != TOMBSTONE &&        // Do not call users' equals call with a Tombstone
       // Do the match the hard way - with the users' key being the loop-
       // invariant "this" pointer.  I could have flipped the order of
       // operands (since equals is commutative), but I'm making mega-morphic
       // v-calls in a reprobing loop and nailing down the 'this' argument
       // gives both the JIT and the hardware a chance to prefetch the call target.
       key.equals(K));          // Finally do the hard match
  }
 
  // --- get -----------------------------------------------------------------
  // Never returns a Prime nor a Tombstone.
  @SuppressWarnings("unchecked")
  @Override
  public TypeV get( Object key ) {
    final Object V = get_impl(this,_kvs,key);
    assert !(V instanceof Prime); // Never return a Prime
    assert V != TOMBSTONE;
    return (TypeV)V;
  }
 
  private static Object get_impl( final NonBlockingHashMap topmap, final Object[] kvs, final Object key ) {
    final int fullhash= hash (key); // throws NullPointerException if key is null
    final int len     = len  (kvs); // Count of key/value pairs, reads kvs.length
    final CHM chm     = chm  (kvs); // The CHM, for a volatile read below; reads slot 0 of kvs
    final int[] hashes=hashes(kvs); // The memoized hashes; reads slot 1 of kvs
 
    int idx = fullhash & (len-1); // First key hash
 
    // Main spin/reprobe loop, looking for a Key hit
    int reprobe_cnt=0;
    while( true ) {
      // Probe table.  Each read of 'val' probably misses in cache in a big
      // table; hopefully the read of 'key' then hits in cache.
      final Object K = key(kvs,idx); // Get key   before volatile read, could be null
      final Object V = val(kvs,idx); // Get value before volatile read, could be null or Tombstone or Prime
      if( K == null ) return null;   // A clear miss
 
      // We need a volatile-read here to preserve happens-before semantics on
      // newly inserted Keys.  If the Key body was written just before inserting
      // into the table a Key-compare here might read the uninitialized Key body.
      // Annoyingly this means we have to volatile-read before EACH key compare.
      // .
      // We also need a volatile-read between reading a newly inserted Value
      // and returning the Value (so the user might end up reading the stale
      // Value contents).  Same problem as with keys - and the one volatile
      // read covers both.
      final Object[] newkvs = chm._newkvs; // VOLATILE READ before key compare
 
      // Key-compare
      if( keyeq(K,key,hashes,idx,fullhash) ) {
        // Key hit!  Check for no table-copy-in-progress
        if( !(V instanceof Prime) ) // No copy?
          return (V == TOMBSTONE) ? null : V; // Return the value
        // Key hit - but slot is (possibly partially) copied to the new table.
        // Finish the copy & retry in the new table.
        return get_impl(topmap,chm.copy_slot_and_check(topmap,kvs,idx,key),key); // Retry in the new table
      }
      // get and put must have the same key lookup logic!  But only 'put'
      // needs to force a table-resize for a too-long key-reprobe sequence.
      // Check for too-many-reprobes on get - and flip to the new table.
      if( ++reprobe_cnt >= reprobe_limit(len) || // too many probes
          K == TOMBSTONE ) // found a TOMBSTONE key, means no more keys in this table
        return newkvs == null ? null : get_impl(topmap,topmap.help_copy(newkvs),key); // Retry in the new table
 
      idx = (idx+1)&(len-1);    // Reprobe by 1!  (could now prefetch)
    }
  }
 
  // --- getk -----------------------------------------------------------------
  // Never returns a Prime nor a Tombstone.
  @SuppressWarnings("unchecked")
  public TypeK getk( TypeK key ) {
    return (TypeK)getk_impl(this,_kvs,key);
  }
 
  private static final Object getk_impl( final NonBlockingHashMap topmap, final Object[] kvs, final Object key ) {
    final int fullhash= hash (key); // throws NullPointerException if key is null
    final int len     = len  (kvs); // Count of key/value pairs, reads kvs.length
    final CHM chm     = chm  (kvs); // The CHM, for a volatile read below; reads slot 0 of kvs
    final int[] hashes=hashes(kvs); // The memoized hashes; reads slot 1 of kvs
 
    int idx = fullhash & (len-1); // First key hash
 
    // Main spin/reprobe loop, looking for a Key hit
    int reprobe_cnt=0;
    while( true ) {
      // Probe table.
      final Object K = key(kvs,idx); // Get key before volatile read, could be null
      if( K == null ) return null;   // A clear miss
 
      // We need a volatile-read here to preserve happens-before semantics on
      // newly inserted Keys.  If the Key body was written just before inserting
      // into the table a Key-compare here might read the uninitialized Key body.
      // Annoyingly this means we have to volatile-read before EACH key compare.
      // .
      // We also need a volatile-read between reading a newly inserted Value
      // and returning the Value (so the user might end up reading the stale
      // Value contents).  Same problem as with keys - and the one volatile
      // read covers both.
      final Object[] newkvs = chm._newkvs; // VOLATILE READ before key compare
 
      // Key-compare
      if( keyeq(K,key,hashes,idx,fullhash) )
        return K;              // Return existing Key!
 
      // get and put must have the same key lookup logic!  But only 'put'
      // needs to force a table-resize for a too-long key-reprobe sequence.
      // Check for too-many-reprobes on get - and flip to the new table.
      if( ++reprobe_cnt >= reprobe_limit(len) || // too many probes
          K == TOMBSTONE ) { // found a TOMBSTONE key, means no more keys in this table
        return newkvs == null ? null : getk_impl(topmap,topmap.help_copy(newkvs),key); // Retry in the new table
      }
 
      idx = (idx+1)&(len-1);    // Reprobe by 1!  (could now prefetch)
    }
  }
 
  // --- putIfMatch ---------------------------------------------------------
  // Put, Remove, PutIfAbsent, etc.  Return the old value.  If the returned
  // value is equal to expVal (or expVal is NO_MATCH_OLD) then the put can be
  // assumed to work (although might have been immediately overwritten).  Only
  // the path through copy_slot passes in an expected value of null, and
  // putIfMatch only returns a null if passed in an expected null.
  static volatile int DUMMY_VOLATILE;
  private static final Object putIfMatch( final NonBlockingHashMap topmap, final Object[] kvs, final Object key, final Object putval, final Object expVal ) {
    assert putval != null;
    assert !(putval instanceof Prime);
    assert !(expVal instanceof Prime);
    final int fullhash = hash  (key); // throws NullPointerException if key null
    final int len      = len   (kvs); // Count of key/value pairs, reads kvs.length
    final CHM chm      = chm   (kvs); // Reads kvs[0]
    final int[] hashes = hashes(kvs); // Reads kvs[1], read before kvs[0]
    int idx = fullhash & (len-1);
 
    // ---
    // Key-Claim stanza: spin till we can claim a Key (or force a resizing).
    int reprobe_cnt=0;
    Object K=null, V=null;
    Object[] newkvs=null;
    while( true ) {             // Spin till we get a Key slot
      V = val(kvs,idx);         // Get old value (before volatile read below!)
      K = key(kvs,idx);         // Get current key
      if( K == null ) {         // Slot is free?
        // Found an empty Key slot - which means this Key has never been in
        // this table.  No need to put a Tombstone - the Key is not here!
        if( putval == TOMBSTONE ) return putval; // Not-now & never-been in this table
        if( expVal == MATCH_ANY ) return null;   // Will not match, even after K inserts
        // Claim the null key-slot
        if( CAS_key(kvs,idx, null, key ) ) { // Claim slot for Key
          chm._slots.add(1);      // Raise key-slots-used count
          hashes[idx] = fullhash; // Memoize fullhash
          break;                  // Got it!
        }
        // CAS to claim the key-slot failed.
        //
        // This re-read of the Key points out an annoying short-coming of Java
        // CAS.  Most hardware CAS's report back the existing value - so that
        // if you fail you have a *witness* - the value which caused the CAS to
        // fail.  The Java API turns this into a boolean destroying the
        // witness.  Re-reading does not recover the witness because another
        // thread can write over the memory after the CAS.  Hence we can be in
        // the unfortunate situation of having a CAS fail *for cause* but
        // having that cause removed by a later store.  This turns a
        // non-spurious-failure CAS (such as Azul has) into one that can
        // apparently spuriously fail - and we avoid apparent spurious failure
        // by not allowing Keys to ever change.
 
        // Volatile read, to force loads of K to retry despite JIT, otherwise
        // it is legal to e.g. haul the load of "K = key(kvs,idx);" outside of
        // this loop (since failed CAS ops have no memory ordering semantics).
        int dummy = DUMMY_VOLATILE;
        continue;
      }
      // Key slot was not null, there exists a Key here
 
      // We need a volatile-read here to preserve happens-before semantics on
      // newly inserted Keys.  If the Key body was written just before inserting
      // into the table a Key-compare here might read the uninitialized Key body.
      // Annoyingly this means we have to volatile-read before EACH key compare.
      newkvs = chm._newkvs;     // VOLATILE READ before key compare
 
      if( keyeq(K,key,hashes,idx,fullhash) )
        break;                  // Got it!
 
      // get and put must have the same key lookup logic!  Lest 'get' give
      // up looking too soon.
      //topmap._reprobes.add(1);
      if( ++reprobe_cnt >= reprobe_limit(len) || // too many probes or
          K == TOMBSTONE ) { // found a TOMBSTONE key, means no more keys
        // We simply must have a new table to do a 'put'.  At this point a
        // 'get' will also go to the new table (if any).  We do not need
        // to claim a key slot (indeed, we cannot find a free one to claim!).
        newkvs = chm.resize(topmap,kvs);
        if( expVal != null ) topmap.help_copy(newkvs); // help along an existing copy
        return putIfMatch(topmap,newkvs,key,putval,expVal);
      }
 
      idx = (idx+1)&(len-1); // Reprobe!
    } // End of spinning till we get a Key slot
 
    // ---
    // Found the proper Key slot, now update the matching Value slot.  We
    // never put a null, so Value slots monotonically move from null to
    // not-null (deleted Values use Tombstone).  Thus if 'V' is null we
    // fail this fast cutout and fall into the check for table-full.
    if( putval == V ) return V; // Fast cutout for no-change
 
    // See if we want to move to a new table (to avoid high average re-probe
    // counts).  We only check on the initial set of a Value from null to
    // not-null (i.e., once per key-insert).  Of course we got a 'free' check
    // of newkvs once per key-compare (not really free, but paid-for by the
    // time we get here).
    if( newkvs == null &&       // New table-copy already spotted?
        // Once per fresh key-insert check the hard way
        ((V == null && chm.tableFull(reprobe_cnt,len)) ||
         // Or we found a Prime, but the JMM allowed reordering such that we
         // did not spot the new table (very rare race here: the writing
         // thread did a CAS of _newkvs then a store of a Prime.  This thread
         // reads the Prime, then reads _newkvs - but the read of Prime was so
         // delayed (or the read of _newkvs was so accelerated) that they
         // swapped and we still read a null _newkvs.  The resize call below
         // will do a CAS on _newkvs forcing the read.
         V instanceof Prime) )
      newkvs = chm.resize(topmap,kvs); // Force the new table copy to start
    // See if we are moving to a new table.
    // If so, copy our slot and retry in the new table.
    if( newkvs != null )
      return putIfMatch(topmap,chm.copy_slot_and_check(topmap,kvs,idx,expVal),key,putval,expVal);
 
    // ---
    // We are finally prepared to update the existing table
    assert !(V instanceof Prime);
 
    // Must match old, and we do not?  Then bail out now.  Note that either V
    // or expVal might be TOMBSTONE.  Also V can be null, if we've never
    // inserted a value before.  expVal can be null if we are called from
    // copy_slot.
 
    if( expVal != NO_MATCH_OLD && // Do we care about expected-Value at all?
        V != expVal &&            // No instant match already?
        (expVal != MATCH_ANY || V == TOMBSTONE || V == null) &&
        !(V==null && expVal == TOMBSTONE) &&    // Match on null/TOMBSTONE combo
        (expVal == null || !expVal.equals(V)) ) // Expensive equals check at the last
      return V;                                 // Do not update!
 
    // Actually change the Value in the Key,Value pair
    if( CAS_val(kvs, idx, V, putval ) ) {
      // CAS succeeded - we did the update!
      // Both normal put's and table-copy calls putIfMatch, but table-copy
      // does not (effectively) increase the number of live k/v pairs.
      if( expVal != null ) {
        // Adjust sizes - a striped counter
        if(  (V == null || V == TOMBSTONE) && putval != TOMBSTONE ) chm._size.add( 1);
        if( !(V == null || V == TOMBSTONE) && putval == TOMBSTONE ) chm._size.add(-1);
      }
    } else {                    // Else CAS failed
      V = val(kvs,idx);         // Get new value
      // If a Prime'd value got installed, we need to re-run the put on the
      // new table.  Otherwise we lost the CAS to another racing put.
      // Simply retry from the start.
      if( V instanceof Prime )
        return putIfMatch(topmap,chm.copy_slot_and_check(topmap,kvs,idx,expVal),key,putval,expVal);
    }
    // Win or lose the CAS, we are done.  If we won then we know the update
    // happened as expected.  If we lost, it means "we won but another thread
    // immediately stomped our update with no chance of a reader reading".
    return (V==null && expVal!=null) ? TOMBSTONE : V;
  }
 
  // --- help_copy ---------------------------------------------------------
  // Help along an existing resize operation.  This is just a fast cut-out
  // wrapper, to encourage inlining for the fast no-copy-in-progress case.  We
  // always help the top-most table copy, even if there are nested table
  // copies in progress.
  private final Object[] help_copy( Object[] helper ) {
    // Read the top-level KVS only once.  We'll try to help this copy along,
    // even if it gets promoted out from under us (i.e., the copy completes
    // and another KVS becomes the top-level copy).
    Object[] topkvs = _kvs;
    CHM topchm = chm(topkvs);
    if( topchm._newkvs == null ) return helper; // No copy in-progress
    topchm.help_copy_impl(this,topkvs,false);
    return helper;
  }
 
 
  // --- CHM -----------------------------------------------------------------
  // The control structure for the NonBlockingHashMap
  private static final class CHM<TypeK,TypeV> {
    // Size in active K,V pairs
    private final ConcurrentAutoTable _size;
    public int size () { return (int)_size.get(); }
 
    // ---
    // These next 2 fields are used in the resizing heuristics, to judge when
    // it is time to resize or copy the table.  Slots is a count of used-up
    // key slots, and when it nears a large fraction of the table we probably
    // end up reprobing too much.  Last-resize-milli is the time since the
    // last resize; if we are running back-to-back resizes without growing
    // (because there are only a few live keys but many slots full of dead
    // keys) then we need a larger table to cut down on the churn.
 
    // Count of used slots, to tell when table is full of dead unusable slots
    private final ConcurrentAutoTable _slots;
    public int slots() { return (int)_slots.get(); }
 
    // ---
    // New mappings, used during resizing.
    // The 'new KVs' array - created during a resize operation.  This
    // represents the new table being copied from the old one.  It's the
    // volatile variable that is read as we cross from one table to the next,
    // to get the required memory orderings.  It monotonically transits from
    // null to set (once).
    volatile Object[] _newkvs;
    private final AtomicReferenceFieldUpdater<CHM,Object[]> _newkvsUpdater =
      AtomicReferenceFieldUpdater.newUpdater(CHM.class,Object[].class, "_newkvs");
    // Set the _next field if we can.
    boolean CAS_newkvs( Object[] newkvs ) {
      while( _newkvs == null )
        if( _newkvsUpdater.compareAndSet(this,null,newkvs) )
          return true;
      return false;
    }
 
    // Sometimes many threads race to create a new very large table.  Only 1
    // wins the race, but the losers all allocate a junk large table with
    // hefty allocation costs.  Attempt to control the overkill here by
    // throttling attempts to create a new table.  I cannot really block here
    // (lest I lose the non-blocking property) but late-arriving threads can
    // give the initial resizing thread a little time to allocate the initial
    // new table.  The Right Long Term Fix here is to use array-lets and
    // incrementally create the new very large array.  In C I'd make the array
    // with malloc (which would mmap under the hood) which would only eat
    // virtual-address and not real memory - and after Somebody wins then we
    // could in parallel initialize the array.  Java does not allow
    // un-initialized array creation (especially of ref arrays!).
    volatile long _resizers; // count of threads attempting an initial resize
    private static final AtomicLongFieldUpdater<CHM> _resizerUpdater =
      AtomicLongFieldUpdater.newUpdater(CHM.class, "_resizers");
 
    // ---
    // Simple constructor
    CHM( ConcurrentAutoTable size ) {
      _size = size;
      _slots= new ConcurrentAutoTable();
    }
 
    // --- tableFull ---------------------------------------------------------
    // Heuristic to decide if this table is too full, and we should start a
    // new table.  Note that if a 'get' call has reprobed too many times and
    // decided the table must be full, then always the estimate_sum must be
    // high and we must report the table is full.  If we do not, then we might
    // end up deciding that the table is not full and inserting into the
    // current table, while a 'get' has decided the same key cannot be in this
    // table because of too many reprobes.  The invariant is:
    //   slots.estimate_sum >= max_reprobe_cnt >= reprobe_limit(len)
    private final boolean tableFull( int reprobe_cnt, int len ) {
      return
        // Do the cheap check first: we allow some number of reprobes always
        reprobe_cnt >= REPROBE_LIMIT &&
        (reprobe_cnt >= reprobe_limit(len) ||
         // More expensive check: see if the table is > 1/2 full.
         _slots.estimate_get() >= (len>>1));
    }
 
    // --- resize ------------------------------------------------------------
    // Resizing after too many probes.  "How Big???" heuristics are here.
    // Callers will (not this routine) will 'help_copy' any in-progress copy.
    // Since this routine has a fast cutout for copy-already-started, callers
    // MUST 'help_copy' lest we have a path which forever runs through
    // 'resize' only to discover a copy-in-progress which never progresses.
    private final Object[] resize( NonBlockingHashMap topmap, Object[] kvs) {
      assert chm(kvs) == this;
 
      // Check for resize already in progress, probably triggered by another thread
      Object[] newkvs = _newkvs; // VOLATILE READ
      if( newkvs != null )       // See if resize is already in progress
        return newkvs;           // Use the new table already
 
      // No copy in-progress, so start one.  First up: compute new table size.
      int oldlen = len(kvs);    // Old count of K,V pairs allowed
      int sz = size();          // Get current table count of active K,V pairs
      int newsz = sz;           // First size estimate
 
      // Heuristic to determine new size.  We expect plenty of dead-slots-with-keys
      // and we need some decent padding to avoid endless reprobing.
      if( sz >= (oldlen>>2) ) { // If we are >25% full of keys then...
        newsz = oldlen<<1;      // Double size, so new table will be between 12.5% and 25% full
        // For tables less than 1M entries, if >50% full of keys then...
        // For tables more than 1M entries, if >75% full of keys then...
        if( 4L*sz >= ((oldlen>>20)!=0?3L:2L)*oldlen )
          newsz = oldlen<<2;    // Double double size, so new table will be between %12.5 (18.75%) and 25% (25%)
      }
      // This heuristic in the next 2 lines leads to a much denser table
      // with a higher reprobe rate
      //if( sz >= (oldlen>>1) ) // If we are >50% full of keys then...
      //  newsz = oldlen<<1;    // Double size
 
      // Last (re)size operation was very recent?  Then double again; slows
      // down resize operations for tables subject to a high key churn rate.
      long tm = System.currentTimeMillis();
      long q=0;
      if( newsz <= oldlen && // New table would shrink or hold steady?
          (tm <= topmap._last_resize_milli+10000 || // Recent resize (less than 10 sec ago)
           (q=_slots.estimate_get()) >= (sz<<1)) )  // 1/2 of keys are dead?
        newsz = oldlen<<1;      // Double the existing size
 
      // Do not shrink, ever
      if( newsz < oldlen ) newsz = oldlen;
 
      // Convert to power-of-2
      int log2;
      for( log2=MIN_SIZE_LOG; (1<<log2) < newsz; log2++ ) ; // Compute log2 of size
      long len = ((1L << log2) << 1) + 2;
      // prevent integer overflow - limit of 2^31 elements in a Java array
      // so here, 2^30 + 2 is the largest number of elements in the hash table
      if ((int)len!=len) {
        log2 = 30;
        len = (1L << log2) + 2;
        if (sz > ((len >> 2) + (len >> 1))) throw new RuntimeException("Table is full.");
      }
 
      // Now limit the number of threads actually allocating memory to a
      // handful - lest we have 750 threads all trying to allocate a giant
      // resized array.
      long r = _resizers;
      while( !_resizerUpdater.compareAndSet(this,r,r+1) )
        r = _resizers;
      // Size calculation: 2 words (K+V) per table entry, plus a handful.  We
      // guess at 64-bit pointers; 32-bit pointers screws up the size calc by
      // 2x but does not screw up the heuristic very much.
      long megs = ((((1L<<log2)<<1)+8)<<3/*word to bytes*/)>>20/*megs*/;
      if( r >= 2 && megs > 0 ) { // Already 2 guys trying; wait and see
        newkvs = _newkvs;        // Between dorking around, another thread did it
        if( newkvs != null )     // See if resize is already in progress
          return newkvs;         // Use the new table already
        // TODO - use a wait with timeout, so we'll wakeup as soon as the new table
        // is ready, or after the timeout in any case.
        //synchronized( this ) { wait(8*megs); }         // Timeout - we always wakeup
        // For now, sleep a tad and see if the 2 guys already trying to make
        // the table actually get around to making it happen.
        try { Thread.sleep(megs); } catch( Exception e ) { }
      }
      // Last check, since the 'new' below is expensive and there is a chance
      // that another thread slipped in a new thread while we ran the heuristic.
      newkvs = _newkvs;
      if( newkvs != null )      // See if resize is already in progress
        return newkvs;          // Use the new table already
 
      // Double size for K,V pairs, add 1 for CHM
      newkvs = new Object[(int)len]; // This can get expensive for big arrays
      newkvs[0] = new CHM(_size); // CHM in slot 0
      newkvs[1] = new int[1<<log2]; // hashes in slot 1
 
      // Another check after the slow allocation
      if( _newkvs != null )     // See if resize is already in progress
        return _newkvs;         // Use the new table already
 
      // The new table must be CAS'd in so only 1 winner amongst duplicate
      // racing resizing threads.  Extra CHM's will be GC'd.
      if( CAS_newkvs( newkvs ) ) { // NOW a resize-is-in-progress!
        //notifyAll();            // Wake up any sleepers
        //long nano = System.nanoTime();
        //System.out.println(" "+nano+" Resize from "+oldlen+" to "+(1<<log2)+" and had "+(_resizers-1)+" extras" );
        //if( System.out != null ) System.out.print("["+log2);
        topmap.rehash();        // Call for Hashtable's benefit
      } else                    // CAS failed?
        newkvs = _newkvs;       // Reread new table
      return newkvs;
    }
 
 
    // The next part of the table to copy.  It monotonically transits from zero
    // to _kvs.length.  Visitors to the table can claim 'work chunks' by
    // CAS'ing this field up, then copying the indicated indices from the old
    // table to the new table.  Workers are not required to finish any chunk;
    // the counter simply wraps and work is copied duplicately until somebody
    // somewhere completes the count.
    volatile long _copyIdx = 0;
    static private final AtomicLongFieldUpdater<CHM> _copyIdxUpdater =
      AtomicLongFieldUpdater.newUpdater(CHM.class, "_copyIdx");
 
    // Work-done reporting.  Used to efficiently signal when we can move to
    // the new table.  From 0 to len(oldkvs) refers to copying from the old
    // table to the new.
    volatile long _copyDone= 0;
    static private final AtomicLongFieldUpdater<CHM> _copyDoneUpdater =
      AtomicLongFieldUpdater.newUpdater(CHM.class, "_copyDone");
 
    // --- help_copy_impl ----------------------------------------------------
    // Help along an existing resize operation.  We hope its the top-level
    // copy (it was when we started) but this CHM might have been promoted out
    // of the top position.
    private final void help_copy_impl( NonBlockingHashMap topmap, Object[] oldkvs, boolean copy_all ) {
      assert chm(oldkvs) == this;
      Object[] newkvs = _newkvs;
      assert newkvs != null;    // Already checked by caller
      int oldlen = len(oldkvs); // Total amount to copy
      final int MIN_COPY_WORK = Math.min(oldlen,1024); // Limit per-thread work
 
      // ---
      int panic_start = -1;
      int copyidx=-9999;            // Fool javac to think it's initialized
      while( _copyDone < oldlen ) { // Still needing to copy?
        // Carve out a chunk of work.  The counter wraps around so every
        // thread eventually tries to copy every slot repeatedly.
 
        // We "panic" if we have tried TWICE to copy every slot - and it still
        // has not happened.  i.e., twice some thread somewhere claimed they
        // would copy 'slot X' (by bumping _copyIdx) but they never claimed to
        // have finished (by bumping _copyDone).  Our choices become limited:
        // we can wait for the work-claimers to finish (and become a blocking
        // algorithm) or do the copy work ourselves.  Tiny tables with huge
        // thread counts trying to copy the table often 'panic'.
        if( panic_start == -1 ) { // No panic?
          copyidx = (int)_copyIdx;
          while( !_copyIdxUpdater.compareAndSet(this,copyidx,copyidx+MIN_COPY_WORK) )
            copyidx = (int)_copyIdx;      // Re-read
          if( !(copyidx < (oldlen<<1)) )  // Panic!
            panic_start = copyidx;        // Record where we started to panic-copy
        }
 
        // We now know what to copy.  Try to copy.
        int workdone = 0;
        for( int i=0; i<MIN_COPY_WORK; i++ )
          if( copy_slot(topmap,(copyidx+i)&(oldlen-1),oldkvs,newkvs) ) // Made an oldtable slot go dead?
            workdone++;         // Yes!
        if( workdone > 0 )      // Report work-done occasionally
          copy_check_and_promote( topmap, oldkvs, workdone );// See if we can promote
 
        copyidx += MIN_COPY_WORK;
        // Uncomment these next 2 lines to turn on incremental table-copy.
        // Otherwise this thread continues to copy until it is all done.
        if( !copy_all && panic_start == -1 ) // No panic?
          return;       // Then done copying after doing MIN_COPY_WORK
      }
      // Extra promotion check, in case another thread finished all copying
      // then got stalled before promoting.
      copy_check_and_promote( topmap, oldkvs, 0 );// See if we can promote
    }
 
    // --- copy_slot_and_check -----------------------------------------------
    // Copy slot 'idx' from the old table to the new table.  If this thread
    // confirmed the copy, update the counters and check for promotion.
    //
    // Returns the result of reading the volatile _newkvs, mostly as a
    // convenience to callers.  We come here with 1-shot copy requests
    // typically because the caller has found a Prime, and has not yet read
    // the _newkvs volatile - which must have changed from null-to-not-null
    // before any Prime appears.  So the caller needs to read the _newkvs
    // field to retry his operation in the new table, but probably has not
    // read it yet.
    private final Object[] copy_slot_and_check( NonBlockingHashMap topmap, Object[] oldkvs, int idx, Object should_help ) {
      assert chm(oldkvs) == this;
      Object[] newkvs = _newkvs; // VOLATILE READ
      // We're only here because the caller saw a Prime, which implies a
      // table-copy is in progress.
      assert newkvs != null;
      if( copy_slot(topmap,idx,oldkvs,_newkvs) )   // Copy the desired slot
        copy_check_and_promote(topmap, oldkvs, 1); // Record the slot copied
      // Generically help along any copy (except if called recursively from a helper)
      return (should_help == null) ? newkvs : topmap.help_copy(newkvs);
    }
 
    // --- copy_check_and_promote --------------------------------------------
    private final void copy_check_and_promote( NonBlockingHashMap topmap, Object[] oldkvs, int workdone ) {
      assert chm(oldkvs) == this;
      int oldlen = len(oldkvs);
      // We made a slot unusable and so did some of the needed copy work
      long copyDone = _copyDone;
      assert (copyDone+workdone) <= oldlen;
      if( workdone > 0 ) {
        while( !_copyDoneUpdater.compareAndSet(this,copyDone,copyDone+workdone) ) {
          copyDone = _copyDone; // Reload, retry
          assert (copyDone+workdone) <= oldlen;
        }
        //if( (10*copyDone/oldlen) != (10*(copyDone+workdone)/oldlen) )
        //System.out.print(" "+(copyDone+workdone)*100/oldlen+"%"+"_"+(_copyIdx*100/oldlen)+"%");
      }
 
      // Check for copy being ALL done, and promote.  Note that we might have
      // nested in-progress copies and manage to finish a nested copy before
      // finishing the top-level copy.  We only promote top-level copies.
      if( copyDone+workdone == oldlen && // Ready to promote this table?
          topmap._kvs == oldkvs &&       // Looking at the top-level table?
          // Attempt to promote
          topmap.CAS_kvs(oldkvs,_newkvs) ) {
        topmap._last_resize_milli = System.currentTimeMillis(); // Record resize time for next check
        //long nano = System.nanoTime();
        //System.out.println(" "+nano+" Promote table to "+len(_newkvs));
        //if( System.out != null ) System.out.print("]");
      }
    }
 
    // --- copy_slot ---------------------------------------------------------
    // Copy one K/V pair from oldkvs[i] to newkvs.  Returns true if we can
    // confirm that the new table guaranteed has a value for this old-table
    // slot.  We need an accurate confirmed-copy count so that we know when we
    // can promote (if we promote the new table too soon, other threads may
    // 'miss' on values not-yet-copied from the old table).  We don't allow
    // any direct updates on the new table, unless they first happened to the
    // old table - so that any transition in the new table from null to
    // not-null must have been from a copy_slot (or other old-table overwrite)
    // and not from a thread directly writing in the new table.  Thus we can
    // count null-to-not-null transitions in the new table.
    private boolean copy_slot( NonBlockingHashMap topmap, int idx, Object[] oldkvs, Object[] newkvs ) {
      // Blindly set the key slot from null to TOMBSTONE, to eagerly stop
      // fresh put's from inserting new values in the old table when the old
      // table is mid-resize.  We don't need to act on the results here,
      // because our correctness stems from box'ing the Value field.  Slamming
      // the Key field is a minor speed optimization.
      Object key;
      while( (key=key(oldkvs,idx)) == null )
        CAS_key(oldkvs,idx, null, TOMBSTONE);
 
      // ---
      // Prevent new values from appearing in the old table.
      // Box what we see in the old table, to prevent further updates.
      Object oldval = val(oldkvs,idx); // Read OLD table
      while( !(oldval instanceof Prime) ) {
        final Prime box = (oldval == null || oldval == TOMBSTONE) ? TOMBPRIME : new Prime(oldval);
        if( CAS_val(oldkvs,idx,oldval,box) ) { // CAS down a box'd version of oldval
          // If we made the Value slot hold a TOMBPRIME, then we both
          // prevented further updates here but also the (absent)
          // oldval is vacuously available in the new table.  We
          // return with true here: any thread looking for a value for
          // this key can correctly go straight to the new table and
          // skip looking in the old table.
          if( box == TOMBPRIME )
            return true;
          // Otherwise we boxed something, but it still needs to be
          // copied into the new table.
          oldval = box;         // Record updated oldval
          break;                // Break loop; oldval is now boxed by us
        }
        oldval = val(oldkvs,idx); // Else try, try again
      }
      if( oldval == TOMBPRIME ) return false; // Copy already complete here!
 
      // ---
      // Copy the value into the new table, but only if we overwrite a null.
      // If another value is already in the new table, then somebody else
      // wrote something there and that write is happens-after any value that
      // appears in the old table.  If putIfMatch does not find a null in the
      // new table - somebody else should have recorded the null-not_null
      // transition in this copy.
      Object old_unboxed = ((Prime)oldval)._V;
      assert old_unboxed != TOMBSTONE;
      putIfMatch(topmap, newkvs, key, old_unboxed, null);
 
      // ---
      // Finally, now that any old value is exposed in the new table, we can
      // forever hide the old-table value by slapping a TOMBPRIME down.  This
      // will stop other threads from uselessly attempting to copy this slot
      // (i.e., it's a speed optimization not a correctness issue).
      while( oldval != TOMBPRIME && !CAS_val(oldkvs,idx,oldval,TOMBPRIME) )
        oldval = val(oldkvs,idx);
 
      return true;
    } // end copy_slot
  } // End of CHM
 
 
  // --- Snapshot ------------------------------------------------------------
  // The main class for iterating over the NBHM.  It "snapshots" a clean
  // view of the K/V array.
  private class SnapshotV implements Iterator<TypeV>, Enumeration<TypeV> {
    final Object[] _sskvs;
    public SnapshotV() {
      while( true ) {           // Verify no table-copy-in-progress
        Object[] topkvs = _kvs;
        CHM topchm = chm(topkvs);
        if( topchm._newkvs == null ) { // No table-copy-in-progress
          // The "linearization point" for the iteration.  Every key in this
          // table will be visited, but keys added later might be skipped or
          // even be added to a following table (also not iterated over).
          _sskvs = topkvs;
          break;
        }
        // Table copy in-progress - so we cannot get a clean iteration.  We
        // must help finish the table copy before we can start iterating.
        topchm.help_copy_impl(NonBlockingHashMap.this,topkvs,true);
      }
      // Warm-up the iterator
      next();
    }
    int length() { return len(_sskvs); }
    Object key(int idx) { return NonBlockingHashMap.key(_sskvs,idx); }
    Object val(int idx) { return NonBlockingHashMap.val(_sskvs,idx); }
    private int _idx;              // Varies from 0-keys.length
    private Object _nextK, _prevK; // Last 2 keys found
    private TypeV  _nextV, _prevV; // Last 2 values found
    public boolean hasNext() { return _nextV != null; }
    @SuppressWarnings("unchecked")
    public TypeV next() {
      // 'next' actually knows what the next value will be - it had to
      // figure that out last go-around lest 'hasNext' report true and
      // some other thread deleted the last value.  Instead, 'next'
      // spends all its effort finding the key that comes after the
      // 'next' key.
      if( _idx != 0 && _nextV == null ) throw new NoSuchElementException();
      _prevK = _nextK;          // This will become the previous key
      _prevV = _nextV;          // This will become the previous value
      _nextV = null;            // We have no more next-key
      Object nV;
      // Attempt to set <_nextK,_nextV> to the next K,V pair.
      // _nextV is the trigger: stop searching when it is != null
      while( _idx<length() ) {  // Scan array
        _nextK = key(_idx++);   // Get a key that definitely is in the set (for the moment!)
        if( _nextK != null &&   // Found something?
            _nextK != TOMBSTONE &&       // Key is not TOMBSTONE
            // Keys can be deleted and not TOMBSTONE, and after deletion had
            // their innards gutted... so you cannot use them to compute their
            // own hash or 'get' a Value.  So get the paired Value directly.
            (nV=val(_idx-1)) != null &&  // Get value
            nV != TOMBSTONE )            // Value is not TOMBSTONE
          { _nextV = (TypeV)nV; break; } // Got it!  _nextK/_nextV is a valid Key/Value
      }                         // Else keep scanning
      return _prevV;            // Return current value.
    }
    public void remove() {
      if( _prevV == null ) throw new IllegalStateException();
      putIfMatch( NonBlockingHashMap.this, _sskvs, _prevK, TOMBSTONE, _prevV );
      _prevV = null;
    }
 
    public TypeV nextElement() { return next(); }
    public boolean hasMoreElements() { return hasNext(); }
  }
  public Object[] raw_array() { return new SnapshotV()._sskvs; }
 
  public Enumeration<TypeV> elements() { return new SnapshotV(); }
 
  // --- values --------------------------------------------------------------
  @Override
  public Collection<TypeV> values() {
    return new AbstractCollection<TypeV>() {
      @Override public void    clear   (          ) {        NonBlockingHashMap.this.clear        ( ); }
      @Override public int     size    (          ) { return NonBlockingHashMap.this.size         ( ); }
      @Override public boolean contains( Object v ) { return NonBlockingHashMap.this.containsValue(v); }
      @Override public Iterator<TypeV> iterator()   { return new SnapshotV(); }
    };
  }
 
  // --- keySet --------------------------------------------------------------
  @SuppressWarnings("unchecked")
  private class SnapshotK implements Iterator<TypeK>, Enumeration<TypeK> {
    final SnapshotV _ss;
    public SnapshotK() { _ss = new SnapshotV(); }
    public void remove() { _ss.remove(); }
    public TypeK next() { _ss.next(); return (TypeK)_ss._prevK; }
    public boolean hasNext() { return _ss.hasNext(); }
    public TypeK nextElement() { return next(); }
    public boolean hasMoreElements() { return hasNext(); }
  }
 
  public Enumeration<TypeK> keys() { return new SnapshotK(); }
 
  @Override
  public Set<TypeK> keySet() {
    return new AbstractSet<TypeK> () {
      @Override public void    clear   (          ) {        NonBlockingHashMap.this.clear   ( ); }
      @Override public int     size    (          ) { return NonBlockingHashMap.this.size    ( ); }
      @Override public boolean contains( Object k ) { return NonBlockingHashMap.this.containsKey(k); }
      @Override public boolean remove  ( Object k ) { return NonBlockingHashMap.this.remove  (k) != null; }
      @Override public Iterator<TypeK> iterator()   { return new SnapshotK(); }
    };
  }
 
 
  // --- entrySet ------------------------------------------------------------
  // Warning: Each call to 'next' in this iterator constructs a new NBHMEntry.
  private class NBHMEntry extends AbstractEntry<TypeK,TypeV> {
    NBHMEntry( final TypeK k, final TypeV v ) { super(k,v); }
    public TypeV setValue(final TypeV val) {
      if( val == null ) throw new NullPointerException();
      _val = val;
      return put(_key, val);
    }
  }
 
  @SuppressWarnings("unchecked")
  private class SnapshotE implements Iterator<Map.Entry<TypeK,TypeV>> {
    final SnapshotV _ss;
    public SnapshotE() { _ss = new SnapshotV(); }
    public void remove() { _ss.remove(); }
    public Map.Entry<TypeK,TypeV> next() { _ss.next(); return new NBHMEntry((TypeK)_ss._prevK,_ss._prevV); }
    public boolean hasNext() { return _ss.hasNext(); }
  }
 
  @Override
  public Set<Map.Entry<TypeK,TypeV>> entrySet() {
    return new AbstractSet<Map.Entry<TypeK,TypeV>>() {
      @Override public void    clear   (          ) {        NonBlockingHashMap.this.clear( ); }
      @Override public int     size    (          ) { return NonBlockingHashMap.this.size ( ); }
      @Override public boolean remove( final Object o ) {
        if( !(o instanceof Map.Entry)) return false;
        final Map.Entry<?,?> e = (Map.Entry<?,?>)o;
        return NonBlockingHashMap.this.remove(e.getKey(), e.getValue());
      }
      @Override public boolean contains(final Object o) {
        if( !(o instanceof Map.Entry)) return false;
        final Map.Entry<?,?> e = (Map.Entry<?,?>)o;
        TypeV v = get(e.getKey());
        return v.equals(e.getValue());
      }
      @Override public Iterator<Map.Entry<TypeK,TypeV>> iterator() { return new SnapshotE(); }
    };
  }
 
  // --- writeObject -------------------------------------------------------
  // Write a NBHM to a stream
  private void writeObject(java.io.ObjectOutputStream s) throws IOException  {
    s.defaultWriteObject();     // Nothing to write
    for( Object K : keySet() ) {
      final Object V = get(K);  // Do an official 'get'
      s.writeObject(K);         // Write the <TypeK,TypeV> pair
      s.writeObject(V);
    }
    s.writeObject(null);        // Sentinel to indicate end-of-data
    s.writeObject(null);
  }
 
  // --- readObject --------------------------------------------------------
  // Read a CHM from a stream
  @SuppressWarnings("unchecked")
  private void readObject(java.io.ObjectInputStream s) throws IOException, ClassNotFoundException {
    s.defaultReadObject();      // Read nothing
    initialize(MIN_SIZE);
    for(;;) {
      final TypeK K = (TypeK) s.readObject();
      final TypeV V = (TypeV) s.readObject();
      if( K == null ) break;
      put(K,V);                 // Insert with an offical put
    }
  }
 
} // End NonBlockingHashMap class