|
|
<!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML//EN"><html><head><title>What's new in this release of LKRhash?</title><link rel=Stylesheet type="text/css" media=all href="./docs/lkr.css"></head>
<body>
<h1>What's new in this release of LKRhash?</h1>
<ul> <li> <a href="#dec2000"> December 2000</a></li> <li> <a href="#2000-07-31"> 2000/07/31</a></li> <li> <a href="#2000-04-24"> 2000/04/24</a></li> <li> <a href="#2000-03-22"> 2000/03/22</a></li> <li> <a href="#1999-11-04"> 1999/11/04</a></li></ul>
<a name="dec2000"><h2>December 2000</h2></a>
<ul> <li> Refactored the code.</li> <li> Public API.</li> <li> C API.</li> <li> Kernel-mode support.</li> <li> Fixed a deadlock bug in the table lock. If you explicitly call <code>Table::ReadLock</code> and <code>Table::WriteLock</code>, you <strong>need</strong> this fix, or you will deadlock under stress.</li> <li> Fixed a bug in table compaction: until now, the table never actually compacted after elements were deleted, as the test for compaction always evaluated to `false'.</li> <li> Changed the name of <code>EqualKeys</code> to <code>CompareKeys</code>, which now returns an <code>int</code> instead of a <code>bool</code>. This is needed for multikeys support. <li> Changed the signature of <code>AddRefRecord</code>. The second parameter is now an <code>LK_ADDREF_REASON</code>, instead of an <code>int</code> whose value is <code>+1</code> or <code>-1</code>. The reason code aids in debugging refcount leaks. If its value is negative, the refcount should be decremented; otherwise the refcount should be incremented. Also, <code>AddRefRecord</code> now returns <code>LONG</code> instead of <code>void</code>: the new value of the reference count.</li> <li> Changed behavior of <code>ApplyIf</code> locking: now locks one subtable at a time, instead of all subtables. Can use <code>Table::ReadLock</code> or <code>Table::WriteLock</code> to retain old behavior.</li> <li> Changed names and signatures of <code>void LKRHashTableInit()</code> and <code>void LKRHashTableUninit()</code>, to <code>BOOL LKR_Initialize(DWORD)</code> and <code>void LKR_Terminate()</code>, respectively.</li> <li> Fixed <code>iterator::operator*()</code> and <code>iterator::operator->()</code>.</li></ul>
<a name="2000-07-31"><h2>2000/07/31</h2></a>
<ul> <li> <strong>Really</strong> fixed the bug in <code>Clear</code> that left certain internal state variables in an inconsistent state. If you later inserted/deleted enough new records, LKRhash would AV. (The fix in the 2000/03/22 release did not work in all cases.)</li></ul>
<a name="2000-04-24"><h2>2000/04/24</h2></a>
<ul> <li> I added support for STL-style iterators <ul> <li> New iterators <strong>do not lock</strong> the table or a bucket chain. In a multithreaded situation, it is <strong>your</strong> responsibility to call <code>WriteLock</code> (or <code>ReadLock</code>) on the table before initializing any iterators, and to call <code>WriteUnlock</code> (or <code>ReadUnlock</code>) when you are finished.</li> <li> The table provides <code>begin()</code> and <code>end()</code> methods. As the compiler isn't quite smart enough to realize that <code>end()</code> always returns a trivial empty iterator, a loop such as<pre> MyTable::iterator iter; for (iter = pTbl->begin(); iter != pTbl->end(); ++iter) ...</pre> is more efficiently expressed as<pre> MyTable::iterator iter; MyTable::iterator iterEnd = pTbl->end(); for (iter = pTbl->begin(); iter != iterEnd; ++iter) ...</pre></li> <li> iterators can be pre- and post-incremented; i.e.: <code>++iter</code> and <code>iter++</code>. Pre-increment is more efficient than post-increment.</li> <li> Table provides a constructor that accepts a range of iterators into another container.</li> <li> Provides an <code>Insert</code> method that returns an iterator, pointing to the newly inserted record, or <code>end()</code> on failure.</li> <li> Provides a <code>Find</code> method that returns an iterator pointing to the record with the passed-in key, or <code>end()</code> on failure.</li> <li> Provides an <code>EqualRange</code> method that returns two iterators describing the range that contain all the records whose keys match the passed-in key. Until full support for multiple, identical keys is added, the range will contain either zero or one record(s).</li> <li> Provides an <code>Erase</code> method that deletes the record pointed to by the iterator. Updates the iterator to point to the next record in the table.</li> <li> Provides an <code>Erase</code> method that takes two iterators, which will delete all the records in the range described by the two iterators.</li> <li> Unlike the old, deprecated iterators (<code>CIterator</code>), more than one iterator can be active at a time. It is best not to call the non-iterator insert/delete methods (<code>InsertRecord</code>, <code>DeleteRecord</code>, <code>DeleteKey</code>) while iterators are open, as the non-iterator methods can rebalance bucket chains, leading to invalid iterators, undercounting, and/or overcounting. This is true even if the table was WriteLocked, before the iterators were initialized. It is best to use the iterator <code>Insert</code> or <code>Erase</code> methods in such a case.</li> <li> The iterators are reference-counted.</li> </ul> </li> <li> I have provided an NTSD/CDB debugger extension, <i>lkrdbg.dll</i>, with one method, <code>!lkrdbg.lkrhash</code>: <ul> <li> <code>!lkrhash -l[0-2] <i>addr</i></code> will dump the hashtable at <code><i>addr</i></code> at verbosity level <code>l</code> (default 0).</li> <li> <code>!lkrhash -g[0-2]</code> will dump ALL hashtables at at verbosity level <code>l</code> (default 0).</li> <li> I have provided an easy-to-use customization mechanism in <i>lkrcust.h</i> to provide custom dumps for different hashtables. It's keyed off the <code>pszName</code> parameter used in the hashtable constructor. You can provide a custom dump routine for the table (to dump whatever other fields you might have added), as well as a custom dump routine for the record class stored by the hashtable. Provided three examples of customization, based on the samples.</li> </ul> </li> <li> Fixed various build issues <ul> <li> All debug code is now bracketed with <code>#ifdef IRTLDEBUG</code> (instead of <code>#ifdef _DEBUG</code>). Currently equivalent, but you can control this in <i>irtldbg.h</i></li> <li> Fix all Unicode build issues. Code is now TCHAR-aware.</li> <li> Added <i>lkrhash.rc</i> to provide a version resource</li> </ul> </li> <li> Turned on <code>LKRhash</code> and <code>HashFn</code> namespaces by default. See <i>readme.txt</i>.</li> <li> Reorganized the samples into their own subtree.</li> <li> Removed more old code that used to be present so that I could test some changes (e.g., stuff bracketed by <code>LKR_OLD_SEGMENT</code>, <code>LKR_SIGS_NODES</code>, etc).</li> <li> Bucket Lock is once again <code>CSmallSpinLock</code>, unless <code>LKR_DEPRECATED_ITERATORS</code> is defined (off by default).</li> <li> Moved a lot of nested classes out of the table classes, to be top-level classes.</li> <li> Better compile-time and run-time assertions.</li></ul>
<a name="2000-03-22"><h2>2000/03/22</h2></a>
<ul> <li> Fixed a bug in <code>Clear</code> that left certain internal state variables in an inconsistent state. If you later inserted/deleted enough new records, LKRhash would AV.</li> <li> Changed <code>BucketLock</code> to <code>CReaderWriterLock3</code> (recursive MRSW lock) to support certain scenarios, such as being able to call <code>FindKey</code> while enumerating with an old-style iterator. Slightly slower, but the speed improvements below more than compensate.</li> <li> Removed the 300-line example from the end of lkrhash.h. Now in hashtest.cpp, bracketed by <code>SAMPLE_LKRHASH_TESTCLASS</code>.</li> <li> Replaced <code>TRACE</code> macro with <code>IRTLTRACE</code> so as not to interfere with other <code>TRACE</code> macros (e.g., MFC's).</li> <li> Added <i>dirs</i> and <i>sources</i> files so that you can build LKRhash with the NT build environment.</li> <li> Added <code>STATIC_ASSERT</code> macro for compile-time assertions. The <code>IRTLASSERT</code> macro is still used for run-time assertions.</li> <li> Removed old code that used to be present so that I could test some changes (e.g., stuff bracketed by <code>LKR_NEWCODE</code>, <code>LKR_MASK</code>, <code>LKR_SUBTABLE</code>, <code>LKR_COMPACT_DELETE</code>, etc).</li> <li> Upped <code>LK_DFLT_MAXLOAD</code> to 6 (<code>NODES_PER_CLUMP</code>) to get better memory usage.</li> <li> Added support for RockAll (not enabled by default)</li></li> <li> Turned <code>CSegment</code> into a concrete base class. Somewhat hacky but faster.</li> <li> Made the locks a little faster, esp. <code>CReaderWriterLock3::IsWriteLocked</code>.</li> <li> Experimented with countdown loops (turned out to be slightly slower).</li> <li> Experimented with bitwise scrambling for subtable index calculation. Faster.</li> <li> Experimented with using a bitwise mask for subtable index calculation. Faster.</li> <li> Removed some inlines from <i>lkrhash.h</i> to improve modularity.</li> <li> Removed unimplemented <code>Print</code> methods.</li> <li> Bracketed global lists of hashtables with <code>LKR_NO_GLOBAL_LIST</code>.</li> <li> Reduced number of subtables to min(1, #CPUs) for <code>LK_SMALL_TABLESIZE</code> Was min(2, #CPUs). Max number of subtables is now 64.</li></ul>
<a name="1999-11-04"><h2>1999/11/04</h2></a>
<ul> <li> New reader-writer locks.</li> <li> Smarter, faster simple spinlocks.</li> <li> compact delete.</li> <li> debugging support.</li> <li> increased default load factor from 4.0 to 5.0 after reducing size of spinlock => reduced memory usage.</li> <li> deprecated CIterator.</li> <li> better error checking.</li> <li> Win64 clean.</li> <li> expose table locks => composition of operations.</li> <li> global list.</li> <li> faster hash scrambling function..</li> <li> won't fail messily in low-memory situations.</li> <li> fixed a race condition in some of the assertions.</li> <li> enhanced test program.</li></ul>
<hr><address></address><!-- hhmts start -->Last modified: Sat Nov 25 20:41:23 2000<!-- hhmts end --></body></html>
|
