| 1 | |
1 | |
|
|
2 | TODO: try unaligned copy again in decompressor |
|
|
3 | TODO: allow size-optimised binaries by avoiding unrolling |
|
|
4 | |
|
|
5 | - switch to a multiplicative hash (developed with Steinar Gunderson), |
|
|
6 | which is faster on modern cpus and compresses a bit better. The old |
|
|
7 | hash function which uses only shifts is still available. |
|
|
8 | - allow user configurable hash table slots, which makes it possible |
|
|
9 | to use e.g. 16 bit offsets for a smaller hashtable (if your data is |
|
|
10 | always < 64kb). |
|
|
11 | - use _WIN32, not WIN32, when testing for windows (fails with bcc), |
|
|
12 | patch by Tamas Tevesz. |
|
|
13 | - try to port to win64+gcc, which is non-posix but has not _int64 |
|
|
14 | (reported by Maciej Adamczyk). |
|
|
15 | |
|
|
16 | 3.6 Mon Feb 7 17:37:31 CET 2011 |
| 2 | - fixed hash calculation in C♯ version (Tiago Freitas Leal). |
17 | - fixed hash calculation in C♯ version (Tiago Freitas Leal). |
| 3 | - unroll copy for small sizes, use memcpy for larger sizes, |
18 | - unroll copy for small sizes, use memcpy for larger sizes, |
| 4 | greatly speeding up decompression in most cases. |
19 | greatly speeding up decompression in most cases. |
| 5 | - finally disable rep movsb - it's a big loss on modern intel cpus, |
20 | - finally disable rep movsb - it's a big loss on modern intel cpus, |
| 6 | and only a small win on amd cpus. |
21 | and only a small win on amd cpus. |
| 7 | - improve c++ compatibility of the code. |
22 | - improve C++ compatibility of the code. |
| 8 | - slightly improve compressor speed. |
23 | - slightly improve compressor speed. |
| 9 | - halved memory requirements for compressor on 64 bit architectures, |
24 | - halved memory requirements for compressor on 64 bit architectures, |
| 10 | which can improve the speed quite a bit on older cpus. |
25 | which can improve the speed quite a bit on older cpus. |
| 11 | |
26 | |
| 12 | 3.5 Fri May 1 02:28:42 CEST 2009 |
27 | 3.5 Fri May 1 02:28:42 CEST 2009 |
