diff options
| author | Fredrik Kuivinen <frekui@gmail.com> | 2011-08-20 22:42:18 (GMT) |
|---|---|---|
| committer | Junio C Hamano <gitster@pobox.com> | 2011-08-21 05:33:58 (GMT) |
| commit | 9eceddeec6ccdbcb0c6ce02a45357ffd5a428f39 (patch) | |
| tree | c30936b59bc81245ed33b7ce8c3b34efe675da9f /grep.h | |
| parent | b95c5ada990ee2afb83af9be5b82d2804618a59d (diff) | |
| download | git-9eceddeec6ccdbcb0c6ce02a45357ffd5a428f39.zip git-9eceddeec6ccdbcb0c6ce02a45357ffd5a428f39.tar.gz git-9eceddeec6ccdbcb0c6ce02a45357ffd5a428f39.tar.bz2 | |
Use kwset in grep
Benchmarks for the hot cache case:
before:
$ perf stat --repeat=5 git grep qwerty > /dev/null
Performance counter stats for 'git grep qwerty' (5 runs):
3,478,085 cache-misses # 2.322 M/sec ( +- 2.690% )
11,356,177 cache-references # 7.582 M/sec ( +- 2.598% )
3,872,184 branch-misses # 0.363 % ( +- 0.258% )
1,067,367,848 branches # 712.673 M/sec ( +- 2.622% )
3,828,370,782 instructions # 0.947 IPC ( +- 0.033% )
4,043,832,831 cycles # 2700.037 M/sec ( +- 0.167% )
8,518 page-faults # 0.006 M/sec ( +- 3.648% )
847 CPU-migrations # 0.001 M/sec ( +- 3.262% )
6,546 context-switches # 0.004 M/sec ( +- 2.292% )
1497.695495 task-clock-msecs # 3.303 CPUs ( +- 2.550% )
0.453394396 seconds time elapsed ( +- 0.912% )
after:
$ perf stat --repeat=5 git grep qwerty > /dev/null
Performance counter stats for 'git grep qwerty' (5 runs):
2,989,918 cache-misses # 3.166 M/sec ( +- 5.013% )
10,986,041 cache-references # 11.633 M/sec ( +- 4.899% ) (scaled from 95.06%)
3,511,993 branch-misses # 1.422 % ( +- 0.785% )
246,893,561 branches # 261.433 M/sec ( +- 3.967% )
1,392,727,757 instructions # 0.564 IPC ( +- 0.040% )
2,468,142,397 cycles # 2613.494 M/sec ( +- 0.110% )
7,747 page-faults # 0.008 M/sec ( +- 3.995% )
897 CPU-migrations # 0.001 M/sec ( +- 2.383% )
6,535 context-switches # 0.007 M/sec ( +- 1.993% )
944.384228 task-clock-msecs # 3.177 CPUs ( +- 0.268% )
0.297257643 seconds time elapsed ( +- 0.450% )
So we gain about 35% by using the kwset code.
As a side effect of using kwset two grep tests are fixed by this
patch. The first is fixed because kwset can deal with case-insensitive
search containing NULs, something strcasestr cannot do. The second one
is fixed because we consider patterns containing NULs as fixed strings
(regcomp cannot accept patterns with NULs).
Signed-off-by: Fredrik Kuivinen <frekui@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Diffstat (limited to 'grep.h')
| -rw-r--r-- | grep.h | 2 |
1 files changed, 2 insertions, 0 deletions
@@ -7,6 +7,7 @@ typedef int pcre; typedef int pcre_extra; #endif +#include "kwset.h" enum grep_pat_token { GREP_PATTERN, @@ -41,6 +42,7 @@ struct grep_pat { regex_t regexp; pcre *pcre_regexp; pcre_extra *pcre_extra_info; + kwset_t kws; unsigned fixed:1; unsigned ignore_case:1; unsigned word_regexp:1; |