Re: Official request: Please make GNU grep the default

From: Tijl Coosemans <tijl_at_coosemans.org>
Date: Thu, 19 Aug 2010 17:34:15 +0200
On Thursday 19 August 2010 15:38:54 Dag-Erling Smørgrav wrote:
> Gabor Kovesdan <gabor_at_FreeBSD.org> writes:
>> I've just committed a patch with the kind help of Dimitry Andric,
>> which gives BSD grep a huge performance boost. The performance is
>> now almost comparable to GNU grep.
> 
> Not quite, as Doug pointed out.  I don't know what benchmark you're
> using, but I'm using a greatly simplified variant of Doug's:
> 
> % time sh -c 'for n in $(jot 1000) ; do /usr/obj/usr/src/usr.bin/grep/grep -q "^xfce4-wm" /usr/ports/INDEX-9 ; done'
> sh -c   13.57s user 7.06s system 99% cpu 20.783 total
> % time sh -c 'for n in $(jot 1000) ; do /usr/obj/usr/src/gnu/usr.bin/grep/grep -q "^xfce4-wm" /usr/ports/INDEX-9 ; done'
> sh -c   7.98s user 7.47s system 100% cpu 15.424 total
> 
> The bottleneck is now in quite an unexpected location:
> 
>   %   cumulative   self              self     total           
>  time   seconds   seconds    calls  ms/call  ms/call  name    
>  38.8       0.03     0.03    12717     0.00     0.00  memchr [5]
>  35.6       0.07     0.03      395     0.08     0.08  _read [6]
>  16.4       0.08     0.01        0  100.00%           _mcount [7]
>   1.7       0.08     0.00    12362     0.00     0.00  memset [9]
>   1.5       0.08     0.00        0  100.00%           .mcount (110)
>   1.5       0.08     0.00        0   43.41%           re_search_internal [8]
>   0.8       0.08     0.00      820     0.00     0.00  memcpy [12]
>   0.6       0.09     0.00    37045     0.00     0.00  free [13]
>   0.6       0.09     0.00    12332     0.00     0.01  grep_fgetln [4]
>   0.6       0.09     0.00        1     0.49    66.27  procfile [3]
>   0.4       0.09     0.00        0  100.00%           re_string_construct_common [26]
>   0.3       0.09     0.00        1     0.25     0.34  _Read_RuneMagi [27]
>   0.1       0.09     0.00      261     0.00     0.00  arena_avail_comp [39]
>   0.1       0.09     0.00      155     0.00     0.00  arena_malloc [24]
>   0.1       0.09     0.00      153     0.00     0.00  arena_bin_malloc_easy [40]
>   0.1       0.09     0.00       54     0.00     0.00  arena_avail_tree_insert [35]
>   0.1       0.09     0.00        5     0.02     0.02  arena_purge [37]
>   0.1       0.09     0.00        3     0.04     0.44  setlocale [10]
>   0.1       0.09     0.00        1     0.12     0.46  __wrap_setrunelocale [21]
>   0.1       0.09     0.00        0   21.76%           re_string_destruct [14]
>   0.1       0.09     0.00        0  100.00%           regexec [38]
> 
> The culprit seems to be the first memchr() in grep_fgetln().  For some
> reason, even with -O2, it is not inlined:
> 
> % echo "disassemble grep_fgetln" | gdb -q -batch -x /dev/stdin /usr/obj/usr/src/usr.bin/grep/grep | grep memchr
> 0x000000000040291e <grep_fgetln+244>:	callq  0x40176c <memchr>
> 0x00000000004029fa <grep_fgetln+464>:	callq  0x40176c <memchr>

The base system gcc doesn't have a built-in version of memchr to inline.

Received on Thu Aug 19 2010 - 13:34:29 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:06 UTC