Re: [PATCH v2 3/6] match_basename: use strncmp instead of strcmp

From: Duy Nguyen <pclouds@gmail.com>
Date: Mar 10, 2013, 10:38 UTC
Message-ID: <CACsJy8A_4SqLu5L6P0PJ78Lwy12fjL7T2p-KbVEVLJmKNqhyRw@mail.gmail.com>
In-Reply-To: <7vy5dvd7yq.fsf@alter.siamese.dyndns.org>

On Sun, Mar 10, 2013 at 2:34 PM, Junio C Hamano <gitster@pobox.com> wrote:

Show 15 quoted lines

> Nguyễn Thái Ngọc Duy  <pclouds@gmail.com> writes:
>
>> strncmp is provided length information which could be taken advantage
>> by the underlying implementation.
>
> I may be missing something fundamental, but I somehow find the above
> does not make any sense.
>
> strcmp(a, b) has to pay attention to NUL in these strings and stop
> comparison.  strncmp(a, b, n) not only has to pay the same attention
> to NUL in the strings, but also needs to stop comparing at n bytes.
>
> In what situation can the latter take advantage of that extra thing
> that it needs to keep track of and operate faster, when n is the
> length of shorter of the two strings?

glibc's C strncmp version does 4-byte comparison at a time when n >=4, then fall back to 1-byte for the rest. I don't know if it's faster than a plain always 1-byte comparison though. There's also the hand written assembly version that compares n from 1..16, not exactly sure how this version works yet.

Show 20 quoted lines

>> diff --git a/dir.c b/dir.c
>> index 9960a37..46b24db 100644
>> --- a/dir.c
>> +++ b/dir.c
>> @@ -610,12 +610,14 @@ int match_basename(const char *basename, int basenamelen,
>>                  int flags)
>>  {
>>       if (prefix == patternlen) {
>> -             if (!strcmp_icase(pattern, basename))
>> +             if (patternlen == basenamelen &&
>> +                 !strncmp_icase(pattern, basename, patternlen))
>>                       return 1;
>
> What happens if you replace this with
>
>                 if (patternlen == baselen &&
>                     !strcmp_icase(pattern, basename, patternlen))
>
> and drop the other hunk and run the benchmark again?
>

        before      after
user    0m0.533s    0m0.522s
user    0m0.549s    0m0.530s
user    0m0.550s    0m0.534s
user    0m0.551s    0m0.545s
user    0m0.556s    0m0.550s
user    0m0.557s    0m0.552s
user    0m0.559s    0m0.554s
user    0m0.564s    0m0.561s
user    0m0.567s    0m0.565s
user    0m0.567s    0m0.565s

-- 
Duy