From: Dmitry Potapov <dpotapov@gmail.com>
Date: Thu, 28 Apr 2011 09:32:56 GMT
Subject: Re: [PATCH] git gc: Speed it up by 18% via faster hash comparisons
Message-ID: <BANLkTin+qb6j8p+kOTEkQ2iK29ZWOsRk-g@mail.gmail.com>
In-Reply-To: <20110427231748.GA26632@elie>

2011/4/28 Jonathan Nieder <jrnieder@gmail.com>:
>
> Hm.  This would be very sensitive to the compiler, since a too-smart
> optimizer could take this loop and rewrite it back to memcmp!  So I
> wonder if it's possible to convey this to the compiler more precisely:
>
>        return memcmp_probably_differs_early(sha1, sha2, 20);
>
> E.g., how would something like
>
>        const unsigned int *start1 = (const unsigned int *) sha1;
>        const unsigned int *start2 = (const unsigned int *) sha2;
>
>        if (likely(*start1 != *start2)) {
>                if (*start1 < *start2)

It can be a problem with unalligned access. So, IMHO, it is
better to use get_be32 here:

        unsigned start1 = get_be32(sha1);
        unsigned start2 = get_be32(sha2);

        if (likely(start1 != start2)) {
                if (start1 < start2)

...

Dmitry