From: brian m. carlson <sandals@crustytoothpaste.net>
Date: Tue, 05 Oct 2021 00:04:13 GMT
Subject: Re: [PATCH] hash: reduce size of algo member of object ID
Message-ID: <YVuWfXqlpZxORS8p@camp.crustytoothpaste.net>
In-Reply-To: <YVq5XCyLDr0SPBzx@coredump.intra.peff.net>

On 2021-10-04 at 08:20:44, Jeff King wrote:
> Oh wait, I'm reading it totally wrong. Adding in the extra 4 bytes
> actually made it _faster_ than not having an algo field. Now I'm
> super-confused. I could believe that it gave us some better alignment,
> but the original struct was 32 bytes. 36 seems like a strictly worse
> number.

My guess is that the increased alignment means that memcpy can perform
much better.  Just because x86 has "fast" unaligned access doesn't mean
it's free; there remains a penalty for that, although other
architectures which support unaligned access have much worse ones.
memcpy and memcmp will perform better when they can use 32-bit chunks to
read without having to process the potentially unaligned pieces at the
beginning and end.

For the record, I have no particular stylistic opinion about whether we
should adopt the proposed patch, but of course if it's faster as it is,
we should probably leave it.
-- 
brian m. carlson (he/him or they/them)
Toronto, Ontario, CA