Re: performance problem: "git commit filename"

From: Junio C Hamano <gitster@pobox.com>
Date: Jan 14, 2008, 18:38 UTC
Message-ID: <7v63xwl0c6.fsf@gitster.siamese.dyndns.org>
In-Reply-To: <alpine.LFD.1.00.0801140902140.2806@woody.linux-foundation.org>

Linus Torvalds <torvalds@linux-foundation.org> writes:

Show 5 quoted lines

> ... I think it would be even better if we 
> just bit the bullet and started looking at having a different in-memory 
> representation from the on-disk one. Possibly not *that* much different: 
> perhaps just keeping a pointer to the on-disk one along with a flags 
> value.

We have two things we currently do that are not about on-disk index file ('-'), and this patch adds another ('+'):

 - Update the work tree file that corresponds to this entry
   (CE_UPDATE);

 - This entry is to be removed (ce_mode == 0);

 + The work tree file that corresponds to this entry is known to
   be unchanged (CE_UPTODATE).

We could introduce "struct in_core_cache_entry" that has these information, indexed and sorted by name, and has a pointer to what we read from the on-disk index.

	struct in_core_cache_entry {
		struct cache_entry *e;
                unsigned is_up_to_date : 1,
                	 to_be_updated : 1,
                         to_be_removed : 1;
	};

The code that iterate over active_cache[] will instead iterate over this. The number of the entries in this array will be the new active_nr.

In the existing code, we reference "ce->name" and "ce->sha1" everywhere. When we check and update flags we do bitops between "ce->ce_flags" and "htons(CE_BLAH)" in many places. Converting them would adds another indirection and be quite painful. But the compiler can reliably spot the places we fail to find, so it at least is not so risky. It will just be a lot of work.