Re: [PATCH 0/3] Teach Git about the patience diff algorithm

From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Jan 2, 2009, 23:05 UTC
Message-ID: <alpine.LFD.2.00.0901021500530.3179@localhost.localdomain>
In-Reply-To: <20090102205208.GA9273@coredump.intra.peff.net>

On Fri, 2 Jan 2009, Jeff King wrote:

Show 29 quoted lines

> 
> Here's another interesting one: d592b315. The commit removes dashes
> from git commands in test scripts. Git says:
> 
>         echo "tag-one-line" >expect &&
> -       git-tag -l | grep "^tag-one-line" >actual &&
> +       git tag -l | grep "^tag-one-line" >actual &&
>         test_cmp expect actual &&
> -       git-tag -n0 -l | grep "^tag-one-line" >actual &&
> +       git tag -n0 -l | grep "^tag-one-line" >actual &&
>         test_cmp expect actual &&
> -       git-tag -n0 -l tag-one-line >actual &&
> +       git tag -n0 -l tag-one-line >actual &&
>         test_cmp expect actual &&
> 
> whereas patience says:
> 
>         echo "tag-one-line" >expect &&
> -       git-tag -l | grep "^tag-one-line" >actual &&
> -       test_cmp expect actual &&
> -       git-tag -n0 -l | grep "^tag-one-line" >actual &&
> -       test_cmp expect actual &&
> -       git-tag -n0 -l tag-one-line >actual &&
> +       git tag -l | grep "^tag-one-line" >actual &&
> +       test_cmp expect actual &&
> +       git tag -n0 -l | grep "^tag-one-line" >actual &&
> +       test_cmp expect actual &&
> +       git tag -n0 -l tag-one-line >actual &&
>         test_cmp expect actual &&

Yeah, the bazaar version clearly is inferior here. But again, I don't think that's actually a patience diff issue, I think it's because the bazaar diff has merged consecutive diff lines too aggressively.

I suspect both patience and the straight Mayers diff (that git uses) actually finds exactly the same differences, and then bazaar has a "merge closeby -/ pairs together if there is just a single unmodified line in between them".

And _that_ is where it should care whether the unmodified line is complex or not. If it's complex, you shouldn't merge it into the -/+ region.

(but yes, it's possible that the bazaar diff uses "uniqueness" as a complexity analysis marker, and while that is somewhat valid, it is _not_ valid enough to be useful. Unique lines tend to be complex, but complex lines are _not_ always unique)

		Linus