From: Junio C Hamano Date: Mon, 26 Jan 2026 17:32:14 GMT Subject: Re: [PATCH] xdiff: re-diff shifted change groups when using histogram algorithm Message-ID: In-Reply-To: <3aeb49dd-8618-42e0-b9f9-6a4fb8065793@gmail.com> Phillip Wood writes: > On 25/01/2026 17:34, Junio C Hamano wrote: >> Phillip Wood writes: >> >>> On 21/01/2026 20:51, Junio C Hamano wrote: >>>> "Yee Cheng Chin via GitGitGadget" writes: >>>> >>>>> @@ -915,6 +919,45 @@ int xdl_change_compact(xdfile_t *xdf, xdfile_t *xdfo, long flags) { >>>>> } >>>>> } >>>>> >>>>> + /* >>>>> + * If this has a matching group from the other file, it could >>>>> + * either be the original match from the diff algorithm, or >>>>> + * arrived at by shifting and joining groups. When it's the >>>>> + * latter, it's possible for the two newly joined sides to have >>>>> + * matching lines. Re-diff the group to mark these matching >>>>> + * lines as unchanged and remove from the diff output. >> >> Also, after reading the first paragraph of the big comment again, it >> makes me wonder if it is saying the same thing as "When histogram is >> being used, we shouldn't bother shifting up and down to join groups, >> as the result will always worse than the fallback", but is it that >> bad? > > Looking at the example in the commit message the result of shifting up > and down and then calling the fallback is better than either the > unshifted diff or shifting without the fallback, so I don't think just > disabling shifting improves things. It would also stop us coalescing > changed lines, for example > > -A A > A -> -A > -B -B > > The indent heuristic seems to assume that we've shifted down as far as > possible before trying it so that would probably get messed up as well. I see. Thanks for a good explanation. > To me the problem is that the histogram diff does not always generate > particularly good diffs (maybe I'm biased - whenever I've tried > switching the default to "histogram" I've always switched back > "patience" fairly quickly after being presented with a diff that I found > hard to comprehend) > > Thanks > > Phillip Thanks.