On Sat, Jun 05, 2021 at 08:56:24PM +0200, Johan Corveleyn wrote: > Hmm, I tried this patch with my "annotate large XML file with deep > history test", but the result isn't the same as with 1.14. I'd have to > investigate a bit more to find out where it took another turn, and > what that diff looks like (and whether or not the "other diff" is a > problem for me).
It would be good to know how many iterations your file needs to produce a diff without a limit. You can use something like SVN_DBG(("p=%d\n", p)) to print this number. Could it help to use the new heuristics only in cases where our existing performance tricks fail? I.e. enforce a limit only when the amount of common tokens between the files is low, and the amount of common prefix/suffix lines relative to the total number of lines is low?