[PATCH 4 of 4] dirstate: speed up case collision checking

[FUJIWARA Katsunori]

At Fri, 22 Jun 2012 13:49:49 -0700,
Joshua Redstone wrote:
> 
> # HG changeset patch
> # User Joshua Redstone <joshua.redstone at fb.com>
> # Date 1340048916 25200
> # Node ID 4bdb6a9bb7ce10f78a8aa1d2b4d69b00e2949839
> # Parent  a97598a0ca30329ad513f8e9fe443ec3b0fc1baa
> dirstate: speed up case collision checking
> 
> scmutil.casecollisionauditor is slow for large repos.  Switch to a faster
> implementation based on reducing the fraction of all files that are scanned
> and caching the results of case conversion lazily rather than precomputing.
> 
> diff --git a/mercurial/cmdutil.py b/mercurial/cmdutil.py
> --- a/mercurial/cmdutil.py
> +++ b/mercurial/cmdutil.py
> @@ -1202,12 +1202,15 @@
>      cca = None
>      abort, warn = scmutil.checkportabilityalert(ui)
>      if abort or warn:
> -        cca = scmutil.casecollisionauditor(ui, abort, wctx)
> +        cca = scmutil.casecollisionauditor(repo.dirstate.sortedfiles())
>      for f in repo.walk(match):
>          exact = match.exact(f)
>          if exact or not explicitonly and f not in repo.dirstate:
> -            if cca:
> -                cca(f)
> +            if cca and cca.collides(f):
> +                msg = _('possible case-folding collision for %s') % f
> +                if abort:
> +                    raise util.Abort(msg)
> +                ui.warn(_("warning: %s\n") % msg)
>              names.append(f)
>              if ui.verbose or not exact:
>                  ui.status(_('adding %s\n') % match.rel(join(f)))

IMHO, to encapsulate initialization/aborting process of the auditor,
arguments for the constructor of casecollisionauditor should be kept
as same as they are: the auditor can get "repo.dirstate.sortedfiles()"
via "wctx._repo".

In fact, I'm just working to add the new auditor to restrict filename
normalization form (e.g. NFC/NFD). It'll prevent users from adding
files of which names are normalized in unexpected normalization form
especially on Mac OS HFS+ environments.

For such feature:

  - auditor needs at least "ui" to get configurations in each
    repositories

  - abort message should be encapsulated in each auditors

----------------------------------------------------------------------
[FUJIWARA Katsunori]                             foozy at lares.dti.ne.jp