[PATCH 3 of 3] revert: account for computed changes in interactive revert (issue5789)

[Yuya Nishihara]

On Tue, 13 Feb 2018 21:46:54 +0100, Denis Laxalde wrote:
> # HG changeset patch
> # User Denis Laxalde <denis at laxalde.org>
> # Date 1518554564 -3600
> #      Tue Feb 13 21:42:44 2018 +0100
> # Node ID a7d28fab177642e028e37b77727603601c3a76cb
> # Parent  4b8c889eb9d0b7ca6883b93dbd476323c94f677f
> # EXP-Topic revert-interactive-pats
> revert: account for computed changes in interactive revert (issue5789)
> 
> When going through _performrevert() in the interactive case, we build a
> matcher with files to revert and pass it patch.diff() for later
> selection of diff hunks to revert. The files set used to build the
> matcher comes from dirstate and accounts for patterns explicitly passed
> to revert ('hg revert -i <file>') and, in case of nonexistent pattern,
> this set is empty (which is expected). Unfortunately, when going through
> patch.diff() with the resulting matcher (for which .always() is True), a
> new changes tuple will be built that completely ignores patterns passed
> by the user. This leads to the situation described in issue5789, where
> one gets prompted about reverting files unrelated to specified patterns
> because they made a typo or so.
> 
> We fix this by building a 'changes' tuple (scmutil.status tuple) from
> information computed during dirstate inspection in cmdutil.revert() and
> pass this to _performrevert() which then hands it to patch.diff(), thus
> avoiding re-computation of a 'changes' tuple when 'match.always is True'.

Perhaps we should instead build an exact matcher from a list of canonical
paths:

  torevert = actions['revert'][0]  # XXX needs to drop excluded_files
  m = scmutil.matchfiles(repo, torevert)