[PATCH/RFC v3 0/6] N-th last checked out branch

Junio C Hamano [off-list ref] writes:

"cd -" is a very good analogy why your "-" shortcut is a short-sighted
convenience feature that is too narrow and not well designed.  "cd -" can
go back, but you cannot say "ls -" to list the contents of the previous
directory.

True.

Another reason is the one level limitation.

It shares that limitation with my mind ;-)

But well, since you all seem to want a more general solution, here's a
draft.

I started poking around in refs.c for the backwards search, but the
two reflog functions had a lot of functionality implemented very
differently.  So 1-4 are just cleanup and adding that backwards
iterator.  Note that 3/6 changes for_each_reflog_ent() to die() when
it finds a format error, unlike the previous behaviour.  No tests
seemed to care, but maybe someone out there relies on having broken
reflogs? ;-)

 * The code read the reflog twice, first to count how many branch
   switching there are and then to locate the N-th entry we are interested
   in, because I was lazy.  We may want an API to enumerate reflog entries
   in reverse.

So this might be settled.

 * The reflog parser only parses "checkout" and not rebase action.  It
   also does not notice "git checkout HEAD^" is not switching to a real
   branch.

I don't handle this, but I'm not sure how much of a problem it is.  It
can correctly detach and re-attach even in cases such as HEAD^.

Ok, now that you mention it, it will probably put you on the parent of
whatever you happened to be on, instead of the earlier value of HEAD^.

     $ git checkout @{-1}

One problem with the syntax is that the assumption that there can only
be a single @{} construct is quite entrenched in sha1_name.c.  So if
we want to support @{-1}@{1}, that'll need some extra work.

@{-1}~2 and similar work fine though.


Other things of note:

I changed the semantics to really only look at checkouts that "did
something".  If you keep saying 'git checkout master' over and over,
those will not count towards the N.

I changed the parser to read the 'checkout: moving from $old' instead
of the 'to $new'.  The above semantics introduced too many extra
conditions for my taste if we wanted to support the pre-1.5.3 reflog
comment format.  (You'd have to check if the $to was changed from the
last [newer] entry, but then there's another border case when you do
'git init; git checkout -b side; git checkout -' because you never
moved to master.)

The @{date} syntax will be marginally slower after the refactoring
because the old parser carefully avoided parsing numbers where it
could.  (I guess it could actually be done as a bisection for an extra
order of magnitude.)

It's far too early here to be sending mail.

Interactive rebase needs a "move hunk" feature.


Thomas Rast (6):
  reflog: refactor parsing and checking
  reflog: refactor log open+mmap
  reflog: make for_each_reflog_ent use mmap
  reflog: add backwards iterator
  sha1_name: implement @{-N} syntax for N-th last checked out
  checkout: implement '@{-N}' and '-' special abbreviations

 Documentation/git-checkout.txt  |    4 +
 Documentation/git-rev-parse.txt |    3 +
 builtin-checkout.c              |   15 ++-
 cache.h                         |    1 +
 refs.c                          |  285 +++++++++++++++++++++++----------------
 refs.h                          |    1 +
 sha1_name.c                     |   79 +++++++++++-
 t/t1505-rev-parse-last.sh       |   71 ++++++++++
 t/t2012-checkout-last.sh        |   50 +++++++
 9 files changed, 387 insertions(+), 122 deletions(-)
 create mode 100755 t/t1505-rev-parse-last.sh
 create mode 100755 t/t2012-checkout-last.sh