[RFD] on removing "git-foo" for builtins from disk

[RFD] on removing "git-foo" for builtins from disk

[Junio C Hamano]

Back in 2005, all "git" subcommands were installed in $PATH as
"git-commit" etc., and people wrote their scripts assuming that they
are all available on their $PATH.  But it then became cumbersome to
have so many "git-foo" commands in /usr/bin/.  Because "git" is
installed in /usr/bin/ and knows to dispatch to 'git-foo' subcommand
when invoked with 'foo' as its first argument, we asked our users to
use "git foo" form and moved "git-foo" commands out of /usr/bin/,
moved to /usr/libexec/git-core/, and gave an escape hatch, which is
to add the output of "git --exec-path" early in their $PATH, to old
scripts (i.e. the scripts needed a one-line fix).

Exectly 12 years ago, we asked the users if they still need these
on-disk binaries [*1*].  Many built-in commands (i.e. subcommands
whose code is in the "git" binary), as long as they are invoked in
"git foo" form, do not need to be installed anywhere on-disk, and
theoretically can be removed, as long as the users and the scripts
they wrote long time ago, following our advise to use the escape
hatch, can somehow be convinced.

Removing /usr/libexec/git-core/git-{add,branch,commit,...} will
allow us to reduce disk footprint on systems that do not support
hardlinks and slightly simplifies the installation procedure.  Some
folks want to do so unconditionally.  The only downside being that
we finally break the promise we made our users 12 years ago.

I do not have a strong opinion either way, but if proponents for
removal can convince others successfully, I do not mind removing the
unnecessary on-disk binaries for built-in commands.

Discuss away.


[References and Footnotes]

*1* https://lore.kernel.org/git/7vprnzt7d5.fsf@gitster.siamese.dyndns.org/