Re: [PATCH 3/6] Makefile: refactor out "ln || ln -s || cp" pattern

From: Jeff King <peff@peff.net>
Date: Mar 30, 2021, 15:20 UTC
Message-ID: <YGNBoljqknW+Kpif@coredump.intra.peff.net>
In-Reply-To: <xmqqsg4dha3i.fsf@gitster.g>

On Mon, Mar 29, 2021 at 03:24:33PM -0700, Junio C Hamano wrote:

Show 34 quoted lines

> Ævar Arnfjörð Bjarmason  <avarab@gmail.com> writes:
> 
> > Refactor out the hard-to-read and maintain "ln || ln -s || cp"
> > pattern.
> >
> > This was initially added in 3e073dc5611 (Makefile: always provide a
> > fallback when hardlinks fail, 2008-08-25), but since then it's become
> > a lot more complex as we've added:
> >
> >  * 3426e34fedd (Add NO_CROSS_DIRECTORY_HARDLINKS support to the
> >    Makefile, 2009-05-11)
> >
> >  * NO_INSTALL_HARDLINKS in 70de5e65e8c (Makefile:
> >    NO_INSTALL_HARDLINKS, 2012-05-02)
> >
> >  * INSTALL_SYMLINKS in ad874608d8c (Makefile: optionally symlink
> >    libexec/git-core binaries to bin/git, 2018-03-13)
> >
> >  * SKIP_DASHED_BUILT_INS 179227d6e21 (Optionally skip linking/copying
> >    the built-ins, 2020-09-21)
> >
> > Each of those commits had to add a new special-case to this code,
> > resulting in quite an unmaintainable mess for adding any sort of new
> > option.
> >
> > Let's use the newly introduced ln-or-cp.sh script instead, note that
> > we only sometimes pass the --no-cross-directory-hardlinks option, per
> > the previous behavior. The target of the "ln -s" is also another
> > special snowflake, but we're careful to carry that forward.
> 
> Nice.  These explicit command-line options to the helper may be
> harder to initially write and maintain than just exporting the
> relevant $(MAKE) macros and using it from the helper, but once
> it works correctly, it is much easier to see what is going on.

Another option is to make "ln-or-cp" itself a target that "make" knows how to build, and bake the values into its "built" version. Besides making the invocations a bit shorter, it also means that the dependency graph is more correct. If a rule invokes ln-or-cp, its behavior will change if $NO_INSTALL_HARDLINKS changes, for example. Telling that to make requires depending on a sentinel file in each such caller (like GIT-BUILD-OPTIONS). Whereas we could do that once for the "ln-or-cp" target, and then everyone who uses it just depends on it.

I had a series a long time ago that moved the whole Makefile in that direction, but I got nervous that it was too disruptive and too non-idiomatic to be worth pursuing. So I offer the alternative here mostly as food for thought. It may not be a good direction (and we already have good-enough solutions, like depending on GIT-BUILD-OPTIONS).

-Peff