Re: [PATCH v4 0/4] RUNTIME_PREFIX relocatable Git

From: Dan Jacques <dnj@google.com>
Date: Nov 29, 2017, 22:38 UTC
Message-ID: <20171129223807.91343-1-dnj@google.com>
In-Reply-To: <87h8tcvlew.fsf@evledraar.booking.com>

Show 10 quoted lines

> In general this whole thing structurally looks good to me with the
> caveats noted in other review E-Mails.
>
> I haven't done anything but skim the details of the "where's my
> executable" C code though, just looked at what it's doing structurally.
>
> I think it makes sense for this to land first ahead of my patch. This is
> an actual feature you need, whereas I just hate our use of MakeMaker,
> but that can wait, unless you're keen to rebase this on my patch. Would
> probably make your whole diff a bit shorter.

I'm not strictly pressed for time here, so we can put this off if that's strategically appropriate. Chromiuum, right now, just manually patches upstream Git with a similar patch set, so it's working and function. This is just an effort to upstream those changes for everyone else to enjoy!

I think that landing in either order is probably okay, so if your RFC goes through pretty quickly I don't mind rebasing, but if this is otherwise good to go in v5, I wouldn't mind landing it and then removing the obsoleted parts during the Makefile update.

Show 6 quoted lines

> The whole converting our absolute to relative paths in the make code is
> unavoidably ugly, but after having an initial knee-jerk reaction to it I
> don't see how it can be avoided. I was hoping most of these paths
> could/would just be a fixed path away from our libexec path, but alas
> due to having had these configurable all along that simplicity seems out
> of reach.

Yeah I spent no small amount of time massaging that code hoping some better formulation would shake out, and this is what I ended up with. UNTIME_PREFIX_PERL is a specialty build with a stronger set of assumptions than the standard installation (RE prefix-relative paths).

> Maybe I asked this before, but I don't see any obvious reason for why
> RUNTIME_PREFIX_PERL is a different thing than RUNTIME_PREFIX as opposed
> to us just doing the right thing for the perl scripts.

I did this to isolate Windows builds from the Perl script changes. Git-for- Windows uses (invented) RUNTIME_PREFIX, but runs its Perl scripts in a MinGW sub-environment which, internally, has the Perl libraries installed at a fixed path, so it doesn't need any special resolution logic.

I support that if Git-for-Windows wants to, a trivial future patch can merge the two, so I opted to play it safe and keep these changes isolated.

===

> We could use $ENV{GIT_EXEC_PATH} instead of FindBin here though, I
> missed that the first time. But that's just a nano-optimization. I just
> wondered whether git wasn't already passing us this info.

Oh good idea - I'll go ahead and do this.

Show 9 quoted lines

> There is one remaining bug here. Git::I18N isn't doing the right thing,
> I installed in /tmp/git and moved to /tmp/git2, and it has:
>
>     our $TEXTDOMAINDIR = $ENV{GIT_TEXTDOMAINDIR} || '/tmp/git/share/locale';
>
> And GIT_TEXTDOMAINDIR is not passed by git (it's only used for the tests
> IIRC). Would need a similar treatment as this. Easiest to just set the
> path we find here in $Git::Whatever and pick it up in $Git::I18N later,
> it's not like anyone uses it outside of git.git.

Good catch! I'm going to see what I can do here.

Show 6 quoted lines

> But that does raise a more general concern for me. Isn't there some way
> we can run the test suite against an installed git (don't remember),
> then build, install, move the dir, and run the tests from the moved dir.
>
> That would have caught this bug, and anything else that may be lurking
> still.

I am not aware of such an option, but I'll take a look again. This sort of thing might just require a reformulation of the test suite in general to make it run against a Git installation instead of the intermediate artifacts. A positive outcome would be the ability to remove the Perl path environment variable hacks ($ENV{...} || ...) and just use production resolution logic, increasing the test surface! But I think that's a bit more work than I have time for at the moment.

===

Show 8 quoted lines

> Noticed after I sent the last E-Mail, this is missing @@INSTLIBDIR@@
> which per my reading of it being initially introduced should be here in
> addition to this relative path.
>
> My reading of the intial patch that added it, as indicated in my patch,
> is that it's the dir we're going to be installing our libs to, so I
> didn't fiddle with it, but I think with your patches we need that dir
> *and* or own $perllibdir.

INSTLIBDIR is used for the standard fixed-prefix header, but not in this one. This one uses a combination of GITEXECDIR and PERLLIBDIR. PERLLIBDIR is effectively the prefix-relative part of INSTLIBDIR, so it's here in spirit!

Thanks for taking the time to review! I'll go ahead and see what I can do RE your comments.

Cheers, -Dan