Re: [PATCH] Makefile: generate 'git' as 'cc [...] -o git+ && mv git+ git'

From: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Date: Mar 8, 2021, 12:38 UTC
Message-ID: <87ft15kegn.fsf@evledraar.gmail.com>
In-Reply-To: <xmqq35x68zob.fsf@gitster.c.googlers.com>

On Sun, Mar 07 2021, Junio C Hamano wrote:

Show 9 quoted lines

> Ævar Arnfjörð Bjarmason  <avarab@gmail.com> writes:
>
>> Change the compilation of the main 'git' binary to not have the CC
>> clobber 'git' in-place. This means that e.g. running the test suite
>> in-place and recompiling won't fail whatever tests happen to be
>> running for the duration of the binary being regenerated.
>
> I am not sure why I would want to support the workflow this is
> trying to help.

Because it also allows me and others to do more testing on patches to git.git.

If I'm working on a patch to e.g. git-fsck I might be doing edit->save->some-tests, where "some-tests" are a subset of the test suite I think is relevant to fsck.

But after doing N commits with passing tests I might notice that some other part of the test suite I didn't expect to have anything to do with fsck broke because I wasn't running that test.

I wasn't running that test because I'm not going to wait 10-15 minutes for it to run while actively editing, but will wait 30s-1m for 10-50 test files to run.

So I can also have the full test suite running in a loop in some side window that'll give me a headsup if the "while do-full-tests; [...]" loop breaks, at which point I'm likely to investigate it sooner than otherwise, and not waste time going down the wrong path.

You can of course do that now, but it requires having a worktree on the side or whatever, which isn't always ideal (sometimes I'd like to have these tests on uncommitted changes).

This change makes it mostly just work.

Show 9 quoted lines

> I do not want to see a patch on this list claiming that "While the
> whole test suite is running, I tweaked the source three times and
> recompiled, so some tests may have used my second iteration while
> others may have used my third and the final version of the code---in
> any case, all tests passed".  And when a patch that was developed
> that way came in, I do not want to hear "Yes, the test suite used
> mixed binaries, but I KNOW the difference between my second and
> third iteration should not matter the parts of the earlier tests
> that used the older iteration".

We have a lot of existing rules in the Makefile that are of the form:

    make thing >thing+ &&
    mv thing+ thing

Where we're not doing the rename dance to avoid clobbering the file we're reading.

As far as I can see all/most of those rules can just be rewritten like e.g. this if this is a use-case we not only don't care about, but would like to go out of our way to break:

    diff --git a/Makefile b/Makefile
    index dfb0f1000fa..8b57c3a5e63 100644
    --- a/Makefile
    +++ b/Makefile
    @@ -2190,2 +2190 @@ config-list.h:
    -       $(QUIET_GEN)$(SHELL_PATH) ./generate-configlist.sh \
    -               >$@+ && mv $@+ $@
    +       $(QUIET_GEN)$(SHELL_PATH) ./generate-configlist.sh >$@