Re: [PATCH 0/11] annotating unused function parameters

From: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Date: Aug 25, 2022, 11:00 UTC
Message-ID: <220825.86edx44lzh.gmgdl@evledraar.gmail.com>
In-Reply-To: <YwCtkwjWdJVHHZV0@coredump.intra.peff.net>

On Sat, Aug 20 2022, Jeff King wrote:

Show 12 quoted lines

> On Fri, Aug 19, 2022 at 10:58:08PM +0200, Ævar Arnfjörð Bjarmason wrote:
>
>> Yes, I spoke too soon, sorry. We still need ((unused)). FWIW the below
>> on top of master and doing:
>
> Right. Using ((deprecated)) is really just a substitute for the variable
> renaming part.
>
> And I agree it works as advertised, though I think I prefer the
> variable-renaming version.
>
> One, it feels like we're abusing the deprecated attribute here. The

Definitely, but structurally it seems like a better pick. I.e. isn't the only problem with it the "deprecated" and its interaction with -Wno-deprecated.

If the exact same feature existed as a "insert-custom-warning", which would work exactly "deprecated" without the default warning "prefix" would you think this would fit perfectly?

> ...
> -Wno-deprecated-declarations to get around _actual_ deprecated warnings
> (e.g., compiling with OPENSSL_SHA1=Yes). And doing so would be cutting
> out half the protection of UNUSED() in that case.

This is mildly annoying, but I don't really think it's a practical issue. We're talking about running this without -Wno-deprecated-declarations in CI, and by default.

For unused parameters it's enough that we're catching them somewhere, or in common compilation settings, we don't need to catch them *everywhere*, do we?

IOW is anyone writing patches where they're testing with -Wno-deprecated-declarations *and* adding unused parameters *and* won't test without -Wno-deprecated-declarations before submitting them, *and* nobody else will catch it?

Show 6 quoted lines

> Likewise, one thing I like about the renaming is that it fails
> compilation regardless of -Werror. So it will be caught in any compile,
> no matter what. And I do automatically compile without DEVELOPER=1 when
> on a detached HEAD, because historical commits often trigger warnings.
> Go back far enough and OPENSSL_SHA1 was the default, which generates
> lots of warnings these days. :)

*nod*, I think this also goes the other way. It's nice to be able to use DEVOPTS=no-error to "get past" various minor issues. I consider an unused parameter as being a minor issue. E.g. when ad-hoc cherry-picking something to test on an older version it can be annoying to have to make larger changes when a DEVOPTS=no-error would do.

> And finally, I actually prefer the parentheses of:
>
>   static int register_ref(const char *refname, const struct object_id *oid,
> 			  int UNUSED(flags), void *UNUSED(cb_data))

...and now to the real reason for the follow-up. You/Junio were CC'd, but this is breaking coccinelle, see: https://lore.kernel.org/git/220825.86ilmg4mil.gmgdl@evledraar.gmail.com/

So, bikeshedding-wise I don't care to argue the point. But between odd syntax highlighting and now one analysis tool barfing on it it's a bit more than a bikeshed :)