Re: [PATCH 0/2] -Wuninitialized

From: Jeff King <peff@peff.net>
Date: Mar 20, 2018, 04:32 UTC
Message-ID: <20180320043225.GB13302@sigill.intra.peff.net>
In-Reply-To: <8f6d8e2e-aba4-128c-f17c-6c1b9c12436c@ramsayjones.plus.com>

On Mon, Mar 19, 2018 at 05:53:09PM +0000, Ramsay Jones wrote:

Show 8 quoted lines

> This series removes all 'self-initialised' variables (ie. <type> var = var;).
> This construct has been used to silence gcc '-W[maybe-]uninitialized' warnings
> in the past [1]. Unfortunately, this construct causes warnings to be issued by
> MSVC [2], along with clang static analysis complaining about an 'Assigned value
> is garbage or undefined'. The number of these constructs has dropped over the
> years (eg. see [3] and [4]), so there are currently only 6 remaining in the
> current codebase. As demonstrated below, 5 of these no longer cause gcc to
> issue warnings.

Great. I'm happy to see these going away, and thanks for all the careful digging.

Show 16 quoted lines

> If we now add a patch to remove all self-initialization, which would be the
> first patch plus the obvious change to 'saved_namelen' in read-cache.c, then
> note the warnings issued by various compilers at various optimization levels
> on several different platforms [5]:
> 
>                     O0      O1      O2      O3      Os       Og
>  1) gcc 4.8.3   |   -      1,20     1    1,18-19  1-4,21-23  1,5-17
>  2) gcc 4.8.4   |   -      1,20     1       1     1-4,21-23  1,5-8,10-13,15-16 
>  3) clang 3.4   |   -       -       -       -        -       n/a 
>  4) gcc 5.4.0   |   -       1       1       1     1,3-4,21   1,5-8,10-13,16-16
>  5) clang 3.8.0 |   -       -       -       -        -       n/a 
>  6) gcc 5.4.0   |   -       1       1       1       1-4     1,5-17 
>  7) clang 3.8.0 |   -       -       -       -        -       n/a 
>  8) gcc 6.4.0   |   -       1       1    1,18-19    1,4     1,5-17
>  9) clang 5.0.1 |   -       -       -       -        -        -
> 10) gcc 7.2.1   |   -       1       1       1       1,4     1,5-17

So I guess this could create headaches for people using DEVELOPER=1 on as ancient a compiler as 4.8.4, but most other people should be OK. I think I can live with that as a cutoff, and the Travis builds should work there.

(And if we do the detect-compiler stuff from the other nearby thread, somebody who cares can even loosen the warnings for those old gcc versions).

I'm neglecting anybody doing -O3 or -Os here, but IMHO those are sufficiently rare that the builder can tweak their own settings.

I wonder if people use -Og, though? I don't (I usually do -O0 for my edit-compile-debug cycle).

-Peff