[PATCH 00/28] iio: zero init stack with { } instead of memset()

[Andy Shevchenko]

On Thu, Jun 12, 2025 at 3:12 PM Nicolas Frattaroli
<nicolas.frattaroli at collabora.com> wrote:

> I thought I'd chime in as someone uninvolved because this seemed
> interesting.

Welcome! Other opinions on such a topic are always appreciated.

> On Thursday, 12 June 2025 11:17:52 Central European Summer Time Pavel Machek wrote:
> >
> > > Jonathan mentioned recently that he would like to get away from using
> > > memset() to zero-initialize stack memory in the IIO subsystem. And we
> > > have it on good authority that initializing a struct or array with = { }
> > > is the preferred way to do this in the kernel [1]. So here is a series
> > > to take care of that.
> >
> > 1) Is it worth the churn?
> >
> > 2) Will this fail to initialize padding with some obscure compiler?
>
> as of right now, the only two C compilers that are supported are
> GCC >= 8.1, and Clang >= 13.0.1. If anyone even manages to get the kernel
> to finish a build with something else, I think the compiler not
> implementing the C standard correctly is the least of their worries.
>
> My bigger worry is that = { } is only guaranteed to be as correct as
> memset on C23, and the kernel's standard right now is C11. For that
> reason alone, I don't think memset should be moved away from for now,
> unless someone can verify that every GCC release >= 8.1 and every
> Clang release >= 13.0.1 does the right thing here regardless.
>
> >
> > 3) Why do you believe that {} is the preffered way? All we have is
> > Kees' email that explains that = {} maybe works in configs he tested.
>
> = { } is guaranteed to work in C23, as per the standard, but again we're
> not on C23.
>
> The reason to prefer this is likely that it's easier for static analysis
> to see the struct as initialised, but that's me making assumptions here.
>
> A more human-centric argument is that once we're on a C standards version
> where = { } is guaranteed to be correct, then = { } is much more obviously
> correct to a reader than a memset with a value and a size somewhere later
> in the code. This argument is evident from the number of patches in this
> series where the memset and the declaration are not in the same hunk.
> That's the kind of stuff that keeps me awake at night, sweating profusely.

While all you said seems true and I agree with, the pedantism here is
not needed as in the Linux kernel we have {} used for ages in tons of
code and if something went wrong with that we should have had bug
reports already. Are you aware of such? Personally I haven't heard
even one related to this. But if you know, I am really more than
interested to read about (please, give pointers to such a discussion).

-- 
With Best Regards,
Andy Shevchenko