Re: [RFC] [PATCH] v4l2: use unsigned rather than enums in ioctl() structs

["Rémi Denis-Courmont" <remi@remlab.net>]

Le mercredi 11 avril 2012 22:53:23 Mauro Carvalho Chehab, vous avez écrit :
> > But you have been royally ignoring that rule when it
> > comes to extending existing enumerations:
> The existing enumerations can be extended, by adding new values for unused
> values, otherwise API functionality can't be extended.

Yes. That's why they should be some unsigned type.

> Yet, except
> for a gcc bug (or weird optimize option), I fail to see why this would
> break the ABI.

>From the perspective of the compiler, this is a feature not a bug. In C and 
C++, loading or storing a value in an enumerated variable whereby the value is 
not a member of the enumeration is undefined. In other words, the compiler can 
assume that this does not happen, and optimize it away.

> If this is all about some gcc optimization with newer gcc versions, then
> maybe the solution may be to add some pragmas at the code to disable such
> optimization when compiling the structs with enum's at videodev2.h.

Maybe the Linux kernel could be specifically compiled to prevent GCC from 
range-optimizing enumerations. But as -fno-jump-table only disables one of 
several potential range optimizations, I doubt this is even possible.

Regardless, you cannot require all of Linux userspace to rely on an hypothetic 
non-standard GNU C extension. Thus extending enums remains a silent userspace 
ABI break in any case.

-- 
Rémi Denis-Courmont
http://www.remlab.net/
http://fi.linkedin.com/in/remidenis