Re: [kbuild-all] Re: include/linux/compiler_types.h:315:38: error: call to '__compiletime_assert_536' declared with attribute error: BUILD_BUG_ON failed: offsetof(struct can_frame, len) != offsetof(struct canfd_frame, len) || offsetof(struct can_frame, data) != offsetof(struc...

[Oliver Hartkopp <socketcan@hartkopp.net>]

On 23.03.21 12:36, Rasmus Villemoes wrote:
> On 23/03/2021 08.45, Oliver Hartkopp wrote:
> 
>> IMO we facing a compiler problem here - and we should be very happy that
>> the BUILD_BUG_ON() triggered an issue after years of silence.
>>
>> I do not have a good feeling about what kind of strange effects this
>> compiler issue might have in other code of other projects.
>>
>> So I would explicitly suggest NOT to change the af_can.c code to work
>> around this compiler issue.
>>
>> Let the gcc people fix their product and let them thank all of us for
>> detecting it.
> 
> I'm sure you'd be eligible for a full refund in case this was a bug in
> gcc. It is not. It's a pretty clear ABI requirement for (at least some
> flavors of) ARM:
> 
> https://stackoverflow.com/questions/43786747/struct-layout-in-apcs-gnu-abi
> 
> and more directly from the horse's mouth:
> 
> https://developer.arm.com/documentation/dui0067/d/arm-compiler-reference/c-and-c---implementation-details/structures--unions--enumerations--and-bitfields
> 
> Field alignment
> 
>      Structures are arranged with the first-named component at the lowest
> address. Fields are aligned as follows:
> 
>          A field with a char type is aligned to the next available byte.
> 
>          A field with a short type is aligned to the next even-addressed
> byte.
> 
>          Bitfield alignment depends on how the bitfield is declared. See
> Bitfields in packed structures for more information.
> 
>          All other types are aligned on word boundaries.
> 
> That anonymous union falls into the "All other types" bullet.
> 
> __packed is the documented and standard way to overrule the
> compiler's/ABI's layout decisions.

So why is there a difference between

gcc version 10.2.0

and

gcc version 10.2.1 20210110 (Debian 10.2.1-6)

https://lore.kernel.org/linux-can/20210323073437.yo63wreqnubbeqby@pengutronix.de/

??

Would this mean that either STRUCTURE_SIZE_BOUNDARY or the command line 
option -mstructure_size_boundary=<n>

are set differently?

https://stackoverflow.com/questions/43786747/struct-layout-in-apcs-gnu-abi/43829053#43829053

I'm not a compiler expert but this does not seem to be consistent.

Especially as we only have byte sizes (inside and outside of the union) 
and "A field with a char type is aligned to the next available byte."

The union is indeed aligned to the word boundary - but the following 
byte is not aligned to the next available byte.

Regards,
Oliver