Re: RFA: fix gcc/20371

[Joern RENNECKE <joern dot rennecke at st dot com>]

Richard Henderson wrote:

>     It looks generally ok. I wish someone who knew about and/or cared
>
> about ms bitfield issues would double-check it.  But if you can't
> find such an individual, it's fine by me.
>  
Thanks for the review. I don't know anyone who specializes in this,
so I checked the patch in now.
Most of the issues really arose from unforseen interactions between MS 
bitfields
and the packed / aligned attributes, so that is mainly a matter of 
defining what this
should do.  The way I made packed bitfields interact with MS bitfields 
is basically
that packed bitfields are packed as if the MS bitfield insanity was not 
present, but
MS bitfields still adhere to their peculiar rules, treating a packed 
bitfield either as
part of a run if it fits in underlying type size and alignment, or as a 
non-bitfield if
it doesn't.
It would be posible to get higher packing if a packed bitfield was 
re-interpreted as
having a different underlying type size to agree with an following MS 
bitfield, and
where a packed bit field is not suitably aligned, pretend that only the 
latter part of it
forms the start of a run if it straddles a boundary.  It would also 
possible to include
packed non-bitfields in a run... but this would all greatly complicate 
the code.

If we want to allow users to mix packed and non-packed structure members 
in an
efficient way, it would be better to provide an attribute to switch off 
MS bitfield
rules for a given structure.  (IIRC there is already a port where this 
is controlled
with a machine dependent attribute).