Re: [DVB patch 17/51] flexcop: add big endian register definitions

[Johannes Stezenbach <js@linuxtv.org>]

Andrew Morton wrote:
> Johannes Stezenbach <js@linuxtv.org> wrote:
> >
> > From: Patrick Boettcher <pb@linuxtv.org>
> > 
> > Add big-endian register definitions for running on a PowerPC.
> > (Thanks to Paavo Hartikainen for testing.)
> > 
> > ...
> > +	struct {
> > +		u32 dma_address0                   :30;
> > +		u32 dma_0No_update                 : 1;
> > +		u32 dma_0start                     : 1;
> > +	} dma_0x0;
> >...
> > +
> > +	struct {
> > +		u32 dma_0start                     : 1;
> > +		u32 dma_0No_update                 : 1;
> > +		u32 dma_address0                   :30;
> > +	} dma_0x0;
> 
> Oh dear.  This is a good demonstration of the downside of trying to use
> compiler bitfields to represent hardware registers.  I have vague memories
> of writing BFINS and BFEXT in 3c59x to stomp this problem.
> 
> I don't think there's any guarantee that the code you have there will work
> on all architectures/compiler versions btw.
> 
> Also...  The code appears to be assuming that BE architectures will
> bit-reverse their bitfields.  Is that right?  I'd expect them to only
> byte-reverse them?

Probably the code should use __BIG_ENDIAN_BITFIELD /
__LITTLE_ENDIAN_BITFIELD instead of __BIG_ENDIAN / __LITTLE_ENDIAN?
Anyway, the comment from the CVS commit suggests that it was tested.

I completely agree that this code is ugly as hell.
It was the obvious, simple fix to make the driver work on
PowerPC (and a few users happy), though. Rewriting the
driver not to use bitfields seems to be quite a bit of work.
Blame me for not paying enough attention when the initial
flexcop driver was submitted ;-(

Johannes