> : Note, the patch to softfloat.h is to show you what needs to be done.  It 
> : should really be dressed up in #ifdefs to check the compiler version is at 
> : least 2.8 (so that 2.7 won't emit warnings on the unrecognized attribute).
> 
> Actually, I have one better - we assert the presence of format(__kprintf__),
> a native-gcc hack, by defining __KPRINTF_ATTRIBUTE__ in the cpp spec.  
> Something like that would be useful here.
> 

I'm not convinced this is the correct approach.  It is not so bad for 
kprintf, because if the attribute is not available, all that happens is 
that you loose potential warnings.  But for return_in_regs, the way your 
code gets compiled will change.  Personally, I think it would be a bad 
idea to deliberately go in a direction where the only compiler that can 
ever compile the kernel & other netbsd sources is the one maintained by 
members of this list (For a start, it would make starting new ports more 
difficult).

Secondly, I don't like littering CPP with gazillions of pre-defined 
pre-processor options describing the compiler's capability.  It would be 
much cleaner if this could just be worked out where needed from the 
standard definitions (compiler major and minor release numbers).