Re: kernel coding style for if ... else which cross #ifdef

[Jeremy Fitzhardinge <jeremy@goop.org>]

H. Peter Anvin wrote:
> That's a very defeatist stance, and quite frankly bogus.

But <stamp foot> coding is *hard*.

> Doing it as a flag day event is not really practical, which is why we 
> need a new set of symbols.  However, at that point we can discourage 
> continuing use of the CONFIG_ symbols and deprecate them over time. 
> It's not like we're talking about user-space-visible interfaces here!

Well, I'm thinking more along the lines of:

   1. We introduce this <whatever> mechanism
   2. Hundreds of people pop out of the woodwork thinking "this looks
      more fun than tweaking whitespace"
   3. They produce one-hundred trillion "convert #ifdef to if()" patches
   4. We have one-hundred trillion^2 followup "fix build with this
      .config" patches

3 might be enough to finally drive Andrew out of the kernel business, 
but 4 definitely would be.

The whole point is to try and get config-invariant build breakages, so 
that we become less dependent on lots of randconfig testing.  If the 
definedness of the KCONFIG_ constants is still dependent on a particular 
.config, then we're not really making all that much progress.

If we move to having a single unified kernel config rather than per-arch 
ones (as Sam mentioned), then we can be sure of generating a complete 
list of all config variables, and we can explicitly define them.  But if 
we don't move to that state more or less simultaneously with using 
KCONFIG_ constants, then we should do it in the defeatest way so we can 
make forward progress with minimal regression.

    J