Re: The current proposal
Le mardi 24 février 2009 à 19:49 -0600, Corliss, George a écrit :
> In my opinion, policy-based design is FAR too heavy weight for the standard
> we are considering. This is not a language standard.
>
> Ideally, much of our standard will be implemented in hardware. I share the
> desire of Professor Kulisch that it be in hardware, although I do not share
> his view that the standard should require hardware. If hardware
> implementation is a goal, policy-based design is too much.
>
> Whatever we do must be SIMPLE to understand, to implement, and to use.
> Unfortunately, those three sometimes conflict.
Perhaps I shouldn't have used the word "policy". Call them whatever you
want, but the fact is that they are part of simple standards and they
are actually implemented in hardware. So it isn't as much as you think
it is.
Let's take an example close to us: the IEEE-754R standard. Consider the
floating-point addition. The standard actually mandates it to be
governed by six (!) orthogonal policies. The standard prefers to call
them "attributes", but conceptually they are no different from policies:
- rounding direction,
- target precision and range,
- behavior in case of overflow,
- behavior in case of underflow,
- behavior in case of inexact result,
- behavior in case of invalid operation.
So there isn't a single floating-point addition, there are tons of them,
and hardware is mostly happy with them.
I hope it's now a bit clearer what I meant by policies, and why I feel
it's a viable way of doing things. Perhaps we won't have to use them in
the end; but right now it is premature to say that they are incompatible
with a hardware implementation.
Best regards,
Guillaume