[Gzz] Cg, preliminary analysis

[Tuomas Lukka]

I've been looking at Cg tonight, as a front-end for ARB vertex programs;
some positive and negative findings:

        + Being able to use expressions and many temporaries
          and have the Cg compiler do the register assignments
          is great. Life's much easier.

        + Library: e.g. cos() implemented there.

        + Produces relatively good code for most cases
        
        - The compiler is buggy; a very simple program failed; luckily,
          the bug was such that I could work around it by munging
          the resulting text

        - Of course, the proprietarity is not good; even the arbvp1
          back-end is not public. So whatever happens with bugs, 
          we won't be using this as our platform unless that changes.
        
        - The conversion/cast rules are sometimes a little too strict.

We do want some system that can at least do register assigning,
but I'm not quite sure which level of abstraction we want.

        Tuomas