[PATCH] Improve SH literal pool generation.

[Joern Rennecke]

> This had been suggested Joern Rennecke.
> http://gcc.gnu.org/ml/gcc/2003-04/msg01453.html
> and some of his earlier messages.
> Also see these comments in sh.c (add_constant)
> 2659 /* ??? This stuff should be done at the same time that we shorten
> branches.
> 2660    As it is now, we must assume that all branches are the maximum size,
> and
> 2661    this causes us to almost always output constant pools sooner than
> 2662    necessary.  */
> 
> The strategy is more or less as suggested in the link.

I was really thinking of doing a full-blown branch shortening at the
same time as constant pool allocation and branch splitting, also
putting the constants from branch spiltting into constant pools.
And all of that in machine-independent code.