ADDRESS_COST cleanups

[Jan Hubicka]

> Jan Hubicka writes:
> 
>  > I've rewrite it to function using decompose_address, but I've hit problem in
>  > cse, that calls address_cost even for badly formed addresses.  It looks like
>  > calling of problems here (also because currently cse will ignore correctly
>  > formed ones when the badly formed one appears cheaper), so I've made new
>  > function address_cost that calls ADDRESS_COST only for correct addresses and
>  > gives high cost to the incorrect ones.
> 
> On many targets the address cost is dependent upon the choice of the
> other operands within the instruction.  I wonder if we should add an
> additional piece of context, such as the current insn, when
> ADDRESS_COST is called?
OK, I don't think this is actually problem for CSE, and I believe this
information is known in loop as well, so I will make a patch.  This will
require changes to all machine descriptions, so it will be quite long, can
we do that independently on this bit?

Honza
> 
> Michael.