Re: register hint usefulness

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

    Local spills are absolutely necessary for correctness in that example. 

Yes, if we decide we need to support that: in the past, we've just
simply said we don't, which seems reasonable to me.

    Of necessity I would assume local spilling is always going to be more
    expensive than global spilling.  However, that's not to say that we
    should not work to make things more efficient than they currently are.

Well, order_regs_for_reload loops through all pseudos.  Since both the
number of reloads and pseudos is a linear function of the size of
a function, this is quadratic.  It dominates the compilation time of
large functions, most notably with -O0, but even with -O2.

I doubt this sort of optimization ever makes sense, but certainly not at -O0!