Re: Regclass tweak 3

[Jan Hubicka <hubicka at atrey dot karlin dot mff dot cuni dot cz>]

>   > When source does not die, this is loss IMO, because two colidding registers
>   > end up in preferrence for single class, thats catastrofic for single
>   > register classes in i386, and wrong for other cases anyway (because move
>   > is not saved).
> The code in question is only used when the preferred class has a single
> register element.  So your logic is correct in this particular circumstance.
> 
> However, your logic is not correct in the general case since most of the
> time register classes have more than one element.

Are you sure that this is a win?  We emit copy insn in any case (because we
have two colliding registers, so we must copy the source) Your point is
probably that reg-reg copy is cheaper when it is done inside single class, but
this is not true in all cases and ought to be modelled well by the constraints
that ought to result in correct costs out of scan_one_insn.

If I am missing something here, please try to give me some example of code.

Thanks..
Honza