Re: Reduce startup cost of compiler (patch 2)

[Michael Matz <matz@suse.de>]

Hi,

On Fri, 3 Aug 2007, Rask Ingemann Lambertsen wrote:

> On Tue, Jul 31, 2007 at 05:44:49PM +0200, Jan Hubicka wrote:
> 
> > Well, regclass is not able to work out all details (it always assumes
> > worst case for union classes), but in your case it seems you got
> > something wrong.
> > ALL_REGS even if it contains hard regs such as stack pointers and
> > virtual frame pointers, they are fixed and should not cause much harm to
> > you. Can you send a bit more details how your classes/costs are set up?
> 
>    The comment I have for REGISTER_MOVE_COST says: "Discourage regclass.c
> from using ALL_REGS as alternate class." So the issue is not about ALL_REGS
> as the preferred class but that regclass would choose e.g. GENERAL_REGS as
> preferred class and ALL_REGS as alternate class where it should choose MEM
> as alternate class.

Well, the question was why do you think it was bad for regclass to choose 
ALL_REGS as alternate class?  It may have been not optimal perhaps, but 
from the thread it sounds as if it was a correctness issue for you.  If 
that is the case then this means you have a real bug somewhere else, as 
regclass theoretically could choose random classes as preferred and 
alternate class, and everything still works (with many reloads).  It's 
only an optimization hint, not a correctness measure.

OTOH if it was really only about optimization (i.e. choosing ALL_REGS is 
a bad decision) then adjusting the move cost was correct, but you didn't 
really have to use 65535 :-)  It would have been enough to make the cost 
higher than for memory moves.


Ciao,
Michael.