[prev in list] [next in list] [prev in thread] [next in thread] 

List:       gcc-patches
Subject:    Re: Reduce startup cost of compiler (patch 2)
From:       Jan Hubicka <jh () suse ! cz>
Date:       2007-07-31 15:44:49
Message-ID: 20070731154449.GA11467 () kam ! mff ! cuni ! cz
[Download RAW message or body]

> On Tue, Jul 31, 2007 at 05:52:16AM +0200, Jan Hubicka wrote:
> > 
> > Hi,
> > the limitation was alwas there
> 
>    COSTS_N_INSNS (60000) = 240000 used to work fine. :-)

Well, it depends on definition of the word "work" :)
It a sense of not ICEing it works, in a sense of giving meaningful cost
model it does not.
> 
> > - regclass used value of 65535 as an
> > infinity for impossible combinations.  This also means that you should
> > not need to discougrate ALL_REGS and similar classes since regstack is
> > supposed to work this out.
> 
>    It's a good thing you say "supposed to". Regclass tends to prefer
> ALL_REGS over MEM and ALL_REGS tends to include such things as virtual
> hard regs. Returning a high REGISTER_MOVE_COST for ALL_REGS keeps regclass
> clear of ALL_REGS.

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?
> 
> >   if the @samp{mov@var{m}} pattern's constraints do not allow such copying.
> > + The maximal allowed value is 65535.
> >   @end defmac
> 
>    Well, it's 65534, isn't it?

Hmm, if we can find good reason for register move cost to trick the
infinity cost as you want to do, we can probably relax the assert and
document 65535 as infinity.

Honza
> 
> -- 
> Rask Ingemann Lambertsen
[prev in list] [next in list] [prev in thread] [next in thread]