Re: Optimising redundant compares

[Richard Earnshaw <rearnsha at arm dot com>]

> 
> 
> > -----Original Message-----
> > From: Jim Wilson [mailto:wilson@specifixinc.com] 
> > Sent: 26 September 2003 07:56
> > To: jbeniston@compxs.com
> > Cc: gcc@gcc.gnu.org
> > Subject: Re: Optimising redundant compares
> > 
> > 
> > Jon Beniston wrote:
> > > Yep, my port is very similar. The only apparent significant 
> > difference 
> > > is that I was using a fixed hard register, where as I 
> > believe the PPC 
> > > allows CC regs to be pseduos.
> > 
> > The PPC has 8 condition code registers.  Using pseudos lets 
> > us allocate 
> > values to the 8 available cc registers.  If you only have one, then 
> > using a hard register should be fine.
> 
> After playing around for a while, it would appear that if I use a hard
> register, then GCC will not attempt to optimise the compare
> instructions. However, if I use a pseudo it will, but it will also try
> lots more optimisations that would mean I have to write rules for
> spilling and restoring the CC reg which I don't want to do at this point
> in time. Does this sound correct?
> 

Quite believable.  GCC is generally reluctant to do any optimizations that 
will extend the life of a hard register.  The reason being that it can 
make register allocation impossible later on.

Maybe the rule could be relaxed in the case of a hard register that is 
never a target for allocation to pseudo registers (in that case you know 
that extending the life won't affect register allocation).  I'm not sure 
how easy that would be, but I think it would help in this case.

R.