Re: Compiler slowdowns (was Re: [PATCH] Optimize manual byte swap implementations v2)

[Vladimir Makarov <vmakarov at redhat dot com>]

Paolo Bonzini wrote:
> Bradley Lucier wrote:
>   
> > OK.
> >
> > Well, I'm seeing an -O1 compile (with some extra passes thrown in, like
> > -fschedule-insn2) go from 168 seconds to 228.17 seconds in four months. 
> > Without a nail to pound down.
> >     
>
> The obvious culprit here is IRA, but I may be wrong.
>
>   
I believe you are wrong.  I've just check trunk right before removing 
the old RA on x86_64 native and x86_64->ppc64 cross toolchain built with 
--enable-checking=release.  I got the following gcc compilation times on 
the test

       old       IRA
native  108.77s   103.19s
ppc     160.90s   160.69s
> > That's not directly related to -O2 times or whether bswap optimization
> > is worth 1% compile time, but I'm guessing that a number of similar
> > discussions went on (at least internally) about other patches over those
> > four months, and we've ended up with a big difference.  (In the PR I
> > noted a 13% difference just between September 19 and September 26 of
> > last year.)
> >     
>
> I think if you can nail down the day in that week and file a separate
> PR, something *will* be done.
>
>   
> > I don't ever quite remember seeing such a rate of increase in compile
> > time while I've been watching gcc development.
> >     
>
> I agree on this.
>   
It might be true but I would not do such conclusion on this extreme test 
which is quite different from most programs.