Re: PING fwprop and pr/19653

[Bernd Schmidt <bernds_cb1 at t-online dot de>]

Paolo Bonzini wrote:
> > I found one instance of the problem in some leftover output files:
> >
> > @@ -8253,12 +8159,11 @@
> >         if cc jump L$L$1478;
> > -       R0 = B [P5+2] (Z);
> > -       cc =R0==0;
> > +       R1 = B [P5+2] (Z);
> > +       cc =R1==0;
> >         if !cc jump L$L$1478 (bp);
> >         R0 = B [P4+2] (Z);
> >         cc =R0==0;
> > -       R1 = 0 (X);
> >         R0 = 1 (X);
> >         if cc R0 =R1; /* movsicc-2a */
> >         jump.s L$L$1501;
> >
> > Before is fwprop, after is fwprop + expensive-cse.
> As a first guess, I would think that you are testing three 
> inter-basic-block CSE passes (path following in CSE, and the two GCSE 
> passes).  fwprop adds a second GCSE pass to mimic at least part of what 
> CSE path following did.  Testing fwprop + expensive-cse means getting 
> path following + two GCSE passes.

Ok.  As far as I remember, the first file was compiled with a 
fwprop-patched compiler, and in the second I just turned cse-skip-blocks 
and cse-follow-jumps back on.

> I made some experiments with a reduced testcase:

Thanks.  This is helpful and I'll look into it, but let me just quickly 
point out that I didn't mean to require that level of detail work.  If 
you can demonstrate, with numbers, that on a couple of targets 
cse-skip-blocks and cse-follow-jumps have no measurable effect after 
fwprop is applied, then I'm all in favour given the effects on compile-time.


Bernd