[Bug middle-end/35545] tracer pass is run too late

["xinliangli at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=35545

--- Comment #19 from davidxl <xinliangli at gmail dot com> ---
(In reply to Jan Hubicka from comment #18)
> > WE can try some internal benchmarks with this change too.
> 
> That would be very welcome.  Tracer used to be quite useful pass in old days,
> doing 1.6% on -O3+FDO SPECint (for 1.4% code size cost) that put it very
> close
> to the inliner (1.8%) that however needed 12.9% of code size.
> 
> http://www.ucw.cz/~hubicka/papers/amd64/node4.html (see aggressive
> optimization table)
> 
> I do not think it was ever resonably returned for GIMPLE. Martin, do you have
> any current scores? Part of benefits came from lack of global optimizers, but
> I think tail duplication has a potential in modern compiler, too.
> 
> Perhaps we should try to do that and given that we now have tail merging
> pass,
> consider getting it useful without FDO, too.
> 
> Honza

I saw small improvement across most of the benchmarks with this change on
sandybridge machines with FDO. The geomean improvement is about 0.3%. The
baseline compiler is Google gcc-49 compiler, and the test compiler is Google
gcc-49 with the patch.