[Bug target/56200] New: queens benchmark is faster with -O0 than with any other optimization level

[abel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56200

             Bug #: 56200
           Summary: queens benchmark is faster with -O0 than with any
                    other optimization level
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: abel@gcc.gnu.org
                CC: amonakov@gcc.gnu.org
            Target: x86_64-unknown-linux-gnu


Created attachment 29346
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29346
queens benchmark

While preparing some teaching slides I have noticed that the queens benchmark
from Aburto generates slower code with -O1/O2/O3 than with plain -O0.  I've
removed all timing routines from the attached file, compile and run it with
./a.out 30.

For 4.7.1 of OpenSUSE 12.2, we have:

$ time ./o0 30 >/dev/null
user    0m10.017s

$ time ./o1 30 >/dev/null
user    0m11.824s

$ time ./o2 30 >/dev/null
user    0m10.388s

$ time ./o3 30 >/dev/null  //this one is with -march=native
user    0m11.065s

For today's 4.8, 

$ time ./o048 30 >/dev/null
user    0m9.780s

$ time ./o148 30 >/dev/null
user    0m12.590s

$ time ./o248 30 >/dev/null //this one is with -march=native
user    0m10.487s

$ time ./o348 30 >/dev/null //this one is with -march=native
user    0m10.850s

Perf shows that in the -O1 versions of above we have the new lea that has a lot
of counter hits and lots more of branch mispredictions.