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

[ysrumyan at gmail dot com]

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

Yuri Rumyantsev <ysrumyan at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ysrumyan at gmail dot com

--- Comment #7 from Yuri Rumyantsev <ysrumyan at gmail dot com> 2013-02-06 14:04:36 UTC ---
Here is my 2 cents.

I assume that we must not put 2 branches to the same target into the same fetch
line since x86 branch predictor uses IP without last 4 bits as hash function.
It means that 2 consequtive branches to the same targets have the same
prediction history, i.e. they both either taken or not-taken but we can suppose
that for
   if (cond1 || cond2)
     goto L;

one of condition can be true with more than 50% probability. If we are at the
2nd branch (i.e. 1st branch was not-taken) we must assume that it will be taken
and it must be put to another fetch line.