[Bug c++/69564] [5/6 Regression] lto and/or C++ make scimark2 LU slower

law at redhat dot com gcc-bugzilla@gcc.gnu.org
Fri Apr 1 04:32:00 GMT 2016


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69564

Jeffrey A. Law <law at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |law at redhat dot com

--- Comment #26 from Jeffrey A. Law <law at redhat dot com> ---
In response to c#24.  Coalescing's primary sort key is highly dependent on the
edge frequencies -- which is generally what we want.

The highest edge frequency represents the highest cost for any particular copy
implied by a PHI node.

This obviously falls down if there are > 1 lower frequency copies that can't be
coalesced due to a higher frequency copy that does coalesce and the sum of the
lower frequency copy costs are greater than the sum of the single higher
frequency copy.

One might be able to model that by looking at the conflict lists, but before
going down that path, it'd obviously be best to know if that scenario is the
root cause of the slowdown.


More information about the Gcc-bugs mailing list