This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Merging mem-ssa into mainline


On 12/6/06, Diego Novillo <dnovillo@redhat.com> wrote:
Daniel Berlin wrote on 12/06/06 19:17:

> Uh, 119145 was before the rechanging of the alias analysis stuff.
>
The only effect of your new PTA changes is that I should now be able to
test PR15855 again (notice that it wasn't included in these numbers).

The speedups in PTA are only a side-effect of the other aliasing
speedups I have.  As I said, many other passes got faster because
changes in operand scanner and SSA renaming are much more pervasive than
changes to a single pass such as PTA.

I am not saying the compile time speedups are not compelling.  I'm not
saying it has no runtime performance impact.

I am simply  saying the changes i made to PTA have a serious affect on
the may-alias sets used to represent variables in all cases.  In
particular, i had reverted the change that made us link together every
nonlocal store and load (which in C++, is basically all of them :P).

If you look at something like
http://www.suse.de/~gcctest/c++bench/tramp3d/split-run.html
you'll notice the *runtime* (not compile time) improved from 27s to
22s (a 22% improvement) as a result.
This is why i said you may want to retest your *runtime* performance
numbers to make sure they are still good in the face of this change.

That is all.
Nothing more, nothing less.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]