This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Statistics infrastructure, updated
On Tue, May 13, 2008 at 2:32 AM, Richard Guenther <rguenther@suse.de> wrote:
> On Mon, 12 May 2008, Seongbae Park (???, ???) wrote:
>
> > I think it would be more useful if we convert
> > prop_stats::{num_const_prop, num_copy_prop, num_dce,num_pre_folded}
> > and pre_stats::{constified, pa_insert, eliminations} into debug counters
> > (pre_stats::insertions and phis seem to be purely statistics
> > still overhead of using debug counter is low enough that it probably
> > won't matter
> > to simply use it for statistics)
> > - that way, we can have both statistics and debugging aid.
> > Making debug counter dump per-pass-invocation statistics is straightforward,
> > and I can easily make it dump per-function statistics as well.
>
> If you cannot convert things like pre_stats::insertions (that would
> require the ability to dynamically register new counters) then I don't
> see how this would simplify the patch - it would only add to it.
>
> The patch tries to handle per-function statistics (what -fdump-*-*-stats
> does), per pass statistics per function (-fdump-statistics) and per
> compilation-unit statistics (-fdump-statistics-stats). Adding a helper
> to be able to break in gdb with a condition like curr_stat ("foo") == 5
> would be easy. I have never used (nor did I know of) the debug counter
> stuff, so I have no opinion on its usefulness.
>
> So, show us/me the patch you are proposing ;)
>
> Richard.
Attached is the (half-cooked) patch - it will currently only print
per-function-per-pass
but the rest is the matter of figuring out when to take the base
snapshot to compare against.
I've converted pre_stats to the proposed scheme as an example.
Seongbae