Re: [PATCH] add statistics counting to postreload, copy-rename, and math-opts

[Richard Guenther <richard.guenther@gmail.com>]

On Wed, Apr 13, 2011 at 8:43 PM, Nathan Froyd <froydnj@codesourcery.com> wrote:
> On Wed, Apr 13, 2011 at 11:07:15AM +0200, Richard Guenther wrote:
>> On Tue, Apr 12, 2011 at 5:09 PM, Nathan Froyd <froydnj@codesourcery.com> wrote:
>> > Granted, but that fact should still be recorded.  The situation we have
>> > today, for something like:
>> >
>> > func1: statistic for "statx" was 0
>> >  - nothing is recorded in the statistics table
>> > func2: statistic for "statx" was 0
>> >  - nothing is recorded in the statistics table
>> > func3: statistic for "statx" was 0
>> >  - nothing is recorded in the statistics table
>> > ...
>> >
>> > and so forth, is that at the end of the day, the dump file won't even
>> > include any information about "statx".  If you had some func7387 where
>> > "statx" was non-zero, you could infer that nothing else happened in the
>> > previous 7386 functions.  For the case where a pass is truly useless on
>> > a TU, it's hard to figure out from the statistics dump alone.  And I'd
>> > argue that it's useful to see explicitly that the pass only helped in 1
>> > out of 7387 functions, rather than trying to infer it from missing data.
>>
>> I always use statistics-stats (thus, overall stats, not per function).  The
>> per function ones omit zero counts during dumping on purpose
>> (to make the dump smaller).
>
> I didn't know about statistics-stats (or didn't realize that's what the
> code was trying to do), that's useful.  And it looks like all the
> statistics dumping things omit zero counts on purpose, not just the
> per-function ones.
>
> But that has no bearing on the point above: zero counts are not even
> *recorded* today.  E.g. if you apply the patch upthread, grab a random C
> file, compile it with -O2/3 -fdump-statistics/-stats, and examine the
> dump file, you might not even know that new statistics counters have
> been added.  Taking out the checks to avoid printing zero counts doesn't
> help either, because the data simply doesn't get recorded.  This
> infrastructure makes it somewhat difficult to figure out, in an
> automated way from the dump file alone, whether passes are actually
> doing anything.

Oh, ok - now I understand what you are saying.  Yes, that's by design
and hard to fix - counters are "registered" on their first bump.

> Enough grousing.  I'm assuming turning on accumulation and dumping of
> zero counts always would be frowned upon; would it be acceptable to turn
> accumulation and dumping of zero counts if -details is given?

Well, you'd have to fix the registration problem ...

But I don't see zero counts as very important, if not to verify that the
statistics code is placed at a good spot.

Richard.

> -Nathan
>