Re: Check that there are no missing probabilities

[Richard Biener <richard dot guenther at gmail dot com>]

On Fri, Oct 13, 2017 at 9:27 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Fri, Oct 13, 2017 at 09:06:55PM +0200, Jan Hubicka wrote:
>> For EH we should set it to profile_probability::zero () because we know it is unlikely
>> path.   I will take a look.
>
> With the
>
> --- gcc/cfghooks.c.jj   2017-10-13 18:27:12.000000000 +0200
> +++ gcc/cfghooks.c      2017-10-13 19:15:11.444650533 +0200
> @@ -162,6 +162,7 @@ verify_flow_info (void)
>               err = 1;
>             }
>           if (profile_status_for_fn (cfun) >= PROFILE_GUESSED
> +             && (e->flags & (EDGE_EH | EDGE_ABNORMAL | EDGE_FAKE)) == 0
>               && !e->probability.initialized_p ())
>             {
>               error ("Uninitialized probability of edge %i->%i", e->src->index,
>
> hack x86_64-linux and i686-linux bootstrapped fine, but I see still many
> graphite related regressions:
>
> /home/jakub/src/gcc/gcc/testsuite/gcc.dg/graphite/id-16.c:15:1: error: Uninitialized probability of edge 41->17
> /home/jakub/src/gcc/gcc/testsuite/gcc.dg/graphite/id-16.c:15:1: error: Uninitialized probability of edge 44->41
> /home/jakub/src/gcc/gcc/testsuite/gcc.dg/graphite/id-16.c:15:1: error: Uninitialized probability of edge 36->21
> /home/jakub/src/gcc/gcc/testsuite/gcc.dg/graphite/id-16.c:15:1: error: Uninitialized probability of edge 29->36
> /home/jakub/src/gcc/gcc/testsuite/gcc.dg/graphite/id-16.c:15:1: error: Uninitialized probability of edge 32->29
> during GIMPLE pass: graphite
> dump file: id-16.c.150t.graphite
> /home/jakub/src/gcc/gcc/testsuite/gcc.dg/graphite/id-16.c:15:1: internal compiler error: verify_flow_info failed
> 0xafac1a verify_flow_info()
>         ../../gcc/cfghooks.c:268
> 0xf2a624 checking_verify_flow_info
>         ../../gcc/cfghooks.h:198
> 0xf2a624 cleanup_tree_cfg_noloop
>         ../../gcc/tree-cfgcleanup.c:901
> 0xf2a624 cleanup_tree_cfg()
>         ../../gcc/tree-cfgcleanup.c:952
> 0x162df85 graphite_transform_loops()
>         ../../gcc/graphite.c:422
> 0x162f0c0 graphite_transforms
>         ../../gcc/graphite.c:447
> 0x162f0c0 execute
>         ../../gcc/graphite.c:524
>
> So probably graphite needs to be tweaked for this too.

graphite does

  if (changed)
    {
      cleanup_tree_cfg ();
      profile_status_for_fn (cfun) = PROFILE_ABSENT;
      release_recorded_exits (cfun);
      tree_estimate_probability (false);

so it runs into CFG cleanup running before it properly resets counts.

I wonder if we shouldn't simply get rid of the explicit checking calls in
cfg cleanup...  or if the profile checking should happen somewhere
else.

I'd also appreciate a better way for doing the above.  Shouldn't we
end up with a proper initialization on all edges as we just split
existing ones and use create_empty_if_region_on_edge and
create_empty_loop_on_edge?

Ah, those use make_edge as well.

The tree_estimate_probablility call above should be ideally
replaced with sth like "propagate-SESE-entry-probability".

Richard.

>         Jakub