[Bug gcov-profile/111559] [14 regression] ICE when building Python with PGO

["slyfox at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111559

--- Comment #6 from Sergei Trofimovich <slyfox at gcc dot gnu.org> ---
Uninitialized value comes from `ipa_merge_profiles()` for our `rule1_same()`
alias and `rule1()` functions:

  // in gcc/ipa-icf.cc:

  else if (create_alias)
    {
      alias->icf_merged = true;

      /* Remove the function's body.  */
      ipa_merge_profiles (original, alias);
      // ...

If I comment out `ipa_merge_profiles (original, alias);` call to leave
`original` as is then failure does not happen. Which means at least
`original`'s profile is fine.

Tracing through `ipa_merge_profiles()` we generate uninitialized probalility
profile when divide by zero at:

  // in gcc/ipa-utils.cc

void
ipa_merge_profiles (struct cgraph_node *dst,
                    struct cgraph_node *src,
                    bool preserve_body)
    // ...

      /* TODO: merge also statement histograms.  */
      FOR_ALL_BB_FN (srcbb, srccfun)
        {
          unsigned int i;

          if (copy_counts) { /* snip: ireelevant */ }
          else
            {
              for (i = 0; i < EDGE_COUNT (srcbb->succs); i++)
                {
                  edge srce = EDGE_SUCC (srcbb, i);
                  edge dste = EDGE_SUCC (dstbb, i);
                  dste->probability =
                    dste->probability * dstbb->count.ipa ().probability_in
                                                 (dstbb->count.ipa ()
                                                  + srccount.ipa ())
                    + srce->probability * srcbb->count.ipa ().probability_in
                                                 (dstbb->count.ipa ()
                                                  + srccount.ipa ());
                }
              dstbb->count = dstbb->count.ipa () + srccount.ipa ();
            }
        }
    // ...

Here `dstbb->count.ipa () + srccount.ipa ()` is zero.

This assert should expose it as well:

--- a/gcc/ipa-utils.cc
+++ b/gcc/ipa-utils.cc
@@ -651,13 +651,15 @@ ipa_merge_profiles (struct cgraph_node *dst,
                {
                  edge srce = EDGE_SUCC (srcbb, i);
                  edge dste = EDGE_SUCC (dstbb, i);
+
+                 profile_count den = dstbb->count.ipa () + srccount.ipa ();
+                 gcc_assert(den.nonzero_p());
+
                  dste->probability =
                    dste->probability * dstbb->count.ipa ().probability_in
-                                                (dstbb->count.ipa ()
-                                                 + srccount.ipa ())
+                                                (den)
                    + srce->probability * srcbb->count.ipa ().probability_in
-                                                (dstbb->count.ipa ()
-                                                 + srccount.ipa ());
+                                                (den);
                }
              dstbb->count = dstbb->count.ipa () + srccount.ipa ();
            }

If we attach `gdb` it agrees we exercise these edges 0 times.

    (gdb) call dstbb->count.debug()
    0 (precise)
    (gdb) call srccount.ipa ().debug()
    0 (precise)

For comparison we are trying to clobber `always` probability with `undefined`:

    (gdb) call dste->probability.debug()
    always

What edge is that?

    (gdb) call debug_edge(srce)
    edge (bb_3, bb_4)

__attribute__((noinline))
void rule1 ()
{
  int p.0_1;

  <bb 2> [count: 2]:
  p.0_1 = p;
  if (p.0_1 != 0)
    goto <bb 3>; [0.00%]
  else
    goto <bb 4>; [100.00%]

  <bb 3> [count: 0]:
  edge ();

  <bb 4> [count: 2]:
  return;

}

`always` should valid for `bb_3->bb_4`. But for our data input it's `never`.