[Bug ipa/102581] [12 Regression] ice in forced_merge, at ipa-modref-tree.h:352 with -fno-strict-aliasing and -O2 since r12-3202-gf5ff3a8ed4ca9173

["hubicka at ucw dot cz" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102581

--- Comment #7 from Jan Hubicka <hubicka at ucw dot cz> ---
Hi,
the problem is that we assume that merge is symmetric (merging a to b
succeeds if and only if merging b to a succeeds). There was one
symetrical path missing in the (fancy and bit ugly) logic on what we can
merge losslessly.

I also added sanity check for this assumption.  So bugs like this should
fire more often.

Honza

diff --git a/gcc/ipa-modref-tree.h b/gcc/ipa-modref-tree.h
index 6a9ed5ce54b..61b4b16dbed 100644
--- a/gcc/ipa-modref-tree.h
+++ b/gcc/ipa-modref-tree.h
@@ -234,6 +234,17 @@ struct GTY(()) modref_access_node
                      record_adjustments);
              return true;
            }
+         if (known_size_p (a.size)
+             && (!known_size_p (size) || known_lt (size, a.size)))
+           {
+             if (((known_size_p (max_size) || known_size_p (a.max_size))
+                  && !known_eq (max_size, a.max_size))
+                  || !known_eq (offset1, aoffset1))
+               return false;
+             update (new_parm_offset, offset1, size, max_size,
+                     record_adjustments);
+             return true;
+           }
          /* If sizes are same, we can extend the interval.  */
          if ((known_size_p (size) || known_size_p (a.size))
              && !known_eq (size, a.size))
@@ -618,6 +629,7 @@ private:
            found = true;
          if (!found && n->merge (*a, false))
            found = restart = true;
+         gcc_checking_assert (found || !a->merge (*n, false));
          if (found)
            {
              accesses->unordered_remove (i);