[Bug tree-optimization/111519] [13/14 Regression] Wrong code at -O3 on x86_64-linux-gnu since r13-455-g1fe04c497d

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111519

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Reverting

@@ -5089,8 +5123,9 @@ strlen_pass::handle_store (bool *zero_write)
          return false;
        }

-      if (storing_all_zeros_p
-         || storing_nonzero_p
+      if (storing_nonzero_p
+         || storing_all_zeros_p
+         || (full_string_p && lenrange[1] == 0)
          || (offset != 0 && store_before_nul[1] > 0))
        {
          /* When STORING_NONZERO_P, we know that the string will start

fixes the issue.  We somehow compute a lenrange of { 0, 0, 1 } for
d = _8 but also set storing_all_zeros_p to false.

It seems that strlen_pass::count_nonzero_bytes happily skips intermediate
stores when it handles SSA names here:

      gimple *stmt = SSA_NAME_DEF_STMT (exp);
      if (gimple_assign_single_p (stmt))
        {
          exp = gimple_assign_rhs1 (stmt);
          if (!DECL_P (exp)
              && TREE_CODE (exp) != CONSTRUCTOR
              && TREE_CODE (exp) != MEM_REF)
            return false;
          /* Handle DECLs, CONSTRUCTOR and MEM_REF below.  */

we have

    _8 = *_7;
...
    *_7 = prephitmp_89;
    d = _8;

when we recursively process a memory reference we need to verify there
are no intermediate aliases, the simplest thing would be to pass down
a virtual use.  Trying some prototype.