[Bug tree-optimization/101419] [9/10/11/12 Regression] collapsing memset() calls can break __builtin_object_size()

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101419

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
--- gcc/tree-object-size.c.jj   2021-01-04 10:25:39.911221618 +0100
+++ gcc/tree-object-size.c      2021-07-12 11:28:51.328120222 +0200
@@ -1393,6 +1393,11 @@ pass_object_sizes::execute (function *fu
                        {
                          tree tem = make_ssa_name (type);
                          gimple_call_set_lhs (call, tem);
+                         if (object_size_type == 1)
+                           {
+                             ost = build_zero_cst (TREE_TYPE (ost));
+                             gimple_call_set_arg (call, 1, ost);
+                           }
                          enum tree_code code
                            = object_size_type == 1 ? MIN_EXPR : MAX_EXPR;
                          tree cst = build_int_cstu (type, bytes);
works, but unfortunately only for __builtin_object_size (, 1).
When I did that also for 3 (to turn it into 2), it breaks several
gcc.dg/builtin-object-size*.c tests.
0 is defined as an upper bound of the object size, subobjects not taken into
account
1 is defined similarly, but subobjects are taken into account
2 is defined as a lower bound of the object size, subobjects not taken into
account
3 is 2 with subobjects
-D_FORTIFY_SOURCE={1,2} typically only uses 0 and 1 modes, 2 and 3 are rare,
and are useful e.g. when somebody wants to find out if the object size is exact
or not (for exact case the upper and lower bounds are equal).