https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98499
--- Comment #3 from Sergei Trofimovich <slyfox at gcc dot gnu.org> ---
`--param=modref-max-depth=0` makes the bug disappear.
Looking at `-fdump-tree-all` the harmful optimization happens at `107.fre3`
where:
```
if (&MEM[(struct string *)&D.2237 + 48B]._M_local_buf != _17)
goto <bb 3>; [0.00%]
else
goto <bb 4>; [100.00%]
<bb 3> [count: 0]:
__builtin_trap ();
```
is converted into an unconditional `__builtin_trap ();`. `-fdump-tree-all-all`
says it's a constant fold:
```
Value numbering stmt = if (&MEM[(struct string *)&D.2237 + 48B]._M_local_buf !=
_17)
Applying pattern match.pd:4920, gimple-match.c:3373
marking known outgoing edge 2 -> 3 executable
Block 1: BB4 found not executable
...
```
It looks like it's constant-folded into always-true instead of expected
always-false. `match.pd:4920` is a `pta`:
```
/* Simplify pointer equality compares using PTA. */
(for neeq (ne eq)
(simplify
(neeq @0 @1)
(if (POINTER_TYPE_P (TREE_TYPE (@0))
&& ptrs_compare_unequal (@0, @1))
{ constant_boolean_node (neeq != EQ_EXPR, type); })))
```
`036t.ealias` and `043t.modref1` says `--param=modref-max-depth=0` has the
effect on how many things escape from main(), but I don't understand the
output.