https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108358
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Both -O2 and -O1 can optimize away the call to foo. For -Os we end up with
<bb 2> [local count: 118111600]:
l = *.LC0;
goto <bb 4>; [100.00%]
<bb 3> [local count: 955630225]:
bar31_ ();
i.0_1 = i;
_2 = i.0_1 + 1;
i = _2;
<bb 4> [local count: 1073741824]:
i.1_3 = i;
if (i.1_3 != 0)
goto <bb 3>; [89.00%]
else
goto <bb 5>; [11.00%]
<bb 5> [local count: 118111600]:
j = l;
m$h$b_18 = j.h.b;
if (m$h$b_18 != 0)
goto <bb 7>; [67.00%]
else
goto <bb 6>; [33.00%]
<bb 6> [local count: 38976828]:
foo ();
but FREs looking through the j = l copy causes us to fail disambiguating
against the cycle because of 'abort_on_visited' which I think is a bit
overeager here.