https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81088
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
And the issue is that
associate:
...
/* Only do something if we found more than two objects. Otherwise,
nothing has changed and we risk infinite recursion. */
if (ok
&& (2 < ((var0 != 0) + (var1 != 0)
+ (con0 != 0) + (con1 != 0)
+ (lit0 != 0) + (lit1 != 0)
+ (minus_lit0 != 0) + (minus_lit1 != 0))))
{
bool any_overflows = false;
if (lit0) any_overflows |= TREE_OVERFLOW (lit0);
if (lit1) any_overflows |= TREE_OVERFLOW (lit1);
...
/* Don't introduce overflows through reassociation. */
if (!any_overflows
&& ((lit0 && TREE_OVERFLOW_P (lit0))
|| (minus_lit0 && TREE_OVERFLOW_P (minus_lit0))))
return NULL_TREE;
fails to see the overflow when associating
(1(OVF) - (int) s) + 2147483647
the 1(OVF) is from (int)0xffffffffU which is implementation defined but
IIRC the FEs require the TREE_OVERFLOW flag to be set.
IMHO this any_overflows code is simply bogus. To avoid dropping useful
association split_tree should drop pre-existing TREE_OVERFLOW.
