https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122271
Jeffrey A. Law <law at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Last reconfirmed| |2025-12-08
CC| |dbarboza at ventanamicro dot
com
Status|UNCONFIRMED |NEW
--- Comment #2 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Lots of places this could probably be implemented. In the first case we want
to realize that we have a PHI feeding a comparison:
# c_1 = PHI <a_3(D)(2), 0B(3)>
_7 = c_1 != 0B;
We don't really care about the value in a_3, just whether it is nonzero in both
paths. I would expect ranger to globally record its range as being nonzero.
In second case:
b_3 = *a_2(D);
tt_5 = b_3 != 0;
c_6 = tt_5 ? a_2(D) : 0B;
tt_7 = c_6 != 0B;
Same basically applies there, we just happened to have already optimized away
the PHI.
I think if we got the first case if-converted into the second form, then we
could focus on exploiting the nonzero status of "a" to simplify further rather
than trying to handle it in two different paths.
