https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116845
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |NEW
Summary|gcc.dg/pr109393.c test |gcc.dg/pr109393.c test
|fails |fails on ilp32 targets (and
| |maybe others)
Assignee|pinskia at gcc dot gnu.org |unassigned at gcc dot
gnu.org
--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
In the ILP32 case we have:
j.0_1 = (sizetype) j_11(D);
_2 = j.0_1 + 1073741823;
_3 = _2 * 4;
k_12 = j_11(D) + -1;
k.1_6 = (unsigned int) k_12;
_7 = k.1_6 * 4;
1073741823 is -1u/4
But the upper 2 bits of -1u/4 does not matter as they get pushed out by the *4.
so we need another canonicalization here.
(a + CST0) * powerof2 -> (a + CST0`) * CST1 sign extend the last N bits for
CST0.
I have other things going on so I am not going to implement this but this
should get someone started here. and it shows the problem of canonicalization
dealing with powerof2 multiplies and mod math :).
