https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117556
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Ever confirmed|0 |1
Last reconfirmed| |2024-11-13
Status|UNCONFIRMED |ASSIGNED
CC| |rsandifo at gcc dot gnu.org
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
So for aarch64 we use vector([2,2]) long int, with a group size of four this
means the division is pos = [0, 4] / nunits = [2, 2]. This is similar to
PR117558 where this kind of division is too naiive for VLA.
The desired result is of course vec_entry == 0, vec_index == 0 in this case
(slp_index is zero, num_scalar == group_size == 4). The VF is [1, 1] in
case this helps.
int num_scalar = SLP_TREE_LANES (slp_node);
int num_vec = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
poly_uint64 pos = (num_vec * nunits) - num_scalar + slp_index;
and num_vec is VF * SLP_TREE_LANES / nunits, so
pos = VF * num_scalar - num_scalar + slp_index;
where we know slp_index < num_scalar. For VF [1, 1] pos simplifies to
slp_index?
With single-lane SLP we end up doing load-lanes where I think the code
will be confused anyway - we have to look at the permute node. Fixing
that vectorizes the testcase as expected.
So mine.
