https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91272
Bug ID: 91272
Summary: [SVE] Use fully-masked loops for CLASTB reductions
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
Target: aarch64-linux-gnu
Tests like clastb_6.c show that we don't yet support CLASTB
reductions in fully-masked/predicated loops. E.g.: the main
loop is:
.L3:
ld1w z1.s, p0/z, [x0, x2, lsl 2]
addpl x1, x2, #2
incw x2
fcmlt p1.s, p0/z, z1.s, z3.s
cmp w2, w3
clastb s0, p1, s0, z1.s
bls .L3
This loop operates on full vectors only and relies on a scalar
loop to handle the rest.
We should instead support fully-masked loops by ANDing the
comparison result in a CLASTB reduction with the loop mask.
I think this means:
* making vectorizable_condition apply vect_get_loop_mask
for reductions. (There might be cases we want to do this
for normal conditions as well as for reductions, but that's
separate work).
* relaxing the LOOP_VINFO_CAN_FULLY_MASK_P handling in
vectorizable_reduction to account for the above.
