https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115895
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
Last reconfirmed| |2024-07-12
Ever confirmed|0 |1
Status|UNCONFIRMED |ASSIGNED
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
<bb 3> [local count: 505088130]:
# vectp_sp.25_67 = PHI <vectp_sp.25_68(5), sp_10(D)(2)>
# vectp_mprr_2.31_75 = PHI <vectp_mprr_2.31_76(5), &mprr_2(2)>
# ivtmp_79 = PHI <ivtmp_80(5), 15(2)>
# loop_mask_69 = PHI <_82(5), { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1, -1, -1 }(2)>
# loop_mask_77 = PHI <_84(5), { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1 }(2)>
vect__5.27_70 = .MASK_LOAD (vectp_sp.25_67, 32B, loop_mask_69);
vectp_sp.25_71 = vectp_sp.25_67 + 18446744073709551568;
vect__5.28_72 = VEC_PERM_EXPR <vect__5.27_70, vect__5.27_70, { 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }>;
vect__5.29_73 = VEC_PERM_EXPR <vect__5.27_70, vect__5.27_70, { 2, 2, 2, 2, 2,
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2 }>;
vect__6.30_74 = VEC_PACK_TRUNC_EXPR <vect__5.28_72, vect__5.29_73>;
.MASK_STORE (vectp_mprr_2.31_75, 512B, loop_mask_77, vect__6.30_74);
So indeed here the .MASK_LOAD overreads. We are still peeling for gaps, but
not enough.
In theory we should be able to adjust the mask - that might be cheaper than
shortening the access, which is what we do without masking since masking off
the gap would also allow to avoid peeling for gaps(?)
<bb 3> [local count: 441952113]:
# vectp_sp.25_70 = PHI <vectp_sp.25_71(5), sp_10(D)(2)>
# vectp_mprr_2.31_79 = PHI <vectp_mprr_2.31_80(5), &mprr_2(2)>
# ivtmp_82 = PHI <ivtmp_83(5), 0(2)>
_72 = MEM <uint128_t> [(int *)vectp_sp.25_70];
_73 = {_72, 0, 0, 0};
vect__5.27_74 = VIEW_CONVERT_EXPR<vector(16) int>(_73);
I'll leave this open for a bit - the easiest fix is to disable partial
vectors when we do the gap optimization and not perform it for variable-length
vectors (I'm not sure all the may/must checks are on the safe side). I think
we want to use smaller vectors for the load in a more concious way, possibly
the SLP work of lowering vectors will help here.