https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115528
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rsandifo at gcc dot gnu.org
--- Comment #22 from Richard Biener <rguenth at gcc dot gnu.org> ---
Creating dr for aa[_13]
analyze_innermost: success.
base_address: &aa
offset from base address: (ssizetype) ((sizetype) i_35 * 8)
constant offset from base address: -8
step: 32
base alignment: 16
base misalignment: 0
offset alignment: 8
step alignment: 32
base_object: aa
Access function 0: {{0, +, 1}_1, +, 4}_2
but somehow we computed dr_info->misalignment of zero and thus we classify
this as dr_aligned. Previously the gap code disqualified SLP use and
avoided this latent issue.
t.f:14:12: note: inner step divides the vector alignment.
t.f:14:12: missed: misalign = 0 bytes of ref aa[_13]
it's a bit odd - vect_compute_data_ref_alignment does
innermost_loop_behavior *drb = vect_dr_behavior (vinfo, dr_info);
but vect_dr_behavior in this case returns &STMT_VINFO_DR_WRT_VEC_LOOP
(stmt_info) because the access is in the inner loop:
if (loop_vinfo == NULL
|| !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info))
return &DR_INNERMOST (dr_info->dr);
else
return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info);
with respect to the outer loop the access is computed misaligned through
step_alignment
$8 = {base_address = <addr_expr 0x7ffff683a140>,
offset = <integer_cst 0x7ffff68256c0>, init = <integer_cst 0x7ffff68256c0>,
step = <integer_cst 0x7ffff68256d8>, base_alignment = 16,
base_misalignment = 0, offset_alignment = 128, step_alignment = 8}
and misaligned with respect to the inner loop via offset_alignment:
$12 = {base_address = <addr_expr 0x7ffff68342c0>,
offset = <nop_expr 0x7ffff683a000>, init = <integer_cst 0x7ffff68309d8>,
step = <integer_cst 0x7ffff6830a50>, base_alignment = 16,
base_misalignment = 0, offset_alignment = 8, step_alignment = 32}
but we seem to only look at offset_alignment of one and step_alignment of
the other data piece. It seems to me the innermost behavior should
cover the outer step via offset_alignment (but maybe I'm wrong).
Still this seems to be a quite old issue. It seems to me we have to
consider both for the nested case, thus the 'else' arm should be
unconditional and use 'drb'.
I'm testing a patch.
