Hello,
Should we try to get this refactoring in stage 3 or park it for next
stage 1?
Cheers,
Andre
On 10/12/2019 13:36, Andre Vieira (lists) wrote:
Hi,
This patch aims at refactoring the vectorizer code to no longer need to
reset DR_OFFSET for epilogue vectorization and instead keep track of
DR_OFFSET changes per dr_vec_info and just update it as needed. I added
a member of type 'tree' called 'offset' to dr_vec_info, which keeps
track of the changes to the data_reference's offset per dr_vec_info and
thus per instance of loop vectorization. To get the current loop's
DR_OFFSET I introduced a function 'get_dr_vinfo_offset' which will add
the dr_vec_info's offset to either the data_reference's innermost offset
or the offset of the 'innermost_loop_behavior' returned by
'vect_dr_behavior' depending on whether 'get_dr_vinfo_offset's second
parameter 'check_outer' is true. This is to mimic the behavior of using
the outer loop relative 'innermost_loop_behavior' in
'vect_create_addr_base_for_vector_ref'.
I regression tested vect.exp on aarch64 and x86_64. I also regression
tested libgomp on aarch64 and x86_64, no changes, but there were quite a
few test failures with the commit I based this patch on...
Is this OK for trunk or shall I wait until libgomp stabilizes?
2019-12-10 Andre Vieira <andre.simoesdiasvie...@arm.com>
* tree-vect-data-refs.c (vect_create_addr_base_for_vector_ref):
Use
get_dr_vinfo_offset
* tree-vect-loop.c (update_epilogue_loop_vinfo): Remove
orig_drs_init
parameter and its use to reset DR_OFFSET's.
(vect_transform_loop): Remove orig_drs_init argument.
* tree-vect-loop-manip.c (vect_update_init_of_dr): Update the
offset
member of dr_vec_info rather than the offset of the associated
data_reference's innermost_loop_behavior.
(vect_update_init_of_dr): Pass dr_vec_info instead of
data_reference.
(vect_do_peeling): Remove orig_drs_init parameter and its
construction.
* tree-vect-stmts.c (check_scan_store): Replace use of DR_OFFSET
with get_dr_vinfo_offset.
(vectorizable_store): Likewise.
(vectorizable_load): Likewise.
