On Tue, 6 May 2025, [email protected] wrote:
>
> This removes the non-SLP paths from vectorizable_conversion and in the
> process eliminates uses of 'ncopies' and removes
> 'get_initial_def_for_reduction'. Not done anything about
> STMT_VINFO_VECTYPE as I'm not sure about its uses in
> vectorizable_reduction in relation to the vectype_in/out.
That's fine I guess. Further cleanup is always possible.
- unsigned int group_size = 1, k;
+ unsigned int group_size, k;
I guess it's clearer to also move the declaration to where the
unconditional init now happens?
- for (j = 0; j < ncopies; j++)
+ for (j = 0; j < 1; j++)
{
tree new_def = copy_ssa_name (def);
phi = create_phi_node (new_def, exit_bb);
- if (j)
+ if (0)
def = gimple_get_lhs (STMT_VINFO_VEC_STMTS (rdef_info)[j]);
guess the next in the series will elide the loop and replace the
other use of 'j'.
@@ -7102,7 +7027,6 @@ vectorize_fold_left_reduction (loop_vec_info
loop_vinfo,
tree reduc_var = gimple_phi_result (reduc_def_stmt);
int vec_num = vec_oprnds0.length ();
- gcc_assert (vec_num == 1);
not sure if that's the one I complained about earlier.
- vec_initial_defs.create (ncopies);
- for (i = 0; i < ncopies; ++i)
+ vec_initial_defs.create (1);
+ for (i = 0; i < 1; ++i)
vec_initial_defs.quick_push (vec_initial_def);
likewise.
- the BB of the related stmt is inside this loop. */
+ the BB of the related stmt is inside this loop.
+ TODO: Do we still need this now that
+ 'get_initial_def_for_reduction' has been removed? */
I think so, there's a "copy" for SLP called
get_initial_defs_for_reduction, so possibly just change the reference
in the comment.
Richard.
> gcc/tree-vect-loop.cc | 127 +++++++----------------------------------
> gcc/tree-vect-stmts.cc | 4 +-
> gcc/tree-vectorizer.h | 5 +-
> 3 files changed, 25 insertions(+), 111 deletions(-)
>
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
