"Yangfei (Felix)" <felix.y...@huawei.com> writes:
>> On June 30, 2020 4:23:03 PM GMT+02:00, Richard Sandiford
>> <richard.sandif...@arm.com> wrote:
>> >Richard Biener <rguent...@suse.de> writes:
>> >> So it seems odd to somehow put in the number of vectors... so to me
>> >> it would have made sense if it did
>> >>
>> >> possible_npeel_number = lower_bound (nscalars);
>> >>
>> >> or whateveris necessary to make the polys happy. Thus simply elide
>> >> the vect_get_num_vectors call? But it's been very long since I've
>> >> dived into the alignment peeling code...
>> >
>> >Ah, I see what you mean. So rather than:
>> >
>> > /* Save info about DR in the hash table. Also include peeling
>> > amounts according to the explanation above. */
>> > for (j = 0; j < possible_npeel_number; j++)
>> > {
>> > vect_peeling_hash_insert (&peeling_htab, loop_vinfo,
>> > dr_info, npeel_tmp);
>> > npeel_tmp += target_align / dr_size;
>> > }
>> >
>> >just have something like:
>> >
>> > while (known_le (npeel_tmp, nscalars))
>> > {
>> > …
>> > }
>> >
>> >?
>>
>> Yeah.
>
> Not sure if I understand correctly. I am supposing the following check in
> the original code is not necessary if we go like that.
>
> 1822 if (unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo)))
>
> Is that correct?
I think we still need it. I guess there are two choices:
- make nscalars default to npeel_tmp before the “if” above.
- put the loop inside the “if” and add a single call to
vect_peeling_hash_insert as a new “else”.
Thanks,
Richard
