On 19/01/2022 11:04, Richard Biener wrote:
> On Tue, 18 Jan 2022, Andre Vieira (lists) wrote:
>
>> On 14/01/2022 09:57, Richard Biener wrote:
>>> The 'used_vector_modes' is also a heuristic by itself since it registers
>>> every vector type we query, not only those that are used in the end ...
>>>
>>> So it's really all heuristics that can eventually go bad.
>>>
>>> IMHO remembering the VF that we ended up with (maybe w/o unrolling)
>>> for each analyzed vector_mode[] might be really the easiest thing to do,
>>> that should make it easy to skip those modes where the VF is larger
>>> or equal as the VF of the main loop for the purpose of epilogue
>>> vectorization.  Likewise those vector_mode[] that failed analysis can
>>> be remembered (with -1U VF for example).
>>>
>>> Richard.
>> I liked the caching suggestion, so here it is. Sorry for the delay, wanted to
>> post this after pushing the vect unroll which was waiting on some retesting
>> for the rebase.
> LGTM.
>
> Thanks,
> Richard.
>
>> gcc/ChangeLog:
>>
>>          PR 103997
>>          * tree-vect-loop.c (vect_analyze_loop): Fix mode skipping for
>> epilogue
>>          vectorization.


Thanks! Committed the following patch. I updated the comment above the 
last change as I realized it still described the old behaviour.