On 6/2/25 2:23 AM, Pengfei Li wrote:
Hi all,
I would like to bring attention back to this patch:
https://inbox.sourceware.org/gcc-patches/20250508164950.5646-1-pengfei....@arm.com/
The patch improves auto-vectorization for loops with known small trip counts by
introducing a new target hook for subvector selection.
I fully understand Richard Biener's concern about introducing new target hooks -
that it's preferable to avoid adding new hooks unless necessary. However, as
explained in our earlier responses, in this specific case it seems difficult to
be completely avoided, as existing hooks don't have the ability to me to express
which kind of vector (scalable or fixed-length) is preferred.
So I'd like to check if there's any additional suggestions for improving this
patch, or other alternative directions I can explore. I'd be happy to respin or
adjust the patch depending on your feedback.
Thanks again for your time and all your inputs.
We very much want to avoid introducing target dependencies in gimple as
a design goal. While we have made exceptions, they are not to be taken
lightly.
I'll defer to Richi, Richard, Tamar, etc who know the vectorizer far
better than I.
Jeff
