On Thu, Nov 6, 2025 at 8:58 AM Robin Dapp <[email protected]> wrote:
>
> Grml, this is turning into a rabbit hole.
>
> The initial plan was to "just" support the widening shift optab for riscv.
> This patch here was supposed to be the first step.  The next patch adds 
> support
> for n-n (SLP style) widening ops.
>
> But I'm realizing now that we should actually have all the checking and
> fallback options that vectorizable_shift has, just also inside
> vectorizable_conversion if the op is a widening shift...  Not sure this is
> worth at it the end of stage 1 when we can easily synthesize the widening 
> shift
> on RTL level in the backend.
>
> Of course the long-term plan still is to move most of our RTL combinations
> (single widening ops, double widening ops, narrowing ops) earlier in the
> pipeline, making IFNs out of them but shift seems like the worst choice ;)

Yes, shifts are of the ugliest variant with all possible combinations of
scalar vs. vector, constant vs. non-constant present/not present on targets ...

Richard.

>
> --
> Regards
>  Robin
>

Reply via email to