On Fri, 3 Nov 2023, Andre Vieira (lists) wrote:
> Hi,
>
> The current codegen code to support VF's that are multiples of a simdclone
> simdlen rely on BIT_FIELD_REF to create multiple input vectors. This does not
> work for non-constant simdclones, so we should disable using such clones when
> the VF is a multiple of the non-constant simdlen until we change the codegen
> to support those.
>
> Enabling SVE simdclone support will cause ICEs if the vectorizer decides to
> use a SVE simdclone with a VF that is larger than the simdlen. I'll be away
> for the next two weeks, so cant' really discuss this further.
> I initially tried to solve the problem, but the way
> vectorizable_simd_clone_call is structured doesn't make it easy to replace
> BIT_FIELD_REF with the poly-suitable solution right now of using
> unpack_{hi,lo}.
I think it should be straight-forward to use unpack_{even,odd} (it's
even/odd for VLA, right? If lo/hi would be possible then doing
BIT_FIELD_REF would be, too? Also you need to have multiple stages
of unpack/pack when the factor is more than 2).
There's plenty of time even during stage3 to address this.
At least your patch should have come with a testcase (or two).
Is there a bugreport tracking this issue? It should affect GCN as well
I guess.
Richard.
> Unfortunately I only found this now as I was adding further
> tests for SVE :(
>
> gcc/ChangeLog:
>
> * tree-vect-stmts.cc (vectorizable_simd_clone_call): Reject simdclones
> with non-constant simdlen when VF is not exactly the same.
