"Robin Dapp" <rdapp....@gmail.com> writes:
>> Could you walk me through the failure in more detail? It sounds
>> like can_duplicate_and_interleave_p eventually gets to the point of
>> subdividing the original elements, instead of either combining consecutive
>> elements (the best case), or leaving them as-is (the expected fallback
>> for SVE). But it sounds like those attempts fail in this case, but an
>> attempt to subdivide the elements succeeds. Is that right? And if so,
>> why does that happen?
>
> Apologies for the very late response.
>
> What I see is that we start with a base_vector_type vector([1,1]) long int
> and a count of 2, so ELT_BYTES = 16.
> We don't have a TI vector mode (and creating a single-element vector by
> interleaving is futile anyway) so the first attempt fails.
> The type in the second attempt is vector([1,1]) unsigned long but this
> is rejected because of
>
> && multiple_p (GET_MODE_NUNITS (TYPE_MODE (vector_type)),
>
> │
> 2, &half_nelts))
>
> Then we try vector([2,2]) unsigned int which "succeeds". This, however,
> eventually causes the ICE when we try to build a vector with 0 elements.
Ah, ok, thanks.
How about going for a slight variation of your original patch. After:
nvectors *= 2;
add:
/* We need to be able to to fuse COUNT / NVECTORS elements together. */
if (!multple_p (count, nvectors))
return false;
OK like that if it works.
Richard
> Maybe another option would be to decline 1-element vectors right away?
>
> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> index eac16e80ecd..d3e52489fa8 100644
> --- a/gcc/tree-vect-slp.cc
> +++ b/gcc/tree-vect-slp.cc
> @@ -427,7 +427,9 @@ can_duplicate_and_interleave_p (vec_info *vinfo, unsigned
> int count,
> tree *permutes)
> {
> tree base_vector_type = get_vectype_for_scalar_type (vinfo, elt_type,
> count);
> - if (!base_vector_type || !VECTOR_MODE_P (TYPE_MODE (base_vector_type)))
> + if (!base_vector_type
> + || !VECTOR_MODE_P (TYPE_MODE (base_vector_type))
> + || maybe_lt (GET_MODE_NUNITS (TYPE_MODE (base_vector_type)), 2))
> return false;
>
> Regards
> Robin
