Richard Biener <rguent...@suse.de> writes:
> This implements patterns combining vector element insertion of
> vector element extraction to a VEC_PERM_EXPR of both vectors
> when supported. Plus it adds the more generic identity transform
> of inserting a piece of itself at the same position.
>
> Richard - is there anything I can do to make this SVE aware?
> I'd need to construct an identity permute and "insert" into
> that permute that element from the other (or same) vector.
> I suppose for most element positions that won't work but
> at least inserting at [0] should? I'm mostly struggling
> on how to use vec_perm_builder here when nelts is not constant,
> since it's derived from vec<> can I simply start with
> a single pattern with 1 stride and then insert by using []?
I guess for SVE we still want to know that the range is safe
for all VL, so after dropping the is_constant check, we'd
want something like:
{
poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (type);
unsigned int min_nelts = constant_lower_bound (nelts);
}
(if (...
&& at + elemsz <= min_nelts)
In theory (hah) it should then just be a case of changing the
vec_perm_builder constructor to:
vec_perm_builder sel (nelts, min_nelts, 3);
and then iterating over min_nelts * 3 instead of nelts here:
> + for (unsigned i = 0; i < nelts; ++i)
> + sel.quick_push (i / elemsz == at
> + ? nelts + elem * elemsz + i % elemsz : i);
So as far as the encoding goes, the first min_nelts elements are arbitrary
values, and the following two min_nelts sequences form individual linear
series.
This ought to be work for both SVE and non-SVE, although obviously
there's a bit of wasted work for non-SVE.
(And thanks for asking :-))
Richard
