Hi,
On 2019-06-06 13:39:50 -0400, Tom Lane wrote:
> Lastly, I don't think adding additional allowed widths of pass-by-value
> types would be cost-free, because it would add cycles to the inner loops
> of the tuple forming and deforming functions.
I'm not sure I quite buy that.
I think that we have branches over a fixed number of lengths is largely
unnecessary. att_addlength_pointer() doesn't care - it just uses the
length. And I think we should just consider doing the same for
fetch_att(). E.g. by using memcpy().
That'd also have the advantage that we'd not be *forced* to rely
alignment of byval types. The only reason we actually need that is the
heaptuple-to-struct mapping for catalogs. Outside of that we don't have
pointers to individual byval tuples, and waste a fair bit of padding due
to that.
Additionally we'd get rid of needing separate versions for SIZEOF_DATUM
!= 8/not.
> (No, I don't believe that JIT makes that an ignorable concern.)
Obviously not.
Greetings,
Andres Freund