On Thu, Sep 22, 2022 at 1:01 AM Nathan Bossart <
nathandbossart@gmail.com> wrote:
>
> On Wed, Sep 21, 2022 at 01:17:21PM +0700, John Naylor wrote:
>
> > In short, this code needs to be lower level so that we still have full
> > control while being portable. I will work on this, and also the related
> > code for node dispatch.
>
> Is it possible to use approach #2 here, too? AFAICT space is allocated for
> all of the chunks, so there wouldn't be any danger in searching all them
> and discarding any results >= node->count.
Sure, the caller could pass the maximum node capacity, and then check if the returned index is within the range of the node count.
> Granted, we're depending on the
> number of chunks always being a multiple of elements-per-vector in order to
> avoid the tail path, but that seems like a reasonably safe assumption that
> can be covered with comments.
Actually, we don't need to depend on that at all. When I said "junk" above, that can be any bytes, as long as we're not reading off the end of allocated memory. We'll never do that here, since the child pointers/values follow. In that case, the caller can hard-code the size (it would even happen to work now to multiply rt_node_kind by 16, to be sneaky). One thing I want to try soon is storing fewer than 16/32 etc entries, so that the whole node fits comfortably inside a power-of-two allocation. That would allow us to use aset without wasting space for the smaller nodes, which would be faster and possibly would solve the fragmentation problem Andres referred to in
While on the subject, I wonder how important it is to keep the chunks in the small nodes in sorted order. That adds branches and memmove calls, and is the whole reason for the recent "pg_lfind_ge" function.
--