Hi,
On 2019-01-27 08:03:17 +0100, Greg Stark wrote:
> I assume you already considered and rejected having a fixed size null
> bitmap followed by a variable size array of datums. That seems like it
> would be denser and work better with cpu cache.
It'd be more expensive to access individually (offset calculation +
masks, ~5 insn, not fully pipelineable), it'd not guarantee that the
null bit and datum are on the same cacheline, you could not pass the
null-bit to various functions accepting a bool*, you could not pass the
new equivalent NullableDatums to other functions (like both the past and
current solution allow).
Greetings,
Andres Freund