Re: Making empty Bitmapsets always be NULL

On Fri, 30 Jun 2023 at 14:11, Yuya Watari <watari.yuya@gmail.com> wrote:
> I have tested these versions with installcheck. Since the planning
> times obtained by installcheck vary each time, it is important to run
> it repeatedly and examine its distribution. I ran installcheck 100
> times for each version. The following tables and the attached figure
> show the results. From these results, we can conclude that the v4
> patch has no regression in the installcheck test. It seems to be
> slightly (0.31-0.38%) faster than the master. The difference between
> v4 and v4 with 0th word check is not so clear, but v4 may be faster.

I did the same on the AMD 3990x machine and an Apple M2 Pro machine.
On the M2 over the 100 runs v4 came out 1.18% faster and the 3990x was
1.25% faster than master. I've plotted the results in the attached
graphs.

Looking over the patch again, the only thing I'm tempted into changing
is to add Asserts like: Assert(a ==  NULL || a->words[a->nword - 1] !=
0) to each function just as extra reassurance that nothing
accidentally leaves trailing empty words.

If nobody else wants to take a look, then I plan to push the v4 + the
asserts in the next day or so.

David