Horton

Thread

Commits

  1. Use standard compare_exchange loop style in ProcArrayGroupClearXid().

  1. ProcArrayGroupClearXid() compare-exchange style

    Noah Misch <noah@leadboat.com> — 2019-10-15T03:53:48Z

    ProcArrayGroupClearXid() has this:

	while (true)
	{
		nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);

		...

		if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
										   &nextidx,
										   (uint32) proc->pgprocno))
			break;
	}

This, from UnpinBuffer(), is our more-typical style:

		old_buf_state = pg_atomic_read_u32(&buf->state);
		for (;;)
		{
			...

			if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
											   buf_state))
				break;
		}

That is, we typically put the pg_atomic_read_u32() outside the loop.  After
the first iteration, it is redundant with the side effect of
pg_atomic_compare_exchange_u32().  I haven't checked whether this materially
improves performance, but, for style, I would like to change it in HEAD.

  2. Re: ProcArrayGroupClearXid() compare-exchange style

    Amit Kapila <amit.kapila16@gmail.com> — 2019-10-15T09:30:01Z

    On Tue, Oct 15, 2019 at 9:23 AM Noah Misch <noah@leadboat.com> wrote:
>
> ProcArrayGroupClearXid() has this:
>
>         while (true)
>         {
>                 nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
>
>                 ...
>
>                 if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
>                                                                                    &nextidx,
>                                                                                    (uint32) proc->pgprocno))
>                         break;
>         }
>
> This, from UnpinBuffer(), is our more-typical style:
>
>                 old_buf_state = pg_atomic_read_u32(&buf->state);
>                 for (;;)
>                 {
>                         ...
>
>                         if (pg_atomic_compare_exchange_u32(&buf->state, &old_buf_state,
>                                                                                            buf_state))
>                                 break;
>                 }
>
> That is, we typically put the pg_atomic_read_u32() outside the loop.  After
> the first iteration, it is redundant with the side effect of
> pg_atomic_compare_exchange_u32().  I haven't checked whether this materially
> improves performance, but, for style, I would like to change it in HEAD.
>

+1.  I am not sure if it would improve performance as this whole
optimization was to reduce the number of attempts to acquire LWLock,
but definitely, it makes the code consistent.

-- 
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com