On 2017-11-11 12:54:29 -0500, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2017-11-11 12:41:40 -0500, Tom Lane wrote:
> >> Yeah, changing MAXALIGN is out of the question. I'm thinking about
> >> another flag bit for MemoryContextAllocExtended. Do we need to think
> >> about other use-cases besides int128? Should we just force 16-byte
> >> alignment on all architectures, or does it need to be platform-specific?
>
> > I'm not sure we want to
> > a) Rely on one alignment being enough for everybody.
> > b) Additionally burden already hot code paths with a growing number of
> > alignment flag tests, and the necessary math.
>
> Well, (a) we could have more flag bits later if there are other use-cases
> that require even stricter alignment
Right, that's my point. It's not a scalable interface.
> and (b) I do not believe that MemoryContextAllocExtended is a hot code
> path at present; there are not enough call sites.
It's not that hot, true.
> > How about a MemoryContextAllocAligned(context, size, alignto, flags)
> > that passes on most flags but adds the necessary overhead to size, and
> > padds the result appropriately?
>
> This'd result in hard-wiring the alignment requirement at call sites,
> which I think might not be a great idea.
Well, it'd require using a macro specifying the alignment (or if we were
using a modern C dialect alignof()), not the alignment spelled out.
> For example, one plausible future use-case is "align on cacheline
> boundaries".
Unless we go to threads that doesn't seem *that* likely, given what
palloc's used for.
> I think that would be better served by a flag like
> MCXT_ALLOC_ALIGN_CACHELINE than by having the callers demand a
> specific numeric alignment value --- it'd be a lot easier to make the
> alignment match actual hardware requirements if it were being inserted
> at one specific place.
Specifying MCXT_ALLOC_ALIGN_CACHELINE rather than CACHELINE_ALIGNMENT
doesn't seem to make it meaningfully harder to adjust. Or are you
thinking of probing the hardware?
Greetings,
Andres Freund
--
Sent via pgsql-bugs mailing list (pgsql-bugs@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-bugs
