On 2026-Apr-02, Andres Freund wrote:
> I got pretty annoyed this cycle with how much rebuilding a simple change to
> bufmgr.h triggers (due to hacking on it a lot). I started to write a series
> to improve that, but didn't get around to posting that yet due to encountering
> the issue 771fe0948ca fixed while improving the situation.
>
> During that I encountered the objectaddress.h include, as part of which I then
> complained about the issue in [1].
Ah yeah, I noticed and was annoyed by that too.
> After the attached patches, a change to bufmgr.h triggers rebuilding 213
> files, before it was 323. Not perfect, but better.
>
> Would be nice to get rid of the bufmgr.h includes in access/nbtree.h and such,
> but it looks like that'd be a bit more work.
>
> I included your catalog/publication.h in it, as my version had surprisingly
> extensive bitrot...
I ran each patch individually under headerscheck and a full tree
compile; they all pass for me. Also, each change is sensible on its own.
Looking at what else includes bufmgr.h, I think the minimum it can
reduced to is compiling 157 files when you change bufmgr.h, per the
patches attached here. Most of them are direct inclusions, so reducing
further is tough. The only one we could blame is xlogutils.h, but it
needs the ReadBufferMode enum, so in order to do better, we'd have to
split bufmgr.h in two.
(A "fun" one is logicalproto.h being included by walreceiver.h, in turn
being included by slot.h, in turn being included by logical.h, in turn
being included by decode.h.)
--
Álvaro Herrera 48°01'N 7°57'E — https://www.EnterpriseDB.com/
