Hi,
On 2021-11-05 14:13:38 -0400, Robert Haas wrote:
> On Fri, Nov 5, 2021 at 1:20 PM Andres Freund <andres@anarazel.de> wrote:
> > Yes. Optimally we'd do JIT caching across connections as well. One of the
> > biggest issues with the costs of JITing is actually parallel query, where
> > we'll often recreate the same JIT code again and again. For that you really
> > can't have much in the way of pointers...
>
> Well that much is clear, and parallel query also needs relative
> pointers in some places for other reasons, which reminds me to ask you
> whether these new relative pointers can't reuse "utils/relptr.h"
> instead of inventing another way of do it. And if not maybe we should
> try to first change relptr.h and the one existing client
> (freepage.c/h) to something better and then use that in both places,
> because if we're going to be stuck with relative pointers are all over
> the place it would at least be nice not to have too many different
> kinds.
Hm. Yea, that's a fair point. Right now the "allocno" bit would be a
problem. Perhaps we can get around that somehow. We could search for
allocations by the offset, I guess.
> > > It's a pretty annoying problem, really. Somehow it's hard to shake the
> > > feeling that there ought to be a better approach than relative
> > > pointers.
> >
> > Yes. I don't like it much either :(. Basically native code has the same issue,
> > and also largely ended up with making most things relative (see x86-64 which
> > does most addressing relative to the instruction pointer, and binaries
> > pre-relocation, where the addresses aren't resolved yed).
>
> Yes, but the good thing about those cases is that they're handled by
> the toolchain. What's irritating about this case is that we're using a
> just-in-time compiler, and yet somehow it feels like the job that
> ought to be done by the compiler is having to be done by our code, and
> the result is a lot of extra notation. I don't know what the
> alternative is -- if you don't tell the compiler which things it's
> supposed to assume are constant and which things might vary from
> execution to execution, it can't know. But it feels a little weird
> that there isn't some better way to give it that information.
Yes, I feel like there must be something better too. But in the end, I think
we want something like this for the non-JIT path too, so that we can avoid the
expensive re-creation of expression for every query execution. Which does make
referencing at least the mutable data only by offset fairly attractive, imo.
Greetings,
Andres Freund