On Mon, Sep 19, 2022 at 4:53 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Thomas Munro <thomas.munro@gmail.com> writes:
> > On Mon, Sep 19, 2022 at 3:39 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> >> I also note that our existing code in this area would break pretty
> >> thoroughly on such a machine, so this isn't making it worse.
>
> > Yeah, I don't expect it to be a practical problem on any real system
> > (that is, I don't expect any real calling convention to transfer a
> > struct T * argument in a different place than void *). I just wanted
> > to mention that it's a new liberty.
>
> No, it's not, because the existing coding here is already assuming that.
> The walker callbacks are generally declared as taking a "struct *"
> second parameter, but expression_tree_walker et al think they are
> passing a "void *" to them. Even if a platform ABI had some weird
> special rule about how to call functions that you don't know the
> argument list for, it wouldn't fix this because the walkers sure do know
> what their arguments are. The only reason this code works today is that
> in practice, "void *" *is* ABI-compatible with "struct *".
True.
> I'm not excited about creating a demonstrable opportunity for bugs
> in order to make the code hypothetically more compatible with
> hardware designs that are thirty years obsolete. (Hypothetical
> in the sense that there's little reason to believe there would
> be no other problems.)
Fair enough.