Thread: Re: BUG #16920: Can't compile PostGIS with MingW64 against PostgreSQL 14 head

Hi,

On 2021-03-15 12:12:59 -0400, Tom Lane wrote:
> Although I remain worried about this being an ABI break, I don't think
> we are locked into it until we get to beta, or maybe even RC stage.

Could it make sense to define sigjmp_buf as a union over the potentially
needed implementations? That'd allow us to switch back without an ABI
break  if we discover a problem with the gcc approach.  And it might
even allow to backpatch this, if we cared enough, since I assume the
mingw jmp_buf is larger than intptr_t[5].

Greetings,

Andres Freund

Andres Freund <andres@anarazel.de> writes:
> On 2021-03-15 12:12:59 -0400, Tom Lane wrote:
>> Although I remain worried about this being an ABI break, I don't think
>> we are locked into it until we get to beta, or maybe even RC stage.

> Could it make sense to define sigjmp_buf as a union over the potentially
> needed implementations? That'd allow us to switch back without an ABI
> break  if we discover a problem with the gcc approach.

No, it'd still be an ABI break, because the setjmp and the longjmp calls
have to use the same implementation.  Ain't gonna work if elog.c tries
to throw via mingw's longjmp() while some extension contains a PG_TRY
that uses __builtin_setjmp().  Nor vice versa.

            regards, tom lane

Hi,

On 2021-03-19 20:37:17 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2021-03-15 12:12:59 -0400, Tom Lane wrote:
> >> Although I remain worried about this being an ABI break, I don't think
> >> we are locked into it until we get to beta, or maybe even RC stage.
> 
> > Could it make sense to define sigjmp_buf as a union over the potentially
> > needed implementations? That'd allow us to switch back without an ABI
> > break  if we discover a problem with the gcc approach.
> 
> No, it'd still be an ABI break, because the setjmp and the longjmp calls
> have to use the same implementation.  Ain't gonna work if elog.c tries
> to throw via mingw's longjmp() while some extension contains a PG_TRY
> that uses __builtin_setjmp().  Nor vice versa.

Yea, I momentarily forgot that we can't easily wrap setjmp in a
function... I guess we could just make sigsetjmp set a flag that
indicates which longjmp to use, but that's probably more complication
than the issue / mingw warrants.

- Andres