On Sun, Jun 08, 2014 at 06:04:46PM -0400, Tom Lane wrote:
> Noah Misch <noah@leadboat.com> writes:
> > Since src/makefiles/Makefile.win32 does not set BUILDING_DLL for src/port,
> > PGDLLIMPORT is set improperly for code to be linked directly into the backend.
> > Makefile.win32 does set BUILDING_DLL for src/common. (Makefile.cygwin has the
> > same discrepancy, though I haven't checked whether it causes an actual build
> > failure there. The MSVC build system sets BUILDING_DLL for both src/port and
> > src/common files.) This affects any reference to a data symbol from src/port.
> > The fix is straightforward enough: cause Makefile.{win32,cygwin} to treat
> > src/port like src/common.
>
> I wonder whether these cases shouldn't distinguish between the "frontend"
> and "backend" builds of src/port/ and src/common/. In particular, it
> seems odd that we're getting this type of failure in the backend build.
Good question; they need not distinguish. !BUILDING_DLL means that the code
being compiled will access backend symbols through dynamic linking to
postgres.exe. Server modules, such as plpgsql, build with !BUILDING_DLL, and
normal backend code builds with BUILDING_DLL. The setting is irrelevant for
"frontend"/non-server code, which does not link to backend symbols at all.
> > Second, src/template/{win32,cygwin} completely replaces LDFLAGS, so overriding
> > LDFLAGS on the "configure" command line is ineffective. Those scripts should
> > instead add to the existing LDFLAGS, like other templates do for CPPFLAGS.
> > Several other templates completely override CFLAGS; that's undesirable for the
> > same reason. I don't have ready access to those affected configurations, so
> > I'm leaving the CFLAGS damage alone.
>
> +1 for doing something about CFLAGS while we're at it.
Scratch that; configure.in has logic to discard template script CFLAGS changes
if the user had specified CFLAGS.
--
Noah Misch
EnterpriseDB http://www.enterprisedb.com