Re: FreeTDS port to VMS V9.x on x86?

On 7/3/25 3:07 PM, Dan Cross wrote:

In article <1046hc5$62th$2@dont-email.me>,
Arne Vajhøj  <arne@vajhoej.dk> wrote:

On 7/3/2025 1:40 PM, Dan Cross wrote:

You refer to `_PATH_DEVNULL` but do not `#include <paths.h>`, as
required by POSIX.

>
This must support non *nix systems (as an example
VMS !!) - config.h is expected to provide that with the rest.

 If copyright dates are anything to judge by, `<paths.h>` has
been a thing since 1989, but I was wrong in that it is not
actually in POSIX: it is a BSD extension, though common.
 In general, it is good practice and good hygiene in C programs
to `#include` the header files that are documented to define the
symbols and types that you use in your program, instead of
relying on transitive includes to populate things ambiantly via
e.g. `config.h`.
 Regardless, in that case, you shouldn't use `_PATH_DEVNULL`.
Note the leading `_`: that signifies a reserved identifier,
which is not something you should be defining yourself.  If you
want to punt this to e.g. config.h, better would be to define a
new name (say, `TDS_PATH_DEVNULL`) and use that.

The code ya'll are talking about replacing:
https://github.com/FreeTDS/freetds/blob/Branch-1_5/src/replacements/vasprintf.c
does the following:
#if HAVE_PATHS_H
#include <paths.h>
#endif /* HAVE_PATHS_H */
. . .
#ifndef _PATH_DEVNULL
#define _PATH_DEVNULL "/dev/null"
#endif
FreeTDS is relying on autoconf to sort out whether paths.h is available
and will only include it if HAVE_PATHS_H is defined in config.h, and
thus only defines _PATH_DEVNULL if it's known not to exist. Obviously
one should avoid making up one's own reserved identifiers and clobbering
documented and well-known names, but that's not really the same thing as
providing an implementation for one that's known to be missing. This
fallback approach could of course be added to Arne's implementation if
desired; to me it seems a lot cleaner than defining yet another macro,
which would need to be defined in terms of _PATH_DEVNULL when it exists
but otherwise not.
The problem for VMS is that the workaround doesn't work.  Or at least it
didn't when I was first porting FreeTDS 20+ years ago because the CRTL
at the time did not recognize '/dev/null' as a valid path.  I believe
recentish CRTLs (maybe VMS 8.x and later?) do have special case code
that translates it to the native null device.  But that wasn't
available, so I added the following line to the template from which
config.h is generated on VMS:
#define _PATH_DEVNULL "_NLA0:"
The main advantage of this approach is that it worked, and it did so
with a one-line change to a file I was already maintaining. I didn't
have to sprinkle '#ifdef __VMS' in files that already existed and were
already working on other platforms. I didn't have to create a lot of new
files, the presence of which in the repository I couldn't test without
autoconf/automake, which I didn't have access to at the time.  All of
which meant upstream maintainers were far more likely to accept my changes.
I'm sure I'll get a lecture about everything I did wrong, but at least
now the context is out there for why it was reasonable for Arne to
depend on prior art, i.e., the presence of _PATH_DEVNULL in config.h.