Re: %2$, %1$ gettext placeholder replacement is not working under Win32

Greetings,
for a couple of days I have been hacking on src/port/snprintf.c.
With Magnus' help I have managed to implement argument replacement
in snprintf(). The code is very crude and not quite optimised,
any suggestions will be more than welcome.

Here is what I did:

1. I renamed snprintf() to pg_snprintf(), vsnprintf() to pg_vsnprintf()
and introduced pg_printf() that calls pg_vsnprintf().

2. After running configure I manually added snprintf.o to src/Makefile.global's
LIBOBJ declaration and -lpgport to Makefile.shlib's DLLWRAP declaration

3. To make sure these functions are used everywhere I introduced the following
lines at the end of src/include/c.h:

#define snprintf pg_snprintf
#define vsnprintf pg_vsnprintf
#define printf pg_printf

4. I introduced a volatile static char[] variable in snprintf.c code so I can
grep executables for this string and be sure that it is included.

5. Before running regression test I always ran make install, apparently because
libpq is read from /usr/local/.

During compilation the following warnings were reported:
../../../src/include/utils/elog.h:121: warning: `pg_printf' is an
unrecognized format function type

which is perfectly fine because we replace printf with pg_printf and
gcc's format() does
know anything about it.

On Linux, PostgreSQL passed regression tests with flying colours and
prints messages
with %n$ just fine. On win32: int8, timestamp, timestamptz, abstime, horology,
constraints, vacuum, and many others failed. To check my code, I
reverted snprintf.c
to the original one from CVS and forced win32 port to use these
functions and it
fails in same places. After examining regression tests diff I came to
conclusion that problem is in fmtnum() function when it operates with
particularly
long integers.

In snprintf() file I changed only and only dopr() function, neve touching
fmtnum(), fmtstr() or fmtfolat().

I would like to kindly ask these questions:

1. Am I on the right to implement %n$ ? Can it be accepted?
2. Why would we not just take FreeBSD's vfprintf() and use it instead?
3. What is wrong with fmtnum() on Win32?
4. What %m format string is used for? And where is it handled. Do I
need to implement it?

I am attaching my version if snprintf.c (because it is too different
from the original to
make a patch) and regression.diff of a failed Win32 regression test
produced wither with
my or with original snprintf.c.

Best regards,
Nicolai Tufar