From: Tom Lane <tgl@sss.pgh.pa.us>
To: pgsql-hackers@lists.postgresql.org
Subject: Residual cpluspluscheck issues
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <20517.1559494394.1@sss.pgh.pa.us>
Content-Transfer-Encoding: quoted-printable
Date: Sun, 02 Jun 2019 12:53:14 -0400
Message-ID: <20518.1559494394@sss.pgh.pa.us>
Precedence: bulk

cpluspluscheck's expanded coverage is now passing cleanly for me on
the macOS laptop I was testing it with at PGCon.  But on returning
home, I find there's still some issues on some other boxes:

* On Linux (at least Fedora and RHEL), I get variants of this:

/usr/include/arpa/inet.h:84: error: declaration of 'char* inet_net_ntop(in=
t, const void*, int, char*, size_t) throw ()' throws different exceptions
/home/postgres/pgsql/src/include/port.h:506: error: from previous declarat=
ion 'char* inet_net_ntop(int, const void*, int, char*, size_t)'

That's because /usr/include/arpa/inet.h declares it as

extern char *inet_net_ntop (int __af, const void *__cp, int __bits,
                            char *__buf, size_t __len) __THROW;

and of course when a C++ compiler reads that, __THROW will expand as
something nonempty.

One possible fix for that is to teach configure to test whether
arpa/inet.h provides a declaration, and not compile our own declaration
when it does.  This would require being sure that we include arpa/inet.h
anywhere we use that function, but there are few enough callers that
that's not much of a hardship.

Alternatively, we could rename our function to pg_inet_net_ntop to
dodge the conflict.  This might be a good idea anyway to avoid
confusion, since our function doesn't necessarily recognize the same
address-family codes that libc would.

* On FreeBSD 12, I get

/home/tgl/pgsql/src/include/utils/hashutils.h:23:23: warning: 'register' s=
torage
      class specifier is deprecated and incompatible with C++17
      [-Wdeprecated-register]
extern Datum hash_any(register const unsigned char *k, register int keylen=
);
                      ^~~~~~~~~
/home/tgl/pgsql/src/include/utils/hashutils.h:23:56: warning: 'register' s=
torage
      class specifier is deprecated and incompatible with C++17
      [-Wdeprecated-register]
extern Datum hash_any(register const unsigned char *k, register int keylen=
);
                                                       ^~~~~~~~~
/home/tgl/pgsql/src/include/utils/hashutils.h:24:32: warning: 'register' s=
torage
      class specifier is deprecated and incompatible with C++17
      [-Wdeprecated-register]
extern Datum hash_any_extended(register const unsigned char *k,
                               ^~~~~~~~~
/home/tgl/pgsql/src/include/utils/hashutils.h:25:11: warning: 'register' s=
torage
      class specifier is deprecated and incompatible with C++17
      [-Wdeprecated-register]
                                                           register int ..=
.
                                                           ^~~~~~~~~

which I'm inclined to think means we should drop those register keywords.

(The FreeBSD box shows another couple of complaints too, but I think
the fixes for those are uncontroversial.)

Comments?

			regards, tom lane