Re: [PORTS] Core dump with initdb on Irix 6.2

You must take care of a bug with gcc-2.8.1 on irix: there is a wrong
padding of 4 bytes structures when calling MipsPro C compiled
functions (such as the whole irix libc) by gcc-compiled functions (see
http://reality.sgi.com/ariel/freeware/gcc-2.8.1-notes.html
paragraph "Known bugs"). You will run into this bug when the
postgres backend call semctl(), (in src/backend/storage/ipc/).

As a quick and dirty workaround (please tell me if you find a better
solution), I modified my /usr/include/sys/sem.h and
/usr/include/arpa/inet.h :

/usr/include/sys/sem.h :

#ifndef _KERNEL
#if defined(sgi) && defined(__GNUC__) && defined(_ABIN32)
#define semctl gnu_semctl
#else
extern int semctl (int, int, int, ...);
#endif
extern int semget (key_t, int, int);
extern int semop (int, struct sembuf *, size_t);
#endif

/usr/include/arpa/inet.h :

#if defined(sgi) && defined(__GNUC__) && defined(_ABIN32)
#define inet_lnaof gnu_inet_lnaof
#define inet_netof gnu_inet_netof
#define inet_ntoa gnu_inet_ntoa
#else
extern in_addr_t inet_lnaof(struct in_addr);
extern in_addr_t inet_netof(struct in_addr);
extern char * inet_ntoa(struct in_addr);
#endif

and I (gcc-) compiled a library with :

workaround.h :

#if !defined(sgi) || !defined(__GNUC__) || !defined(_ABIN32)
ERROR !defined(sgi) || !defined(GNUC) || !defined(_ABIN32)
#endif

char *gnu_inet_ntoa(struct in_addr in);
unsigned long int gnu_inet_lnaof(struct in_addr in);
unsigned long int gnu_inet_netof(struct in_addr in);
int gnu_semctl(int semid, int semnun, int cmd, union semun arg);

workaround.c :

#if !defined(sgi) || !defined(__GNUC__) || !defined(_ABIN32)
ERROR !defined(sgi) || !defined(GNUC) || !defined(_ABIN32)
#endif

#include <sys/types.h>

struct in_addr {

unsigned long int s_addr;
};

char *gnu_inet_ntoa(struct in_addr in) {

return inet_ntoa( (int64_t)in.s_addr << 32 );
}

unsigned long int gnu_inet_lnaof(struct in_addr in) {

return inet_lnaof( (int64_t)in.s_addr << 32 );

}

unsigned long int gnu_inet_lnaof(struct in_addr in) {

return inet_lnaof( (int64_t)in.s_addr << 32 );

}

unsigned long int gnu_inet_netof(struct in_addr in) {

return inet_netof( (int64_t)in.s_addr << 32 );

}

union semun {
int val; /* used for SETVAL only */
struct semid_ds *buf; /* for IPC_STAT and IPC_SET */
ushort *array; /* used for GETALL and SETALL */
};

int gnu_semctl(int semid, int semnun, int cmd, union semun arg) {

return semctl(semid, semnun, cmd, (int64_t)arg.val << 32 );

}

Then, when compiling with gcc-2.8.1 in n32 binary with irix,
semctl() is changed in gnu_semctl(), and, of course, the final
link must include the workaround library.

PostgreSQL doesn't use inet_xtoy (does it ?). These functions
(semctl() and inet_xtoy) seem to be the only ones in the irix
libc that make problems.

I hope this will help,

Rémi.

Thomas Lockhart writes:
> > After switching to gcc 2.8.1 (thanks to Remi), I've managed to get
> > initdb and postmaster running now - woo who!!
> > However, my excitement was short live. After examining make log, I
> > find that psql doesn't compile. Hence createdb will stall.
> > Here's the message
> > ld32: FATAL 12: Expecting n32 objects:
> > /usr/local/lib/libreadline.a(readline.o) is o32.
> > How shouild I fix this? I'm now using gcc 2.8.1 My environment
> > variable SGI_ABI is set to -n32.
>
> It looks like libreadline (which you installed earlier) was built
> using o32, whatever that is, and that is different from n32. Try
> rebuilding libreadline or build all of postgres with using o32.
>
> Another option is to explicitly disable using readline; look in
> config.h or figure out how to disable it using configure.
>
> Good luck.
>
> - Tom
>
> --
> Thomas Lockhart lockhart(at)alumni(dot)caltech(dot)edu
> South Pasadena, California
>