Crash bug in 8.2.3 on Solaris 10/Sparc

Hi,

we have found that psql in PostgreSQL 8.2.3
has problems connecting to the server
running on Solaris 10/Sun SPARC.

$ uname -a
SunOS dev-machine 5.10 Generic_118833-36 sun4u sparc SUNW,Sun-Fire-V440

It seems that somehow the system provided
GCC 3.4.3 miscompiles timestamptz_send()
and it segfaults. The default function looks like this:

Datum
timestamptz_send(PG_FUNCTION_ARGS)
{
TimestampTz timestamp = PG_GETARG_TIMESTAMPTZ(0);
StringInfoData buf;

pq_begintypsend(&buf);
#ifdef HAVE_INT64_TIMESTAMP
pq_sendint64(&buf, timestamp);
#else
pq_sendfloat8(&buf, timestamp);
#endif
PG_RETURN_BYTEA_P(pq_endtypsend(&buf));
}

GDB indicates crash at the last line.
No matter how I unrolled the function calls,
the indicated crasher line was always the one
before:

pq_sendfloat8(&buf, timestamp);

I must be a stack corruption somehow.
I also unrolled pq_sendfloat8() so the function looks like this:

Datum
timestamptz_send(PG_FUNCTION_ARGS)
{
TimestampTz timestamp = PG_GETARG_TIMESTAMPTZ(0);
StringInfoData buf;
bytea *byteap;
union
{
float8 f;
int64 i;
} swap;
uint32 n32;
pq_begintypsend(&buf);
#ifdef HAVE_INT64_TIMESTAMP
pq_sendint64(&buf, timestamp);
elog(NOTICE, "timestamptz_send() HAVE_INT64_TIMESTAMP after
pq_sendint64");
#else
swap.f = (float8)timestamp;
elog(NOTICE, "timestamptz_send() int64: %lld", swap.i);
/* High order half first, since we're doing MSB-first */
#ifdef INT64_IS_BUSTED
/* don't try a right shift of 32 on a 32-bit word */
n32 = (swap.i < 0) ? -1 : 0;
elog(NOTICE, "timestamptz_send() INT64_IS_BUSTED high 32: %d", n32);
#else
n32 = (uint32) (swap.i >> 32);
elog(NOTICE, "timestamptz_send() high 32: %d", n32);
#endif
n32 = htonl(n32);
elog(NOTICE, "timestamptz_send() htonl high 32: %d", n32);
appendBinaryStringInfo(&buf, (char *) &n32, 4);

/* Now the low order half */
n32 = (uint32) swap.i;
elog(NOTICE, "timestamptz_send() low 32: %d", n32);
n32 = htonl(n32);
elog(NOTICE, "timestamptz_send() htonl low 32: %d", n32);
appendBinaryStringInfo(&buf, (char *) &n32, 4);

elog(NOTICE, "timestamptz_send() pq_sendfloat8");
#endif
byteap = (bytea *) buf.data;
elog(NOTICE, "timestamptz_send() buf->data = %p", byteap);
Assert(buf.len >= VARHDRSZ);
VARATT_SIZEP(byteap) = buf.len;
PG_RETURN_BYTEA_P(byteap);
}

Th crashing line according to GDB is now the elog() call after:

swap.f = (float8)timestamp;

This is a simple explicit type cast which shouldn't cause problems,
however it is the one that somehow corrupts something on the stack
and causes the segfault upon entering the function at the next
statement.

As a workaround, we recompiled PostgreSQL 8.2.3 with
--enable-integer-datetimes
and the client can connect to the server now, after initdb.

I tried to exercise calling timestamptz_send() but creating a table
with float8 field, INSERTing and SELECTing works, too.
Both textual and binary COPY FROM and COPY TO work, too.
Either these exercises didn't call pq_sendfloat8() or it
doesn't cause problems elsewhere, only in timestamptz_send().

--
----------------------------------
Zoltán Böszörményi
Cybertec Geschwinde & Schönig GmbH
http://www.postgresql.at/