Re: Horology Test Case Fails with Sun Compiler

Tom,

I have been doing some more digging on this issue by comparing a build
of 8.0.3 and 8.1.4.

Whilst doing a build of 8.0.3 with the Sun compiler with these options:

./configure --without-readline --without-zlib CC=/opt/SUNWspro/bin/cc

Postgres fails over 60% of the test cases. In an effort to determine the
problem, I included --enable-debug. This broke the compiler in
backend/storage/lmgr/s_lock.c with the following error:

--------------Cut here------------
/opt/SUNWspro/prod/bin/fbe: "/tmp/yabeAAAHVaasb", line 279: error: can't
compute difference between symbols in different segments
Failure in /opt/SUNWspro/prod/bin/fbe, status = 0x100
Fatal Error exec'ing /opt/SUNWspro/prod/bin/fbe
cc: acomp failed for s_lock.c
--------------Cut here------------

I could not get beyond this so I could not run any test cases:

After attempting other compiler options, I am now using the following
for 8.0.3:

./configure --without-readline --without-zlib CC=/opt/SUNWspro/bin/cc
CFLAGS=-fast

However, the horology test case fails.

This morning I did another build with 8.1.4, using a variety of options.

./configure --without-readline --without-zlib CC=/opt/SUNWspro/bin/cc
--enable-debug

The compiler did not break and all test cases passed. So, I looked at
the code, comparing the 8.0.3 and 8.1.4 versions of s_lock.c. I think
the problem in 8.0.3 was caused by these two assembler directives
indicated below in the file. The 8.1.4 version of this file does not
have these lines.

--------------Cut here----------------
#if defined(__sparc__) || defined(__sparc)
/*
* sparc machines not using gcc
*/
static void
tas_dummy() /* really means:
extern int tas(slock_t
*
*lock); */
{

#ifdef SUNOS4_CC
asm(".seg \"data\"");
asm(".seg \"text\"");
#else
asm(".section \"data\""); <===================
asm(".section \"text\""); <===================
#endif

asm("_tas:");

/*
* Sparc atomic test and set (sparc calls it "atomic
load-store")
*/
asm("ldstub [%r8], %r8");
asm("retl");
asm("nop");
}
#endif /* __sparc || __sparc__ */
#endif /* not __GNUC__ */
#endif /* HAVE_SPINLOCKS */
--------------Cut here----------------

I commented these lines out for 8.0.3 and compiled with --enable-debug.
The compiler did not break and all test cases passed.

I notice that the tas_dummy() function is static and not called
anywhere. It appears that its only purpose is to break the compiler!

After making this change I notice that the -fast option has been causing
the test cases to fail.

Regards,
Match

-----Original Message-----
From: Grun, Match (TH USA)
Sent: Friday, June 09, 2006 1:14 PM
To: 'Tom Lane'
Subject: RE: [PORTS] Horology Test Case Fails with Sun Compiler

Tom,

Thank you for this information.

I noted that that reltime() was marked as deprecated in the 8.0.3
documentation. However, the equivalent section in the documentation did
not indicate reltime() was deprecated in 8.1.4. Perhaps you should
notify the remaining developers?

Even though reltime() is deprecated, the 8.0.3 build seems to be good.

BTW: I was not using --enable-integer-datetimes. I tried this option but
it made no difference to the test cases.

Note that I am doing a build with "-fast" for 8.0.3. If I leave this
option out, over 60% of the test cases fail. If do not use "-fast" but
use "--enable-debug", the compiler breaks! (fem complains of an error).
This seems to indicate that there may be a problem with the compiler. I
need to follow up with Sun on the compiler issue.

Also, I did a trial build and "make check" for 8.1.4 with the Sun
compiler. We will not be using that version for a while. Someone has
added another SQL test case to "interval":

---------------Cut here-----------------------------------------------
*** ./expected/interval.out Tue Oct 25 17:13:07 2005
--- ./results/interval.out Thu Jun 8 22:43:49 2006
***************
*** 218,224 ****
select avg(f1) from interval_tbl;
avg
-------------------------------------------------
! @ 4 years 1 mon 9 days 28 hours 18 mins 23 secs
(1 row)

-- test long interval input
--- 218,224 ----
Select avg(f1) from interval_tbl;
avg
-------------------------------------------------
! @ 4 years 1 mon 10 days 4 hours 18 mins 23 secs
(1 row)

-- test long interval input
---------------Cut here-----------------------------------------------

Thanks for your help. I will let you know if I make any more progress on
this problem.

Regards,
Match

-----Original Message-----
From: Tom Lane [mailto:tgl(at)sss(dot)pgh(dot)pa(dot)us]
Sent: Thursday, June 08, 2006 6:11 PM
To: Grun, Match (TH USA)
Cc: pgsql-ports(at)postgresql(dot)org
Subject: Re: [PORTS] Horology Test Case Fails with Sun Compiler

<Match(dot)Grun(at)thomson(dot)com> writes:
> After building Postgres 8.0.3, Solaris 8, Sun compiler v 5.7, I get
> the following diff with the horology test case:

Seems like either interval_reltime() or reltime2tm() (both in
src/backend/utils/adt/nabstime.c) must be screwing up. I've never seen
a similar report so I think you've got a broken compiler, but feel free
to look for a workaround ...

reltime is a deprecated type anyway so you could just ignore it, but the
$64 question here is whether it is a compiler bug and if so whether it
also hit any places that're more important. The code in those two
functions is not all that different from code elsewhere in the datetime
support.

BTW, are you using --enable-integer-datetimes? That changes this code
quite a bit.

regards, tom lane