Regression failure for floats

I am seeing the following regression failure for current CVS. On my OS,
BSD/OS 4.3, it seems once you hit Infinity, you can't negate it.
/usr/include/math.h has:

/* Generate an overflow to create +Inf; the multiply shuts up gcc 1 */
#define HUGE_VAL (1e250*1e250) /* IEEE: positive infinity */

and our float4in code has:

else if (strcasecmp(num, "-Infinity") == 0)
val = -HUGE_VAL;

Seems that doesn't work for me.

---------------------------------------------------------------------------

*** ./expected/float4.out Sat Mar 13 23:07:10 2004
--- ./results/float4.out Sat Mar 13 23:10:30 2004
***************
*** 58,65 ****

SELECT ' -INFINiTY '::float4;
float4
! -----------
! -Infinity
(1 row)

-- bad special inputs
--- 58,65 ----

SELECT ' -INFINiTY '::float4;
float4
! ----------
! Infinity
(1 row)

-- bad special inputs

======================================================================

*** ./expected/float8.out Sat Mar 13 23:07:10 2004
--- ./results/float8.out Sat Mar 13 23:10:30 2004
***************
*** 58,65 ****

SELECT ' -INFINiTY '::float8;
float8
! -----------
! -Infinity
(1 row)

-- bad special inputs
--- 58,65 ----

SELECT ' -INFINiTY '::float8;
float8
! ----------
! Infinity
(1 row)

-- bad special inputs

--
Bruce Momjian | http://candle.pha.pa.us
pgman(at)candle(dot)pha(dot)pa(dot)us | (610) 359-1001
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073

Bruce Momjian <pgman(at)candle(dot)pha(dot)pa(dot)us> writes:
> I am seeing the following regression failure for current CVS. On my
> OS, BSD/OS 4.3, it seems once you hit Infinity, you can't negate it.

Actually, I suspect the problem is that isinf() on your platform
returns 1 for any infinity (rather than -1 for negative infinity and 1
for positive infinity). Some existing code in float4out() and
float8out() assumed that a positive return from isinf() indicated a
positive infinity, which is not per C99.

Anyway, Tom and I worked through this issue, and a couple other
portability problems with the recent float changes, via private
email. The current patch is attached -- Tom hasn't yet gotten back to
me on whether this fixes the problem for him on HPUX, but it fixes my
OS X box.

-Neil

Bruce Momjian <pgman(at)candle(dot)pha(dot)pa(dot)us> writes:
> I am seeing the following regression failure for current CVS. On my OS,
> BSD/OS 4.3, it seems once you hit Infinity, you can't negate it.
> /usr/include/math.h has:
> /* Generate an overflow to create +Inf; the multiply shuts up gcc 1 */
> #define HUGE_VAL (1e250*1e250) /* IEEE: positive infinity */
> and our float4in code has:
> else if (strcasecmp(num, "-Infinity") == 0)
> val = -HUGE_VAL;
> Seems that doesn't work for me.

No, the bug is that Neil assumed isinf() would tell him the sign of an
infinity. I believe he's about to commit a patch for this and some
other bogosities.

regards, tom lane

Neil Conway <neilc(at)samurai(dot)com> writes:
> The current patch is attached -- Tom hasn't yet gotten back to me on
> whether this fixes the problem for him on HPUX, but it fixes my OS X
> box.

Patch applied to CVS.

-Neil