Обсуждение: NAN code
I have changed the NAN code, so if NAN is not defined, we just use 'num'
as NAN, since we only get here if num is NAN. Seems like a good fix for
platforms that can't assign a NAN to a variable.
---------------------------------------------------------------------------
if (NUMERIC_IS_NAN(num)) { result = (float32)palloc(sizeof(float32data));
#ifdef NAN *result = NAN;
#else *result = num;
#endif return result; }
-- Bruce Momjian | http://www.op.net/~candle maillist@candle.pha.pa.us | (610)
853-3000+ If your life is a hard drive, | 830 Blythe Avenue + Christ can be your backup. | Drexel Hill,
Pennsylvania19026
>
> I have changed the NAN code, so if NAN is not defined, we just use 'num'
> as NAN, since we only get here if num is NAN. Seems like a good fix for
> platforms that can't assign a NAN to a variable.
>
> ---------------------------------------------------------------------------
>
> if (NUMERIC_IS_NAN(num))
> {
> result = (float32)palloc(sizeof(float32data));
> #ifdef NAN
> *result = NAN;
> #else
> *result = num;
> #endif
> return result;
> }
And I where so happy to have finally found that damned NAN
define :-}
Anyway - that above is absolutely wrong.
num is of type Numeric, which is a pointer to struct
NumericData! So the above should result in some compiler
warning (in fact it type casts a pointer into a double value
- and that will give unpredictable results).
Seems that only isnan() is defined as part of Posix. But not
a definition that can force a NAN. So we have to find a
portable way to define the value NaN for double and float.
Does anybody know of such a way?
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#======================================== jwieck@debis.com (Jan Wieck) #
>
> I have changed the NAN code, so if NAN is not defined, we just use 'num'
> as NAN, since we only get here if num is NAN. Seems like a good fix for
> platforms that can't assign a NAN to a variable.
>
> ---------------------------------------------------------------------------
>
> if (NUMERIC_IS_NAN(num))
> {
> result = (float32)palloc(sizeof(float32data));
> #ifdef NAN
> *result = NAN;
> #else
> *result = num;
> #endif
> return result;
> }
And I where so happy to have finally found that damned NAN
define :-}
Anyway - that above is absolutely wrong.
num is of type Numeric, which is a pointer to struct
NumericData! So the above should result in some compiler
warning (in fact it type casts a pointer into a double value
- and that will give unpredictable results).
Seems that only isnan() is defined as part of Posix. But not
a definition that can force a NAN. So we have to find a
portable way to define the value NaN for double and float.
Does anybody know of such a way?
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#======================================== jwieck@debis.com (Jan Wieck) #
orion.SAPserv.Hamburg.dsh.de!wieck@sapserv.debis.de writes: > Seems that only isnan() is defined as part of Posix. But not > a definition that can force a NAN. So we have to find a > portable way to define the value NaN for double and float. > Does anybody know of such a way? Bruce changed it to defining NAN as (0.0/0.0) if NAN isn't already defined. This works on FreeBSD. I don't know if it will break on other systems that don't have NAN already.
> num is of type Numeric, which is a pointer to struct > NumericData! So the above should result in some compiler > warning (in fact it type casts a pointer into a double value > - and that will give unpredictable results). > > Seems that only isnan() is defined as part of Posix. But not > a definition that can force a NAN. So we have to find a > portable way to define the value NaN for double and float. > Does anybody know of such a way? See my later postings. 0.0/0.0 seems to do it. -- Bruce Momjian | http://www.op.net/~candle maillist@candle.pha.pa.us | (610) 853-3000+ If your life is a hard drive, | 830 Blythe Avenue + Christ can be your backup. | Drexel Hill, Pennsylvania19026
> > Seems that only isnan() is defined as part of Posix. But not
> > a definition that can force a NAN. So we have to find a
> > portable way to define the value NaN for double and float.
> > Does anybody know of such a way?
> Bruce changed it to defining NAN as (0.0/0.0) if NAN isn't already
> defined. This works on FreeBSD. I don't know if it will break on other
> systems that don't have NAN already.
The NaN concept is not portable on every platform, since it is an IEEE
754 floating point feature. Older machines like VAX don't have it, and
older compilers don't support it cleanly. The divide-by-zero technique
to get NaN is likely to result in a signaled FPE, where we want a quiet
assignment of a special value (NaN on those platforms for which it is
available).
Here is the code we already have in Postgres to support the date/time
data types; you'll see NAN, HUGE_VAL, and DBL_MAX used for special
values:
#ifdef NAN
#define DT_INVALID (NAN)
#else
#define DT_INVALID (DBL_MIN+DBL_MIN)
#endif
#ifdef HUGE_VAL
#define DT_NOBEGIN (-HUGE_VAL)
#define DT_NOEND (HUGE_VAL)
#else
#define DT_NOBEGIN (-DBL_MAX)
#define DT_NOEND (DBL_MAX)
#endif
#define DT_CURRENT (DBL_MIN)
#define DT_EPOCH (-DBL_MIN)
#define DATETIME_INVALID(j) {j = DT_INVALID;}
#ifdef NAN
#define DATETIME_IS_INVALID(j) (isnan(j))
#else
#define DATETIME_IS_INVALID(j) (j == DT_INVALID)
#endif
Of course, for historical reasons the date/time stuff has these special
reserved values which may not be necessary for the numerics. If
necessary, I would suggest something similar for the numeric stuff, or
perhaps on platforms without NAN and no way to generate it then just
throw an error.
As an aside, I worked on this a few years ago on my Alpha-dunix boxes
(for converting back and forth between VAX and IEEE floating point
binary values) and found that for the DEC compiler you had to use a
particular compiler flag to keep NaN as an unsignaled value. Otherwise
the call to isnan() failed when the value was put on the stack to set up
the function call to isnan() :/
- Tom
Bruce Momjian wrote:
>
> > Seems that only isnan() is defined as part of Posix. But not
> > a definition that can force a NAN. So we have to find a
> > portable way to define the value NaN for double and float.
> > Does anybody know of such a way?
>
> See my later postings. 0.0/0.0 seems to do it.
Seen them. Just that I'm a little in doubt if this construct
couldn't generate a SIGFPE on all of our supported
platform/compiler combos. Still think we should add autoconf
stuff to search for a NAN definition and only fallback to the
above if that fails.
While searching for the NAN definition I've noticed too that
our float4/float8 datatypes can output 'NaN', but do not
parse them back. They simply elog(ERROR, ...) if you try to
use 'NaN' as an input string for a floating point attribute.
Shouldn't all input functions be able to parse back any
possible result of the corresponding output function? As of
now, I cannot imagine a construct (except a user defined C
function), that could result in a float8-NaN value stored in
the database. But as soon as it happens, the database
wouldn't any longer be dump/reloadable!
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#======================================== jwieck@debis.com (Jan Wieck) #
> While searching for the NAN definition I've noticed too that
> our float4/float8 datatypes can output 'NaN', but do not
> parse them back.
Yeah, I just noticed that a few minutes ago too.
I've got some patches which fix this, at least for machines with NaN
support (and for HUGE_VAL/Infinity too). We'll need some more code to
handle non-IEEE machines...
- Tom
>
> > While searching for the NAN definition I've noticed too that
> > our float4/float8 datatypes can output 'NaN', but do not
> > parse them back.
>
> Yeah, I just noticed that a few minutes ago too.
>
> I've got some patches which fix this, at least for machines with NaN
> support (and for HUGE_VAL/Infinity too). We'll need some more code to
> handle non-IEEE machines...
The final might be that for machines/compilers with NaN
support, NAN is
#define NAN (0.0/0.0)
or whatever is the correct thing for it (best choice is still
the definition from a system header like nan.h if available),
while for those who cannot support it it'll be
#define NAN unsupported_nan_request()
double unsupported_nan_request(void)
{
elog(ERROR, "unsupported request for NaN as floating point value");
return 0.0;
}
Jan
--
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me. #
#======================================== jwieck@debis.com (Jan Wieck) #