Signaling NaNs

Signaling NaNs

[Marc Nieper-Wißkirchen]

[-- Attachment #1: Type: text/plain, Size: 590 bytes --]

Hi,

I have been searching through the list of modules but haven't been able to
find it:  Does Gnulib offer a way to store a signaling NaN in a memory
location (if supported by the platform) and some way to check a memory
location whether it contains a signaling NaN?

The forthcoming C2x standard will have FLT_SNAN, DBL_SNAN, and LDBL_SNAN,
but before that there doesn't seem to be a portable way to get signaling
NaNs.

In case the following problem is easier: What I really need is a bit
pattern for a double that won't be returned by the usual floating-point
functions.

Thanks,

Marc

[-- Attachment #2: Type: text/html, Size: 1332 bytes --]

Re: Signaling NaNs

[Bruno Haible]

Hi Marc,

> I have been searching through the list of modules but haven't been able to
> find it:  Does Gnulib offer a way to store a signaling NaN in a memory
> location (if supported by the platform)

It doesn't, because quiet NaNs are easier to work with.

> The forthcoming C2x standard will have FLT_SNAN, DBL_SNAN, and LDBL_SNAN,
> but before that there doesn't seem to be a portable way to get signaling
> NaNs.

... indeed, when you have a non-optimizing compiler, how to prevent the
compiler from generating instructions that produce a floating-point exception
earlier than desired?

> and some way to check a memory
> location whether it contains a signaling NaN?

Why would you need that? If you are using floating-point operations
(such as addition, square root, etc.) the signaling NaN will produce an
exception, as desired. Whereas if you are implementing some extra floating-
point operation by looking at the precise bit pattern (using integer
arithmetic), the standards [1][2] tell you which bit pattern to look for.

> In case the following problem is easier: What I really need is a bit
> pattern for a double that won't be returned by the usual floating-point
> functions.

When you look at [1] and [2], all bit patterns have a meaning.

So, you could
  - either use one of the many particular quiet NaN values, and hope that
    no other code produces it, or
  - (like what the IA64 CPU does in hardware) work with 65-bit units,
    where the 65th bit means "uninitialized", or
  - reserve a couple of extra bits for a type tag on every value, like
    what some Lisp implementations do (e.g. GNU clisp).
I think valgrind uses one of the latter two techniques as well.

Bruno

[1] https://en.wikipedia.org/wiki/Single-precision_floating-point_format
[2] https://en.wikipedia.org/wiki/Double-precision_floating-point_format

Re: Signaling NaNs

[Marc Nieper-Wißkirchen]

[-- Attachment #1: Type: text/plain, Size: 3003 bytes --]

Hi Bruno,

thanks for replying so quickly!

My intention is neither to feed the signaling NaN into floating-point
operations nor to cause it an exception to be raised.  What I really want
to do is to model a type whose value is either a floating-point number
(including infinities and the NaNs returned by the floating-point
functions) or a sentinel value. Furthermore, I want this to fit into the
size of a double.

The reason why I thought of using the bit pattern of a signaling NaN is
that such a signaling NaN wouldn't be returned by the usual floating-point
functions. A quiet NaN, which wouldn't either, would also work for my
purpose.  This would correspond to the first of your three suggestions.  So
I could reformulate my question to: "Is there a way to produce a (quiet)
NaN that won't occur as a result of the C library functions?" This seems to
be possible on most architectures but will need specific code for some
architectures, which is why I thought of Gnulib.

Thanks,

Marc

Am Do., 9. Dez. 2021 um 22:49 Uhr schrieb Bruno Haible <bruno@clisp.org>:

> Hi Marc,
>
> > I have been searching through the list of modules but haven't been able
> to
> > find it:  Does Gnulib offer a way to store a signaling NaN in a memory
> > location (if supported by the platform)
>
> It doesn't, because quiet NaNs are easier to work with.
>
> > The forthcoming C2x standard will have FLT_SNAN, DBL_SNAN, and LDBL_SNAN,
> > but before that there doesn't seem to be a portable way to get signaling
> > NaNs.
>
> ... indeed, when you have a non-optimizing compiler, how to prevent the
> compiler from generating instructions that produce a floating-point
> exception
> earlier than desired?
>
> > and some way to check a memory
> > location whether it contains a signaling NaN?
>
> Why would you need that? If you are using floating-point operations
> (such as addition, square root, etc.) the signaling NaN will produce an
> exception, as desired. Whereas if you are implementing some extra floating-
> point operation by looking at the precise bit pattern (using integer
> arithmetic), the standards [1][2] tell you which bit pattern to look for.
>
> > In case the following problem is easier: What I really need is a bit
> > pattern for a double that won't be returned by the usual floating-point
> > functions.
>
> When you look at [1] and [2], all bit patterns have a meaning.
>
> So, you could
>   - either use one of the many particular quiet NaN values, and hope that
>     no other code produces it, or
>   - (like what the IA64 CPU does in hardware) work with 65-bit units,
>     where the 65th bit means "uninitialized", or
>   - reserve a couple of extra bits for a type tag on every value, like
>     what some Lisp implementations do (e.g. GNU clisp).
> I think valgrind uses one of the latter two techniques as well.
>
> Bruno
>
> [1] https://en.wikipedia.org/wiki/Single-precision_floating-point_format
> [2] https://en.wikipedia.org/wiki/Double-precision_floating-point_format
>
>
>
>

[-- Attachment #2: Type: text/html, Size: 4498 bytes --]

Re: Signaling NaNs

[Bruno Haible]

Marc Nieper-Wißkirchen wrote:
> I could reformulate my question to: "Is there a way to produce a (quiet)
> NaN that won't occur as a result of the C library functions?"

There are 2^51 possible quiet NaNs [1]. I would guess that the hardware and
libc functions produce only a small portion of this space (because what
would be the point of encoding additional information in the result of
0.0 / 0.0, sqrt(-31.2), acos(-1.4), etc.?). But I don't really know.
You would need to find out yourself.

Bruno

[2] https://en.wikipedia.org/wiki/Double-precision_floating-point_format