[cfp-interest 3657] Re: C2Y's canonicalise() in AnnexF - F.10.9.7

[Damian McGuckin]

I suggested:

>>> F.10.9.7#1 The canonicalize functions
>>>
>>> - canonicalize (cx, x) returns a quiet NaN in the object pointed to by cx
>>> and raises the "invalid" floating-point exception when the object pointed
>>> to by x is a signaling NaN .
>>> - canonicalize (cx, x) returns a quiet NaN in the object pointed to by cx
>>> when the object pointed to by x is a quiet NaN .

On Tue, 14 Oct 2025, Jim Thomas wrote:
>
> Not a rewording. Clause 7 does not require the function to be 
> successful: simply returning 1 would be a valid implementation. F.10.9.7 
> requires the function to store a canonical representation, per ISO/IEC 
> 60559, and return 0.
>
> The suggestion above omits the recommendation that signaling NaN input 
> result in a quiet NaN "which should be the canonical version of that 
> signaling NaN made quiet?.

I believe you are saying that I need to add words to the affect that:

a)	F.10.9.7 always stores a canonical represention per ISO/IEC 60559,
 	i.e. it must always be successful,

b)	the recommendation that signaling NaN input result in a quiet NaN
 	"which should be the canonical version of that signaling NaN made
 	quiet".

If the pointer is invalid, the routine will fail so I cannot see how I can 
uncategorically say (a). What about a qualification in a second paragraph 
like:

 	In the event that cx is a valid pointer, ISO/IEC 60559 mandates
 	that the canonical representation MUST be returned.

As far as (b) is concerned, I read Section 7 (see below) and I thought 
that this has been said already. If Section 7 does not say this, should 
not that recommendation belong in Section 7, not Annex F? And what words 
do you suggest?

Thanks - Damian

This is Section 7:

The canonicalize functions attempt to produce a canonical version of
the floating-point representation in the object pointed to by the
argument x , as if to a temporary object of the specified type, and
store the canonical result in the object pointed to by the argument cx
. 287) If the input *x is a signaling NaN, the canonicalize functions
are intended to store a canonical quiet NaN. If a canonical result is
not produced the object pointed to by cx is unchanged.

This is Annex F

The canonicalize functions produce 451) the canonical version of the
representation in the object pointed to by the argument x . If the input
*x is a signaling NaN, the "invalid" floating-point exception is raised
and a (canonical) quiet NaN (which should be the canonical version of
that signaling NaN made quiet) is produced.  For quiet NaN, infinity,
and finite inputs, the functions raise no floating-point exceptions.

Looking at Annex F:

> The canonicalize functions produce the canonical version of the
> representation in the object pointed to by the argument x.

This is already said in Section 7, albeit qualified by the word "attempt"

> If the input *x is a signaling NaN, the "invalid" floating-point exception
> is raised ...

This is normative - why say it again.

> For quiet NaN, infinity, and finite inputs, the functions raise no
> floating-point exceptions.

This is normative - why say it again.

Thanks - Damian