[cfp-interest 4026] Re: C FP Paper results in the 2026/11/16 WG14 meeting

Damian McGuckin damianm at esi.com.au
Tue Nov 17 22:45:08 PST 2026


On Tue, 17 Nov 2026, RAJAN BHAKTA wrote:

> N3877: Sign of 0 returned by fromfp
>     Cranmer: LLVM and libc return -0.
>     Ballman: Michael Jones from LLVM libc said that this seems fine but not
> sure about the -0 to +0 change.
>     
>     ^Decision poll: Put N3877 (sign of 0 from fromfp) into C2Y?
>       4+8/0+0/0+4. 12/0/4. Goes into C2Y.
> 
> N3962: Clarification of requirements in Annex H
>     ^Decision poll: Put N3962 (Annex H fixes) into C2Y?
>       3+9/0+0/0+2. 12/0/2. Goes into C2Y.
>    
> N3943: Annex F: Special cases rework
>     Thiago: Using math symbol for infinite. Should we use INF instead of
> NaN?

The use of mathematics was deliberate.

>     Banham: The "y a quiet NaN" should be worded better. Add "is".

As Jim pointed out a while ago, "y a quiet NaN" is mathematical. If you 
want more colloquial, add "is". What is the consensus.

>     ^Joseph: [SC22WG14.38450] comments for the paper.
> ??????1. Additional statements are added for atan2 and atan2pi, "atan2 (?y, x) and -atan2
> (y, x) are equivalent" and "atan2pi (?y, x) and -atan2pi (y, x) are
> equivalent", described as "[Not a new requirement - see F.10.1#3]".  But
> such equivalence for an odd function is only true in certain rounding modes,
> not for rounding upward or downward.

I have simply copied what C23 says. What rule do we want to enforce?

> ??????2. ilogb is 7.12.7.8 not 7.12.6.8 (the edits change a correct 
> subclause number to an incorrect one).

Sorry, fat fingers. Fixed/
> 
>     ^Opinion poll: Want something like N3943 (Annex F special cases rework)
> with Josephs comments addressed in C2Y?
>       2+14/0+0/1+3. 16/0/4. Clear direction.

GOod to hear.
>     
> N3942: Annex F: Introduction changes
>     ^Joseph:: SC22WG14.38428 has my comments
> ??????1. I don't think the new table headings, describing them as "Mandated" /
> "Recommended" operations, are accurate - aren't the maximum / minimum
> operations actually recommended ones (albeit ones deliberately listed in the
> first table and required to give exact results)?  (Likewise the changes to
> text in F.3 about what each table contains.)

I will change. Should I say

 	Mandated operations + Minimum/Maximum operations

> ??????2. There's a reference to DEC_RADIX, but no such macro exists.

It will soon.  Please see my draft paper on the Wiki.

> ??????3. Thesecond reference to convertToIntegerTiesToEven should be
> convertToIntegerExactTiesToEven.

Will fix.

> ??????4. "(of a real floating type)" in reference to FLT_RADIX isn't 
> correct since the decimal floating types are also real floating types; 
> probably "(of a standard floating type)" is meant.

Will fix.

> ??????5. I'm not convinced by the changes in F.10.1 to refer to table 
> F.4 instead of table F.1 - I think all those references were correctly 
> about the main operation binding table, not about the subsequent one 
> (F.4 after these changes) for operations that are not required to be 
> correctly rounded, and so should not be changed.

Needs more thought.

>     ^Opinion poll: Want something like N3942 (Annex F introduction rework)
> in C2Y?
>       17/0/2. Clear direction.

Thanks.

>     Joseph: For printf SNAN, the ambiguity is for the convertFormat vs copy
> operation as well.
>     Cranmer: The IEEE functions convert to character don't give enough
> information either.
>     Ballman: I don't want SNaN to be handled differently from varargs.
>     Cranmer: vararg has float converted to double due to default argument
> promotions.

I would like a bit more explanation please.

> N3947: nexttoward and nextafter
>     ^CFP: Look at WG14 message 38263
> (https://www.open-std.org/jtc1/sc22/wg14/38263).
>     Joseph: Message 38263.
>     Joseph: For a macro, what happens? And for evaluation formats, it talks
> about casts but not clear what happens with extra range and precision.
>     Ballman: Will we break users by doing this?
>     Rajan: Possibly. But this is fixing a wrong result in the standard.
>     Joseph: Not sure this is forward progress since it loses the removal of
> excess range and precision is a problem.

I would like a bit more explanation please.

Thanks - Damian


More information about the cfp-interest mailing list