[cfp-interest 3702] Comments on proposal for Annex F special cases

[Jim Thomas]

For Action

> Group: Review Damian's coming Annex F changes and study pow and others which weren't covered in the meeting, for lack of time.


below are some comments and suggestions for

https://wiki.edg.com/pub/CFP/WebHome/C2Y-F-SpecialCases-20251030.pdf


F.10.2.4#1 and F.10.2.11#1
-0-: Change minus symbol to hyphen. It’s part of a C expression.

NOTE: 
C doesn’t use a colon after “NOTE”. 
“Likewise“ doesn’t seem helpful.

F.10.2.12#1 and F.10.2.13#1
 -2-: Remove “all”. There’s no “all” in “for even integers"

F.10.4.7#2 
Per your suggestion in offline email discussion, change to
   NOTE: No floating point exception is raised by frexp(x, p) for all x (including NaN).
Annex F generally refers to arguments by the parameter name in the prototype, which for frexp is “value”. I don’t have a suggestion for how to do that here.

F.10.5.2#1 
Fred asked about adding compound(0,n) returns 1. What other exact results, that 754 doesn’t mention because they are implied by correct rounding, are missing in Annex F? This is a separate issue, better not included in here. 

F.10.5.4#1 
Suggest:
-3- hypot(+/-inf, y) returns +inf for all y (including NaN).
-4- hypot(x, NaN) returns a NaN for -inf < x < +inf.
It follows the thinking: hypot(+/-inf, y) returns +inf for all y (including NaN), but there’s the usual NaN behavior for other first arguments. The following would also work:
-3- hypot(x, +/-inf) returns +inf for all x (including NaN).
-4- hypot(NaN, y) returns a NaN for -inf < y < +inf.

F.10.5.5#1 
All cases: Remove “is” after “for”, or change “for” to “if” or “when”.
-4- Given the explicit style, maybe use “+inf” instead of “inf”, e.g. in 0 < y < +inf.
-5- Change “y > 0” to match -4-.’s-7- Why “+1" instead of “1”?
-17- Change “y > 0” to match -16-.

F.10.5.6#1
Use +/-inf to match +/-0 cases and reduce number of new cases. E.g.
pown(+/-inf,n) returns +/-inf … odd n > 0
…
Or expand +/-0 cases to match new +/-inf ones.
Maybe add a NOTE with the deleted statement of equivalence. Only if there are not a lot of similar cases.

F.10.7.3#1
Maybe delete “for any rounding direction mode”.

F.10.7.9#!
I believe the Annex F specs for roundeven, round, trunc, ceil and floor should all be the same (except for the examples and references to examples). trunc seems to be the one out of line. Similar considerations might apply in 7.12.10, but I think that should be a separate investigation.

F.10.7.10#1
The “or” makes it seem like they’re the same function. Prefer “and” (and plural verbs).

F.10.8.1#1 and F.10.8.2#1
-1- saying "for y < 0 or y > 0” might avoid wondering about NaNs.

F.10.9.3#1
-1- Prefer combined exceptions (per current).
-1- Change to “and the result is infinite.”
-2- Change to “and the result is subnormal or zero.”

F.10.9.7#1
The rewrite loses the requirement that the functions successfully produce a canonical result (for valid pointers).
-1- The quiet NaN returned must be the canonical version of the input quiet NaN.
-3- “the canonicalized version of the signaling NaN that was made quiet” sounds like a signaling NaN. It's intended to be the canonical version of the quiet NaN resulting from quieting the signaling NaN input. 
-3- I believe the footnote was intended to apply generally, not just to quieting a signaling NaN. Fred?
-4- Because the function specification covers signaling NaNs explicitly, it would be better to qualify all references to “NaN”.

General comments follow:

Unchanged cases need to be checked for the prescribed form.

For Rationale for a WG14 proposal, I’d suggest casting it as improving Annex F by using a consistent form for special cases. Avoid subjective terms like “confusing" and “cluttered" and phrases like “is now easier to read and understand”. Claim more consistent punctuation, not "punctuation corrections” - there might be corrections, but we haven’t (to my knowledge) consulted ISO guidelines regarding punctuation.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20251106/89603cee/attachment.htm>