[Cfp-interest 1352] WG14 IEEE 754-C binding meeting minutes 2019/07/17
Rajan Bhakta
rbhakta at us.ibm.com
Wed Jul 17 10:09:58 PDT 2019
Attendees: Rajan, Jim, Fred, Mike, David H., Ian,
New agenda items:
None.
Carry over action items:
Fred: Give a new version of the SNAN initialization paper (as per
CFP1316). - Done.
Last meeting action items:
Jim: Create a link to the 250 draft into the references section in the
C FP wiki. - Done.
Rajan: Forward the IEEE article to WG14 once David H sends it out to
us. - Not done.
David: OK to send the draft even though it is not final and has been
changed.
Jim: Draft a slide deck and a proposal based on CFP1331. - Partially
done. Slide deck to carry over. Proposal done.
Jim: Draft a note to warn about CFP1315's rounding of negative
constants issue. - Done.
New action items:
Jim: Draft a slide deck based on CFP1331.
Jim: Add wording to the CFP1331 proposal make it clear this is for
particular evaluation methods, and submit the document to W14.
Jim: Add a statement as to why there is a second name for log1p as a
footnote as a new proposal.
Fred: Submit CFP 1340 to WG14.
Jim: Reword CFP 1337 as per action item discussion in the 2019/07/17
meeting.
Jim: Send CFP 1349 to the WG14 reflector as the CFP position.
Fred: Write a paper to make the range error for small nonzero x
consistent for expm1, logp1, log10p1's other base functions.
Jim: Add in the normalized discussion from Fred (CFP 1341, CFP 1342)
to the agenda for next meeting.
Next Meeting(s):
Wednesday, August 21st, 2019, 11:00 EST, 8:00 PST, 4PM UTC
Same teleconference number.
Please notify the group if this time slot does not work.
Discussion:
754 revision:
Should be done. Likely to be published on Monday. Draft 263 seems to
be good.
Likely one more meeting to discuss the background document,
repository and maintenance, hand over to any future group for a future
revision.
Mike will send the 263 pdf for posting on the CFP wiki.
Jim: Specifying identities for math functions was something I wanted
to ask the 754 group. Will bring it up there.
C++ Liaison:
Nothing new.ok
C2X integration:
Draft with TS 1, 2, and 4a:
http://wiki.edg.com/pub/CFP/WebHome/all-20190708.pdf
Part 3 – as annex, given to Jens:
http://wiki.edg.com/pub/CFP/WebHome/n2405.pdf
Part 4b - Looking as an updated TS.
Part 5a,b,c,d – Discuss later. Part 5d is a TS update.
Action item details:
Fred: Give a new version of the SNAN initialization paper (as per
CFP1316). See Fred’s CFP 1340.
*Fred: Submit this paper to WG14
Jim: Draft a slide deck and a proposal based on CFP1331.
http://wiki.edg.com/pub/CFP/WebHome/C2x_proposal_-_TS_18661-5abc-20190709.pdf
Mike: Where it says evaluation to _Decimal64 is it greater than or
equal to or only _Decimal64?
Jim: It is about _Decimal{32,64} being _Decimal64 while larger is
to the larger ones like _Decimal128 -> _Decimal128.
Jim: Requires support for these evaluation methods. We could add
this.
Fred: Or just say required to have _Decimal32 evaluated as
_Decimal64.
Jim: But that doesn't say what _Decimal64 is evaluated to.
Jim: Do we need to make it clearer that we are referring to
particular evaluation methods?
Fred/Mike: Yes.
Ian: Does the result have to be rounded correctly to _Decimal32?
Jim/Fred: No.
Jim: This does not preclude other evaluation methods. They have to
have at least this one, but can have others.
*AI*: Jim: Add wording to the CFP1331 proposal make it clear this is
for particular evaluation methods and submit to WG14.
Jim: Draft a note to warn about CFP1315's rounding of negative
constants issue. See Jim’s CFP 1337.
There is one correction in the last two sentences to have precision
to be precision and range.
Rajan: There could be other rounding modes so we should not list
them as the only ways. Perhaps say "rounding modes such as ..." for each
of the same result and different result cases. Also the part of exact
results should be a separate note or a separate paragraph.
Fred: We should also not assume correct rounding.
Other issues
Fred’s WG 14 papers (WG14 email thread “N2380: printf of NaN()”) - See
Jim’s CFP 1349
*Jim: Send CFP 1349 to the WG14 reflector as the CFP position.
Issues raised by Jens
Approach is define prefixes and reserve names under those prefixes.
Jim/Ian: The different styles with existing C18 names and these
new ones with similar functionality is not good.
WANT macros are not sufficient since the names could already be in
the library independently of the user source WANT macro definition.
Jim: How much is this a real problem? Is it not solved by the
compiler and linker people?
Jim: Is there a namespace issue with the operators and types for
your library Mike?
Mike: Not that I've come across
Jim: Do language providers have application test suites that would
fail if these identifiers were added to math.h?
David H: It is hard to add 1700 identifiers.
Ian: Ideally you'd want some major customers involved as well.
Should we have a clear statement of the pros and cons of the WANT
macros?
Fred: For GCC, I can specify a flag that works on the compiler and
linker and automatically links in all the libraries for me.
Jim: Was there consideration of allowing includes with finer
granularity.
Rajan: Yes, but no desire for it in WG14 when we discussed it in
London.
Naming of correctly rounded math functions.
No real issue with using cr_, we only did it cr to match existing
C standard text.
Fred: Prefer the _.
We can support this if Jens proposes it.
Obsolescing log1p - See Jim’s CFP 1348
Ian: I got an email complaining about this as it would break their
application and they don't have source code that can be recompiled.
Jim: It would be good to add a statement as to why there is a
second name as a footnote.
Specifying more special cases for math functions, e.g., periodicity
for half-revolution trig functions.
Perhaps as recommended practice.
Issues with +/-0, NaNs, etc. and other identities.
Fred: I would like having the identities being added and requiring
sin(30 degrees) being exactly a half.
Jim: How would we come up with a list?
Ian: We could ask Robert Enenkel at IBM.
David H: I don't think we want to copy the entire IEEE standard
into the C standard. C still caters to non-IEEE implementations.
Jim: All IEEE says is to correctly round, and if you can't do
that, this doesn't help.
Keep on the agenda since this right now is QoI.
Putting the half-revolution trig functions into their own subclause.
No issue with this.
Range error may occur if nonzero x is too small for expm1.
Fred: We should address this.
Rajan: It is covered under the general statement about allowing
other range errors, but we can make it clearer or consistent with the
other exp*10* functions.
Result: Look into doing something for this to make it more
consistent.
Fred: exp10m1 has finite while exp2m1 doesn't. It is only a
problem for a positive large number, not just any large number.
Rajan: Related to DR40?
Range error may occur if nonzero x is too small for logp1 and
log10p1.
Result: Look into doing something for this to make it more
consistent.
*Fred: Write a paper to make the range error for small nonzero x
consistent for expm1, logp1, log10p1's other base functions.
Action items from WG14 London meeting:
C FP: Give 18661 part 4a (not reduction functions) for inclusion
into C2X (http://wiki.edg.com/pub/CFP/WebHome/n2401.pdf)
Done.
C FP: Put N2309 into TS 18661-4 and C2X (
http://wiki.edg.com/pub/CFP/WebHome/n2401.pdf)
Done.
TS DR13: Move to C2X (C FP action item)
Done.
TS DR16: Move to C2X (C FP action item).
Done.
TS DR20-25: Move to C2X (C FP action item).
Done.
CFP compendium - See Jim’s CFP 1332
Plan looks good.
Other items:
CFP 1341: FP_NORMAL.
Besides normalized finite numbers, FP_NORMAL may have other kinds
such as unnormals.
Fred: DFP has a lot of finite numbers that are not normalized.
Jim: There is a problem with the term normalized. I may have been
better saying normal. Unnormals may resolve to sub-normals if it was
attempted to be normalized. There can be unnormalized representations of
normal numbers.
David H: Maybe the footnote needs to say unnormal numbers can be
FP_NORMAL or FP_SUBNORMAL depending on their value.
Fred: Will go back to look at the test around normal numbers and
come back here.
*AI: Add in the normalized discussion from Fred to the agenda for next
meeting.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20190717/f4b3b76b/attachment-0001.html
More information about the Cfp-interest
mailing list