raw March 18 Minutes
IEEE 754r Meeting minutes March 17, 2004
Sun, Santa Clara
Attendees (partial list): Affiliation
Dan Zuras Self
David Hough Sun
Eric Schwartz IBM (phone)
Alex Liu Sun
Joe Darcy Sun (host)
Jim Thomas HP
Mike Cowlishaw IBM (phone)
Ivan Godard OOTBC
Jeff Kidder Intel (minute taker)
Fred Tydeman Self
David Bindel UC Berkeley
Jason Riedy UCB
W Kahan UC Berkeley
Note taker: Jeff Kidder
Called to order 1:15
Would like to extend schedule through August.
Would like to confirm we have come to a
consensus on NaNs.
What are we going to do about extended.
Fixed mode declarations (ahead of Alt exception)
Jim, Jeff, Prof. Kahan, and Jason presented.
Substantial discussion ensued.
The weight of history and opinion seemed
to weigh against the merged QNaN approach.
There seemed to be stronger support for
specifying non-commutative NaN propagation.
Joe argued that we should provide an informative
appendix to record the desired end user capabilities.
Where is the killer ap for the specific features.
There are four or five papers. A few rounding
mode references. Should list related papers.
Where is xblas using the C99 features?
Jason pointed out that the language community
uses a different language to discuss the things
Use it or lose it.
Portability is equiv.
Any piece of code that uses the contents of a
NaN payload is operating outside the spec.
David's thesis is there are no portable way to
take advantage of signaling NaNs.
We could take the "shall" out of the body and put
shall not raise Invalid on SNaNs in Appendix N.
April 15, 10 am - 5 pm. Clem Meas - Arithmetica Redwood City
Draft review - 10-12 (with a hard stop at Noon)
NaN Issue - Jeff
Extended - Jim
Language Issues - Jason
Prior - Ivan
May 12,13 1-5, Jeff Kidder - Intel Santa Clara
June 16,17 1-5 Jim - HP Cupertino
July 21,22 1-5 Michael Parks - Sun Santa Clara
August 18,19 1-5 David Hough - Sun Menlo Park
Dan will ask for a 2 quarter extension.
The PAR for what we are doing is 754.
1. Leave NaNs as is in the body standard
2. Shall deliver one of the arguments, will have payload functions, and
In the body: you should return a NaN with the same payload. The
standard doesn't say which NaN should be produced. The standard doesn't
say which sign a resulting NaN will have. Extract and insert functions
shall be defined. Payload will be defined.
In appendix N: Payload shall be those one of the inputs. The standard
does not say which will be returned.
Yea - 7; No - 7; Abs -
Without the insert/extract.
Yea - 12; No - 4; Abs -
C99 has mechanisms to insert and extract.
Eric's objection: a bit manipulation. This bit manipulation would be
expected in the FP unit. Don't want to see bit type of operations put
into the standard.
Changing the sign bits.
SNaN Option: Deprecate Signaling NaNs
What about casts? ... [dinner ensued]