Hi martin
open a ticket so that we do not lose this point.
Stef
[...]
The current implementation of #sign is
self > 0 ifTrue: [^ 1].
(self < 0 or: [((self at: 1) bitShift: -31) = 1]) ifTrue: [^ -1].
^ 0
I'd propose factoring this into two simpler methods:
sign
self > 0 ifTrue: [^ 1].
self < 0 ifTrue: [^ -1].
^ 0
maybe self isNan ifTrue [^-1 raisedTo: self signBit], or the standard
tells it should be 0 too?
This addition makes sense to me.
The standards help a *little* and suggest on alternative.
ANSI Smalltalk acts like NaNs don't exist (basically, it lets
implementers do what they want for those values).
ANSI Smalltalk relies on ISO/IEC 10967 of 1994 for numerics (even when
it *cannot* rely on 10967, it does, sigh). That spec only defines a
"sign" operation for floats with a definite numeric value, and says that
operations are *permitted* to accept infinities and NaNs, but does not
require this, nor say what the answer should be.
The 2007 update of 10967 is somewhat more helpful. It replaces the
"sign" operation with one called "signum" which returns 1, -1, or NaN.
It returns 1 for positive zero and positive infinity, and -1 for
negative zero and negative infinity. If given a qNaN, it returns a qNaN.
If given a signaling NaN, it returns a qNaN but notifies the application
of an "invalid" situation.
what is a qNan?
If we extend this to Smalltalk, I *think* this would have us answer qNaN
for qNaN, and for a sNaN signal a resumable exception (a Notification,
perhaps?) that if resumed would answer qNaN. This also seems to make
some sense -- NaNs are supposed to propagate, and asking the just plain
"sign" (as opposed to signBit or isSignMinus) of a NaN is more or less
nonsense. But if you prefer (-1 raisedTo: self signBit) I won't complain.
That means that we'd have to refactor most (all?) senders of sign...
Certainly should audit senders of sign, but I'd expect that most senders
don't really care about what the sign of -0.0 or NaNs are. The answer
would remain the same for all other receivers.
Regards,
-Martin
--
Using Opera's mail client: http://www.opera.com/mail/
