Issue 23563 in dart: double.operator-() is unstable

[da...@googlecode.com]

Status: Triaged
Owner: ----
Labels: Type-Defect Priority-Unassigned Area-VM

New issue 23563 by ajoh...@google.com: double.operator-() is unstable
https://code.google.com/p/dart/issues/detail?id=23563

The following:

print(identical(double.NAN, -double.NAN));

will print true at first, but false when optimized.

Note that the initial behavior is unexpected, as the compare is bitwise but
the sign-bit is not changed.


--
You received this message because this project is configured to send all
issue notifications to this address.
You may adjust your notification preferences at:
https://code.google.com/hosting/settings

[da...@googlecode.com]

Updates:
Status: Accepted
Owner: srd...@google.com

Comment #1 on issue 23563 by srd...@google.com: double.operator-() is
unstable
https://code.google.com/p/dart/issues/detail?id=23563

(No comment was entered for this change.)

[da...@googlecode.com]

Updates:
Cc: gbr...@google.com ko...@google.com

Comment #2 on issue 23563 by srd...@google.com: double.operator-() is
unstable
https://code.google.com/p/dart/issues/detail?id=23563

It seems to me that similarly to 0.0, unary minus must treat NaN-s
specially so that double.Nan & -double.NaN are not identical

[da...@googlecode.com]

Comment #3 on issue 23563 by ko...@google.com: double.operator-() is unstable
https://code.google.com/p/dart/issues/detail?id=23563

Quotes directly from the IEEE 754-1985 standard
http://homepages.math.uic.edu/~jan/mcs471/Lec3/ieee754.pdf
(which is the only official full text version I have found so far):

"Every operation involving one or two input NaNs, none of them signaling,
shall signal no exception but, if a floating-point result is to be
delivered,
shall deliver as its result a quiet NaN, which should be one of the input
NaNs."

This seems to require -x to "one of the input NaNs", and thus, it must be
exactly x.

Furthermore, it states: "This standard does not interpret the sign of an
NaN"

[da...@googlecode.com]

Updates:
Cc: floi...@google.com

Comment #4 on issue 23563 by ajoh...@google.com: double.operator-() is
unstable
https://code.google.com/p/dart/issues/detail?id=23563

+floitsch

I'm not sure how to interpret it, but it says 'should', not 'must'. Also,
the spec does not contain information about negate, only add, minus, mul,
div and mod.

[da...@googlecode.com]

Comment #5 on issue 23563 by ko...@google.com: double.operator-() is unstable
https://code.google.com/p/dart/issues/detail?id=23563

Scratch all that, there's a newer version of the standard, which goes into
more detail.

http://754r.ucbtest.org/drafts/archive/2006-10-04.pdf

"
7.5.1 Sign operations
7.5.1.0
Implementations shall provide the following homogeneous quiet-computational
sign operations for all
supported non-storage floating-point formats. They might propagate
non-canonical encodings. They are
performed as if on strings of bits, treating floating-point numbers and
NaNs alike, and hence signal no
exception.
...
negate(x) copies a floating-point operand x to a destination in the same
format, reversing the sign.
0–x is not the same as –x or negate(x).
"

Note that very last comment... so I'm still not sure what is
mandated/allowed for "-x" when x is a quiet NaN.

[da...@googlecode.com]

Comment #6 on issue 23563 by ajoh...@google.com: double.operator-() is
unstable
https://code.google.com/p/dart/issues/detail?id=23563

So basically copy and swap the sign bit. To me that sounds like optimized
version is correct so that returning a NaN with the sign-bit reversed is
the right solution.

[da...@googlecode.com]

Comment #7 on issue 23563 by srd...@google.com: double.operator-() is
unstable
https://code.google.com/p/dart/issues/detail?id=23563

The agreement is that identical(double.NaN, -double.NaN) should always
return false.

[da...@googlecode.com]

Comment #8 on issue 23563 by srd...@google.com: double.operator-() is
unstable
https://code.google.com/p/dart/issues/detail?id=23563

There is a problem of implementing unary- for NaNs. double.isNegative
always returns false for any nan, i.e., it is not possible to implement
unary- in a similar way as for 0.0.

I am puzzled why double.isNegative does not work same as C-s 'signbit'. Any
insight?