0.0 and -0.0 compare equal but have different hash values

Description

0.0 and -0.0 compare as equal but have different hash values:

This causes problems as the equality/hashing assumption is violated.

Cause: The source of this is due to some differences in Java. Java primitive double 0.0 and -0.0 == but the boxed Double is NOT .equals(). See also: http://docs.oracle.com/javase/7/docs/api/java/lang/Double.html#equals%28java.lang.Object%29

Double equality is checked with == in Clojure, which will report true. Hashing falls through to .hashCode(), which returns different values (but is consistent with the .equals() result on the boxed form).

Approach: While there are times when 0.0 and -0.0 being different are useful (see background below), most Clojure users expect these to compare equal. IEEE 754 says that they should compare as equals as well. So the approach to take here is to leave them as equal but to modify the hash for -0.0 to be the same as 0.0 so that `=` and `hash` are consistent. The attached patch takes this approach.

Patch: clj-1860-2.patch

Screened: Alex Miller

Alternative: Make 0.0 != -0.0. This approach affects a much larger set of code as comparison operators etc may be affected. The patch clj-1860-make-equals-false-for-pos-neg-0.0-v1.patch may be one way to implement this approach, and seems fairly small in the quantity of code affected (2 methods).

Background: https://en.wikipedia.org/wiki/Signed_zero

Environment

None

Activity

Show:
Mike Anderson
February 26, 2016, 1:44 AM

I may be late to the party since I have only just seen this, but I have a strong belief that 0.0 and -0.0 should be == but not =.

Reasons:

  • Anyone doing numerical comparison should use ==, so you want the result to be true

  • Anyone doing value comparison should use =, so you want the result to be false because these are different IEE784 double values. This includes set membership tests etc.

i.e. the Java code is doing it right, and we should be consistent with this.

Alex Miller
February 26, 2016, 5:52 AM

I think there is consensus that == should be true, so we can set that aside and focus on =.

The reality is that there is no easy way to compare two collections with == (for example comparing [5.0 0.0 1.0] and [5.0 -0.0 1.0]). This is an actual use case that has been problematic for multiple people. While I grant there are use cases where 0.0 and -0.0 are usefully differentiable, I do not know of a real case in the community where this is the desired behavior, so I would rather err on the side of satisfying the intuition of the larger (and currently affected) population.

Also note that the Java code is doing it BOTH ways (primitive doubles are equal, boxed doubles are not), so I think that's a weak argument.

Alex Miller
February 26, 2016, 2:44 PM

(after sleeping more on this...) It's possible that a better answer here is to expand what can be done with ==. I trust that when Rich looks at this ticket he will have his opinions which may or may not match up to mine and if so, we'll go in a different direction.

Andy Fingerhut
February 27, 2016, 4:15 PM

Attachment clj-1860-make-equals-false-for-pos-neg-0.0-v1.patch dated Feb 27 2016 is a first cut at implementing a change where = returns false when comparing positive and negative 0.0, float or double.

As far as I can tell, there is no notion of positive and negative 0 for BigDecimal, so no change in behavior there.

Alex Miller
November 17, 2016, 10:16 PM

Rich says: "0.0=-0.0, make hash the same"

Completed

Assignee

Unassigned

Reporter

import

Labels

Approval

Ok

Patch

Code and Test

Fix versions

Affects versions

Priority

Minor
Configure