Unexpected behaviour if BigDecimal.new("NaN") is used as a key of hash

[yui-knk]

$ ruby -v
ruby 2.4.0preview3 (2016-11-07 trunk 56661) [x86_64-darwin14]

irb(main):001:0> require "bigdecimal"
=> true
irb(main):002:0>  nan = BigDecimal.new("NaN")
=> #<BigDecimal:7ff5f213bad8,'NaN',9(9)>
irb(main):003:0> h = {nan => 1000}
=> {#<BigDecimal:7ff5f213bad8,'NaN',9(9)>=>1000}
irb(main):004:0> h[BigDecimal.new("NaN")]
=> nil
irb(main):005:0> h[nan]
=> 1000

I expected h[nan] to return nil, because nan.eql? nan returns false.
This is because hash compare pointers first.
https://github.com/ruby/ruby/blob/v2_4_0_preview3/hash.c#L99-L102

[mrkn]

NaN is difficult even if it is a Float.

irb(main):001:0> a = Float::NAN
=> NaN
irb(main):002:0> b = 0.0 / 0.0
=> NaN
irb(main):003:0> c = Float::INFINITY * 0.0
=> NaN
irb(main):004:0> h = { a => 1, b => 2, c => 3 }
=> {NaN=>1, NaN=>2, NaN=>3}
irb(main):005:0> h[a]
=> 1
irb(main):006:0> h[b]
=> 2
irb(main):007:0> h[c]
=> 3

[mrkn]

IMO at least we need to discuss how to define the equivalence of NaN including Float rather than only BigDecimal in bugs.ruby-lang.org.

[yui-knk]

https://bugs.ruby-lang.org/issues/13146 discusses about Float NaN.

[mrkn]

https://bugs.ruby-lang.org/issues/13146 was rejected.