Julia Evansfloating point: NaN and infinity
simonbp@b0rk Though sqrt(-1) = NaN always annoys me, because it is a number, just not a real number
clar fon@b0rk this is actually why it's called NaN; it's not just one number, but 253 of them
Tanquist
Christ van Willegen
llewelly@b0rk
very good.
minor addendum: most programming languages have a specific function like isnan() which returns true if a floating point value is NaN, and related functions like isinf(), isnormal() and so forth.
Ayke van Laethem
Edzer Pebesma@b0rk R has a special NaN, called NA for "not available", it has different meaning and is propagated consistently
Keith Jenkins
Josh "cortex" Millard@kgjenkins @b0rk what's red and invisible? No tomatoes.
Trammell Hudson@b0rk one minor addition to the bit pattern section might be to label the sign bit for +/- Infinity.
@b0rk A maybe less important addition would be to mention that some languages that only have double variables use the large number of NaN's to implement typed pointers via "NaN Boxing".
Bart Groeneveld@th @b0rk I never heard of that. Here is a link for others who are wondering: https://piotrduperas.com/posts/nan-boxing
NaN boxing or how to make the world dynamic - Blog by Piotr Duperas
Have you ever wondered how dynamic typing, like in JS, works under the hood? There are some obvious solutions in this matter, however, there is also one which is more than brilliant.
ednl 🇪🇺@th @b0rk Yes! Perhaps that was explained on a previous page but I missed that info here. Another little thing is: the right edge of the underline of the exponent is a bit imprecise, looks like 10 bits (it is clear in combination with the text, though, and again, this info is probably on a previous page).
Alan@b0rk have you seen this yet? A fun example of using floats in a really weird way https://www.youtube.com/watch?v=5TFDG-y-EHs
NaN Gates and Flip FLOPS
felix (grayscale) 🐺@b0rk pretty cool. might also want to mention somewhere the difference between quiet NaNs and signaling NaNs?
davidlowryduda@b0rk I just noticed that you give complete alttext for these images. That's really wonderful!
@davidlowryduda I never know if it actually helps anyone, but maybe it does!
@b0rk things get _really_ interesting when you try to use floating point numbers as hashmap keys:
* NaN is unequal to itself, so a hashmap should allow putting many NaNs in a hashmap.
* Retrieving a value by NaN key should never return anything (because NaN != NaN).
* +0 and -0 have a different bit pattern, but are still equal so only one should be allowed (important for hashing).
Not all languages are so strict with float semantics in hashmaps, though.
https://research.swtch.com/randhash
Dr. Moritz Lehmann@b0rk still wondering why IEEE wasted a doubling in range for the large number of ambiguous NaN definitions. For FP64 and FP32 this usually is not a problem, but for FP16 it is.
Alexander Gerber@b0rk
My favorite:
"NaNs spread"
- as soon as one gets in, it gets everywhere
> reminds me of other "mechanics" e.g. in social life. 🙂
My favorite:
"NaNs spread"
- as soon as one gets in, it gets everywhere
> reminds me of other "mechanics" e.g. in social life. 🙂
JS Karate@b0rk TIL there are many, many NANs
Stuart Marks@b0rk Nice. Might be worth specifying this applies to IEEE 754. Some computers (VAX, IBM) have different FP formats.
Maybe also mention all comparisons with NaN are always false.
5.0 < NaN: false
5.0 > NaN: false
5.0 == NaN: false
Many people find it surprising that trichotomy doesn’t hold for NaN.