This is usually called NaN-boxing and is often used to implement dynamic languages.
I wonder if IEEE-754 designers anticipated this use case during development. Great article - this kind of "outside-the-normal-use-case" requires a very careful reading of the specifications/guarantees of the language.
IEEE-754 seems to say that NaN payloads are designed to contain "retrospective diagnostic information inherited from invalid or unavailable data and results". While NaN boxing in particular probably wasn't the intention, untouched payloads in general absolutely were.
It's especially useful since everything is 64 bits uniformly.
I appreciate this article a lot more because it contains an iota of benchmarking with an explanation about why this might be more performant. Especially since my first thought was 'wouldn't this require more instructions to be executed?'
The original post seems really weird to me. I would have dismissed it as someone's hobby project, but... that doesn't seem like what it's trying to be.
"More instructions to execute" is not synonymous with "slower".
NaNboxing lets you use less memory on certain use cases. Because memory access is slow and caches are fixed size, this is usually a performance win, even if you have to do a few extra ops on every access.
It's a joke project.
This doesn't work on Firefox, as it normalizes NaNs as they're extracted from ArrayBuffers. Presumably because SpiderMonkey uses NaN-boxing itself, and thus just doesn't have any way to represent actual non-canonical NaN floats.
The spec mandates normalization of NaNs in ArrayBuffers. If other engines do not normalize, I believe it's a bug in those engines!
Chromium & Node (i.e. V8) do seem to just not normalize:
arr = new Float64Array([0]);
arr[0] = new Float64Array(new BigUint64Array([-123n]).buffer)[0];
[...new Uint8Array(arr.buffer)] // [133, 255, 255, 255, 255, 255, 255, 255]Running the following snippets:
f = new Float64Array(1)
dv = new DataView(f.buffer)
dv.setBigInt64(0, -123n)
console.log([...new Uint8Array(f.buffer)])
f = new Float64Array(1)
f[0] = new Float64Array(new BigUint64Array([-123n]).buffer)[0]
console.log([...new Uint8Array(f.buffer)])
f = new Float64Array([NaN])
console.log([...new Uint8Array(f.buffer)])
[255, 255, 255, 255, 255, 255, 255, 133]
[133, 255, 255, 255, 255, 255, 255, 255]
[0, 0, 0, 0, 0, 0, 248, 127]
[ 255, 255, 255, 255, 255, 255, 255, 133 ]
[ 0, 0, 0, 0, 0, 0, 248, 127 ]
[ 0, 0, 0, 0, 0, 0, 248, 127 ]
Dammit Mozilla, first no WebGPU, now this?! /s
I made a garlic nan: https://www.godbolt.org/z/enjv1c7Tf
I love it, how did I not think about this!
Note that there are a lot of exciting new float-related functions in C23, many related to NaN handling, signed zeros, or precision in edge cases. Some have been in glibc for years; others not yet.
The API for `getpayload`/`setpayload`/`setpayloadsig` looks a little funny but it's easy enough to wrap (just consider the edge cases; in particular remember that whether 0 is valid for quiet or signaling NaNs is platform-dependent).
Finally we have a reliable `roundeven`, and the convenience of directly calling `nextup`/`nextdown` (but note that you'll still only visit one of the zeros).
The new `fmaximum` family is confusing, but I think I have it clear:
without 'imum' (legacy): prefer non-NaN
with 'imum': zeros distinguished (-0 < +0)
with '_num': prefer a non-NaN, also signal if applicable
with '_mag': compare as if fabs first
`rootn` and `compoundn` are surprisingly tricky to implement yourself. I still have one testcase I'm not sure why I have to hand-patch, and I'm not sure which implementation choice keeps the best precision (to say nothing of correct rounding mode).
> but note that you'll still only visit one of the zeros
Out of curiosity, isn't this what you'd pretty much universally want? Or maybe I misunderstand the intent here?
I find `nextup/`nextdown` useful for when you want to calculate `{f(nextdown(x)), f(x), f(nextup(x))}` and see if some property holds (or to what degree it holds) for both `x` and its surrounding values. It's annoying that one such surrounding value gets skipped.
That said, quite often you need to handle signed zero specially already, and if you don't the property you're testing might sufficiently differ in the interesting way for the smallest-magnitude subnormals.
Since the first time I had Indian food sometime in 1987 I have always called naan “not a number bread” and no one has ever laughed. I feel like i may have found my people.
This may appeal: https://doubleyoudoubledoubleewe.www.dash.deeohhtee.dash-das...
Four NaNs? Four? That's insane.
I'm curious why we have not-a-number, but not not-a-string, not-a-boolean, not-an-enum, and not-a-mycustomtype.
Because NaNs come from a standardized hardware-supported type, whereas the rest of those are largely language-specific (and you could consider null/nil as a "not-a-*" type for those in applicable languages; and there are languages which disallow NaN floats too, which completes all combinations).
Itanium had a bit for "not a thing" for integers (and perhaps some older hardware from around the time floats started being a thing had similar things), so the idea of hardware support for not-a-* isn't exclusive to floats, but evidently this hasn't caught on; generally it's messy because it needs a bit pattern to yoink, but many types already use all possible ones (whereas floats already needed to chop out some for infinities).
Because IEEE 754 creators wanted to signal non-trapping error conditions for mathematically undefined operations, and they had a plenty of bit patterns to spare. Apparently back in the 70s and 80s in many cases it was preferable for a computation to go through and produce NaNs rather than trapping instantly when executing an undefined operation. I'm not quite sure what the reasoning was exactly.
In early FP machines the floating point processor could not take a trap at a faulting instruction precisely: it could only go bad things. Furthermore for programmers and hardware it can be very expensive. Rather than go through a loop and filter NaN out of results it becomes trap every time and resume and is a pain.
It avoids traps which are really inconvenient.
You can encode not-a-bool, not-a-(utf-8)-string and not-an-enum using one of the invalid bit patterns – that's exactly what the Rust compiler can do with its "niche optimization": https://www.0xatticus.com/posts/understanding_rust_niche/
Not-a-boolean would be something like the much maligned tri-state bool pattern: https://thedailywtf.com/articles/What_Is_Truth_0x3f_
Nana basically means that floating point arithmetic is predicting that your mathematical expression is an "indeterminate form", as in the thing you learn in calculus.
C++ has std::optional for exactly that purpose.
Some common numeric operations can result in non-numbers(eg division by zero - Nan or infinity).
Are there any common string operations with similar behavior?
Out of range substring? Some languages throw an error, others return an empty string. You could return a propagating NaS instead. I don't know what you'd use it for.
Charset translation.
Unicode’s � is basically a symbol for not-a-char.
Because representing infinity is not possible outside of symbolic logic and isn’t encodable in floats. I think it is a simple numerical reason and not a deeper computer reason.
Well, infinity is totally representable with IEEE 754 floats. For example 1.0/0.0 == +inf, -1.0/0.0 == -inf, but 0.0/0.0 == NaN.
A smart compiler should be able to figure out a better value for 0/0, depending on context.
For example:
for i in range(0, 10):
print(i/0.0)
But:
for i in range(-10, 10):
print(i/0.0)
And:
for i in range(-10, 10):
print(i/i)
Well, it could, but that would be against the spec. The hardware implements IEEE 754, most languages guarantee IEEE 754, and transforming code so that 0.0/0.0 doesn't result in NaN would be invalid.
I'm incorrect. there's a spec discussed here
https://www.gnu.org/software/libc/manual/html_node/Infinity-...
You missed not-a-NaN.
I learned about this when trying to decode data from Firefox IndexedDB. (I was extracting Tana data.) Their structured clone data format uses nan-boxing for serialization.
I guess I just don't get it because the before & after cases don't seem to be showing remotely comparable use cases.
In the "before" use-cases, the programmer has accidentally written a bug and gets back a single NaN from a logical operation they performed as a result.
In the "after" cases, the programmer already has some data and explicitly decides to convert it to a NaN encoding. But instead of actually getting back a NaN that is secretly carrying their data, they actually get a whole array of NaN's, which duck type differently, and thus are likely to disappear into another NaN or an undefined if they are propagated through an API.
Like.. I get that the whole thing is supposed to be humorously absurd but... It just doesn't land unless there's something technical I'm missing to connect up the pieces.
It sounds like you might be interested in the Enterprise Edition
The nan bits were intended to help provide more informative float error diagnostics at the language implementation level for non crashing code. I wonder if I’ll ever get to exploring it myself.
Reminds me of using the highest bit(s) of 64-bit ints to stuff auxiliary data into lockfree algorithms. So long as you’re aware of your OS environment you can enable some efficiencies you couldn’t otherwise.
That's why OCaml has 63 bit integers by default: the language itself uses the last bit to help with GC.
That's curious! Does anyone know why the spec was designed to allow so many possible NaN bit patterns? Seems like just one would do!
As long as you want all bit patterns to be Some float, there's not really one bit pattern you can chop out of somewhere (or, three, rather - both infinities, and NaN).
Taking, say, the 3 smallest subnormal numbers, or the three largest numbers, or whatever, would be extremely bad for allowing optimizations/algorithms/correctness checkers to reason about operations in the abstract.
You can put diagnostic information about the cause of the NaN in the bits. IEEE754 doesn’t mandate any format though, leaving it up to the implementation. You can check section 6 of IEEE754:2019 (which also talks about the possibility to use it to extend the standard to wider ranges, more infinities or other stuff)
It's a shame this won't work on VAX.