Julia: Should true == 1?

[benkamphaus]

@vtjnash recommended I write a bug report after encountering this behavior and looking into workarounds.

At present:

true == 1
# true

This seems fairly jarring given that the Bool type seems much more restrictive than in languages like e.g. Python where implicit promotion/deference to parent methods for comparison are typical. For example, this:

if 1
    println("Ok!")
end
# ERROR: TypeError: non-boolean (Int64) used in boolean context

Would also imply to me (not necessarily in any formal sense, though) that the prior equality test should also result in a TypeError.

In this case I want to be able to put both 1 and true in a Set of type Any:

Set([1, true, "true"])
# Set(Any["true",true])

But due to hash/equality semantics, only one will make it in. I could return an alternative implementation of Set backed by an ObjectIdDict for my use case, which is a serialization library that will convey and receive sets from other languages (whose Boolean equality/hash semantics are heterogenous), but wanted to check in to see if this is the desired behavior, or falling out of some implementation detail.

[StefanKarpinski]

It comes from the fact that Bool is essentially just UInt1 – a one-bit unsigned integer. The true == 1 part doesn't really bother me. The principle is that it's ok to use a boolean where a number is expected, but it's not ok to use a non-boolean where a boolean is expected. This is often handy, e.g. when counting the number of values that are true using sum. The Set business is more concerning. We could make Bool a non-numeric type, but that would have other consequences, e.g. it would no longer be possible to have im = Complex(false,true), although we've discussed recently whether that's a good idea.

[eschnett]

In a set you'll also run into problems if you want to add 1 and 1.0 at the same time. You should be able to use isequal (i.e. ===) instead of == to compare values in a set.

How does Set handle NaN? Technically, NaN != NaN, and it should thus keep all of them?

[benkamphaus]

@eschnett Just to be clear, since the Set impl is a HashSet - backed by a Dict's keys with all values as Void, the important detail is how each value hashes:

julia> Set([NaN, NaN])
Set([NaN])

julia> hash(NaN)
0x15d7d083d04ecb90

julia> hash(1)
0x02011ce34bce797f

julia> hash(true)
0x02011ce34bce797f

The Set is the concerning part for me as well @StefanKarpinski (impact is it will require a workaround that impacts usability when edge case is encountered). Maybe I should revise issue title to reflect Set as focus? The behavior of equality came out of the discussion on irc.

[eschnett]

I was confused; of course, isequal(NaN, NaN) is true.

[vtjnash]

note that Char has a similar issue, and also fails for transitive equality with non-integer types 32.0 != ' ' == 32

The true == 1 part doesn't really bother me. The principle is that it's ok to use a boolean where a number is expected, but it's not ok to use a non-boolean where a boolean is expected.

in comparing bool == foo, I think it's arguably expected that foo::Bool

[benkamphaus]

Yeah, I think it's even more surprising with Char in the Set context:

> Set([32, ' '])
# Set(Any[' '])

Probably more plausibly encountered in a scenario like:

> x = " foo"[1]
# ' '
> Set([32, x])
# Set(Any[' '])

[JeffBezanson]

Since Char isn't a subtype of Number (unlike Bool) I agree ' ' == 32 doesn't seem right.

[StefanKarpinski]

The char vs number thing is definitely a bug left over from when Char was a kind of integer. I'm working on a fix.

[StefanKarpinski]

The Char vs. Int comparison thing is surprisingly annoying to fix – we assume that you can compare integers and chars all over the place. I've almost got it done, but it makes me wonder about the change.

[eschnett]

I think it's still the right way to go. Unlike e.g. for real and integer numbers, there is not commonly used abstraction that makes characters a subset of integers. ASCII and UTF-8 are common, but they specify an encoding, not an equivalence. It's easy enough to convert between characters and integers, and if a certain piece of code requires this too often, then I wonder whether there's an abstraction missing.

Most other languages -- Python, Fortran, Mathematica -- don't allow such comparisons either. C and its descendents seem to be in the exception.

[stevengj]

"there is not commonly used abstraction that makes characters a subset of integers" .... Unicode?

[simonster]

In any case, they don't need to be isequal, which would fix the Set issue.

[nalimilan]

Indeed Unicode offers a standard mapping from integer to char, but comparison between these types is still quite confusing. +1 for deprecating it, and requiring people to write Char(32) == ' ' when that's really what they want.

[stevengj]

I don't know... what languages have true character types that don't allow such comparisons? Python doesn't have a character type per se, it only has length-1 strings, and Fortran is similar if I understand it correctly. I don't know about Mathematica, but Mathematica is not particularly well-known for its strength in string processing.

In any case, I feel like discussion of Char == Integer should be in a separate issue.

[Jeffrey-Sarnoff]

" it would no longer be possible to have im = Complex(false,true), although we've discussed recently whether that's a good idea."
Assuming changing that is a good idea, imo the change should still allow
[U]Int +,-,* Bool, Bool +,-,* [U]Int with true working as 1 and false working as 0
by overloading those signatures (as they are so handy, at times).

[StefanKarpinski]

This is decided and consistent in 1.0: yes, true == 1.