Allow reinterpreting singleton types

[Liozou]

Fix #43403. It does not actually allow reinterpreting all 0-sized types, only the immutable ones (aka singleton types), but I think that's what was intended anyway (see also #17149).

I'm not sure whether it should be merged, considering what @vtjnash said in #34865 (comment), but I'm opening the PR in case it is decided it is worth adding.

Since reinterpret is a fairly basic building block, there is also the potential issue of slowing down critical paths. I don't think there is any harm done in this regard since the sizeof(T) == 0 and isdefined(T, :instance) conditions should all be eliminated at compilation I think, but I'm no expert on this.

[JeffBezanson]

Hmm. It's hard to say whether we should allow this. The divide-by-zero error should be avoided at least. But then if we want to allow this, it should be implemented inside bitcast I think.

[Liozou]

To implement it inside bitcast, maybe singleton types should be considered to be primitive types? That's currently explicitly not the case because of the following line

julia/src/julia.h

Line 1217 in 98b485e

jl_datatype_size(v) > 0);

which originates in 62cca31 and I imagine there may be good reasons not to, but I'm no expert on this.

Additionally, I remembered that singleton types were primarily used for function types, and I don't know if giving a meaning to reinterpret(typeof(sum), nothing) really is desirable... Although I'm not sure it would hurt either.

[JeffBezanson]

From triage: bitcast and reinterpret (on non-arrays) don't support structs, so they shouldn't support 0-size structs either. However, reinterpret of arrays does allow it, so that case should work.

[Liozou]

Thank you for the triage and the review! I added a commit with the fix: doing reinterpret(::singletontype, ::notanarray) reverts to being forbidden while reinterpret(::singletontype, ::somesingletonarray) continues to work.

Let me know if there is anything else I can do.

[Liozou]

CI failure seems unrelated.

[vtjnash]

Suggested change

	return T.instance # setindex! is a noop here
	return a # setindex! is a noop here

This function is supposed to return the argument, though looks like several of the cases are wrong.

[Liozou]

You're right, sorry about that! I fixed both instances in e63a8fb.
I also fixed the currently erroneous return value for setindex! on some reinterpretarrays where it would return the parent array:

julia> t = reinterpret(Int, UInt[15])
1-element reinterpret(Int64, ::Vector{UInt64}):
 15

julia> setindex!(t, 7, 1) # should return t instead of t.parent
1-element Vector{UInt64}:
 0x0000000000000007

That's a bugfix (I believe) but it's technically breaking so let me know if you prefer I'd move that to another PR.

the changes were made

[DilumAluthge]

@Liozou Can you rebase on the latest master? That should fix the llvmpasses failures.

[Liozou]

Done, there is a remaining CI failure for the Sockets test but it looks unrelated.

[Seelengrab]

This PR (or one of the two follow ups, not 100% sure) allows reinterpret(Missing, Int8[]), which doesn't make too much sense to me:

julia> a = Int8[]
Int8[]

julia> ms = reinterpret(Missing, a)
0-element reinterpret(Missing, ::Vector{Int8})

julia> push!(a, 1)
1-element Vector{Int8}:
 1

julia> ms
1-element reinterpret(Missing, ::Vector{Int8}):
 missing

This was encountered during a discussion on discourse, about reinterpreting empty views/arrays. As 1.8 has not been released, perhaps this can be decided soon-ish?

Possible solution: disallow reinterpret to singleton types if the source array type is not also a singleton type.