Type is not equal to itself

[KristofferC]

On 1.5

julia> NamedTuple{(:a, :b), Tuple{Missing, Union{}}} == 
       NamedTuple{(:a, :b), Tuple{Missing, Union{}}}
false

(works in 1.4). From test failure in PkgEval (https://github.com/JuliaCI/NanosoldierReports/blob/90885ffd4005a91d0dc4150b8c1453f2f5948eb0/pkgeval/by_hash/194ddda_vs_b5f13c3/logs/FunctionalTables/1.5.0-DEV-f11a73bd9c.log).

[JeffBezanson]

Bisect says c958b2c

[vtjnash]

Curious. The hash computation seems to be working, so I'm not quite sure why we're failing to normalize this:

julia> A = Val{Tuple{Missing, Union{}}}
Val{Tuple{Missing,Union{}}}

julia> B = Val{Tuple{Missing, Union{}}}
Val{Tuple{Missing,Union{}}}

julia> A === B
false

julia> Tuple{Missing, Union{}}.hash
-302523448

[martinholters]

Tuple{Union{}} is not cached:

julia> T1 = Tuple{Union{}}
Tuple{Union{}}

julia> T2 = Tuple{Union{}}
Tuple{Union{}}

julia> pointer_from_objref(T1)
Ptr{Nothing} @0x00007fb946a38850

julia> pointer_from_objref(T2)
Ptr{Nothing} @0x00007fb946e66a10

But

julia> T1.isconcretetype
true

so

julia/src/jltypes.c

Lines 566 to 567 in f65561f

	if (jl_is_concrete_type(tj) || jl_is_concrete_type(kj))
	return 0;

triggers and declares Tuple{Union{}} not equal to the parameter of Val{Tuple{Union{}}} . I think this not the only place where we assume that a type is cached iff it is concrete, so I guess the problem is that Tuple{Union{}} is not cached? Or should isconcretetype by false here? After all, Tuple{Union{}} cannot have instances. Or should the condition here exclude Tuples with Union{} elements explicitly as an exception to the rule?

[martinholters]

This fixes it:

diff --git a/src/jltypes.c b/src/jltypes.c
index 845936ed10..f0b6fc17a3 100644
--- a/src/jltypes.c
+++ b/src/jltypes.c
@@ -1492,7 +1492,7 @@ static jl_tupletype_t *jl_apply_tuple_type_v_(jl_value_t **p, size_t np, jl_svec
     int cacheable = 1;
     for (size_t i = 0; i < np; i++) {
         assert(p[i]);
-        if (!jl_is_concrete_type(p[i]))
+        if (!jl_is_concrete_type(p[i]) && p[i] != jl_bottom_type)
             cacheable = 0;
     }
     return (jl_datatype_t*)inst_datatype_inner(jl_anytuple_type, params, p, np, cacheable, NULL, NULL);

Does that make sense, @vtjnash? Then I can open a PR.

[vtjnash]

After all, Tuple{Union{}} cannot have instances.

Neither can Some{Union{}}, but they're also concrete for the same reason that Some{T} where T<:Union{} becomes concrete, because that TypeVar became degenerate

[JeffBezanson]

Yes, PR it. ❤️