Opaque closure vararg segfault on 1.12

[mhauru]

This:

module MWE

g(xs...) = 1

# Make a callable out of the format that OpaqueClosure expects, where the first
# argument is a Tuple of the captures. `code_ircode` on this yields IR that can
# be fed to the Core.OpaqueClosure constructor, with g as a capture.
(t::Tuple{typeof(g)})(xs...) = t[1](xs...)
ir = Base.code_ircode_by_type(Tuple{Tuple{typeof(g)}, Symbol, Symbol})[1][1]
ir.argtypes[1] = Tuple{typeof(g)}

oc = Core.OpaqueClosure(ir, g; isva=true, do_compile=true)
oc(:a, :b)  # segfault on 1.12

end

segfaults on 1.12.0-rc1 as well as master from last week, but passes fine on 1.11.6.

Has there been an intentional change to opaque closure varargs on 1.12 that I don't understand, or is this a bug? cc @willtebbutt who is likewise confused by this.

[mhauru]

Actually, my MWE is unnecessary complicated, you don't need the captures part at all. This segfaults just the same:

module MWE

g(xs...) = 1

ir = Base.code_ircode_by_type(Tuple{typeof(g), Symbol, Symbol})[1][1]
ir.argtypes[1] = Tuple{}

oc = Core.OpaqueClosure(ir; isva=true, do_compile=true)
oc(:a, :b)  # segfault on 1.12

end

[serenity4]

This seems to work with ir.argtypes[1] = Tuple, which can be used as a workaround for now. However, we should probably have an actual fix, particularly if that worked well in 1.11, unless that just happened to work by accident.

@Keno @topolarity would you have any pointers here? It seems to me like ir.argtypes[1] = Tuple{typeof(g)} is the expected value to use here, but that something may be going wrong with the codegen.

[Keno]

I do think I agree that this is supposed to work as written, so yeah, worth debugging why the interpreter is upset.

[serenity4]

Interestingly, this now returns an error message on 1.13 instead of crashing:

julia> module MWE

       g(xs...) = 1

       ir = Base.code_ircode_by_type(Tuple{typeof(g), Symbol, Symbol})[1][1]
       ir.argtypes[1] = Tuple{}

       oc = Core.OpaqueClosure(ir; isva=true, do_compile=true)
       oc(:a, :b)  # segfault on 1.12

       end
ERROR: Internal Error: Opaque closure with no source at all
Stacktrace:
 [1] top-level scope
   @ REPL[2]:9
 [2] top-level scope
   @ REPL:1

[adienes]

(two) bisections show that this went from
no error --> segfault in #58131, which was then backported (hence why it crashes in 1.12)

and then segfault --> ERROR: Internal Error in #59362

[KristofferC]

I'm gonna remove this from the milestone because it isn't the type of issue to hold up a release.