Fix type instability of closures capturing types

[tkf]

(trying to) fix #23618

This patch "works" in the sense that closures behave as expected in the "bare" julia:

$ usr/bin/julia --sysimage usr/lib/julia/corecompiler.ji
WARNING: Base._start not defined, falling back to economy mode repl.

julia> f = x -> g -> g(x)
getfield(Main, Symbol("#1#3"))()

julia> typeof(f(Int))  # this takes ~10 seconds
getfield(Main, Symbol("#2#4")){Type{Int64}}

julia> f(Int)(f(1))
1

Note that the output of the second statement would be getfield(Main, Symbol("#2#4")){DataType} in current master. In this branch, it captures a Type{Int64} instead of a DataType.

However, make never finishes:

...
Sysimage built. Summary:
Total ─────── 352.542854 seconds
Base: ─────── 147.912140 seconds 41.9558%
Stdlibs: ──── 204.630686 seconds 58.0442%
    JULIA usr/lib/julia/sys-o.a
Generating precompile statements... 34405 generated in 6878.302163 seconds (overhead 5869.884857 seconds)
^C
signal (2): Interrupt

It ran for hours and consumed more than 12 GiB RAM so I just interrupted the compile.

@JeffBezanson @vtjnash Is there any way to make this work (or fix #23618 in any way)?

[KristofferC]

34405 generated in 6878.302163 seconds (overhead 5869.884857 seconds

Oh my...

[JeffBezanson]

Ah, well, there you go --- sometimes we do these awful things with types for a reason :) This looks like a tough problem.

[tkf]

Oh well, that's too bad. I thought it was such an "easy" fix. But it turned out that it completely destroyed the balance that current Julia compiler achieved...

Can we still make it opt-in somehow? Maybe with closures with @inline?

[vtjnash]

replaced by #40985