Lookup bindings in the latest world

[Keno]

Post-partitioning the executing world age matters for bindings lookup. JuliaInterpreter currently does not model world age correctly. That's a bigger issue. However, for now, just look up bindings in the latest world age, which mostly restores the unpartitioned behavior.

[aviatesk]

If invokelatest isn’t specified, in which world do these functions perform binding lookups?

Not just in JuliaInterpreter, but for packages using getglobal, I assumed the default lookup world for getglobal would be set to the latest as I believed it would then be non-breaking.

[Keno]

If invokelatest isn’t specified, in which world do these functions perform binding lookups?

The current TLS age of the executing task.

Not just in JuliaInterpreter, but for packages using getglobal, I assumed the default lookup world for getglobal would be set to the latest as I believed it would then be non-breaking.

No, those cannot be the semantics as all global accesses would then have to be inferred to Any in case some other task replaced the binding under your feet after you compiled it.

[KristofferC]

This was a quite significant performance hit (the lookup_var change).

Before

julia> function mysum(x)
           s = 0
           for v in x
               s += v
           end
           return s
       end;

# Before
julia> @time @interpret mysum(fill(1, 100000))
  0.566310 seconds (10.40 M allocations: 186.914 MiB, 1.29% gc time)
100000

# After
julia> @time @interpret mysum(fill(1, 100000))
  0.912419 seconds (16.20 M allocations: 280.050 MiB, 0.57% gc time)
100000

anyway to do this in a bit better way?

[Keno]

We can optimize the codegen for this pattern