Fixes #432 ...

[goerch]

... by moving invokelatest up the call stack.

[pfitzseb]

Why does this work? And how much of a performance impact does it have?

[goerch]

Why does this work?

I first tried smaller changes to repair the problem, but then some tests failed. I don't have tools available to diagnose the crash in Julia.

And how much of a performance impact does it have?

Doesn't CI tell us that?

[codecov]

Codecov Report

Merging #504 (735c3fc) into master (2030a00) will decrease coverage by 0.03%.
The diff coverage is 66.66%.

@@            Coverage Diff             @@
##           master     #504      +/-   ##
==========================================
- Coverage   87.31%   87.28%   -0.04%     
==========================================
  Files          12       12              
  Lines        2468     2470       +2     
==========================================
+ Hits         2155     2156       +1     
- Misses        313      314       +1     

Impacted Files	Coverage Δ
src/interpret.jl	81.08% <66.66%> (-0.15%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2030a00...735c3fc. Read the comment docs.

[goerch]

Ouch. Any hints how to repair this for Julia 1?

[pfitzseb]

Wait for #499 to be merged and rebase :)

[aviatesk]

Maybe we want to keep this code around (like with commented out), and leave some comment that says ideally we want to figure out the root problem here and remove the added invokelatest.

[goerch]

Done.

[timholy]

I worry this isn't the right fix...it's just kind of a bandaid, and one that will likely hurt performance.

And how much of a performance impact does it have?

Doesn't CI tell us that?

No, CI is often quite bad for checking performance, because the CI servers are always running multiple tasks and you get spurious failures when server-latency gets confounded for a performance regression. You kind of need a dedicated machine.

You can run benchmarks locally in https://github.com/JuliaDebug/JuliaInterpreter.jl/tree/master/benchmark, although as a caution our benchmarks may not be completely thorough.

[pfitzseb]

I worry this isn't the right fix...it's just kind of a bandaid, and one that will likely hurt performance.

Same, which is why I'd like to understand why this works.
That said, invokelatest performance got much better recently, I think.

[goerch]

You can run benchmarks locally in https://github.com/JuliaDebug/JuliaInterpreter.jl/tree/master/benchmark, although as a caution our benchmarks may not be completely thorough.

For main:

Benchmarkresults:
    Package: JuliaInterpreter
    Date: 13 Dec 2021 - 13:34
    Package commit: dirty
    Julia commit: 3bf9d1
    BenchmarkGroup:
        7-element BenchmarkTools.BenchmarkGroup:
          tags: []
          "recursive other 1_000" => Trial(6.007 ms)
          "recursive self 1_000" => Trial(2.848 ms)
          "ccall library" => Trial(82.300 μs)
          "ccall ptr" => Trial(19.900 μs)
          "long function 5_000" => Trial(21.305 ms)
          "throw long 1_000" => Trial(6.804 ms)
          "tight loop 10_000" => Trial(221.751 ms)

For branch:

Benchmarkresults:
    Package: JuliaInterpreter
    Date: 13 Dec 2021 - 13:40
    Package commit: 7c156d
    Julia commit: 3bf9d1
    BenchmarkGroup:
        7-element BenchmarkTools.BenchmarkGroup:
          tags: []
          "recursive other 1_000" => Trial(5.827 ms)
          "recursive self 1_000" => Trial(3.056 ms)
          "ccall library" => Trial(80.000 μs)
          "ccall ptr" => Trial(18.100 μs)
          "long function 5_000" => Trial(21.651 ms)
          "throw long 1_000" => Trial(6.737 ms)
          "tight loop 10_000" => Trial(233.557 ms)

[goerch]

Why does this work?

At least I now understand that there seems to be some kind of world age confusion. Interesting question I've been pondering is how to explicitly track world age.

[KristofferC]

Let's get this in even though we are not sure exactly what is going on. It fixes a problem and doesn't complicate the code much. Thanks @goerch!