Benchmark studycase: split

[fredrikekre]

The interpreter is fast enough that stepping line for line can usually be done without any noticable lag. However, I have found that split seems very slow:

julia> using BenchmarkTools, JuliaInterpreter
       io = IOBuffer()
       for _ in 1:100
           println(io, "abc 123")
       end
       str = String(take!(io))
       f(str) = split(str, '\n')
       @btime f($str)
       @btime @interpret f($str);
  6.385 μs (182 allocations: 6.52 KiB)
  2.540 s (858969 allocations: 38.95 MiB)

The original case was similar to the code above, but also included a loop over all lines splitting into "abc" and "123". I am not sure what the expected slowdown is, but at least this is something that really stood out. I am just posting this since it seems like an interesting case.

[KristofferC]

Looks like a lot of time is spend in evaluate_foreigncall

[KristofferC]

julia> using BenchmarkTools

julia> const s = "foo";

julia> @btime pointer_from_objref(s)
  1.500 ns (0 allocations: 0 bytes)
Ptr{Nothing} @0x00007f6ac35d33f0

julia> @btime @interpret pointer_from_objref(s)
  1.170 ms (503 allocations: 30.91 KiB)
Ptr{Nothing} @0x00007f6ac35d33f0

[KristofferC]

FWIW

julia> @btime @interpret f($str);
  77.870 ms (1007426 allocations: 73.24 MiB)

now

[KristofferC]

julia> @btime @interpret f($str);
  62.169 ms (380271 allocations: 12.45 MiB)

at #237

[KristofferC]

julia>        @btime @interpret f($str);
  40.062 ms (378290 allocations: 11.77 MiB)

at #245

[KristofferC]

julia>        @btime @interpret f($str);
  35.767 ms (378288 allocations: 11.77 MiB)

with #245 and JuliaLang/julia#31541

Commenting out the last_reference uses:

julia> @btime @interpret f($str);
  31.060 ms (375400 allocations: 11.66 MiB)

[KristofferC]

While this is a nice thing to keep seeing improvements here I don't think there is anything special about the performance of split now that warrants this issue to be kept open.