~2x slowdown from 1.5.4 -> 1.6

[quinnj]

Taking the following code, which measures iteration over ChainecVector, I see a ~2x slowdown from Julia 1.5.4 -> 1.6 on macOS.

using SentinelArrays, BenchmarkTools
A = ChainedVector([[j for j = 1:1000] for i = 1:8])
function eachind2(A)
    x = 0
    for y in A
        x += y
    end
    return x
end

1.5.4:

julia> @btime eachind2(A)
  11.285 μs (1 allocation: 16 bytes)
4004000

1.6:

julia> @btime eachind2(A)
  20.675 μs (1 allocation: 16 bytes)
4004000

[oscardssmith]

for my own sanity, I've updated the mwe so the variable names match.

[johnnychen94]

I believe this is related to JuliaLang/julia#38073 (due to my notorious JuliaLang/julia#37829 commit)...

But probably is CPU-dependent? (I can't reproduce it and is not capable of figuring out the root issue)

julia> @btime eachind2($A)
  4.288 μs (0 allocations: 0 bytes) # Julia 1.6.1
  4.285 μs (0 allocations: 0 bytes) # Julia 1.5.4

julia> versioninfo()
Julia Version 1.5.4
Commit 69fcb5745b (2021-03-11 19:13 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin18.7.0)
  CPU: Intel(R) Core(TM) i9-9880H CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-9.0.1 (ORCJIT, skylake)

[quinnj]

Hmmm, ChainedVector iteration doesn't use CartesianIndices at all though, so at first glance, I'd say unrelated?

[KristofferC]

I can not repro on my linux desktop, nor my mac laptop (i7-9750H). Both have pretty much the exact same performance in 1.5.4 and 1.6.1.

[MarcMush]

maybe SentinelArrays or another package is not the same version in your 1.5 and 1.6 installations ?

[KristofferC]

1.2.16 in both cases.

[MarcMush]

I meant @quinnj

[quinnj]

My bad; it turns out I had removed the @inline on the first function Base.iterate(x::ChainedVector) and that somehow totally blew up performance.