Inference: slow for for loop, fast for corresponding while

[timholy]

Pkg.clone("https://github.com/timholy/ArrayIteration.jl.git")
Pkg.checkout("ArrayIteration", "teh/stored")

using ArrayIteration

A = Int[1 3; 2 4]
B = Array{Int}(2, 2)

julia> @time include("/tmp/slow1.jl")
 23.763574 seconds (31.16 M allocations: 1.095 GB, 2.28% gc time)

where slow1.jl is defined as

for (I, a) in sync(index(B, :, 2), value(A, :, 1))
    B[I] = a
end

In contrast, in a fresh session

julia> @time include("/tmp/slow2.jl")
  0.219750 seconds (130.95 k allocations: 5.498 MB)

where slow2 is the manually-expanded while loop version,

iter = sync(index(B, :, 2), value(A, :, 1))
s = start(iter)
while !done(iter, s)
    (I, a), s = next(iter, s)
    B[I] = a
end

julia> versioninfo()
Julia Version 0.5.0-dev+3780
Commit 811fe45* (2016-04-28 09:02 UTC)
Platform Info:
  System: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Core(TM) i7 CPU       L 640  @ 2.13GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Nehalem)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, westmere)

[vtjnash]

are you testing this at the toplevel? what if you put both in a function?

[timholy]

Hmm. Then it's fast. (Thought I'd tested that, but evidently not.)

So may not be an inference thing?

[timholy]

With regards to closing: perhaps it's worth emphasizing that this is a 2x2 matrix, and that 30 seconds is quite a long time to wait for looping over all of its elements. Since compilation is fast when it's in a function, it's also not an inescapable compilation overhead. In other words, this is likely not your run-of-the-mill "avoid globals, put it in a function" problem.

OK to reopen?

[vtjnash]

IIUC what this was doing, it's fast now:

julia> @time for (I, a) in zip(eachindex(view(B, :, 2)), view(A, :, 1))
           view(B, :, 2)[I] = a
       end
  0.000021 seconds (21 allocations: 672 bytes)