Use of global variables to avoid recomputing solutions

[natema]

We currently make use of global variables to avoid recomputing some solutions: see e.g. here:

function crowdingsolution()
    isdefined(@__MODULE__, :_solution_crowding) && return _solution_crowding
    global _solution_crowding = solve(crowding(), (1900, 2300))
    return _solution_crowding
end

After discussing the problem in the Discourse, we might consider using a closure as a more elegant approach:

let isfirstcall = true, output = 0
   global function crowdingsolution()             
       isfirstcall || return output
       isfirstcall = false
       return output = solve(crowding(), (1900, 2300))
   end
end

[natema]

I did a test out of curiosity and it seems that the closure is faster:

julia> function f_global()
           isdefined(@__MODULE__, :_output) && return _output
           global _output = 42
           return _output
       end
f_global (generic function with 1 method)

julia> f_global()
42

julia> @benchmark f_global()
BenchmarkTools.Trial: 10000 samples with 999 evaluations.
 Range (min … max):  12.185 ns … 54.210 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     13.612 ns              ┊ GC (median):    0.00%
 Time  (mean ± σ):   13.818 ns ±  1.758 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

         ▃▂▇▅█▇▄▄▁                                             
  ▂▂▂▄▃▆▇██████████▇▅▅▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▂▂▂▂▂▂▂▁▂▂▂▂ ▃
  12.2 ns         Histogram: frequency by time        19.4 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> let isfirstcall = true, output = 0
          global function f_closure()             
              isfirstcall || return output
              isfirstcall = false
              return output = 42
          end
       end
f_closure (generic function with 1 method)

julia> f_closure()
42

julia> @benchmark f_closure()
BenchmarkTools.Trial: 10000 samples with 1000 evaluations.
 Range (min … max):  1.622 ns … 3.031 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     1.663 ns             ┊ GC (median):    0.00%
 Time  (mean ± σ):   1.661 ns ± 0.060 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

                                       ▂▄▄▄▃█                
  ▁▁▁▁▂▃▄▄▄▅▆▅▄▃▄▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▂▃▃▇▇████████▅▄▃▂▂▂▂▂▂▂▂▂▂ ▃
  1.62 ns        Histogram: frequency by time       1.68 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

[paulobruno]

Wow, it was a big difference!
It shouldn't be hard to change the current code, since we don't have so many solutions.