Fix compilation and interpolation memory leaks

[willow-ahrens]

Fixes #339. Fixes a memory leak in interpolated variables from setup and teardown, continuing what was started in #265. Additionally, adds a cache of already-compiled sampler functions, so that wrapping a @benchmark in a for-loop doesn't compile separate functions each iteration.

Here's an example:

Setup

julia> using BenchmarkTools, Printf

julia> function f(n)
           x = rand(Int, n)
           target = sort(x)
           y = copy(target)
           @belapsed sort!($y) setup=($y == $target || error("Bad sort"); copyto!($y, $x)) evals=1 gctrial=false samples=3
       end
f (generic function with 1 method)

julia> function meminfo_julia()
           GC.gc()
           @printf "GC total:  %9.3f MiB\n" Base.gc_total_bytes(Base.gc_num())/2^20
           @printf "GC live:   %9.3f MiB\n" Base.gc_live_bytes()/2^20
           @printf "JIT:       %9.3f MiB\n" Base.jit_total_bytes()/2^20
           @printf "Max. RSS:  %9.3f MiB\n" Sys.maxrss()/2^20
       end

Main Branch

julia> meminfo_julia()
GC total:    121.857 MiB
GC live:      14.555 MiB
JIT:           0.384 MiB
Max. RSS:    338.000 MiB

julia> times = f.(1594323:1594323+100);

julia> meminfo_julia()
GC total:  10147.292 MiB
GC live:    3695.383 MiB
JIT:           1.550 MiB
Max. RSS:   4152.938 MiB

After Fixing Setup Leaks

julia> meminfo_julia()
GC total:     77.615 MiB
GC live:      14.343 MiB
JIT:           0.165 MiB
Max. RSS:    305.688 MiB

julia> times = f.(1594323:1594323+100);

julia> meminfo_julia()
GC total:  10112.637 MiB
GC live:       9.663 MiB
JIT:           1.326 MiB
Max. RSS:    495.234 MiB

After Fixing Redundant Compilation

julia> meminfo_julia()
GC total:     77.695 MiB
GC live:      14.409 MiB
JIT:           0.165 MiB
Max. RSS:    306.406 MiB

julia> times = f.(1594323:1594323+100);

julia> meminfo_julia()
GC total:  10011.679 MiB
GC live:       8.914 MiB
JIT:           0.416 MiB
Max. RSS:    476.047 MiB

[vchuravy]

Looks fine, but I must admit I don't understand this code anymore.

[willow-ahrens]

It's confusing to need the function to always run at global scope, and requires some involved metaprogramming to get that right. Hopefully this is one step closer to a correct approach which is still backwards-compatible with the older versions.

[LilithHafner]

This is remarkably clever. Brava!