[FR] Cut down performance hit of dynamically generating prefix-symbols each time

[BioTurboNick]

11% of profiler samples in one of my plots are being spent in Symbol(x, y), which constructs a new symbol from two different symbols. Each time this happens, 3 allocations are required.

This code was run 429,042 times, and consuming 72 MiB, in generating my plot. It thus appears to account for 24% of allocations and 23% of allocation amount of generating my plot, between calls to _update_axis and _update_subplot_args.

Can this list of attribute combinations be generated once and then read from, rather than being constructed expensively on the fly?

[BioTurboNick]

Some possible minor improvements:

# current: Symbol(x...) = Symbol(string(x...))
@btime Symbol(:test, :me)
  111.024 ns (3 allocations: 176 bytes)

Symbol(x::Symbol, y::Symbol) = Symbol(string(x, y))
@btime Symbol(:test, :me)
  110.974 ns (3 allocations: 176 bytes) - no change

Symbol(x::Symbol, y::Symbol) = Symbol(string(x), string(y))
@btime Symbol(:test, :me)
  84.826 ns (3 allocations: 96 bytes) - 25% better in time, 45% better in memory space

Implementing the last one cut plot time by 10%, allocation count by 2%, and allocation amount by 25%, though that's presumably a change that needs to go into Base.

[BioTurboNick]

Should this be closed? The change in Julia base helps a lot, but there may still be a way to eliminate many of those allocations.

[t-bltg]

Sorry, though it was fixed in https://github.com/JuliaLang/julia.

[BioTurboNick]

Thanks for your help on this!