nospecialize in displaying SOneTo

[jishnub]

No description provided.

[mateuszbaran]

For some weird reason this makes printing slower?
Current StaticArrays:

julia> @time repr(SOneTo(3))
  0.002636 seconds (20 allocations: 1.133 KiB, 99.22% compilation time)
"SOneTo(3)"

julia> @time repr(SOneTo(4))
  0.004234 seconds (20 allocations: 1.133 KiB, 99.47% compilation time)
"SOneTo(4)"

julia> function Base.show(io::IO, @nospecialize(x::SOneTo))
           print(io, "SOneTo(", length(x)::Int, ")")
       end

julia> @time repr(SOneTo(3))
  0.007384 seconds (1.04 k allocations: 71.046 KiB, 99.67% compilation time)
"SOneTo(3)"

julia> @time repr(SOneTo(4))
  0.000025 seconds (7 allocations: 344 bytes)
"SOneTo(4)"

[jishnub]

I'd say that this PR improves things, as the compilation happens only once, and subsequent calls are fast. Perhaps the first run is slowed down by the nospecialize, but performance is usually not a big concern in displaying values anyway.

In case this matters, the first call may be precompiled using PrecompileTools, in which case no subsequent call will require compilation.

[jishnub]

Gentle bump

[mateuszbaran]

#1187 is now merged so I think we could add printing of SOneTo to compile_workload and it would be completely fine.

[jishnub]

I'm unsure what's the best way to address this, as show(IOBuffer(), SOneTo(1)) doesn't precompile it for displaying to a terminal. What works is show(SOneTo(1)), but this writes to the terminal during precompilation. I've tried to redirect the stdin to devnull, but this leads to an error on my system.

[mateuszbaran]

I think you should redirect stdout, not stdin. Like this:

redirect_stdout(devnull) do
    show(SOneTo(1))
end

in the workload.

[jishnub]

Thanks for catching that! Oddly, adding this doesn't improve the compilation time. I wonder if this is because the display to Base.TTY is not compiled when the output is redirected to devnull, which is a Base.DevNull

[mateuszbaran]

I see. Well, let's merge this then without precompilation.