make broadcast_foreach generated for better performance

[jkrumbiegel]

This got my attention while looking at CairoMakie plotting functions. It seemed that the args... gave the implementation trouble when maping or broadcasting over it, so I changed to a @generated implementation that can correctly unroll these expressions.

This is approximately 100x faster and has no allocations on its own. Imagine that the below is a scatter plot of 1 million points, then at least 7 seconds of that would have been just this function, without any drawing.

julia> f = ((args...,) -> nothing)
#19 (generic function with 1 method)

julia> function test()
           for i in 1:1000000
               Makie.broadcast_foreach(f, 1:3, 1.0:3, 'a':'c')
           end
       end
test (generic function with 1 method)

# before

julia> @time test()
  6.923584 seconds (22.00 M allocations: 1.222 GiB, 3.12% gc time, 0.00% compilation time)

# after

julia> @time test()
  0.069359 seconds

[MakieBot]

Compile Times benchmark

Note, that these numbers may fluctuate on the CI servers, so take them with a grain of salt.

using time

master:  9.75 < 9.98 > 10.18, 0.15+-
pr:      9.61 < 9.91 > 9.99, 0.13+-
speedup: 0.98 < 1.02 > 1.03, 0.02+-
median:  -0.68% => invariant

This PR does not change the using time.

ttfp time

master   24.09 < 24.78 > 25.27, 0.35+-
pr       23.98 < 24.79 > 24.93, 0.32+-
speedup: 0.99 < 1.00 > 1.02, 0.01+-
median:  +0.05% => invariant

This PR does not change the ttfp time.

[asinghvi17]

Nice! Looking forward to seeing how this changes the previously long mesh/surface drawing times in CairoMakie.