strip out old precompilation code

[tlnagy]

AFAIK this isn't necessary anymore in recent versions of Julia. I can't detect any regressions and this dramatically increases precompilation speed:

Master:

julia> @time Base.compilecache("Gadfly")
INFO: Recompiling stale cache file /Users/tamasnagy/.julia/lib/v0.5/Gadfly.ji for module Gadfly.
 32.664419 seconds (310.71 k allocations: 12.912 MB)
"/Users/tamasnagy/.julia/lib/v0.5/Gadfly.ji"

julia> @time include("../test/points.jl")
  6.441102 seconds (8.14 M allocations: 338.343 MB, 5.44% gc time)

julia> @time include("../test/points.jl")
  0.005091 seconds (649 allocations: 56.553 KB)

This PR:

julia> @time Base.compilecache("Gadfly")
INFO: Recompiling stale cache file /Users/tamasnagy/.julia/lib/v0.5/Gadfly.ji for module Gadfly.
  7.503891 seconds (310.71 k allocations: 12.912 MB)
"/Users/tamasnagy/.julia/lib/v0.5/Gadfly.ji"

julia> @time include("../test/points.jl")
  6.620339 seconds (8.52 M allocations: 352.255 MB, 4.56% gc time)

julia> @time include("../test/points.jl")
  0.004820 seconds (649 allocations: 56.553 KB)

[tlnagy]

@bjarthur, @Mattriks, @shashi can you guys checkout this PR and give it a whirl? I wasn't able to get it to break on julia 0.5. (0.6 is broken regardless).

[codecov-io]

Codecov Report

Merging #958 into master will increase coverage by 14.42%.
The diff coverage is n/a.

@@             Coverage Diff             @@
##           master     #958       +/-   ##
===========================================
+ Coverage   65.47%   79.89%   +14.42%     
===========================================
  Files          35       34        -1     
  Lines        4814     3930      -884     
===========================================
- Hits         3152     3140       -12     
+ Misses       1662      790      -872

Impacted Files	Coverage Δ
src/Gadfly.jl	74.13% <ø> (+0.05%)	✅
src/bincount.jl	94.84% <ø> (-1.08%)	❌
src/statistics.jl	90.58% <ø> (-0.2%)	❌
src/geom/subplot.jl	94.15% <ø> (-0.08%)	❌
src/guide.jl	84.18% <ø> (+0.09%)	✅
src/aesthetics.jl	77.44% <ø> (+0.57%)	✅
src/ticks.jl	92.08% <ø> (+0.71%)	✅
src/mapping.jl	20.31% <ø> (+0.78%)	✅
src/geom/label.jl	99.03% <ø> (+0.92%)	✅
... and 4 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f315cfe...c6ade7e. Read the comment docs.

[bjarthur]

i notice a ~15% slow down the first time one draws to a file without src/precompile:

master:

jjulia> @time draw(PDF("/tmp/foo.pdf",4inch,3inch), plot(y=[1,2,3],Geom.point))
 15.890677 seconds (14.52 M allocations: 614.291 MB, 1.47% gc time)

julia> using Gadfly

julia> @time draw(PDF("/tmp/foo.pdf",4inch,3inch), plot(y=[1,2,3],Geom.point))
  0.029691 seconds (21.00 k allocations: 906.945 KB)

julia> @time draw(PDF("/tmp/foo.pdf",4inch,3inch), plot(y=[1,2,3],Geom.point))
  0.028999 seconds (21.00 k allocations: 906.945 KB)

this PR:

julia> @time draw(PDF("/tmp/foo.pdf",4inch,3inch), plot(y=[1,2,3],Geom.point))
 19.093611 seconds (20.67 M allocations: 859.435 MB, 2.48% gc time)

julia> @time draw(PDF("/tmp/foo.pdf",4inch,3inch), plot(y=[1,2,3],Geom.point))
  0.025143 seconds (21.02 k allocations: 907.836 KB)

julia> @time draw(PDF("/tmp/foo.pdf",4inch,3inch), plot(y=[1,2,3],Geom.point))
  0.028570 seconds (21.02 k allocations: 907.836 KB)

as subsequent draws are comparable, i'd say merge it.

should probably reference #921

[tlnagy]

Yeah, Gadfly still struggles with the first draw and we've had issues with that as long as I've worked on this project. @aviks' suggestion might be hitting on the problem in #921

[tlnagy]

Let's do this

[bjarthur]

should src/precompile be removed in Compose as well?

[tlnagy]

Good point, PR here: GiovineItalia/Compose.jl#236. Increases initial precompilation speed by ~3x.

[timholy]

In the old days those files dropped the time-to-first-plot by a factor of 2, and so the time spent precompiling seemed well worth it. (You precompile "once," unless you're a developer, but you run many times.) Sounds like the advantage diminished with time?

[tlnagy]

Looks like it. Gadfly's time-to-first-plot is still embarrassingly slow (#921) and we still haven't been able to track down the problem. This PR only affects this first plot, the rest are basically instantaneous.