performance regression in println

[JeffBezanson]

I just tried the code from #2050 again to see how we're doing. The good news is, we're faster than 0.3 on the original code.

0.3:

jeff@gurren:~/src/julia$ time ~/src/julia-0.3/julia/julia i2050.jl

real    0m29.973s
user    0m26.618s
sys 0m1.860s

0.4:

jeff@gurren:~/src/julia$ time ./julia i2050.jl

real    0m24.495s
user    0m21.923s
sys 0m1.312s

Then I tried using println instead of write and got this:

0.3:

jeff@gurren:~/src/julia$ time ~/src/julia-0.3/julia/julia i2050.jl

real    0m47.285s
user    0m45.708s
sys 0m1.664s

0.4:

jeff@gurren:~/src/julia$ time ./julia i2050.jl

real    1m22.889s
user    1m20.935s
sys 0m1.877s

Pretty massive regression. If I take out the new lock(io) do io I get

jeff@gurren:~/src/julia$ time ./julia i2050.jl

real    0m39.443s
user    0m37.710s
sys 0m1.688s

again faster than 0.3, so very promising. We might need some manual inlining and hacked-in fast paths for this.

cc @amitmurthy

[amitmurthy]

Closed by #10679