Julia 1.2 start-up is slower than 1.1

[KristofferC]

Julia 1.2:

➜  ~ time julia -e ''
julia -e ''  0.22s user 0.10s system 110% cpu 0.287 total

Julia 1.1:

➜  ~ time julia11 -e ''
julia11 -e ''  0.15s user 0.09s system 105% cpu 0.236 total

With the following diff

diff --git a/stdlib/REPL/src/LineEdit.jl b/stdlib/REPL/src/LineEdit.jl
index 0a71488859..3d834d653e 100644
--- a/stdlib/REPL/src/LineEdit.jl
+++ b/stdlib/REPL/src/LineEdit.jl
@@ -1239,6 +1239,9 @@ function write_prompt(terminal, p::Prompt)
     width = write_prompt(terminal, p.prompt)
     write(terminal, Base.text_colors[:normal])
     write(terminal, suffix)
+    if !REPL.Terminals.is_precompiling[]
+        exit(0)
+    end
     return width
 end

I measured the time to get to the prompt on 1.1 vs 1.2:

Julia 1.2:

➜  julia git:(release-1.2) ✗ time ./julia -q   
./julia -q  1.25s user 0.21s system 132% cpu 1.104 total

Julia 1.1:

➜  julia11 git:(release-1.1) ✗ time ./julia  -q
./julia -q  0.65s user 0.19s system 176% cpu 0.476 total

So we lost about 0.6s in time to prompt.

Running 1.2 with --trace-compile I see the following which is not there on 1.1

precompile(Tuple{typeof(Base.copy), Array{UInt8, 1}})
precompile(Tuple{typeof(Base.similar), Array{Base.Grisu.Bignums.Bignum, 1}})
precompile(Tuple{typeof(Base.length), Array{Base.Grisu.Bignums.Bignum, 1}})
precompile(Tuple{typeof(Base.deepcopy_internal), Array{UInt32, 1}, Base.IdDict{Any, Any}})
precompile(Tuple{typeof(Base.copy), Array{UInt32, 1}})

This should only affect the time to reach the banner though. Not sure what is making the prompt be so much slower.

[KristofferC]

Strangely, building Julia 1.2 locally, the extra precompile statements in --trace-compile are gone and startup is faster (but still not really 1.1 levels)

➜  julia git:(release-1.2) ✗ time ./julia -e ''
./julia -e ''  0.18s user 0.10s system 121% cpu 0.232 total

[KristofferC]

Time to prompt repod by:

time ./julia -e 'Base.run_main_repl(true, false, true, true, true)'

on 1.1 this takes 0.6s while on 1.2 it takes 1.1s.

[KristofferC]

So on 1.1 we have the following trace running the above:

with the heaviest functions being:

Basically, nothing is done in true_main and most of the time is spent loading the sysimg.

Contrast this with 1.2:

where true_main now almost takes half a second(!).

We are also compiling stuff like refresh_multi_line which I would have thought the precompile script would already compile.

Comparing the number of precompile statements in 1.1 (1089) vs 1.2 (900) might mean that we are no longer precompiling a bunch of functions. And indeed, looking at a diff of the precompile files between the two versions (http://www.mergely.com/lGgrvfs9/) a bunch of precompile statements for e.g. refresh_multi_lineare now missing.