rmprocs warning when exiting ranges test

[yuyichao]

On current master, running a single ranges test give the following warning.

yuyichao% ./julia -f test/runtests.jl ranges     
     * ranges               in  24.05 seconds
WARNING: rmprocs: process 1 not removed
    SUCCESS

I could only get this warning from running in the build directory though. The warning somehow disappears when I used the installed version...

Bisect result:

git bisect start
# good: [e65022010305e6bbd3ab70327499fdf3bf3c993d] Merge pull request #11338 from yuyichao/export-interp
git bisect good e65022010305e6bbd3ab70327499fdf3bf3c993d
# bad: [4b757af634597baa5e76c54365d86468d8b112b8] Merge pull request #12172 from ArchRobison/adr/subnormal
git bisect bad 4b757af634597baa5e76c54365d86468d8b112b8
# good: [1b9fc3dd24cd555449d9977cbd85421f3865cbd8] Merge pull request #11815 from JuliaLang/mn/kill@mstr
git bisect good 1b9fc3dd24cd555449d9977cbd85421f3865cbd8
# good: [9d9c4d855bf5d53ded16c4d614547235bb5f665b] Merge pull request #12032 from JuliaLang/jcb/deperr
git bisect good 9d9c4d855bf5d53ded16c4d614547235bb5f665b
# good: [46b81031df284592e15744706af4464f1ccf6eb6] Fix error in sparse lu formula in documentation [skip ci]
git bisect good 46b81031df284592e15744706af4464f1ccf6eb6
# good: [3226819e6d1fd5b0113729c302c563b68292c463] Merge pull request #12179 from JuliaLang/tk/bumpuv
git bisect good 3226819e6d1fd5b0113729c302c563b68292c463
# bad: [824846c76d0f970cd7d5eea9154c7f03d5019371] incremental serialization inferface cleanup
git bisect bad 824846c76d0f970cd7d5eea9154c7f03d5019371
# good: [e78535cfb76e128b6e64aa5c216f3b1193b45a5f] make dump/restore sigatomic
git bisect good e78535cfb76e128b6e64aa5c216f3b1193b45a5f
# bad: [d81f6d2fcaa1ec77402e1e9526886441944ebf24] allow proper backref handling of the element that happened to land at index HT_NOTFOUND
git bisect bad d81f6d2fcaa1ec77402e1e9526886441944ebf24
# bad: [22dfad54137180245db7bb630a9a3f19485e10d6] record binding module in new generic functions
git bisect bad 22dfad54137180245db7bb630a9a3f19485e10d6
# bad: [ef3ec3d31a378566481a05d6e0f707fdf78312ae] reimplement require/include to better support precompilation
git bisect bad ef3ec3d31a378566481a05d6e0f707fdf78312ae
# good: [c8119f8140b8cea7e30ac9626e87e4d72885dabc] give the dump serializer the ability to handle a list of modules to serialize, rather than just a single module
git bisect good c8119f8140b8cea7e30ac9626e87e4d72885dabc
# first bad commit: [ef3ec3d31a378566481a05d6e0f707fdf78312ae] reimplement require/include to better support precompilation

First bad commit is ef3ec3d from #8745

ef3ec3d31a378566481a05d6e0f707fdf78312ae is the first bad commit
commit ef3ec3d31a378566481a05d6e0f707fdf78312ae
Author: Jameson Nash <[email protected]>
Date:   Wed Jul 8 15:33:58 2015 -0400

    reimplement require/include to better support precompilation

    the codepaths through these methods were not particularly amenable to adding incremental precompilation.
    the problem is that the cache-load code paths need to be mirrors of their non-cache counterparts.
    this minimizes the code paths to running "include" to one, and similarly for "reload"/"require"
    those are now limited to only handling module names (symbols), rather than arbitrary paths

    this also fixes a number of issues with the incremental serializer that
    have snuck in over time.

:040000 040000 ab81c489f0eeaa92d552e768fa22bccd0b165462 58497e14ff35bb09bdaa57c1a4e1fd1d752e4cc7 M        base
:040000 040000 d2059df9e086e8283c3d0a3e2107dd9176ba9dd4 8d4d5bae755dac88a39ea0f98add36c7c4f0a089 M        src
:040000 040000 59e3903203201328c4a55c94f456c39ae1962e1c 5b7617dc0c58c912fa8f34fd40988e3160d77412 M        test
:040000 040000 4d926ba22c92c3bee0f64f4f116f95532dc3f56e 3ef2b01808dfa2d09851d1aaf652fb27e2038a2f M        ui

CC. @vtjnash .

P.S. I really appreciate that #8745 is actually bisect-able....

Edit: Sorry I pasted the wrong commit SHA for the link just now (the bisect output is correct)

[amitmurthy]

runtests is defined as

function runtests(name)
    @printf("     \033[1m*\033[0m \033[31m%-20s\033[0m", name)
    tt = @elapsed include("$name.jl")
    @printf(" in %6.2f seconds\n", tt)
    nothing
end

prior to this commit, any variables introduced due to the include were visible only in the function, now, an n in test/ranges.jl is overwriting the value of n in

julia/test/runtests.jl

Line 5 in 4b757af

n = 1

.

[yuyichao]

prior to this commit, any variables introduced due to the include were visible only in the function

IIRC, include execute code in global scope. (just like eval) I don't think there's anyway you can execute arbitrary code at runtime in local scope.

[vtjnash]

i think amit meant the reverse: prior to that commit, include couldn't modify any variable in the runtests file because it was inside an (unnecessary) cd expression: ef3ec3d#diff-fce720c43af3c52c862fd7451c7374b8L5

it may be appropriate to put that back into a let expression, although i still want to go back and try to make #9798 work

[yuyichao]

Ah. I see. that makes sense.

[tkelman]

At least last year that cd was necessary. Are you sure removing it didn't break Base.runtests ?

[yuyichao]

FWIW, ./julia test/runtests.jl ranges was broken for a few commit around this but was fixed later.

[tkelman]

edit: #12212

[tkelman]

warning should be gone from a93d1eb (github didn't auto-close with "fixing #"), but parallel tests from outside of the test directory are still broken by the removal of the cd

[yuyichao]

Yeah, confirmed that the warning disappears on current master. Thx.