Explicitly return nothing from runtests

[tkelman]

runtests isn't expected to return anything meaningful, is it? I'm seeing the return type of the numbers.jl test causing a bizarre problem that I can't reproduce outside of AppVeyor. Compare bad build to better build - gets much further. In the bad build, I had a few printlns between the last few tests in numbers.jl, but none after the final test. In the good build, I only had a println on the last line. This oddly fixed an error

exception on 1: ERROR: type: subtype: expected Type{T<:Top}, got (ASCIIString,DataType)
while loading C:\projects\julia\test\runtests.jl, in expression starting on line 34

[JeffBezanson]

This might only be covering another bug (it looks like one of the "mysterious" failures).

But, this change makes sense anyway since sending back the result of include could potentially cause problems, for example if the file defined a type that's not available on other processors.

[tkelman]

Darn, this didn't completely resolve it. I'm trying MSYS instead of Cygwin to see if it's any faster to get built, and it made the same error come back: https://ci-beta.appveyor.com/project/tkelman/julia/build/1.0.247

(Here's how far Cygwin got, times out at 30 minutes in the very final parallel test https://ci-beta.appveyor.com/project/tkelman/julia/build/1.0.235)

[tkelman]

D'oh, it wasn't the change from Cygwin to MSYS, it was something that changed when I rebased with upstream master. Comparing two builds, one timed out at parallel and the other has this mystery type error in numbers:

I'll try bisecting and see which one made the difference. Wasn't that either. Was my fault for not paying close enough attention to the fact that I had switched from running the tests with julia-debug-basic to julia-basic, hoping to save a few seconds of build time. Tests run cleanly with the debug exe, but strange errors happen with the basic exe. Something having to do with gc and jl_profile_is_running in this case - maybe this only happens in CI because the VM's have much less memory than user machines?

[vtjnash]

something having to do with gc and jl_profile_is_running in this case

I was having the same issue recently on a clean build. I finally slept on it for a few days and the problem went away (or Jiahao did something magical to it). It doesn't make sense to me though -- how could jl_profile_is_running do anything that would cause a segfault?

[tkelman]

Good news, deleting sys.dll makes the tests succeed with the basic exe: https://ci-beta.appveyor.com/project/tkelman/julia/build/1.0.296

That file's only good for making Julia startup faster, right?