New pyexpect version is working on Windows as well.
https://pexpect.readthedocs.org/en/stable/overview.html#pexpect-on-windows

I have tested it on Python 3.5.1 and it was working well (expect including more advanced regexp with capture groups, sendline, timeouts).

This looks very promising. One thing that concerns me is this:

pexpect.spawn and pexpect.run are not available on Windows, as they rely on Unix pseudoterminals (ptys). Cross platform code must not use these.

The best possible scenario is that we're not using either of these methods. I'm 99% sure we are using pexpect.spawn though, so I don't think we will be this lucky.

The second best scenario is that all uses of pexpect.spawn can be ported to some other mechanism that pexpect provides. It looks like that method would be pexpect.popen_spawn.PopenSpawn. I don't know what the difference is between pexpect.spawn and pexpect.popen_spawn.PopenSpawn or what features we're missing out on.

We will also have to decide whether we want to support multiple pexpect versions. If 4.0 is the minimum required version for Windows, it might be worth asking other platform owners if they are ok with requiring pexpect 4.0 for their platforms as well. If it is ok, this could simplify some of the logic in the test suite and we won't have to worry about someone using one version of pexpect breaking someone that uses a different version.

I CC'ed the OSX platform owner and the Android platform owner to see if they have any concerns about bumping the required pexpect version.

I'm looking forward to one day seeing the pexpect tests working on Windows.

I believe it makes sense to bump up pexpect version for every platform to 4.0.
Multiple versions of pexpect may introduce additional complexity since theoretically we may get skewed results per platform due differences between pexpect versions?

It looks like LLDB carries its own version of pexpect (under third_party/Python/module/pexpect-2.4). So replacing it with 4.0 shouldn't be a problem?

What is a problem, though, is that LLDB is one of those programs that alter behavior depending on whether stdio is attached to a tty. I tried replacing pexpect.spawn with pexpect.popen_spawn.PopenSpawn (after upgrading pexpect), but that doesn't work because LLDB won't output the "(lldb)" prompt when stdout is a pipe.
One way to fix this is to give LLDB some sort of --force-interactive flag (and probably add File::SetIsInterative() method). Another - to modify all tests to not expect the prompt, though it might make them more fragile...
Thoughts?

mentioned in issue #24826

I tried switching to PopenSpawn for all the tests that don't need interaction or to set terminal width. By the time you do that, you get 1 maybe 2 new tests running on Windows (once you port the example source too). So I don't think pursuing that direction right now is worth it (this issue itself is still valid though, ideally we'd have parity across platforms).

Of the tests marked with llvm.org/pr22274, most are benchmarks and then there were 4 other tests I had to mark as interactive (and therefore not going to run on Windows at all).