this is a bad idea, because running multiple instances of the
interpreter is known not to work before ghc-8.10; but that's the point,
I want to reproduce the bad things on a pre-8.10 compiler so I can
confirm that those bad things no longer occur in 8.10

it seems to trigger all kinds of problems, but not every time. I would
like a more minimal and more reliable test case in order to better
understand the problem and ideally write a test case which only passes
on ghc-8.10 and up

each thread was supposed to fill a separate MVar, but they were
accidentallly both filling in the same one.

the two copies of the test would both try to modify the global signal
handler table and only one of them would succeed.