bug#36609: 27.0.50; Possible race-condition in threading implementation

[dick . r . chiang]

EZ> Is it, really?

Yes.  While I can't claim to be any more knowledgeable, the naivete of someone
protesting thread atomicity based on word-size is stunningly apparent.

EZ> And you can show a GDB backtrace with values of the variable that supports
EZ> that?

Alas I cannot.  My gdb skills are less impressive in the presence of threads,
and heisenbugs of this nature tend to mysteriously disappear when gdb's
leisurely pace gives caches ample time to flush.

I mentioned earlier I could not reproduce my MRE failure when I merely inserted
print statements.

My patch isn't quite as studied as Pip Cet's (my patch merely skirts locking
the glib unless you're main-thread-p), but it does fix OP and, unlike the
current state of the code, doesn't admit glaring theoretical flaws, even
though you seem to think those flaws could never materialize in practice
despite evidence (gdb-less as it is) to the contrary.