bpo-39776 has shown that it is hard to understand the interaction between
ContextVars and threading in embedded scenarios.

I want to understand the code again, so I'm adding back a compile time
option to enable the thread local context that was present prior to
f13f12d.

Also, when I'm debugging things like bpo-39776, I don't want to
switch between Python versions.

New changeset 815280e by Stefan Krah in branch 'master':
bpo-39794: Add --without-decimal-contextvar (bpo-18702)
815280e

New changeset 4d70124 by Stefan Krah in branch '3.8':
[3.8] bpo-39794: Add --without-decimal-contextvar (GH-18702)
4d70124

New changeset c4ca1f8 by Stefan Krah in branch '3.7':
[3.7] bpo-39794: Add --without-decimal-contextvar (GH-18702)
c4ca1f8

While i don't think this was needed in 3.7 and 3.8, you kept the default behavior the same so I don't think it is a problem.

If someone builds an interpreter with this configure flag, does it break compatibility with anything that code may have started to expect as of 3.7?

+1 for the backport. A build option isn't new feature, but it does provide a debugging tool and work-around for people who are having problems.

For the people who find this issue via a search engine:

1. The defaults in 3.7, 3.8 and 3.9 are unchanged (you get
   the ContextVar).

2. ./configure --without-decimal-contextvar enables the exact
   TLS context code from 3.3-3.6. There is no new code.

3. On Windows you need to edit PC\pyconfig.h to enable the TLS
   context.

4. There are extensive tests for the flag in:

   Modules/_decimal/tests/runall-memorydebugger.sh

If someone builds an interpreter with this configure flag, does it break
compatibility with anything that code may have started to expect as of 3.7?

Yes, anything that relies on the implicit context being coroutine-safe
does not work. The only test that expectedly breaks in the stdlib is:

Lib/test/test_asyncio/test_context.py

Basically you can't use operators inside coroutines in the same thread
if both coroutines use a different implicit context.

You can of course replace all operators inside coroutines by the
corresponding context methods:

   # Unsafe with TLS context:
   async def foo():
       with localcontext(Context(prec=10)):
           x = Decimal(1) / 9

   # Safe, just avoid the TLS context:
   async def foo():
       c = Context(prec=10)
       x = c.divide(Decimal(1), 9)

To make things more clear (it is good that Gregory asked):

ONLY use this flag if you don't use coroutines at all
OR control ALL async libraries in your application.

Random async libraries from PyPI may rely on the
ContextVar and silently give incorrect results.

bpo-39776 is fixed, but _pydecimal was also affected, see msg363266.