Please insert the usual caveats around how reliable benchmarking is impossible. (Last month when I tried this with a previous version of the patch, the interpreter that had the patch applied -- and thus was doing slightly *more* work -- was consistently a few percent faster in general, because something something PGO I guess? Compilers are weird.)

That said, I'm not seeing any measurable difference at all with this patch. This isn't surprising: in the default mode (tracking disabled), the extra operations are:

coroutine creation: 'if (predictably_true) { self->a = NULL; self->b = NULL }'

send/throw: 'if (genobject->ob_type == PyCoro_Type && self->a != NULL) { /* not executed, because the condition always fails */ }'

----- benchmark details

I built 3954f6126f (head of my PR branch, labeled "with-unawaited-coroutines" below) and f23746a934 (its immediate parent, on master, labeled "without-unawaited-coroutines" below), with --enable-optimizations, on a Thinkpad T450s running 64-bit Linux. I ran a script (included below) that simply created and executed a trivial coroutine over and over, and measured using perf.

2 runs of each, alternating:

~/src/cpython$ without-unawaited-tracking/install/bin/python3 corobench.py
.....................
coroutine creation/close: Mean +- std dev: 3.90 us +- 0.11 us
~/src/cpython$ with-unawaited-tracking/install/bin/python3 corobench.py   
.....................
coroutine creation/close: Mean +- std dev: 3.90 us +- 0.09 us
~/src/cpython$ without-unawaited-tracking/install/bin/python3 corobench.py
.....................
coroutine creation/close: Mean +- std dev: 3.93 us +- 0.21 us
~/src/cpython$ with-unawaited-tracking/install/bin/python3 corobench.py   
.....................
coroutine creation/close: Mean +- std dev: 3.91 us +- 0.10 us


Script: attached as "corobench.py"