remove_final_measurements leaves clbit unusable

c = QuantumCircuit(2, 5)
c.measure(0, 0)
c.measure(1, 1)
print(c)

     ┌─┐
q_0: ┤M├───
     └╥┘┌─┐
q_1: ─╫─┤M├
      ║ └╥┘
c: 5/═╩══╩═
      0  1

c.remove_final_measurements(inplace=True)
print(c)

q_0:

q_1:

c: 5/

c.measure(0, 0)     # raises error 

Note that as currently defined, remove_final_measurements() is supposed to remove the measurements plus the classical bits. But evidently it does not do that (looking at the two circuits above). Still, the clbit is unusable.

[kevinhartman]

I’m happy to address this while fixing the other issues discussed so far here.

I rewrote the final op determination logic to match the original doc string, while preserving @mtreinish’s bottom-up approach, and resolving a bug that existed in the implementation even before #7039.

Working on fixing other discussed issues with creg removal currently.

[kevinhartman]

While working on this, I noticed that the current behavior of the RemoveFinalMeasurements pass doesn’t match its doc string. Specifically,

Measurements and barriers are considered final if they are followed by no other operations (aside from other measurements or barriers.)

Currently, a measurement or barrier is only removed if it immediately precedes a qubit’s output node in the DAG. Only the final measurement or barrier will be removed in a chain of successive measure and barrier calls.

qr = QuantumRegister(1)
cr = ClassicalRegister(1)
c = QuantumCircuit(qr, cr)

c.measure(0, 0)
c.measure(0, 0)

dag = circuit_to_dag(c)
remove_final_meas = RemoveFinalMeasurements()
remove_final_meas.run(dag)

measure_node_count = [n.op.name for n in dag.op_nodes()].count("measure")

# error, measure_node_count is 1
self.assertEqual(measure_node_count, 0)

What’s the correct behavior here? Perhaps we should add a parameter to support either scenario. In that case, I wonder which should be the default?

I’m happy to address this while fixing the other issues discussed so far here.

@mtreinish @georgios-ts @kdk @ajavadia