Composing circuit with a switch statement causes a bug in the QASM3 exporter

[taalexander]

Environment

• Qiskit Terra version: 0.25.0
• Python version: 3.9
• Operating system: OSX

What is happening?

When composing a circuit with another circuit containing a switch statement and then exporting to OpenQASM 3. I receive an error

KeyError: "'ClassicalRegister(3, 'random')' is not defined in the current context"

This may be related to #9746.

How can we reproduce the issue?

Run

from qiskit import QuantumCircuit, QuantumRegister, ClassicalRegister

qr = QuantumRegister(3)
cr_random = ClassicalRegister(3, "random")
cr_result = ClassicalRegister(3, "result")
qc = QuantumCircuit(qr, cr_random, cr_result)

# build a random state
qc.h(qr)
# measure the state
qc.measure(qr, cr_random)

with qc.switch(cr_random) as case:
    with case(1):
        qc.x(qr[0])
    with case(2):
        qc.x(qr[1])
    with case(3):
        qc.x(qr[0])
        qc.x(qr[1])
    with case(4):
        qc.x(qr[2])
    with case(5):
        qc.x(qr[0])
        qc.x(qr[2])
    with case(6):
        qc.x(qr[1])
        qc.x(qr[2])
    with case(7):
        qc.x(qr)
    with case(case.DEFAULT):
        pass

qc.measure(qr, cr_result)

composed = QuantumCircuit(3, 6)
composed.compose(qc, inplace=True)

from qiskit import qasm3
print(qc_qasm3 := qasm3.dumps(composed, disable_constants=True, experimental=qasm3.ExperimentalFeatures.SWITCH_CASE_V1))

What should happen?

QASM3 should be exported.

Any suggestions?

No response

[taalexander]

I have verified that using append also does not work and results in the same error.

[dieris]

I believe what needs to happen is to remap the ClassicalRegister in the SwitchCaseOp target analogously to the condition in the IfElseOp here

[dieris]

I believe what needs to happen is to remap the ClassicalRegister in the SwitchCaseOp target analogously to the condition in the IfElseOp here

this may be obvious, but I just realized that this only addresses the issue with compose. For append to work, a dual fix is also needed for the DAGCircuit in substitute_node_with_dag to map the target cbits when the operation has only a target instead of a condition.

[jakelishman]

The surface-level problem in compose is not too difficult to fix, but that alone won't be enough in the current form of qasm3-runner to make things work. There's further demons lurking in nested control flow, I'm fairly sure, but that's a further issue. For the time being in append, I think circuit_to_instruction will need to error out if there are ever any classical registers that are used; there's no current way for this to properly make sense as we've got no concept of scoping a classical value.

If I were to patch compose to handle the simplest case, the code as given would then fail in the OQ3 exporter. It's not immediately clear to me how we could / if we should fix that entirely from the Terra side in a way that the QSS compiler can ingest. The initial creation of composed contains one ClassicalRegister(6, name="c"), so when we compose the circuit with two registers onto it, we're going to have overlap between c (in composed) and cr_random (in qc). Qiskit will happily accept this (it's valid in our model), but I don't think qss-compiler supports the OQ3 aliasing semantics needed to represent it.

As a temporary work-around at least, the base of the composition (composed, in this case) could be created as

from qiskit.circuit import QuantumCircuit, Qubit, Clbit

base = QuantumCircuit([Qubit() for _ in [None]*n_qubits], [Clbit() for _ in [None]*n_clbits])

(see #6496 for an issue that's proposing that these semantics should be the default for QuantumCircuit(int, int), but that would be a breaking change in Terra so it can't happen overnight). That may be sufficient for your use cases, though it doesn't fully solve everything.