Performance regression with two-step transpilation #7217

[t-imamichi]

What is the expected enhancement?

#7217 causes significant performance regression on VQE. We should fix the performance regression.

from timeit import timeit

from qiskit import Aer
from qiskit.algorithms import VQE
from qiskit.algorithms.optimizers import COBYLA
from qiskit.circuit.library import RealAmplitudes
from qiskit.opflow import PauliSumOp
from qiskit.utils import QuantumInstance

quantum_instance = QuantumInstance(backend=Aer.get_backend('aer_simulator'), shots=1)
vqe = VQE(ansatz=RealAmplitudes(), optimizer=COBYLA(maxiter=1), quantum_instance=quantum_instance)
op = PauliSumOp.from_list([("ZZ"*5, 1), ("IX"*5, 2)])
print(f'{timeit(lambda: vqe.compute_minimum_eigenvalue(op), number=10)} sec')

fe0abaa

1.525182469 sec

0196754

3.796546701 sec