Transpilation result of RZZ gate by Qiskit 2.0 is less optimal compared with Qiskit 1.4

[t-imamichi]

Environment

• Qiskit version: 2.0.1
• Python version: 3.12.10
• Operating system: macOS 15.5

What is happening?

The transpilation of the RZZ gate in Qiskit 2.0 is less optimal compared to Qiskit 1.4.
For the example below, Qiskit 2.0 produces circuits with two CZ gates, whereas Qiskit 1.4 results in circuits with zero or one CZ gate.

Credit to @kifumi

How can we reproduce the issue?

from math import pi

from qiskit import QuantumCircuit, __version__, generate_preset_pass_manager

print(f"qiskit version: {__version__}")
pm = generate_preset_pass_manager(optimization_level=3, basis_gates=["rz", "cz", "sx", "x"])

for angle in [pi / 2, pi]:
    qc = QuantumCircuit(2)
    qc.rzz(angle, 0, 1)
    t_qc = pm.run(qc)
    print(t_qc)

output

qiskit version: 2.0.1
global phase: π/4

q_0: ─────────────────────────────■────────■─────────────────────────────
     ┌─────────┐┌────┐┌─────────┐ │ ┌────┐ │ ┌─────────┐┌────┐┌─────────┐
q_1: ┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├─■─┤ √X ├─■─┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├
     └─────────┘└────┘└─────────┘   └────┘   └─────────┘└────┘└─────────┘
global phase: 0

q_0: ─────────────────────────────■───────■─────────────────────────────
     ┌─────────┐┌────┐┌─────────┐ │ ┌───┐ │ ┌─────────┐┌────┐┌─────────┐
q_1: ┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├─■─┤ X ├─■─┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├
     └─────────┘└────┘└─────────┘   └───┘   └─────────┘└────┘└─────────┘

qiskit version: 2.1.0.dev0+31c12d8
global phase: π/4

q_0: ─────────────────────────────■────────■─────────────────────────────
     ┌─────────┐┌────┐┌─────────┐ │ ┌────┐ │ ┌─────────┐┌────┐┌─────────┐
q_1: ┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├─■─┤ √X ├─■─┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├
     └─────────┘└────┘└─────────┘   └────┘   └─────────┘└────┘└─────────┘
global phase: 0

q_0: ─────────────────────────────■───────■─────────────────────────────
     ┌─────────┐┌────┐┌─────────┐ │ ┌───┐ │ ┌─────────┐┌────┐┌─────────┐
q_1: ┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├─■─┤ X ├─■─┤ Rz(π/2) ├┤ √X ├┤ Rz(π/2) ├
     └─────────┘└────┘└─────────┘   └───┘   └─────────┘└────┘└─────────┘

What should happen?

Achieves the same quality as Qiskit 1.4

qiskit version: 1.4.3
global phase: 7π/4
             ┌──────────┐
q_0: ──────■─┤ Rz(-π/2) ├───────────
     ┌───┐ │ └──┬───┬───┘┌─────────┐
q_1: ┤ X ├─■────┤ X ├────┤ Rz(π/2) ├
     └───┘      └───┘    └─────────┘
global phase: π/2
     ┌────────┐
q_0: ┤ Rz(-π) ├
     ├────────┤
q_1: ┤ Rz(-π) ├
     └────────┘

Any suggestions?

No response

[mtreinish]

It looks like there was a regression in the consolidate blocks pass. It's not collecting the 2q block for peephole optimization correctly and that is causing it to not be optimized by unitary synthesis later. Looking through the code this regression was introduced by:
#13884 by this line: https://github.com/Qiskit/qiskit/blob/main/qiskit/transpiler/passes/optimization/consolidate_blocks.py#L137

For any KAK gate that's not "cx" the TwoQubitBasisDecomposer stores a sentinel value "USER_GATE" to indicate that the gate object stored in the decomposer is should be used for the synthesis output directly. CXGate is treated specially because there is a pulse efficient decomposition baked into the decomposer. In the case of CZGate or any other supercontrolled gate that name that gets used with that self.decomposer.gate_name field will be USER_GATE. This means that when the consolidate blocks pass goes to check for the number of basis gates in the block it will find 0 because no gate name will match USER_GATE.

This is any easy fix, but I find it a bit more disturbing that nothing in the test suite caught this bug because it's pretty obvious that the pass would not behave correctly for any non-cx KAK gate.