HLS fix to not synthesize instructions already supported (#13417)

* HLS fix to not synthesize instructions already supported

* reno

* fix for control-flow-operations

The control-flow operations such as 'for-loop` are considered to be a
part of 'basis_gates`, yet they need to be recursively synthesized.

* using faster is_control_flow check

* Update test/python/transpiler/test_high_level_synthesis.py

Co-authored-by: Julien Gacon <gaconju@gmail.com>

---------

Co-authored-by: Julien Gacon <gaconju@gmail.com>

qiskit/transpiler/passes/synthesis/high_level_synthesis.py

@@ -814,6 +814,7 @@ def _definitely_skip_node(
             dag._has_calibration_for(node)
             or len(node.qargs) < self._min_qubits
             or node.is_directive()
+            or (self._instruction_supported(node.name, qubits) and not node.is_control_flow())
         ):
             return True
 
@@ -831,15 +832,12 @@ def _definitely_skip_node(
             # If all the above constraints hold, and it's already supported or the basis translator
             # can handle it, we'll leave it be.
             and (
-                self._instruction_supported(node.name, qubits)
                 # This uses unfortunately private details of `EquivalenceLibrary`, but so does the
                 # `BasisTranslator`, and this is supposed to just be temporary til this is moved
                 # into Rust space.
-                or (
-                    self._equiv_lib is not None
-                    and equivalence.Key(name=node.name, num_qubits=node.num_qubits)
-                    in self._equiv_lib.keys()
-                )
+                self._equiv_lib is not None
+                and equivalence.Key(name=node.name, num_qubits=node.num_qubits)
+                in self._equiv_lib.keys()
             )
         )
 

releasenotes/notes/fix-hls-supported-instructions-0d80ea33b3d2257b.yaml

@@ -0,0 +1,8 @@
+---
+fixes:
+  - |
+    Previously the :class:`.HighLevelSynthesis` transpiler pass synthesized an
+    instruction for which a synthesis plugin is available, regardless of
+    whether the instruction is already supported by the target or a part of
+    the explicitly passed ``basis_gates``. This behavior is now fixed, so that
+    such already supported instructions are no longer synthesized.

test/python/transpiler/test_high_level_synthesis.py

@@ -46,6 +46,7 @@
     IGate,
     MCXGate,
     SGate,
+    QAOAAnsatz,
 )
 from qiskit.circuit.library import LinearFunction, PauliEvolutionGate
 from qiskit.quantum_info import Clifford, Operator, Statevector, SparsePauliOp
@@ -664,6 +665,18 @@ def test_synth_fails_definition_exists(self):
             out = hls(circuit)
             self.assertEqual(out.count_ops(), {"u": 1})
 
+    def test_both_basis_gates_and_plugin_specified(self):
+        """Test that a gate is not synthesized when it belongs to basis_gates,
+        regardless of whether there is a plugin method available.
+
+        See: https://github.com/Qiskit/qiskit/issues/13412 for more
+        details.
+        """
+        qc = QAOAAnsatz(SparsePauliOp("Z"), initial_state=QuantumCircuit(1))
+        pm = PassManager([HighLevelSynthesis(basis_gates=["PauliEvolution"])])
+        qct = pm.run(qc)
+        self.assertEqual(qct.count_ops()["PauliEvolution"], 2)
+
 
 class TestPMHSynthesisLinearFunctionPlugin(QiskitTestCase):
     """Tests for the PMHSynthesisLinearFunction plugin for synthesizing linear functions."""