Performance gap of match statements of classes compared with equivalent Python 3.9 code

[tonybaloney]

Im doing some micro-benchmarks of the match statement.

The following comparison shows two functionally equivalent functions of the class matching pattern in the match statement.

class Driver:
 def __init__(self, name, team, **extra):
 self.name = name
 self.team = team
 self.extra = extra

def bench_class_matching_statement():
 drivers = [
 Driver(name="Max Verstappen", team="Red Bull", ),
 Driver(name="Sergio Perez", team="Red Bull", ),
 Driver(name="Charles Leclerc", team="Ferrari", ),
 Driver(name="Lewis Hamilton", team="Mercedes", ),
 ]

 for _ in range(100_000):
 for driver in drivers:
 match driver:
 case Driver(name="Max Verstappen"): desc = f"Max Verstappen, the current world #1"
 case Driver(name=name, team="Ferrari"): desc = f"{name}, a Ferrari driver!! 🐎"
 case Driver(name=name, team=team): desc = f"{name}, a {team} driver."
 case _: desc = "Invalid request"
 # print(desc)


def bench_class_matching_logical():
 drivers = [
 Driver(name="Max Verstappen", team="Red Bull", ),
 Driver(name="Sergio Perez", team="Red Bull", ),
 Driver(name="Charles Leclerc", team="Ferrari", ),
 Driver(name="Lewis Hamilton", team="Mercedes", ),
 ]

 for _ in range(100_000):
 for driver in drivers:
 if not isinstance(driver, Driver):
 desc = "Invalid request"
 elif driver.name == "Max Verstappen":
 desc = f"Max Verstappen, the current world #1"
 elif driver.team == "Ferrari": 
 desc = f"{driver.name}, a Ferrari driver!! 🐎"
 else:
 desc = f"{driver.name}, a {driver.team} driver."
 # print(desc)

Python 3.11 executes bench_class_matching_statement() at 4x the execution time of bench_class_matching_logical()
Python 3.10 executes bench_class_matching_statement() at 2.5x the execution time of bench_class_matching_logical()

Python 3.11b1 is showing a speedup of both functions (which is great).

Pattern matching for sequences and mapping is faster than the equivalent Python code, but for classes it is significantly slower.

[tonybaloney]

CC @markshannon @brandtbucher

[JelleZijlstra]

So just to be clear, both functions are faster in 3.11 than in 3.10, but the bug is that the speedup in one of them is greater than in the other? That hardly seems like a bug.

[tonybaloney]

Sorry, that was unclear.

The bug is that pattern matching of classes per the spec is slower than the equivalent Python 3.9< code.

I had a conversation with Brandt about this the other week. He was working on a patch to improve the performance, so I wanted to capture this issue to track it.

FF to close if this should go somewhere else (mailing lists)

[eendebakpt]

#138915 improved the performance of match. This can probably be closed as a duplicate of (the newer) #138912