gh-118702: Implement vectorcall for BaseException

[vstinner]

• BaseException_vectorcall() now creates a tuple from 'args' array.
• Creation an exception using BaseException_vectorcall() is now a single function call, rather than having to call BaseException_new() and then BaseException_init(). Calling BaseException_init() is inefficient since it overrides the 'args' attribute.
• _PyErr_SetKeyError() now uses PyObject_CallOneArg() to create the KeyError instance to use BaseException_vectorcall().

Micro-benchmark on creating a KeyError on accessing a non-existent dictionary key:

Mean +- std dev: 447 ns +- 31 ns -> 373 ns +- 15 ns: 1.20x faster

• Issue: Implement vectorcall for BaseException to optimize creation of exception instances #118702

[vstinner]

Benchmark:

import pyperf

EMPTY_DICT = {}
INNER_LOOPS = 1024
KEYS = tuple(range(INNER_LOOPS))

def bench_keyerror():
    d = EMPTY_DICT
    for key in KEYS:
        try:
            d[key]
        except KeyError:
            pass

runner = pyperf.Runner()
runner.bench_func('keyerror', bench_keyerror, inner_loops=INNER_LOOPS)

[vstinner]

cc @erlend-aasland @corona10 @serhiy-storchaka

[erlend-aasland]

Nice change, but you broke CI :)

[vstinner]

Nice change, but you broke CI :)

Right, FAIL: testKeywordArgs (test.test_exceptions.ExceptionTests.testKeywordArgs).

It should now be fixed.

[serhiy-storchaka]

I am surprised that switching to vector call makes a difference here. How is it in comparison with the following BaseException_vectorcall implementation?

argstuple = _PyTuple_FromArray(args, PyVectorcall_NARGS(nargsf));
self = type_obj->tp_new(type_obj, argstuple, NULL);
Py_TYPE(self)->tp_init(self, argstuple, NULL);
Py_DECREF(argstuple);

If it is still faster than the current code, then there is something wrong in the non-vectorcall path.

[eendebakpt]

I am surprised that switching to vector call makes a difference here. How is it in comparison with the following BaseException_vectorcall implementation?

argstuple = _PyTuple_FromArray(args, PyVectorcall_NARGS(nargsf));
self = type_obj->tp_new(type_obj, argstuple, NULL);
Py_TYPE(self)->tp_init(self, argstuple, NULL);
Py_DECREF(argstuple);

If it is still faster than the current code, then there is something wrong in the non-vectorcall path.

In current main there is a call to PyTuple_Pack that is not in the vectorcall path. I did not benchmark this, but I suspect that could explain part of the difference

[vstinner]

If it is still faster than the current code, then there is something wrong in the non-vectorcall path.

This change avoids calling type_call() which contains more code.

[vstinner]

Benchmark result with CPU isolation (this PR):

Mean +- std dev: [ref] 264 ns +- 1 ns -> [baseexc] 227 ns +- 1 ns: 1.16x faster

I also measured only the Python/errors.c changes; replace PyTuple_Pack()+PyObject_Call() with PyObject_CallOneArg():

Mean +- std dev: [ref] 264 ns +- 1 ns -> [optim_keyerror] 248 ns +- 2 ns: 1.07x faster

It's faster, but not as fast as this PR.

[serhiy-storchaka]

This change avoids calling type_call() which contains more code.

But was is wrong with type_call()? On what does it spent time besides calling tp_new(), tp_init() and few trivial checks that it is so slower? Or maybe the difference is just the cost of these trivial checks?

[erlend-aasland]

But was is wrong with type_call()? On what does it spent time besides calling tp_new(), tp_init() and few trivial checks that it is so slower? Or maybe the difference is just the cost of these trivial checks?

AFAICS, in _PyObject_VectorcallTstate, a vectorcall function is called directly, avoiding the overhead of _PyObject_MakeTpCall and type_call.

[serhiy-storchaka]

It is not the same. It does not set __context__, and there may be other differences in case when it fails to create an exception object.

[vstinner]

Oh, right. I fixed this regression and added an unit test.

[serhiy-storchaka]

The half of the gain is due to the change in _PyErr_SetKeyError() which is not equivalent to the current code, so it perhaps should be reverted. Addition of BaseException_vectorcall is still beneficial right now, but optimizing _PyObject_MakeTpCall and type_call may reduce the gain even more. And it does not beneficial for exceptions with __new__ or __init__. In the end, it may make the benefit/cost ratio pretty low.

I do not actively oppose this change, I only suggest that there may be place for more general optimization. If you still want to add BaseException_vectorcall, with or without optimizing the other path, you are free to do this.

[vstinner]

The PR is now ready for a review. The main branch is now Python 3.14.

I rebased the PR on the main branch and squashed commit. I fixed the __context__ regression.

Updated benchmark, Python built with gcc -O3, with CPU isolation:

• key_error: Mean +- std dev: [ref] 267 ns +- 1 ns -> [change] 243 ns +- 1 ns: 1.10x faster
• value_error: Mean +- std dev: [ref] 288 ns +- 2 ns -> [change] 258 ns +- 2 ns: 1.12x faster

I added a benchmark on raise ValueError.

+----------------+--------+----------------------+
| Benchmark      | ref    | change               |
+================+========+======================+
| key_error      | 267 ns | 243 ns: 1.10x faster |
+----------------+--------+----------------------+
| value_error    | 288 ns | 258 ns: 1.12x faster |
+----------------+--------+----------------------+
| Geometric mean | (ref)  | 1.11x faster         |
+----------------+--------+----------------------+

Benchmark:

import pyperf

EMPTY_DICT = {}
INNER_LOOPS = 1024
KEYS = tuple(range(INNER_LOOPS))

def bench_key_error(loops):
    range_it = range(loops)
    t0 = pyperf.perf_counter()

    value = "value"
    for _ in range_it:
        d = EMPTY_DICT
        for key in KEYS:
            try:
                d[key]
            except KeyError:
                pass

    return pyperf.perf_counter() - t0


def bench_value_error(loops):
    range_it = range(loops)
    t0 = pyperf.perf_counter()

    value = "value"
    for _ in range_it:
        try: raise ValueError(value)
        except: pass
        try: raise ValueError(value)
        except: pass

        try: raise ValueError(value)
        except: pass
        try: raise ValueError(value)
        except: pass

        try: raise ValueError(value)
        except: pass
        try: raise ValueError(value)
        except: pass

        try: raise ValueError(value)
        except: pass
        try: raise ValueError(value)
        except: pass

        try: raise ValueError(value)
        except: pass
        try: raise ValueError(value)
        except: pass

    return pyperf.perf_counter() - t0

runner = pyperf.Runner()
runner.bench_time_func('key_error', bench_key_error, inner_loops=INNER_LOOPS)
runner.bench_time_func('value_error', bench_value_error, inner_loops=10)

[vstinner]

@serhiy-storchaka:

The half of the gain is due to the change in _PyErr_SetKeyError() which is not equivalent to the current code, so it perhaps should be reverted.

I fixed the code.

Addition of BaseException_vectorcall is still beneficial right now,

10% to 12% faster to create exceptions sound appealing since raising KeyError is a common operation.

I ran benchmark on functions using METH_VARVARGS. KeyError is the clear winner in number of calls when running ./python -m test test_builtin.

but optimizing _PyObject_MakeTpCall and type_call may reduce the gain even more. And it does not beneficial for exceptions with new or init. In the end, it may make the benefit/cost ratio pretty low.

I don't see any simple optimization opportunity, do you?

I do not actively oppose this change, I only suggest that there may be place for more general optimization.

The optimization applies to 32 built-in exceptions:

• ArithmeticError
• AssertionError
• BaseException
• BufferError
• BytesWarning
• DeprecationWarning
• EOFError
• EncodingWarning
• Exception
• FloatingPointError
• FutureWarning
• GeneratorExit
• ImportWarning
• IndexError
• KeyError
• KeyboardInterrupt
• LookupError
• NotImplementedError
• OverflowError
• PendingDeprecationWarning
• PythonFinalizationError
• RecursionError
• ReferenceError
• ResourceWarning
• RuntimeError
• RuntimeWarning
• StopAsyncIteration
• SyntaxWarning
• SystemError
• TypeError
• UnicodeError
• UnicodeWarning
• UserWarning
• ValueError
• Warning
• ZeroDivisionError

[vstinner]

Merged, thanks for reviews.