python / python/cpython

`METH_METHOD` calling convention is now not so efficient

Open
#123,500 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

performance topic-C-API type-bug
Dominant language
Python
Stars
77.2k
Forks
35.9k
PR merge metrics
PR metrics pending

Description

Bug report

Bug description:

There are callables implemented with the METH_METHOD|METH_FASTCALL signature in C. They can be 5%-15% less efficient than using only METH_FASTCALL (or METH_O) with a PyType_GetModuleByDef function call.

For example, I measured the difference on Windows PGO builds by duplicating functions:

  • CDataType_from_buffer_copy() in _ctypes.c, which is not called when profiling:

    from timeit import timeit
    setup = """if 1:
        import ctypes
        buf = bytearray(16)
        cls = ctypes.c_char * len(buf)
    """
    # with a warmup
    for _ in range(2):
        # METH_METHOD|METH_FASTCALL (as-is)
        r0 = timeit(s0 := f'cls.from_buffer_copy (buf)', setup)
    
        # METH_FASTCALL (no `defining_class`) + PyType_GetModuleByDef
        r1 = timeit(s1 := f'cls.from_buffer_copy1(buf)', setup)
    
    print(s0, r0, 1 + (1 - r0 / r0))
    print(s1, r1, 1 + (1 - r1 / r0))
    
    cls.from_buffer_copy (buf) 0.15552800190635024 1.0
    cls.from_buffer_copy1(buf) 0.13187471489945893 1.1520837837364741
    
  • dec_mpd_qquantize() in _decimal.c profiled with 6800 calls (unfair?):

    # legacy (as-is)
    d1.quantize (d2) 0.1694609627971658 1.0
    
    # METH_METHOD|METH_FASTCALL (`defining_class`) + _PyType_GetModuleState
    d1.quantize1(d2) 0.1408861404022900 1.168621857938327
    
    # METH_FASTCALL (no `defining_class`) + PyType_GetModuleByDef
    d1.quantize2(d2) 0.1258157708973158 1.257553074049807
    
    Script (expand)
    from timeit import timeit
    setup = """if 1:
        from _decimal import Decimal
        d1,d2 = Decimal(1.414), Decimal('0.01')
    """
    for _ in range(2):
        r0 = timeit(s0 := f'd1.quantize (d2)', setup)
        r1 = timeit(s1 := f'd1.quantize1(d2)', setup)
        r2 = timeit(s2 := f'd1.quantize2(d2)', setup)
    
    print(s0, r0, 1 + (1 - r0 / r0))
    print(s1, r1, 1 + (1 - r1 / r0))
    print(s2, r2, 1 + (1 - r2 / r0))
    

Observations:

  • The number of arguments had little to do with this.
  • The gaps seem to be consistent as long as they are equally (un)exercised.
  • The same goes for non-PGO builds and builtin modules (e.g. _sre), where the impacts may be less significant.
CPython versions tested on:

CPython main branch

Operating systems tested on:

Windows

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reproducing the reported benchmarks for CDataType_from_buffer_copy() in _ctypes.c and dec_mpd_qquantize() in _decimal.c, comparing METH_METHOD|METH_FASTCALL with METH_FASTCALL and the module lookup calls shown. Read the METH_METHOD and PyType_GetModuleByDef entry points, then define and validate an accepted approach whose performance is no worse than the alternatives across the cited cases.

Written by the indexing model from the issue text.

Assessment

Tech stack
c, python
Domain
backend, performance
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.