gh-150816: Speed up inspect.signature() for Python functions

[gaborbernat]

inspect.signature() builds a Signature whose every Parameter goes through the public Parameter constructor, which validates the name and the kind on each call. When the parameters come from a function's own code object, the names are already valid identifiers and the kinds are already the canonical constants, so that validation re-checks facts that are guaranteed to hold. Signature introspection runs constantly across the ecosystem — web frameworks resolving view and dependency arguments, click building commands, pytest wiring fixtures, serialization and validation layers — usually once per function and often at import time while an application is assembled.

This adds an internal Parameter._from_valid_args() used only by _signature_from_function, the trusted caller that reads parameters from a code object. It populates the instance directly and skips the redundant checks. The one case that still needs the constructor's handling, a comprehension's implicit .0 argument, is detected and deferred to it. The public Parameter() constructor and all of its validation are untouched, so every other caller behaves exactly as before.

Taking the signature of 400 real callables collected from popular packages (requests, click, jinja2, sqlalchemy, werkzeug, flask and others) improves from 1.46 ms to 1.16 ms, 26% faster.

Benchmark	base	patched
inspect.signature over 400 real callables	1.46 ms	1.16 ms: 26% faster

Benchmark (pyperf)

Run base vs patched by swapping Lib/inspect.py on the same interpreter. The script collects callables from installed third-party packages when present and always includes several stdlib modules, so it runs with no extra installs.

import inspect, importlib, pyperf

names = ["requests", "click", "jinja2", "sqlalchemy", "werkzeug", "flask",
         "urllib3", "rich.console", "json", "argparse", "logging", "http.client",
         "email.message", "configparser", "xml.etree.ElementTree", "subprocess"]
funcs = []
for name in names:
    try:
        mod = importlib.import_module(name)
    except ImportError:
        continue
    for attr in dir(mod):
        obj = getattr(mod, attr, None)
        if inspect.isfunction(obj):
            funcs.append(obj)
        elif inspect.isclass(obj):
            funcs.extend(m for m in vars(obj).values() if inspect.isfunction(m))
good = []
for f in funcs:
    try:
        inspect.signature(f); good.append(f)
    except (ValueError, TypeError):
        pass
good = good[:400]

runner = pyperf.Runner()
runner.metadata["n_callables"] = len(good)
runner.bench_func("inspect.signature over %d callables" % len(good),
                  lambda: [inspect.signature(f) for f in good])

Resolves #150816.

• Issue: Speed up inspect.signature() for Python functions #150816

[skirpichev]

In order to keep the commit history intact, please avoid squashing or amending history and then force-pushing to the PR. Reviewers often want to look at individual commits. When the PR is merged, everything will be squashed into a single commit.

[skirpichev]

Looks good, with few comments.

CC @sobolevn

[gaborbernat]

Good point — I benchmarked both: inlining the recast and deferring the rare .0 comprehension case to __init__ perform identically, so I've kept the simpler version that doesn't duplicate __init__'s logic, as you suggested.

[gaborbernat]

Good catch, thank you. You're right that _signature_from_function also builds signatures for function-like objects (is_duck_function), whose code object's co_varnames can be arbitrary strings, so skipping the name validation there was wrong.

Fixed in the latest commit: the _from_valid_args fast path is now used only for real Python functions, where co_varnames are guaranteed valid identifiers. Function-like objects go through the validating Parameter() constructor, so a duck object with an invalid parameter name raises ValueError exactly as before. Added a check for that case; test_inspect passes and the benchmark is unchanged (1.3x on real functions).

[skirpichev]

@gaborbernat, please stop force-pushing.

[gaborbernat]

@gaborbernat, please stop force-pushing.

Sorry, I think those were my comments from yesterday, I added separate commit today.