Lazily compute lineno of tracebacks for efficient exception handling

[kumaraditya303]

By lazily computing tb_lineno of tracebacks, we can avoid parsing the line and column table and avoid a lot of overhead that comes with accessing line number information.

See faster-cpython/ideas#407 for discussion.

[kumaraditya303]

PR #95237 computes line number lazily.

This microbenchmark shows a 25% speedup when handling exception:

import pyperf
from pyperf import Runner


def bench(loops):
    x = object()
    t1 = pyperf.perf_counter()
    for i in range(loops):
        try:
            x.y
        except AttributeError:
            pass
    t2 = pyperf.perf_counter()
    return t2 - t1


runner = Runner()
runner.bench_time_func('bench exception handling', bench)

Benchmark results:

Benchmark	main	patch
bench exception handling	614 ns	491 ns: 1.25x faster

This is on Linux release build without PGO and LTO.

[kumaraditya303]

PR #95237 makes tb_lineno of PyTracebackObject lazy. C extension modules must call PyTraceback_GetLineNumber to get the line number.

@scoder @da-woods Can you confirm if Cython will be affected by this change? Thanks!

See also discourse discussion

[da-woods]

I did a quick search of Cython and couldn't find any use of tb_lineno so I don't think it's a problem. Even if it was, 3.12 is a while off so there would be plenty of time to find a solution.

[gpshead]

Results from a pyperformance run would be more informative than the microbenchmark when it comes to deciding about the deprecation timeline exemption asked for in python/steering-council#136.

[kumaraditya303]

Another benchmark speedup by 50% on Linux gcc with PGO build:
Benchmark:

import pyperf
from pyperf import Runner


def bench(loops):
    x = {1: "foo", 2: "baz"}
    t1 = pyperf.perf_counter()
    for _ in range(loops):
        try:
            x["fooz"]
        except KeyError:
            pass
    t2 = pyperf.perf_counter()
    return t2 - t1


runner = Runner()
runner.bench_time_func('bench exception handling', bench)

Results

Benchmark	base	patch
bench exception handling	297 ns	196 ns: 1.51x faster

[kumaraditya303]

Results from a pyperformance run would be more informative than the microbenchmark when it comes to deciding about the deprecation timeline exemption asked for in python/steering-council#136.

pyperformance has no benchmarks which use exception handling in hot paths. The speedup if any would be dominated by the noise in the benchmark itself. Also moving pyperformance needle by even 1% isn't easy, the accumulation of these kinds of targeted improvements will result in 1% or more but just one change isn't going to do it.