gh-159446: Release the GIL less in FileIO open and readall

[cmaloney]

When FileIO is doing the open use one GIL release for both the open and fstat. Passing a file descriptor or using an opener still does the fstat separately.

readall always took the GIL at least twice: Once to read the data and a second time through the loop for the EOF. On small files the GIL release and reacquire was a significant portion of time. Move so the readall loop checks for signals just as often but only releases the GIL once before I/O and re-acquires after. _Py_read_noraise() is the platform read() factored out of _Py_read() for callers that release the GIL themselves, like _Py_write_noraise().

Reading a small file with FileIO(path).readall() + close(), PGO+LTO, 4 interleaved rounds of 10 processes:

Benchmark	base	this change
fileio_readall	3.68 us	3.60 us: 1.02x faster

Callgrind, instructions per call:

Build	base	this change
PGO+LTO	10,646	9,842
-O3, no PGO	14,154	13,303

Benchmarks

bm_fileio_readall.py:

import io
import pyperf

def fileio_readall():
    with io.FileIO(__file__) as f:
        f.readall()

runner = pyperf.Runner()
runner.bench_func("fileio_readall", fileio_readall)

for round in 1 2 3 4; do  # alternate the order between rounds
    base/python bm_fileio_readall.py -p 10 -q --append base.json
    change/python bm_fileio_readall.py -p 10 -q --append change.json
done
python -m pyperf compare_to base.json change.json --table

callgrind_readall.py:

import io
import sys

for _ in range(int(sys.argv[1])):
    with io.FileIO(__file__) as f:
        f.readall()

# Instructions per call: (Ir(2000) - Ir(0)) / 2000
for n in 0 2000; do
    PYTHONHASHSEED=0 valgrind --tool=callgrind --callgrind-out-file=cg.$n \
        ./python -S callgrind_readall.py $n
done

Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com

---

This has a NEWS for both the changes together.

• Issue: Reduce unnecessary work in FileIO #159446