Use `PyBytesWriter` API in `io.RawIOBase.readall`

[maurycy]

Feature or enhancement

Proposal:

There's new PyBytesWriter() API:

• https://peps.python.org/pep-0782/

There's an ongoing effort of introducing it in various places:

• https://github.com/python/cpython/pulls?q=is%3Apr+pybyteswriter+is%3Aclosed

With its PyList_Append

cpython/Modules/_io/iobase.c

Line 1003 in 2ca3c85

r = PyList_Append(chunks, data);

and PyBytes_Join:

https://github.com/python/cpython/blob/2ca3c85054b231505c5e6f5f3209eb5f63942dce/Modules/_io/iobase.c#L1010C14-L1010C26

the io.RawIOBase.readall looks like the ideal candidate for PyBytesWriter.

Notably, io.RawIOBase is used internally across many modules:

https://github.com/search?q=repo%3Apython%2Fcpython+io.RawIOBase+path%3A%2F%5ELib%5C%2F%2F&type=code

and some even call readall explicitly sometimes:

cpython/Lib/gzip.py

Line 545 in 2ca3c85

return self.readall()

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

Linked PRs

• gh-140135: use PyBytesWriter in io.RawIOBase.readall; 3.98x faster #140139

[cmaloney]

cc: @vstinner, @emmatyping . This a very similar optimization to gh-139877.

This has definitely shown up in user profiles although most the time FileIO.readall (which is an implementation of RawIOBase.readall) ends up getting called

I like this as an optimization although to me is an intermediate state. I really want to update this to use bytearray + readinto + take_bytes() which does the same thing with less allocations + copies (self.read() allocates and returns a new bytes which is copied into the PyBytesWriter here). I don't think that should block this optimization though. Changing io.RawIOBase to defaulting to .readinto from .read is unfortunately complicated to do without breaking compatibility or making infinite loops.