gh-158585: Optimize bytes object creation

[vstinner]

Add bytes_alloc() helper function. bytes_alloc() and _PyBytes_FromStringAndSize() have less conditional branches than previous code, and so should be a little bit faster.

• Rename existing bytes_alloc() to bytes_type_alloc().
• Replace _PyBytes_FromSize(calloc=1) with _PyBytes_FromSizeZero().
• Add set_ob_shash_unsafe(): similar to set_ob_shash() but don't use an atomic operation on Free Threading. Use this new function on newly allocated bytes objects and in bytes_resize_inplace().
• Replace PyBytes_FromStringAndSize(NULL, size) with bytes_alloc(size).
• Use bytes_alloc() in PyBytes_FromString() and _PyBytes_Repeat().
• Add _PyBytes_FromStringAndSize(): similar to PyBytes_FromStringAndSize(), but str must not be NULL.
• Replace PyBytes_FromStringAndSize() with _PyBytes_FromStringAndSize().
• Test that PyBytes_FromString() and PyBytes_FromStringAndSize() return singletons for 0 or 1 bytes.

• Issue: C API: Rework and enhance the PyBytesWriter implementation #158585

[vstinner]

Results of PR gh-158665 benchmark: Mean +- std dev: [ref] 34.8 ns +- 0.4 ns -> [change] 32.2 ns +- 0.5 ns: 1.08x faster.

The change makes PyBytesWriter_Finish() 1.08x faster, it saves 2.6 nanoseconds.

[vstinner]

Oh. A side effect of this change is that b'abc' * 0 now returns the empty string singleton. I added a test for that.

[vstinner]

I also ran a benchmark on PyBytes_FromStringAndSize("abc", 3) and PyBytes_FromString("abc"): there is no significant different on performance.

        PyObject *bytes = PyBytes_FromStringAndSize("abc", 3);
        if (bytes == NULL) {
            return NULL;
        }
        Py_DECREF(bytes);

and:

        PyObject *bytes = PyBytes_FromString("abc");
        if (bytes == NULL) {
            return NULL;
        }
        Py_DECREF(bytes);

[vstinner]

Ooops, I pushed "Add static inline _Py_NewReferenceInline() function" change by mistake. This time, I used git push --force to remove this commit. It shouldn't be part of this PR.

[eendebakpt]

@vstinner Auto merge?

[eendebakpt]

There are some corner cases that hit an assert in debug builds. Reproducer:

import sys

class B(bytes):
    pass

def rt(x):          # keep the compiler from constant-folding
    return x

empty = rt(b'')
print(sys.version)

cases = {
    "B() + B()":                  lambda: B() + B(),
    "B() + bytearray()":          lambda: B() + bytearray(),
    "B() + memoryview(b'')":      lambda: B() + memoryview(b''),
    "b'' + B()  (control)":       lambda: empty + B(),
}
for name, fn in cases.items():
    r = fn()                      # debug build: aborts here with
                                  # "Assertion failed: size >= 1 ... bytesobject.c"
    print(f"{name:28} -> {r!r:6} is b'': {r is empty}")

Note: returning the singleton more often is a nice-to-have, if we return another empty bytes in some cases that would be fine for me.

[vstinner]

@vstinner Auto merge?

I consider that the PR is ready to be merged, so yes, I enabled auto merged. After I saw your comment with a regression, I disabled auto merge.

There are some corner cases that hit an assert in debug builds. Reproducer: (...)

Oh, well spotted! I misread if (va.len == 0 && PyBytes_CheckExact(b)) { in _PyBytes_Concat(), I missed the PyBytes_CheckExact() part of the condition.

It seems like this case wasn't tested at all by the Python test suite. I fixed bug and I added tests (test_concat_cpython).

Note: returning the singleton more often is a nice-to-have, if we return another empty bytes in some cases that would be fine for me.

The PR only changes bytes * int to return the empty string singleton if the number is zero. It's more a side effect of the implementation rather than a deliberate optimization :-)