Benchmark:

import pyperf
runner = pyperf.Runner()
for list_size in ('10', '10**3', '10**6'):
    runner.timeit(
        f'short items: {list_size}',
        setup=f"data=[b'x']*{list_size}; sep=b'.'",
        stmt="sep.join(data)")

for list_size in ('10', '10**2', '10**4'):
    runner.timeit(
        f'long items: {list_size}',
        setup=f"data=[b'x' * 100]*{list_size}; sep=b'.'",
        stmt="sep.join(data)")

Results on Fedora 44 with CPU isolation:

Benchmark hidden because not significant (1): long items: 10**4

See also #158693 which optimizes PyUnicode_Join().

Merged, thanks for your review @eendebakpt.

While investigating this I found there are larger gains possible in case of exact bytes (the common case?)

Do you mean not avoiding Py_buffer or something else? If you have an idea, please propose a pull request with a benchmark and put me in copy. I will review it :-)