Nice simplification. I agree that relaxed loads should be fine - in the worst case there will be some blocking or duplication of work in code paths for cache misses.

Just to be extra sure, I wrote a little test script which doesn't produce any race reports on a TSan-instrumented build of 3.14:

import numpy as np
from numpy.testing._private.utils import run_threaded


dt = np.dtype(('i4', [('a', '<i2'), ('b', '<i2')]))


def worker(b):
    b.wait()
    hash(dt)


run_threaded(worker, 500, pass_barrier=True)

I'd like someone else who knows numpy's build infrastructure on Windows to weigh in about the new experimental build flag.

Nice, this is worth trying indeed.

The /experimental:c11atomics was added in VS 2022 version 17.5 Preview 2, which I believe corresponds to toolchain version 19.35. Right now the minimum we support is 19.20, so before merging we should bump that version number on line 28 of the top-level meson.build file to 19.35. That way we'll get a clear error message if the installed compiler is too old.

Once that is changed, this LGTM I think.