As mentioned in the comments to the issue

It may just be for warning suppression (e.g., https://stackoverflow.com/questions/24477228/stdc-version-not-defined-in-c11).

and indeed, forcing c90 with gcc on ubuntu 22.04 shows it is not defined:

$ echo | gcc -std=c11 -dM -E - | grep __STDC_VERSION__
#define __STDC_VERSION__ 201112L
$ echo | gcc -std=c99 -dM -E - | grep __STDC_VERSION__
#define __STDC_VERSION__ 199901L
$ echo | gcc -std=c90 -dM -E - | grep __STDC_VERSION__
$

We use it without checking existence in numpy/_core/src/common/npy_atomic.h, all the other uses (and in Cython-generated code) have #if defined() guard. Maybe we could defensively add a guard to npy_atomic.h as well, even though I think we require c99 or higher.

I think it is expected (and correct) that __STDC_VERSION__ is not (automatically) defined for -std=c90. If there is a warning or error issued with -std=c90 then either newer C standard is required to compile or there is a bug in a header. Either in system one, or in numpy.

Should I update this PR to add the defined() guard to the npy_atomic.h file? Or is somewhere specified that C99 (or newer) is needed?

Should I update this PR to add the defined() guard to the npy_atomic.h file?

That would be great, yes. All the other occurrences have one.

Should I update this PR to add the defined() guard to the npy_atomic.h file?

That would be great, yes. All the other occurrences have one.

Done.

Thanks @mtelka