I see one potential argument to resolve this differently: The locale setup in the test harness seems to have been adapted almost verbatim from libc++'s, so libc++'s test harness suffers from the same bug. If we resolve this issue in the same way as libc++, it might make it a little bit easier to sync changes from upstream in the future.

But I guess it's also a valid argument that the test setups have accumulated so many differences that another minor one doesn't matter.

I'm open to a "better" fix, I just don't know what it would look like. Our feature-setup logic is very different.

If you think they have diverged too much, then there isn't much point in trying to keep the code as closely aligned as possible. Your code certainly looks cleaner.

The MSVC STL test harness also appears to be missing some locale-related compile flags that were recently added to libcxx tests, like this one in std\localization\locale.categories\category.monetary\locale.money.get\locale.money.get.members\get_long_double_fr_FR.pass.cpp:

// ADDITIONAL_COMPILE_FLAGS: -DFR_MON_THOU_SEP=%{LOCALE_CONV_FR_FR_UTF_8_MON_THOUSANDS_SEP}

The necessary logic for these flags was added to the original locale processing code in libc++'s features.pyhere.

We should add some equivalent logic to features.py as well (but probably not in this PR).

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.