assigned to @ldionne

Here's a reproducer for one of these failing tests. I suspect they are all the same failure under the hood:

$ clang++ repro.cpp -g -std=c++2a -nostdinc++ -I libcxx/include -isysroot $(xcrun --show-sdk-path) -I libcxx/test/support -L build/lib -nostdlib++ -lc++ -Wl,-rpath,$PWD/build/lib -o a.out && ./a.out

$ cat repro.cpp
#include
#include
#include
#include
#include "test_macros.h"
#include "test_iterators.h"

#include "platform_support.h" // locale name macros

typedef std::money_get<char, input_iterator<const char*> > Fn;

struct my_facet : Fn {
    explicit my_facet(std::size_t refs) : Fn(refs) {}
};

int main(int, char**) {
    std::ios ios(0);
    std::string loc_name(LOCALE_ru_RU_UTF_8);
    ios.imbue(std::locale(ios.getloc(),
                          new std::moneypunct_byname<char, false>(loc_name)));
    ios.imbue(std::locale(ios.getloc(),
                          new std::moneypunct_byname<char, true>(loc_name)));
    ios.imbue(std::locale(ios.getloc(),
                          new std::moneypunct_byname<wchar_t, false>(loc_name)));
    ios.imbue(std::locale(ios.getloc(),
                          new std::moneypunct_byname<wchar_t, true>(loc_name)));

    const my_facet f(1);
    {   // zero, showbase
        std::string v = "0,00 RUB ";
        showbase(ios);
        typedef input_iterator<const char*> I;
        long double ex;
        std::ios_base::iostate err = std::ios_base::goodbit;
        I beg(v.data());
        I end(v.data() + v.size());
        I iter = f.get(beg, end, true, ios, err, ex);
        assert(iter.base() == v.data() + v.size());
        assert(err == std::ios_base::eofbit);
        assert(ex == 0);
        noshowbase(ios);
    }

    return 0;
}

I also ran into this while testing on FreeBSD. It appears that the problem is that the space after "RUB " is not consumed on Apple and FreeBSD.

I can upload a patch for this shortly. Would you prefer

a)

#if defined(FreeBSD) || defined(APPLE)
// FreeBSD's and Apple's libc do not consume the space.
assert(iter.base() == v.data() + v.size() - 1);
assert(err == std::ios_base::goodbit);
#else
assert(iter.base() == v.data() + v.size());
assert(err == std::ios_base::eofbit);
#endif

Or b)

Changing e.g. "-1 234 567,89 RUB " to "-1 234 567,89 RUB". I'm not sure if the space is important, but removing it fixes the test for me.

The fr_FR tests seem to use "1 234 567,89 EUR" without the trailing space, so I guess that means omitting it is fine?

I think I would do b), so that the tests are as streamlined as possible.

TBH, I don't understand why there's subtle differences like:

#if !defined(APPLE_FIXME)
{   // zero, showbase
    std::string v = "0,00 RUB ";
    showbase(ios);
    typedef input_iterator<const char*> I;
    long double ex;
    std::ios_base::iostate err = std::ios_base::goodbit;
    I iter = f.get(I(v.data()), I(v.data() + v.size()),
                                        true, ios, err, ex);
    assert(iter.base() == v.data() + v.size());
    assert(err == std::ios_base::eofbit);
    assert(ex == 0);
    noshowbase(ios);
}
#endif
{   // negative one, showbase
    std::string v = "-0,01 RUB ";
    typedef input_iterator<const char*> I;
    long double ex;
    std::ios_base::iostate err = std::ios_base::goodbit;
    I iter = f.get(I(v.data()), I(v.data() + v.size()),
                                        true, ios, err, ex);
    assert(iter.base() == v.data() + 6);
    assert(err == std::ios_base::goodbit);
    assert(ex == -1);
}

Why do we check for eofbit when parsing 0,00 RUB, but goodbit when parsing -0,01 RUB?

I would try to remove the trailing space and see whether that still works. If you upload a patch on Phabricator, that'll also run the tests on Linux and macOS automatically, so you can experiment that way.

Thanks for looking into it.. these locale APIs are terrible. :-)

I ran into this issue, while poking those test files, and I went ahead and tried to fix it - see https://reviews.llvm.org/D120316.

I also agree that we should remove the trailing space there; that matches other similar tests. This test seemed to be unusually untested in the sense that it was XFAILed in essentially all configurations in our current CI except for Apple platforms, and on Apple platforms, these bits were ifdeffed out.

On top of that, I made https://reviews.llvm.org/D120317 which makes the test a bit more flexible, so that it passes on Glibc, FreeBSD and Windows too.

This issue should be fixed now by 5333732.