[libcxx] [test] Generalize the conditions for testing bitcasts between long double, double and int128

[mstorsjo]

Summary

MSVC targets also have a 64 bit long double, as do MinGW targets on ARM.
This hasn't been noticed in CI because the MSVC configurations there run
with _LIBCPP_HAS_NO_INT128 defined.

This avoids assuming that either __int128_t or double is equal in size to
long double. i386 MinGW targets have sizeof(long double) == 10, which
doesn't match any of the tested types.

Unfortunately, this relies on using nonstandard SIZEOF_* defines.

[Mordante]

It seems we already relay on __SIZEOF_* in our code, including in <bitset> so I don't see that as a big issue.
LGTM modulo one minor remark.

libcxx/test/std/numerics/bit/bit.cast/bit_cast.pass.cpp

233

In __config there's this definition

#ifndef __SIZEOF_INT128__
#define _LIBCPP_HAS_NO_INT128
#endif

So this test can be removed.

[• Quuxplusone]

libcxx/test/std/numerics/bit/bit.cast/bit_cast.pass.cpp

233

LGTM mod this comment as well. Also, I don't think you should be "silently hiding" if __SIZEOF_DOUBLE__ is not defined. In the unlikely event that it's not defined, I think we want to be told about it via -Wundefined-macro-is-zero or whatever that warning's name is.

So when this is addressed, you'll just have

#if __SIZEOF_LONG_DOUBLE__ == __SIZEOF_DOUBLE__
        test_roundtrip_through<double, false>(i);
#endif
#if defined(__SIZEOF_INT128__) && __SIZEOF_LONG_DOUBLE__ == __SIZEOF_INT128__
        test_roundtrip_through<__int128_t, false>(i);
        test_roundtrip_through<__uint128_t, false>(i);
#endif

[mstorsjo]

Simplify the ifdefs, as suggested

libcxx/test/std/numerics/bit/bit.cast/bit_cast.pass.cpp

for (long double i : {0.0l, 1.0l, -1.0l, 10.0l, -10.0l, 1e10l, 1e-10l, 1e100l, 1e-100l,
                           3.141592653589793238462643383279502884197169399375105820974944l,
                           std::nanl(""),
                           std::numeric_limits<long double>::signaling_NaN(),
                           std::numeric_limits<long double>::quiet_NaN()}) {
         // Note that x86's `long double` has 80 value bits and 48 padding bits.
         test_roundtrip_through_nested_T<false>(i);
         test_roundtrip_through_buffer<false>(i);
 
-        // On arm64 on Apple platforms, long double is just double, so we don't
-        // test against int128, but instead against double itself. Otherwise,
-        // we test against int128 if we have those types available.
-#if defined(__aarch64__) && defined(__APPLE__)
-#   define LONG_DOUBLE_IS_DOUBLE
-#endif
-
-#if defined(LONG_DOUBLE_IS_DOUBLE)
+#if defined(__SIZEOF_LONG_DOUBLE__) && defined(__SIZEOF_DOUBLE__) && __SIZEOF_LONG_DOUBLE__ == __SIZEOF_DOUBLE__
         test_roundtrip_through<double, false>(i);
-#elif !defined(_LIBCPP_HAS_NO_INT128)
+#endif
+#if !defined(_LIBCPP_HAS_NO_INT128)
+#if defined(__SIZEOF_LONG_DOUBLE__) && defined(__SIZEOF_INT128__) && __SIZEOF_LONG_DOUBLE__ == __SIZEOF_INT128__
         test_roundtrip_through<__int128_t, false>(i);
         test_roundtrip_through<__uint128_t, false>(i);
 #endif
+#endif
     }
 
     return true;
 }
 
 // TODO: There doesn't seem to be a way to perform non-trivial correctness
 //       tests inside constexpr.
 constexpr bool basic_constexpr_test() {