[libcxx] [test] Clarify the condition for long double hex formatting (#135334)

This test currently hardcodes which environments have got 80 bit long
doubles on x86_64 with a suitable printf formatting; convert the
architecture check into a check specifically for 80 bit long doubles.

Not all x86_64 configurations do have 80 bit long doubles (e.g. 
Android doesn't), and i386 configurations can also have 80 bit long
doubles, compatible with this test.

Also clarify the exact reasons for why specific OSes such as FreeBSD are
skipped for these tests, even though they have 80 bit long doubles.

Author: mstorsjo

libcxx/test/std/localization/locale.categories/category.numeric/locale.nm.put/facet.num.put.members/put_long_double.hex.pass.cpp

@@ -1846,11 +1846,25 @@ void test1() {
 void test2() {
   std::locale lc = std::locale::classic();
   std::locale lg(lc, new my_numpunct);
-#if (defined(__APPLE__) || defined(TEST_HAS_GLIBC) || defined(__MINGW32__)) && defined(__x86_64__)
-  // This test is failing on FreeBSD, possibly due to different representations
-  // of the floating point numbers.
-  // This test is failing in MSVC environments, where long double is equal to regular
-  // double, and instead of "0x9.32c05a44p+27", this prints "0x1.26580b4880000p+30".
+#if (defined(__APPLE__) || defined(TEST_HAS_GLIBC) || defined(__MINGW32__)) && defined(TEST_LONG_DOUBLE_IS_80_BIT)
+  // This test assumes that long doubles are x87 80 bit long doubles, and
+  // assumes one specific way of formatting the long doubles. (There are
+  // multiple valid ways of hex formatting the same float.)
+  //
+  // FreeBSD does use x87 80 bit long doubles, but normalizes the hex floats
+  // differently.
+  //
+  // This test assumes the form used by Glibc, Darwin and others, where the
+  // 64 mantissa bits are grouped by nibble as they are stored in the long
+  // double representation (nibble aligned at the end of the least significant
+  // bits). This makes 1.0L to be formatted as "0x8p-3" (where the leading
+  // bit of the mantissa is the higest bit in the 0x8 nibble), and makes
+  // __LDBL_MAX__ be formatted as "0xf.fffffffffffffffp+16380".
+  //
+  // FreeBSD normalizes/aligns the leading bit of the mantissa as a separate
+  // nibble, so that 1.0L is formatted as "0x1p+0" and __LDBL_MAX__ as
+  // "0x1.fffffffffffffffep+16383" (note the lowest bit of the last nibble
+  // being zero as the nibbles don't align to the actual floats).
   const my_facet f(1);
   char str[200];
   {

libcxx/test/support/test_macros.h

@@ -544,4 +544,8 @@ inline Tp const& DoNotOptimize(Tp const& value) {
 #  define TEST_LONG_DOUBLE_IS_DOUBLE
 #endif
 
+#if defined(__LDBL_MANT_DIG__) && __LDBL_MANT_DIG__ == 64
+#  define TEST_LONG_DOUBLE_IS_80_BIT
+#endif
+
 #endif // SUPPORT_TEST_MACROS_HPP