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

[Martin Storsjö via libcxx-commits]

https://github.com/mstorsjo created https://github.com/llvm/llvm-project/pull/135334

This test currently hardcodes which environments have got 80 bit long doubles on x86_64; add a check for the actual size of the long doubles as well.

This allows waiving this part of the test, if we have x86_64 setups in any of these environments, configured for a nonstandard size of long doubles.

Also clarify the exact reasons for why specific OSes such as FreeBSD are skipped for these tests.

>From 298c2c6e4ce60373455e9afd7b733f48625969b2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Martin=20Storsj=C3=B6?= <martin at martin.st>
Date: Wed, 12 Mar 2025 22:44:09 +0200
Subject: [PATCH] [libcxx] [test] Clarify the condition for long double hex
 formatting

This test currently hardcodes which environments have got 80 bit
long doubles on x86_64; add a check for the actual size of
the long doubles as well.

This allows waiving this part of the test, if we have x86_64
setups in any of these environments, configured for a nonstandard
size of long doubles.

Also clarify the exact reasons for why specific OSes such as
FreeBSD are skipped for these tests.
---
 .../put_long_double.hex.pass.cpp              | 25 +++++++++++++++----
 1 file changed, 20 insertions(+), 5 deletions(-)

diff --git a/libcxx/test/std/localization/locale.categories/category.numeric/locale.nm.put/facet.num.put.members/put_long_double.hex.pass.cpp b/libcxx/test/std/localization/locale.categories/category.numeric/locale.nm.put/facet.num.put.members/put_long_double.hex.pass.cpp
index b9b02f5ea7b67..51863a30445f6 100644
--- a/libcxx/test/std/localization/locale.categories/category.numeric/locale.nm.put/facet.num.put.members/put_long_double.hex.pass.cpp
+++ b/libcxx/test/std/localization/locale.categories/category.numeric/locale.nm.put/facet.num.put.members/put_long_double.hex.pass.cpp
@@ -1846,11 +1846,26 @@ 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(__x86_64__) &&                  \
+    __LDBL_MANT_DIG__ == 64
+  // 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];
   {