Fix atomic_ref<16 bytes>::is_always_lock_free

stl/inc/atomic

@@ -1656,7 +1656,7 @@ _INLINE_VAR constexpr bool _Is_always_lock_free = _TypeSize <= 2 * sizeof(void*)
 #else // ^^^ _ATOMIC_HAS_DCAS / !_ATOMIC_HAS_DCAS vvv
 template <size_t _TypeSize>
 _INLINE_VAR constexpr bool _Is_always_lock_free = _TypeSize <= sizeof(void*);
-#endif // _ATOMIC_HAS_DCAS
+#endif // ^^^ !_ATOMIC_HAS_DCAS ^^^
 #endif // ^^^ break ABI ^^^
 
 template <class _Ty, bool _Is_lock_free = _Is_always_lock_free<sizeof(_Ty)>>
@@ -2179,7 +2179,7 @@ public:
         return sizeof(_Ty) <= 2 * sizeof(void*);
 #else // ^^^ _ATOMIC_HAS_DCAS / !_ATOMIC_HAS_DCAS vvv
         return sizeof(_Ty) <= sizeof(void*) || (sizeof(_Ty) <= 2 * sizeof(void*) && __std_atomic_has_cmpxchg16b());
-#endif // _ATOMIC_HAS_DCAS
+#endif // ^^^ !_ATOMIC_HAS_DCAS ^^^
     }
 #endif // ^^^ break ABI ^^^
 
@@ -2343,11 +2343,16 @@ public:
 
     atomic_ref& operator=(const atomic_ref&) = delete;
 
-    static constexpr bool is_always_lock_free = _Is_always_lock_free<sizeof(_Ty)>;
-
     static constexpr bool _Is_potentially_lock_free =
         sizeof(_Ty) <= 2 * sizeof(void*) && (sizeof(_Ty) & (sizeof(_Ty) - 1)) == 0;
 
+    static constexpr bool is_always_lock_free =
+#if _ATOMIC_HAS_DCAS
+        _Is_potentially_lock_free;
+#else // ^^^ _ATOMIC_HAS_DCAS / !_ATOMIC_HAS_DCAS vvv
+        _Is_potentially_lock_free && sizeof(_Ty) <= sizeof(void*);
+#endif // ^^^ !_ATOMIC_HAS_DCAS ^^^
+
     static constexpr size_t required_alignment = _Is_potentially_lock_free ? sizeof(_Ty) : alignof(_Ty);
 
     _NODISCARD bool is_lock_free() const noexcept {
@@ -2359,7 +2364,7 @@ public:
         } else {
             return __std_atomic_has_cmpxchg16b() != 0;
         }
-#endif // _ATOMIC_HAS_DCAS
+#endif // ^^^ !_ATOMIC_HAS_DCAS ^^^
     }
 
     void store(const _Ty _Value) const noexcept {

tests/std/tests/P0019R8_atomic_ref/env.lst

@@ -2,3 +2,6 @@
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 RUNALL_INCLUDE ..\usual_20_matrix.lst
+RUNALL_CROSSLIST
+PM_CL=""
+PM_CL="/D_STD_ATOMIC_ALWAYS_USE_CMPXCHG16B=1"

tests/std/tests/P0019R8_atomic_ref/test.cpp

@@ -373,6 +373,26 @@ void test_incomplete_associated_class_all() { // COMPILE-ONLY
 }
 #endif // ^^^ no workaround ^^^
 
+void test_gh_4472() {
+    struct two_pointers_t {
+        void* left;
+        void* right;
+    };
+
+    alignas(std::atomic_ref<two_pointers_t>::required_alignment) two_pointers_t two_pointers;
+
+    static_assert(std::atomic_ref<two_pointers_t>::required_alignment == sizeof(two_pointers_t));
+
+#ifdef _WIN64
+    static_assert(std::atomic_ref<two_pointers_t>::is_always_lock_free == _STD_ATOMIC_ALWAYS_USE_CMPXCHG16B);
+#else
+    static_assert(std::atomic_ref<two_pointers_t>::is_always_lock_free);
+#endif
+
+    // we expect tests to run on DCAS machine. Win8+ require that
+    assert(std::atomic_ref<two_pointers_t>(two_pointers).is_lock_free());
+}
+
 int main() {
     test_ops<false, char>();
     test_ops<false, signed char>();
@@ -425,4 +445,5 @@ int main() {
     test_ptr_ops<long*>();
 
     test_gh_1497();
+    test_gh_4472();
 }

tests/std/tests/usual_20_matrix.lst

@@ -35,6 +35,6 @@ PM_CL="/clr /MD /std:c++20"
 PM_CL="/clr /MDd /std:c++20"
 PM_CL="/BE /c /EHsc /MD /std:c++20 /permissive-"
 PM_CL="/BE /c /EHsc /MTd /std:c++latest /permissive-"
-PM_COMPILER="clang-cl" PM_CL="-fno-ms-compatibility -fno-delayed-template-parsing -Wno-unqualified-std-cast-call /EHsc /std:c++20 /permissive- /MD --start-no-unused-arguments"
-PM_COMPILER="clang-cl" PM_CL="-fno-ms-compatibility -fno-delayed-template-parsing -Wno-unqualified-std-cast-call /EHsc /std:c++latest /permissive- /MTd /fp:strict --start-no-unused-arguments"
-PM_COMPILER="clang-cl" PM_CL="-fno-ms-compatibility -fno-delayed-template-parsing -Wno-unqualified-std-cast-call /EHsc /std:c++latest /permissive- /MT /fp:strict -fsanitize=undefined -fno-sanitize-recover=undefined --start-no-unused-arguments"
+PM_COMPILER="clang-cl" PM_CL="-fno-ms-compatibility -fno-delayed-template-parsing -Wno-unqualified-std-cast-call /EHsc /std:c++20 /permissive- /MD --start-no-unused-arguments -mcx16"
+PM_COMPILER="clang-cl" PM_CL="-fno-ms-compatibility -fno-delayed-template-parsing -Wno-unqualified-std-cast-call /EHsc /std:c++latest /permissive- /MTd /fp:strict --start-no-unused-arguments -mcx16"
+PM_COMPILER="clang-cl" PM_CL="-fno-ms-compatibility -fno-delayed-template-parsing -Wno-unqualified-std-cast-call /EHsc /std:c++latest /permissive- /MT /fp:strict -fsanitize=undefined -fno-sanitize-recover=undefined --start-no-unused-arguments -mcx16"

tests/utils/stl/test/features.py

@@ -58,6 +58,7 @@ def getDefaultFeatures(config, litConfig):
         DEFAULT_FEATURES.append(Feature(name='edg'))
         DEFAULT_FEATURES.append(Feature(name='arch_avx2'))
         DEFAULT_FEATURES.append(Feature(name='x64'))
+        DEFAULT_FEATURES.append(Feature(name='cx16'))
 
     elif litConfig.target_arch.casefold() == 'arm'.casefold():
         DEFAULT_FEATURES.append(Feature(name='arch_vfpv4'))

tests/utils/stl/test/tests.py

@@ -314,6 +314,8 @@ def _parseFlags(self, litConfig):
                 foundCRT = True
             elif flag[1:] == 'analyze:plugin':
                 afterAnalyzePlugin = True
+            elif flag[1:] == 'D_STD_ATOMIC_ALWAYS_USE_CMPXCHG16B=1':
+                self.requires.append('cx16')
 
         if not foundStd:
             self._addCustomFeature('c++14')