[libcxx-commits] [libcxx] [libc++][In progress] Floating Point Atomic (PR #67799)

[Louis Dionne via libcxx-commits]

================
@@ -136,6 +140,114 @@ struct atomic<_Tp*>
     atomic& operator=(const atomic&) volatile = delete;
 };
 
+#if _LIBCPP_STD_VER >= 20
+template <class _Tp> requires is_floating_point_v<_Tp>
+struct atomic<_Tp>
+    : public __atomic_base<_Tp>
+{
+  private:
+    // The builtin __cxx_atomic_fetch_add does not work for
+    // long double on some platforms with fp80 type
+    // There is no way on the libc++ side to test whether it is
+    // ok to use the builtin for a certain type.
+    // Therefore, we do not use the builtin here
+    // For more details, see 
+    // lib/Sema/SemaChecking.cpp function IsAllowedValueType
+    // LLVM Parser does not allow atomicrmw with x86_fp80 type.
+    // if (ValType->isSpecificBuiltinType(BuiltinType::LongDouble) &&
+    //    &Context.getTargetInfo().getLongDoubleFormat() ==
+    //        &llvm::APFloat::x87DoubleExtended())
+    // For more info
+    // https://reviews.llvm.org/D53965
----------------
ldionne wrote:

We should file an issue against LLVM to request that the builtin works for `long double` regardless of the underlying representation, and link it here.

https://github.com/llvm/llvm-project/pull/67799