Fix limit behavior of atan2 under -ffast-math

Fix #1231

Author: serge-sans-paille

include/xsimd/arch/common/xsimd_common_trigo.hpp

@@ -354,8 +354,14 @@ namespace xsimd
         XSIMD_INLINE batch<T, A> atan2(batch<T, A> const& self, batch<T, A> const& other, requires_arch<common>) noexcept
         {
             using batch_type = batch<T, A>;
+#ifdef __FAST_MATH__
+            const batch_type q = abs(self / other);
+            const batch_type q_p = abs(other / self);
+#else
             const batch_type q = abs(self / other);
-            const batch_type z = detail::kernel_atan(q, batch_type(1.) / q);
+            const batch_type q_p = 1. / q;
+#endif
+            const batch_type z = detail::kernel_atan(q, q_p);
             return select(other > batch_type(0.), z, constants::pi<batch_type>() - z) * signnz(self);
         }
 

test/test_xsimd_api.cpp

@@ -502,9 +502,17 @@ struct xsimd_api_float_types_functions
     }
     void test_atan2()
     {
-        value_type val0(0);
-        value_type val1(1);
-        CHECK_EQ(extract(xsimd::atan2(T(val0), T(val1))), std::atan2(val0, val1));
+        {
+            value_type val0(0);
+            value_type val1(1);
+            CHECK_EQ(extract(xsimd::atan2(T(val0), T(val1))), std::atan2(val0, val1));
+        }
+
+        {
+            value_type val0(1);
+            value_type val1(0);
+            CHECK_EQ(extract(xsimd::atan2(T(val0), T(val1))), std::atan2(val0, val1));
+        }
     }
     void test_atanh()
     {