feat: moved cast<T1, T2> to architecture-specific definitions

Author: madhav-madhusoodanan

crates/intrinsic-test/src/arm/config.rs

@@ -8,6 +8,14 @@ pub const POLY128_OSTREAM_DECL: &str = r#"
 std::ostream& operator<<(std::ostream& os, poly128_t value);
 std::ostream& operator<<(std::ostream& os, float16_t value);
 #endif
+
+// T1 is the `To` type, T2 is the `From` type
+template<typename T1, typename T2> T1 cast(T2 x) {{
+  static_assert(sizeof(T1) == sizeof(T2), "sizeof T1 and T2 must be the same");
+  T1 ret{{}};
+  memcpy(&ret, &x, sizeof(T1));
+  return ret;
+}}
 "#;
 
 pub const POLY128_OSTREAM_DEF: &str = r#"

crates/intrinsic-test/src/common/gen_c.rs

@@ -122,24 +122,6 @@ pub fn write_mod_cpp<T: IntrinsicTypeDefinition>(
 
     writeln!(w, "{}", forward_declarations)?;
 
-    writeln!(
-        w,
-        r#"
-// T1 is the `To` type, T2 is the `From` type
-template<typename T1, typename T2> T1 cast(T2 x) {{
-  if (std::is_convertible<T2, T1>::value) {{
-      return x;
-  }} else if (sizeof(T1) == sizeof(T2)) {{
-    T1 ret{{}};
-    memcpy(&ret, &x, sizeof(T1));
-    return ret;
-  }} else {{
-    assert("T2 must either be convertable to T1, or have the same size as T1!");
-  }}
-}}
-"#
-    )?;
-
     for intrinsic in intrinsics {
         create_c_test_function(w, intrinsic)?;
     }

crates/intrinsic-test/src/x86/config.rs

@@ -269,6 +269,19 @@ std::ostream& operator<<(std::ostream& os, __m512i value) {
     return os;
 }
 
+// T1 is the `To` type, T2 is the `From` type
+template<typename T1, typename T2> T1 cast(T2 x) {{
+  if (std::is_convertible<T2, T1>::value) {{
+      return x;
+  }} else if (sizeof(T1) == sizeof(T2)) {{
+    T1 ret{{}};
+    memcpy(&ret, &x, sizeof(T1));
+    return ret;
+  }} else {{
+    assert("T2 must either be convertable to T1, or have the same size as T1!");
+  }}
+}}
+
 #define _mm512_extract_intrinsic_test_epi8(m, lane) \
     _mm_extract_epi8(_mm512_extracti64x2_epi64((m), (lane) / 16), (lane) % 16)