diff --git a/erfa/ufunc.c.templ b/erfa/ufunc.c.templ
index e0e2729..e8e18b8 100644
--- a/erfa/ufunc.c.templ
+++ b/erfa/ufunc.c.templ
@@ -147,21 +147,6 @@ static inline void copy_to_eraLDBODY(char *ptr, npy_intp s, npy_intp n,
  *    not contiguous.
  */
 
-{%- macro inner_loop_steps_and_copy(arg, arg_name) %}
-  {%- for i in range(arg.ndim or 1) %}
-    npy_intp is_{{ arg_name }}{{ i }} = *steps++;
-  {%- endfor %}
-  {#- /* copy should be made if buffer not contiguous;
-         note: one can only have 1 or 2 dimensions */ #}
-  {%- if arg.ndim == 2 %}
-    int copy_{{ arg_name
-        }} = (is_{{ arg_name }}1 != sizeof({{ arg.ctype }}) ||
-              is_{{ arg_name }}0 != {{ arg.shape[1] }} * sizeof({{ arg.ctype }}));
-  {%- else %}
-    int copy_{{ arg_name }} = (is_{{ arg_name }}0 != sizeof({{ arg.ctype }}));
-  {%- endif %}
-{%- endmacro %}
-
 {%- for func in funcs %}
 
 static void ufunc_loop_{{ func.pyname }}(
@@ -213,20 +198,8 @@ static void ufunc_loop_{{ func.pyname }}(
     {%- if arg.ctype == 'eraLDBODY' %}
     int nb = (int)dimensions[0];  /* Refuse to worry about INT_MAX */
     {%- endif %}
-    {%- if arg.signature_shape != '()' -%}
-    {{ inner_loop_steps_and_copy(arg, arg.name) }}
-    {%- endif %}
-   {%- endfor %}
-   {%- for arg in func.args_by_inout('inout') %}
-    {%- if arg.signature_shape != '()' -%}
-    {{ inner_loop_steps_and_copy(arg, arg.name + '_in') }}
-    {%- endif %}
-   {%- endfor %}
-   {%- for arg in func.args_by_inout('inout|out') %}
-    {%- if arg.signature_shape != '()' -%}
-    {{ inner_loop_steps_and_copy(arg, arg.name) }}
-    {%- endif %}
    {%- endfor %}
+    {{ func.inner_loop_steps_and_copies() | indent(4) }}
    {#- /* if needed, allocate memory for contiguous eraLDBODY copies */ #}
    {%- if func.user_dtype == 'dt_eraLDBODY' %}
     if (copy_b) {
diff --git a/erfa_generator.py b/erfa_generator.py
index 140cc93..d19ff23 100644
--- a/erfa_generator.py
+++ b/erfa_generator.py
@@ -196,6 +196,23 @@ def signature_shape(self):
                 return "(3, 3)"
         return super().signature_shape
 
+    def inner_loop_steps_and_copy(self, name_suffix: str = "") -> str | None:
+        if self.signature_shape == "()":
+            return None
+        name = self.name + name_suffix
+        lines = [f"npy_intp is_{name}{i} = *steps++;" for i in range(self.ndim or 1)]
+        # copy should be made if buffer not contiguous;
+        # note: one can only have 1 or 2 dimensions
+        lines.append(
+            f"int copy_{name} = (is_{name}0 != sizeof({self.ctype}));"
+            if self.ndim <= 1
+            else (
+                f"int copy_{name} = (is_{name}1 != sizeof({self.ctype}) ||\n"
+                f"          is_{name}0 != {self.shape[1]} * sizeof({self.ctype}));"
+            )
+        )
+        return "\n".join(lines)
+
 
 class StatusCode(Variable):
     def __init__(self, ctype: str, doc: FunctionDoc, funcname: str) -> None:
@@ -369,6 +386,14 @@ def generate_python_body(self) -> str:
         lines.append(f"return {ret_val}")
         return "\n".join(lines)
 
+    def inner_loop_steps_and_copies(self) -> str:
+        in_only = [a.inner_loop_steps_and_copy() for a in self.args_by_inout("in")]
+        inout = [
+            a.inner_loop_steps_and_copy("_in") for a in self.args_by_inout("inout")
+        ]
+        out = [a.inner_loop_steps_and_copy() for a in self.args_by_inout("inout|out")]
+        return "\n".join(filter(None, in_only + inout + out))
+
 
 class Constant: