moved some preprocessing code inside a branch so it only executes if

theano/gradient.py

@@ -672,30 +672,30 @@ def _populate_grad_dict(var_to_node_to_idx,
 
             inputs = node.inputs
 
-            # Each Op's grad function requires inputs and output_grads
-            # If the Op destroys any input, but the grad expression uses it,
-            # then chances are the resulting graph will have a dependency
-            # cycle. We avoid this cycle by passing (symbolic) copies of
-            # each destroyed input.
-            try:
-                dinputs = [node.inputs[x[0]] for x in
-                        node.op.destroy_map.values()]
-            except AttributeError:
-                dinputs = []
-
-            def try_to_copy_if_needed(var):
-                if var in dinputs and hasattr(var, 'copy'):
-                    return var.copy()
-                return var
-
-            inputs = [try_to_copy_if_needed(ipt) for ipt in inputs]
-
             output_grads = [access_grad_cache(var) for var in node.outputs]
 
             if False in [isinstance(g.type, DisconnectedType)
                     for g in output_grads]:
                 # Some outputs of this op are connected to the cost so we must
-                # call the ops grad method
+                # call the op's grad method
+
+                # Each Op's grad function requires inputs and output_grads
+                # If the Op destroys any input, but the grad expression uses it,
+                # then chances are the resulting graph will have a dependency
+                # cycle. We avoid this cycle by passing (symbolic) copies of
+                # each destroyed input.
+                try:
+                    dinputs = [node.inputs[x[0]] for x in
+                            node.op.destroy_map.values()]
+                except AttributeError:
+                    dinputs = []
+
+                def try_to_copy_if_needed(var):
+                    if var in dinputs and hasattr(var, 'copy'):
+                        return var.copy()
+                    return var
+
+                inputs = [try_to_copy_if_needed(ipt) for ipt in inputs]
 
                 input_grads = node.op.grad(inputs, output_grads)