apply vector substitutions in all ast nodes

pystencils/cpu/vectorization.py

@@ -429,8 +429,12 @@ def insert_vector_casts(ast_node, instruction_set, loop_counter_symbol, default_
                 arg.condition_expr = visit_expr(arg.condition_expr, default_type)
                 visit_node(arg, substitution_dict, default_type)
             elif isinstance(arg, ast.ForLoop): # ForLoop is intended not to be vectorized but to set up small finite loops within kernels
-                    visit_node(arg.body, substitution_dict, default_type)
+                visit_node(arg.body, substitution_dict, default_type)
             else:
                 visit_node(arg, substitution_dict, default_type)
+        node = fast_subs(
+            node, substitution_dict,
+            skip=lambda e: isinstance(e, (ast.SympyAssignment, ast.Conditional, ast.ForLoop, ast.ResolvedFieldAccess))
+        )
 
     visit_node(ast_node, {}, default_float_type)