Vectorizer did not seem to properly handle some loops that are smaller than the vector width.

pystencils/cpu/vectorization.py

@@ -172,6 +172,8 @@ def vectorize_inner_loops_and_adapt_load_stores(ast_node, assume_aligned, nontem
             assert len(loop_nodes) in (0, 1, 2)  # 2 for main and tail loop, 1 if loop range divisible by vector width
             if len(loop_nodes) == 0:
                 continue
+            if len(loop_nodes) == 1 and loop_nodes[0].stop - loop_nodes[0].start < vector_width:
+                continue
             for loop in loop_nodes:
                 loop.parent = loop_node.parent
             loop_node = loop_nodes[0]