e56aabf205
. Register blocking sizes are better handled by the cache size heuristics. The current code introduced very small blocks, for instance for 9x9 matrix, thus killing performance.