Semi-Automatic Composition of Loop Transformations for Deep Parallelism and Memory Hierarchies