Compile-time GPU memory access optimizations