Once an optimization attempt is "interrupted", it enters a dummy data-copying phase, in order to maintain static fori_loop bounds.
This copying phase seems somewhat under-performant. We should just compute an interrupted flag and do the copying at once afterwards, rather than in the loop.