Comment author: vnicollet

My suggestion would be to replace the internal raise_undefined in:

Obj.set_field (Obj.repr blk) 0 raise_undefined;

with a function that remembers the thread it was created in and only raises Undefined within its owner thread : otherwise, it blocks until the value is computed by the original thread and then recursively performs the equivalent of Lazy.force on the value again.

This creates light overhead (since the closure that contains the thread identifier has to be created when evaluating the lazy block) but avoids having to wrap the entire force_lazy_block in a critical section (only the initial set_field has to be wrapped).

Comment author: @damiendoligez

I think the documentation should be updated to indicate that Lazy.t is not threadsafe.

I'd rather change the documentation to indicate that nothing is thread-safe except when explicitly noted.

Comment author: @damiendoligez

I have documented the fact that Lazy.force is not thread-safe. Keeping this PR as a feature wish in case we find a good idea to make it thread-safe without too much overhead.

Comment author: @alainfrisch

Marking as suspended. No planned work on this front.