Sylvain Laperche

Yes `forkserver` is notoriously slower (that's probably why Python is moving to `spawn` instead of `forkserver` for macOS). Using the `dodo.py` file given in the first message and applying your...

Ha yes, indeed: moving `action` outside (and thus making it a normal function instead of a closure) solve the issue. I'll do the test on the macOS of my coworker...

I've just tried your patch on macOS and it works as expected.

Looks good to me :slightly_smiling_face: Maybe add a check if someone set `fork` on mac (using the command line argument) and raise an error or print a warning in that...

I see, it makes sense. It could be an issue if you have **a lot** of paths, then the cost of recreating all the Path objects may be noticeable (that...

No problem :slightly_smiling_face: , I'll submit a PR for that :) (probably next week).

Maybe we can do something like what have been done with Go boilerplate: have a lint step ensuring that the codegen was run and the resulting files commited.

For threading we can use what the stdlib provides, see http://www.cplusplus.com/reference/multithreading/

I think that's an extra external dependency, so I would rather stay with the stdlib at first (except if there is a killer feature of OpenMP).

I think it's too early to merge things now. I was developing my stuff inside a "simd" subdirectory to avoid conflict with the current SIMD implementation. Since it's still a...