Building sys.ji takes very long

[ViralBShah]

It would be great if all the individual .j files are compiled separately, and then somehow combined into sys.ji. Even a one line modification triggers sys.ji to be built from scratch. Would be nice if it was faster.

[StefanKarpinski]

Realistically, only Jeff can do this. I'm also not sure if it's even possible. My understanding of the way this works is that incremental compilation is a non-starter, but Jeff is going to have to comment on that.

[JeffBezanson]

The problem is inlining. Or when f calls g, even if g is not inlined, f's inferred type depends on information about g. It's very true that allowing separate compilation is a desirable feature of a type system, but we don't have it.

Actually, there is a way to rebuild sys.ji faster. The idea is to "edit" the image by starting julia with your current system image, then loading your changed files, then saving the system image again. This will work for changes that not too many things depend on. Eventually I can implement something that traverses dependencies and makes sure everything that needs to gets re-compiled. This is the way to do it.

[StefanKarpinski]

That's a cool idea and would be really awesome. We need some sort of dependency tracking in order for methods to get recompiled when other methods that they depend on change anyway. Would implementing that have the desired effect? Currently, when you redefine a method, you then have to also redefine any methods that used that function.

[JeffBezanson]

Yes, I think that is the real issue here. I need to do it at some point anyway. Once that is done, the way to build sys.ji will be to start julia with the current sys.ji, then call load on files newer than sys.ji, then save the image again. It will only take as long as necessary, depending how deep your changes go. That would be perfect.

[ViralBShah]

Yes, this solution is perfect. Perhaps github needs to provide a +1 button to give a thumbs up...

[ViralBShah]

With the base library stabilising, this is not as important. Closing as wontfix.