nimforum mirror

Araq (orginal) [2023-11-02T17:59:16+01:00] view original

Upcoming major versions of Nim will be based on NIR, an intermediate format that is lower level than Nim's AST:

NIR has its own type system that maps cleanly to a C++-like language but remains easy to traverse and interpret. The type system is basically C's without C's array decay design bug. However, inheritance is supported directly as it helps type based alias analysers. Likewise there is a clear separation between aliased and unique pointers and between pointers to a single element and pointers to arrays.

Both simple to interpret and simple to translate to C(++).

Allows for all the optimizations and analysis passes that we need.

Implemented so far:

The type graph and the mapping from Nim types to NIR types.

The NIR packed AST and the mapping from the Nim AST to NIR's AST.

Load/store NIR to disk.

Work in progress:

A new VM that offers a REPL which has superb interop with native code.

A new C(++) code generator.

Now why am I writing this? Because it's a survey. Where should the immediate focus be:

[ ] REPL.

[ ] Not-nil checking and static array index bounds checking.

[ ] New C(++) backend.

[ ] NIR-to-AST mapping ("back to AST for backwards compat") so that we can use this technology to finally give us IC.

[ ] Better borrow checker to refine the "view types" implementation.

jmgomez (orginal) [2023-11-02T18:17:38+01:00] view original

All are awesome features.

If I had to pick one it will be REPL but closely followed by New C(++) backend as I feel the latter will incentivize the creation of new backends, especially if the compiler doesnt need to be forked.

juancarlospaco (orginal) [2023-11-02T18:27:05+01:00] view original

I would try to use the big refactor opportunity to clean out and remove/deprecate/improve experimental and legacy stuff like:

That alone kills a ton of bugs, and incompatible optional features (like the 2 not-nil checkings), and reduces the footprint of the work needed to make all that work again on the new C backends, reduce the footprint before you have to re-build on top, no one is really using all that "experimentals" in production.

Maybe just run nim --fullhelp and check what can be improved/removed, ordering by easiest to hardest IMHO.

matkuki (orginal) [2023-11-02T18:27:15+01:00] view original

[x] NIR-to-AST mapping ("back to AST for backwards compat") so that we can use this technology to finally give us IC.

Would be my vote.

tommo (orginal) [2023-11-02T18:29:09+01:00] view original

I vote for AST mapping if it could bring us IC sooner. I'm happy about the general status of current Nim. IC is the last piece to the real productivity Nim yet to show.

jtv (orginal) [2023-11-02T18:56:04+01:00] view original

Incremental compilation first, REPL second. The rest is all fine but would be much lower down my list.

Niminem (orginal) [2023-11-02T20:32:42+01:00] view original

+1 for IC -> REPL -> then everything else. Mostly because I don't know what the hell this other stuff is ;)

domry (orginal) [2023-11-02T20:55:28+01:00] view original

I believe they are all incredible features and especially looking forwards view-types but I have to vote for tooling first and foremost, so IC it is. inim and nim secret serve me well already, but I can see the merits of an official REPL that can finally interop with C /C++ so that's my second vote :). i would vote view types for my third vote and the rest can follow naturally depending on the situation

Isofruit (orginal) [2023-11-02T21:02:15+01:00] view original

I'd personally find a REPL most fun, but I think adding features that benefit larger projects first like IC might be worth it more. And also of course I'm always in favour of something that makes an existing codebase more stable and easier to maintain, so juancarlos' suggestion (assuming it is actionable, I don't know anything about compilerdev) would also have my support.

Therefore my vote is split somewhere between IC and that.

bpr (orginal) [2023-11-02T21:27:45+01:00] view original

All the proposed features are very valuable.

I agree with @juancarlospaco that this seems like a good time for a cleanup, but, since that wasn't on your list, I'll say incremental compilation is the most important thing to me.

I hadn't even been expecting a REPL for Nim in the near term, and I quite desire a good one. I'm not sure I'd make it second though. Tough call!

Araq (orginal) [2023-11-28T09:56:12+01:00] view original