RFC: better ntuple implementation without @generated

[vtjnash]

instead of requiring Val to force constant propagation through dispatch, this allows type-inference to directly reason about the higher-level functionality of ntuple.

this is an example of one mechanism that could be used for avoiding needing a @generated function (and, as a bonus, a usage of Val in this case). this is not to pick on ntuple (added by @timholy in ae0ce5f), but to open a larger discussion about having more efficient code generation and runtime behaviors though more direct coupling.

ping @JeffBezanson @carnaval who I think might already have ideas for evolving this

[timholy]

Does all this have to go here rather than in inference.jl?

[vtjnash]

yes, that's part of the strategy here. there are two ntuple functions (Core.Inference and Base), and inference needs to know about both.

[timholy]

I love the idea of this, and ntuple has needed some love for some time. My addition was basically a bandaid.

[JeffBezanson]

+1 to replacing generated functions with simpler approaches. However we should not do this by manually adding t-functions; they are too hard to write and only improve type information and not necessarily the generated code.

Discussed with @vtjnash and concluded that (in addition to @inline) we should have @pure, plus something similar to @pure but indicating that the function only depends on argument types, making it very easy to evaluate at compile time. This would generalize the hacks currently in place for functions like typejoin.

[simonster]

Ref #414

[StefanKarpinski]

I'm unconvinced that more than one purity annotation is necessary – one that tells the compiler that the function can be evaluated at any time, including not at all if only the result type is needed and it can be determined without evaluation. It's probably fine for that function to actually have side effects since that makes it easy to do things like print inside of the function to see when the compiler actually decided to call it.

[vtjnash]

I believe in m you conversation with Jeff we noted that the compiler may detect various addition properties (such as a function the operates only on the type domain) and use that to supplement the user annotation.

[StefanKarpinski]

Ref #13555

[JeffBezanson]

Yes probably just @pure is sufficient.

[vtjnash]

sufficient for what? ntuple isn't pure.

[JeffBezanson]

Sufficient for type domain functions.