make printing of typealias env shorter using {<:T} syntax, when possible

[vtjnash]

Inspired by #35710, including taking the tests (though not any of the implementation) from there. Unlike the original PR, this also prefers the short form (eliding the TypeVar name), even in case where it is not a gensym.

julia> struct A{B,C,D,E,F,G} end

julia> A{U, <:Any, T, Int, S, <:T} where {S, T<:S, U<:Integer}
A{<:Integer, <:Any, T, Int64, S, <:T} where {S, T<:S}

Github: this closes #35710 and fixes #34887

[stev47]

Nice, I must have missed the has_typevar intrinsic back when I tried to do this. Avoiding additional io context keys using the mutable wheres is probably good too. It seems solid from my quick tests.

It would be nice if Ref{T} where T and Ref{X} where X both printed consistently, though I guess this requires non-trivial changes in how you build orig for show_typeparams.

[stev47]

maybe reduce nesting: p isa TypeVar || continue?

[vtjnash]

there is an else clause, so this is the right structure

[vtjnash]

oh, I see this is the other loop. The reason is still the same though (keeping the general structure of the two loops in common)

[stev47]

can merge with Union{} branch

[vtjnash]

I agree those should do the same (as did CI). I've used the new jl_types_egal function to handle that now, so let's see how CI fares now.

[JeffBezanson]

I think we should put back the gensym heuristic; i.e. still print where name for non-gensyms. It's nice to print something egal to the original input when possible, and less likely to break tests that might exist out there.

[vtjnash]

Okay, top example is now:

julia> struct A{B,C,D,E,F,G} end

julia> A{U, <:Any, T, Int, S, <:T} where {S, T<:S, U<:Integer}
A{U, <:Any, T, Int64, S, <:T} where {S, T<:S, U<:Integer}

with added gensym (#) test.

[vtjnash]

Thinking to merge this tomorrow, since I haven't heard any concerns with this now

[StefanKarpinski]

I actually prefer the more aggressively normalized original form. It's easier to see what's happening, IMO. Parameter names aren't significant after all, so eliminating where possible seems good.

[vtjnash]

There's some cases were readability was slightly lower. Like hiding parameter names to functions too–they are not a significant part of the call, but they can be useful. However, as Jeff mentioned, it is also slightly more preferable to print an egal copy when feasible. I still think we should merge this now, but partly because we can still easily make that choice later after getting some experience with this.

[StefanKarpinski]

Fair enough, although I think we should probably make a decision before 1.7 to minimize the number of disruptive changes to printing. Although otoh maybe it's good to get people to stop testing exact printing of types.

[vtjnash]

We've made more drastic changes than this is in 1.5 and 1.6 already, so hopefully that's already having an influence 😬

[NHDaly]

Cool, thanks! :)