remove piracy in hcat, vcat, hvcat

[ericphanson]

closes #512

[odow]

I don't know that I like this approach.

Wouldn't it be better to force Convex.hcat and Convex.vcat?

Alternatively, this might work:

Base.hcat(args::AbstractExpr...) = HcatAtom(args...)

function Base.hcat(args::AbstractExprOrValue...)
    if all(Base.Fix2(isa, Value), args)
        return Base.cat(args..., dims = Val(2))
    end
    return HcatAtom(map(arg -> convert(AbstractExpr, arg), args)...)
end

[ericphanson]

I agree, this is unsatisfying. However: Convex.hcat doesn't allow use of the [A B] implicit hcat syntax which is often used in these problems. And I think

function Base.hcat(args::AbstractExprOrValue...)
    if all(Base.Fix2(isa, Value), args)
        return Base.cat(args..., dims = Val(2))
    end
    return HcatAtom(map(arg -> convert(AbstractExpr, arg), args)...)
end

is the same piracy as we are doing now, no? It's the method signature that matters for that, not what happens inside

edit: or are you saying removing the hcat(args::Value...) = Base.cat(args..., dims = Val(2)) method might be sufficient to fix precomilation?

[KristofferC]

Just FYI, these type of combinatorial method definition can lead to bad load times for the package (JuliaDiff/ReverseDiff.jl#226).

[ericphanson]

Yeah, I noticed that when I initially chose n=32 which never finished precompiling 😂. Do you have any ideas for how this can be improved?

I think the "right" fix is to be able to express "Varags{Union{Value,AbstractExpr}} but at least 1 AbstractExpr" in the type system, but I am guessing that isn't so easy to do. (I guess JuliaLang/julia#39878 is the issue for that).

with 57 methods for the three functions, load times "feel" ok to me, but increasing n blows up the number of methods very quickly.

[odow]

is the same piracy as we are doing now, no?

Sort of, but to a lesser extent. Since AbstractExprOrValue is basically Union{Number,AbstractVector,AbstractExpr}, it should take lower precedence than the Union{Number,AbstractVector} method in Base.