Constructor with trivial assert much slower than without

[cstjean]

This code takes ~2μs

struct Myo{C}
    c::C
    function (::Type{Myo{C}}){C}(c)
        if 2 == 2   # expansion of @assert 2==2
            nothing
        else 
            throw(Base.AssertionError("2 == 2"))
        end
        return new{C}(c)
    end
end
foo() = Myo{Type{Int}}(Int)
@benchmark foo()

It's 200X slower than the same code with

• Myo{Int}(2)
• Taking out the if
• Replacing the if with the equivalent if 2 != 2 throw(...) end

Meta-question: I know compiler code is full of heuristics, so is it useful to the Julia team to file an issue like this?

[JeffBezanson]

Yes, keep the issues coming! :) Always better to know about a possible problem.

[JeffBezanson]

In Myo{Type{Int}}(Int), the field that points to Int is a reference, so the object is heap allocated. Myo{Int}(2) is stack allocated. However it looks like we are also failing to remove the type check that Int isa Type{Int}, which seems to be a real codegen issue.

[JeffBezanson]

Another thing happening here: the constructor is being specialized for Myo{Type{Int}}(::Type), causing a dynamically-dispatched convert call that takes most of the time. You can work around this by adding a static parameter for the argument (function (::Type{Myo{C}}){C,T}(c::T)).

Internally, we should be able to handle this by speculatively inlining the matching identity conversion method, along with the needed type assertion.

[JeffBezanson]

This is probably the same dispatch performance issue as #21370, except for convert instead of constructors. convert also has a large number of methods specialized on different types.

[KristofferC]

julia> @btime foo()
  6.809 ns (1 allocation: 16 bytes)

looks fixed to me.