How to avoid needing CSE?

[dpsanders]

julia> @variables x y;

julia> function f(x, y)
       a = x + y
       return a + a^2
       end
f (generic function with 1 method)

julia> f(x, y)
(x() + y()) + (x() + y()) ^ 2

How can we prevent x + y being expanded each time that the intermediate a is used? This seems to be a bad idea, since it's giving more work for compiler later to do CSE. (Is this already implemented somehow?)

[HarrisonGrodin]

This is not yet implemented, but it will be soon. Once Terms.jl is finished, we will have an API for defining terms with custom behaviors and representations; then, the CSE-like behavior will happen inside the term itself, likely in the form of treating terms as DAGs and using "graph rewriting".

[asprionj]

What's the status here? There's been a lot going on w.r.t. components (i.e. connected systems); how do those developments relate to terms, CSE (and possibly automatic calculation of those intermediate values from the solution)?

[ChrisRackauckas]

The component system makes it so there's a lot of calculations that are automatically reused, but not within a single expression. This is still something that could exist, but it's not the biggest deal in the grand scheme of things.

[ChrisRackauckas]

CSE is a build_function Symbolics.jl thing now, and I wouldn't worry about it on the model building side since it rejiggers the equations into an optimized form.