`Bug` Quadratic(?) compiler perf for huge nested expressions

[dsyme]

While investigating #12420 I noticed that one particular part of our optimizer seems quadratic for very large nested expressions made of nothing but function/method applications, e.g. M(M(M(M.....)))) and the kind of code that comes out of computation expressions.

The problem is already noted in our code, see permalink

This isn't critical but we should fix this sometime. It appears to be re-traversing expressions determining free variables.

[T-Gro]

I will check this one.

The behaviour is definitely still present.
A simple single-file program consisting of a linear chain of primitive calls (like f(f(f(f(.. ) has the following timing differences between the --optimize- and --optimize+ switches: