Tail Recursion Optimization

[CAG2Mark]

Adds a pass to the IR which optimizes mutually tail-recursive and mutually tail-recursive modulo cons functions.

The full pass first optimizes modulo cons functions to be tail recursive, then optimizes those functions to be tail recursive.

This PR consists of the following features

IR Field Assignment

Adds an assignment primitive to the IR. Given x: CtorApp and val: TrvialExpr, we now have the primitive

x.field = val in ...

Also modified the interpreter to no longer assume references to constructors are referentially transparent.

Tail Recursion Modulo Cons

Extends the idea in this paper to optimize strongly connected components of tail recursive functions, of which some calls may be within a constructor (guarded recursion).

Supports both modulo-cons calls and regular tail calls within the same strongly connected component.

Rewrite mutually tail recursive functions

Suppose we have functions:

def f(n) = e1
def g(m) = e2

where e1 and e2 tail-call f or g. This function will be rewritten a single function:

def f_g(branch, n, m) =
  let join j(branch, m, n) = 
    if branch == 0 then e1'
    else e2'
  j(branch, m, n)

def f(n) = f_g(0, n, dummy)
def g(n) = f_g(1, dummy, m)

where e1' and e2' are e1 and e2 with their tail calls to f and g replaced with jumps to j.

In general, strongly connected components of mutually tail recursive functions will be optimized in this way. All join points used in each strongly connected component will be rewritten and merged into the same join point.

[CAG2Mark]

TODO:

• Make sure variable names in the stack frame do not clash
• Make sure the merged function's name does not clash with other functions
• Substitute dummy variables (may need extension to the IR). See here
• Propagage @tailrec annotations to the IR.

[CAG2Mark]

OK, should be good to merge. There's just one trivial merge conflict in DiffTests.scala to handle

[LPTK]

OK, should be good to merge.

Great!

There's just one trivial merge conflict in DiffTests.scala to handle

So... do you intend to ever fix these conflicts? 🤔

[CAG2Mark]

So... do you intend to ever fix these conflicts? 🤔

I figured doing it now will make the diff a bit harder to review. Once everything else is sorted, I'll fix the conflicts.

[LPTK]

Amazing change, thanks! Should we merge now? (I'll do it.)

[CAG2Mark]

Go ahead