Note [Join points]

In Core, a *join point* is a specially tagged function whose only occurrences
are saturated tail calls. A tail call can appear in these places:

  1. In the branches (not the scrutinee) of a case
  2. Underneath a let (value or join point)
  3. Inside another join point

We write a join-point declaration as
  join j @a @b x y = e1 in e2,
like a let binding but with "join" instead (or "join rec" for "let rec"). Note
that we put the parameters before the = rather than using lambdas; this is
because it's relevant how many parameters the join point takes *as a join
point.* This number is called the *join arity,* distinct from arity because it
counts types as well as values. Note that a join point may return a lambda! So
  join j x = x + 1
is different from
  join j = \x -> x + 1
The former has join arity 1, while the latter has join arity 0.

The identifier for a join point is called a join id or a *label.* An invocation
is called a *jump.* We write a jump using the jump keyword:

  jump j 3

The words *label* and *jump* are evocative of assembly code (or Cmm) for a
reason: join points are indeed compiled as labeled blocks, and jumps become
actual jumps (plus argument passing and stack adjustment). There is no closure
allocated and only a fraction of the function-call overhead. Hence we would
like as many functions as possible to become join points (see OccurAnal) and
the type rules for join points ensure we preserve the properties that make them
efficient.

In the actual AST, a join point is indicated by the IdDetails of the binder: a
local value binding gets 'VanillaId' but a join point gets a 'JoinId' with its
join arity.

For more details, see the paper:

  Luke Maurer, Paul Downen, Zena Ariola, and Simon Peyton Jones. "Compiling
  without continuations." Submitted to PLDI'17.

  https://www.microsoft.com/en-us/research/publication/compiling-without-continuations/