Note [Which transformations are innocuous]

At one point (Jun 18) I wondered if some transformations (ticks)
might be  "innocuous", in the sense that they do not unlock a later
transformation that does not occur in the same pass.  If so, we could
refrain from bumping the overall tick-count for such innocuous
transformations, and perhaps terminate the simplifier one pass
earlier.

But alas I found that virtually nothing was innocuous!  This Note
just records what I learned, in case anyone wants to try again.

These transformations are not innocuous:

*** NB: I think these ones could be made innocuous
          EtaExpansion
          LetFloatFromLet

LetFloatFromLet
    x = K (let z = e2 in Just z)
  prepareRhs transforms to
    x2 = let z=e2 in Just z
    x  = K xs
  And now more let-floating can happen in the
  next pass, on x2

PreInlineUnconditionally
  Example in spectral/cichelli/Auxil
     hinsert = ...let lo = e in
                  let j = ...lo... in
                  case x of
                    False -> ()
                    True -> case lo of I# lo' ->
                              ...j...
  When we PreInlineUnconditionally j, lo's occ-info changes to once,
  so it can be PreInlineUnconditionally in the next pass, and a
  cascade of further things can happen.

PostInlineUnconditionally
  let x = e in
  let y = ...x.. in
  case .. of { A -> ...x...y...
               B -> ...x...y... }
  Current postinlineUnconditinaly will inline y, and then x; sigh.

  But PostInlineUnconditionally might also unlock subsequent
  transformations for the same reason as PreInlineUnconditionally,
  so it's probably not innocuous anyway.

  One annoying variant is this.  CaseMerge introduces auxiliary bindings
     let b = b' in ...
  This takes another full run of the simplifier to elimiante.  But if
  the PostInlineUnconditionally, replacing b with b', is the only thing
  that happens in a Simplifier run, that probably really is innocuous.
  Perhaps an opportunity here.

KnownBranch, BetaReduction:
  May drop chunks of code, and thereby enable PreInlineUnconditionally
  for some let-binding which now occurs once

EtaExpansion:
  Example in imaginary/digits-of-e1
    fail = \void. e          where e :: IO ()
  > etaExpandRhs
    fail = \void. (\s. (e |> g) s) |> sym g      where g :: IO () ~ S -> (S,())
  > Next iteration of simplify
    fail1 = \void. \s. (e |> g) s
    fail = fail1 |> Void# -> sym g
  And now inline 'fail'

CaseMerge:
  case x of y {
    DEFAULT -> case y of z { pi -> ei }
    alts2 }
  > CaseMerge
    case x of { pi -> let z = y in ei
              ; alts2 }
  The "let z=y" case-binder-swap gets dealt with in the next pass