(see Fig. 6 p.21)
I believe that by applying ANF, you're baking in either CBN or CBV at the time of writing the compiler.
If I've read it right, the author is suggesting that CBPV evaluates (ordinary) values before passing them, but does not evaluate functions (even though they are also values!) before passing them.
I believe the first happens at compile time, and the second happens after the function is passed.
1 f x =
2 let xx = x * x in
3 (\y -> y + xx)
4
5 main = do
6 let clo = f 2
7 let applied = clo 3
8 print applied
The closure is defined on line 3, and instantiated on line 6.xx is known to be the captured variable at compile time. But in a lazy setting, it's fine for the closure to be passed around before the value of xx is known:
f x =
let xx = trace "Evaluating xx" (x * x) in
trace "Returning closure" (\y -> y + xx)
main = do
let clo = f 2
let applied = clo 3
print applied
> Returning closure> Evaluating xx
> 7
That does remind of another thing that CBPV glosses over though - it is only a fusion of call-by-name and call-by-value. In particular it doesn't model call-by-need particularly well, there is no notion of sharing thunks that are duplicated during evaluation. Actually call-by-need is sort of arbitrary, generally one is better off using "an implementation-specific evaluation strategy somewhere between CBN and CBV", which again is not modeled well by CBPV, as the choice of CBN or CBV evaluation is encoded in the IR so cannot be easily changed by the compiler.
If you squint at CBPV, there is exactly one primitive that sequences computation, `M to x. N`. In Haskell this is just the monadic bind `M >>= \x -> N` or `do { x <- M; N}`. If I had to bet, I would place my bet on monads, not CBPV.