recursive pretty print

[enricozb]

Fixes issues with segfault after printing: #360.
Also it doesn't overload the ERA node.

Of course a large enough net will still overflow the stack. On my machine, this example bend code:

def create(n):
  bend x = 0:
    when x < n:
      str = List/Cons { head: 1, tail: fork(x + 1) }
    else:
      str = List/Nil
  return str

def main():
  return create(37)

Which previously segfaulted at create(37), now segfaults somewhere between create(50000) and create(100000).

[VictorTaelin]

You just removed the stack? That's a regression, in CUDA this will cause the compiler to be unable to estimate the stack size of the pretty print function, which can have harsh consequences on performance elsewhere and cause segfaults in unrelated code, making us unable to use the pretty printer for debugging purposes. Have you tested using it inside the interaction kernel on the GPU?

[enricozb]

I didn't test pretty print within any of the cuda code, didn't think of that as a use case. We can have a stacked version for debugging and use this for the output on the host then?

[HigherOrderBot]

Perf run for 6170de3:

compiled
========

file            runtime         main            (local)       
==============================================================
sort_bitonic    c                        3.93s           3.50s
                cuda                     0.24s           0.23s
--------------------------------------------------------------
sum_rec         c                        1.43s           1.47s
                cuda                     0.13s           0.13s
--------------------------------------------------------------
sum_tree        c                        0.12s           0.13s
                cuda                     0.09s           0.07s
--------------------------------------------------------------
tuples          c                        2.85s           3.87s
                cuda                   timeout         timeout
--------------------------------------------------------------

interpreted
===========

file            runtime         main            (local)       
==============================================================
sort_bitonic    c                        4.89s           4.23s
                cuda                     0.23s           0.23s
                rust                   timeout         timeout
--------------------------------------------------------------
sum_rec         c                        1.82s           1.73s
                cuda                     0.13s           5.00s
                rust                    13.90s          13.47s
--------------------------------------------------------------
sum_tree        c                        0.21s           0.24s
                cuda                     0.09s           0.08s
                rust                     0.87s           0.82s
--------------------------------------------------------------
tuples          c                        3.00s           2.61s
                cuda                   timeout         timeout
                rust                     3.76s           3.75s
--------------------------------------------------------------