Add ability to limit the number of generated atoms

[Maria-12648430]

As mentioned in #290, the usage of the atom generator (explicit or implicit) can lead to exhaustion of the atom table. While I understand that PropEr does not want to restrict itself to just afew atoms per se, it would be very useful if a user (like me) was given the choice to limit this if needed.

[Maria-12648430]

When you do add some tests, you will discover that the code for the non-infinity case cannot run on older Erlang/OTP versions (e.g. 21 and 22 - I believe) because persistent_term is not available there. So, this code cannot be included as is as long as PropEr still supports these Erlang/OTP versions...

For that case, I suggest we put in a conditional that falls back to always using infinity (and ignore the setting) when used with an older OTP version that does not support persistent_term. Would that be acceptable?

EDIT: persistent_term was introduced in OTP 21.2.

[kostis]

EDIT: persistent_term was introduced in OTP 21.2.

OK - so then it's not that bad as I feared. In that case, yes, the conditional you suggest will do.

[Maria-12648430]

I'll retract this PR for the time being. Having tried to make the existing tests work, I realized that it changes the behavior of the current atom generator to a considerable degree, especially when it comes to shrinking, which is currently tightly related to list and integer shrinking.

However, as stated in #290 and the follow-up discussion in the Erlang Forums, @juhlig's and my general opinion is that...

• Generating atoms without restraint is, obviously, dangerous (and no, the non-garbage-collection behavior of atoms in Erlang itself won't change anytime soon).
• Any atom is basically as good, or bad for that matter, as any other. There is no difference in complexity based on the length or the number that makes up any of its characters, between an empty atom and an atom composed of 255 255s.
• There are the well-known and universally used atoms like undefined, ok, error, infinity, true, false and whatnot; they are the most likely to cause trouble, but it is very unlikely that one of them is created by an atom generator that produces just random strings and turns them into atoms. For example, the probability that the two-character atom ok is generated by a generator operating at size 2 is about 0.0015%.
• Conversely, there are legions of atoms being generated where any of them is as (un-)likely to be problematic as any other.

For those reasons, it is our opinion that the current way of atom generation is, to put it bluntly, pretty useless, and that it needs to be reconsidered and redesigned. Which is no easy task, because in a realm wherein anything is as good or bad as anything else, what can you reduce to and shrink on?