[falsissime]

Explicit checking of the appropriate instantiations is something for more or less official interfaces, not for every internal predicate. For those official checking via must_be/2, can_be/2 and the library(si) (all in Scryer) is often all you need. (The SWI-version conflates must_be/2 and can_be/2).

But you could save your beloved cut with minimal effort by using dif_si/2 (voir https://stackoverflow.com/a/20238931/772868) just before the cut. In this manner you would get instantiation errors for exactly the cases you cannot handle.

[YeGoblynQueenne]

>> Explicit checking of the appropriate instantiations is something for more or less official interfaces, not for every internal predicate.

But there's no formal concept of "interface" in Prolog, and the module system sucks (or doesn't exist, depending on implementation) and many programs don't ever use it anyway, so instantiations errors are a possible issue at every level of a program. For me at least it's a constant headache knowning how a predicate must be called (hence my practice of documenting the structure of inputs and outputs carefully).

I really don't like the ad-hoc "type" checking in Prolog (as in integer/1 etc). Prolog doesn't have types, because pre-Church logic doesn't have types (and post-Church logic is a mess; typed mess, but still a mess). And yet Prolog terms (in the sense of a functor followed by comma-separated terms in parentheses) are implicitly typed, under unification: T is the type of all terms that unify with T. So making sure that terms are correctly instantiated is the most prudent thing to do when checking the "type" of a term. Unfortunately that adds too much clutter and I personally resent havng to do it. I have seen various attempts to solve that conundrum over the years, for example there's a package for SWI-Prolog that adds Hindley-Milner type checking (brrr). I don't like any of them.

It's a bit weird how that is the kind of thing that Prolog gets wrong that I rarely hear as an actual criticism, as opposed to complaints about the cut, or the imperfect declarative-ness.

I guess all this would be solved in practice by the addition of a static type system, checked at compile time, which would completely change the language of course. Incidentally, I'm currently working on an ancient (30 years old) Visual Prolog project; Visual Prolog is exactly that, a compiled Prolog with a static type system. I feel that while it sure helps catch some errors, it contrives to suck all the energy and life out of Prolog. Prolog is like parcour: it lets you fly around town, jumping and rolling like Spiderman. So sometimes you get a lamppost en plein dans la figure. So what. Who needs type safety anyway?

(She says while changing yet another bloody bandage)

>> But you could save your beloved cut with minimal effort by using dif_si/2 (voir https://stackoverflow.com/a/20238931/772868) just before the cut. In this manner you would get instantiation errors for exactly the cases you cannot handle.

Thanks for the pointer. I had a quick look. I think there's something similar in SWI-Prolog's Picat-style matching with the "=>" operator that is an alternative to ":-". I'm not sure exactly what it does (something about "steadfastness", that you also mentioned, and that is a new term for me) but I prefer to not use libraries that significantly change standard Prolog syntax, for fear of backwards compatibility.

[falsissime]

> But there's no formal concept of "interface" in Prolog

Any predicate you write for someone else has an interface. Someone else includes you in a couple of weeks. All predicates that are not truly auxiliary ones which are more like basic blocks of command oriented programming languages - COPLs.

> instantiations errors are a possible issue at every level of a program

That is not the worst that can happen. Nor is non-termination the worst. The worst is if an unexpected failure or success happens that looks like a legitimate answer. However, for many, non-termination is seen as being worse than such real errors. The reason is that there are only very few Prolog systems that reliably catch resource errors or have reliable timeouts. Well, to be true, there is one system. And the others most often abort.

> (hence my practice of documenting the structure of inputs and outputs carefully).

> I really don't like the ad-hoc "type" checking in Prolog (as in integer/1 etc).

integer/1 is one of the orignal sins of DEC10 which remplaced the errors of Prolog I by silent failure. And everybody then followed suit. Mindlessly. There is library(si) that offers integer_si/1 to make things much cleaner.

As for your remarks on type systems for Prolog, one feature they do not provide is a dynamic conversion between regular Prolog and the typed part. There was an attempt some time ago but unfortunately the individual left for FP.

> Who needs type safety anyway?

It seems there are two different objectives: type safety and instantiation safety. Both are often confounded.

> SWI-Prolog's Picat-style matching with the "=>" operator that is an > alternative to ":-".

This approach does not distinguish between instantiation and type errors. Take the sum_list/2 from https://www.swi-prolog.org/pldoc/man?section=ssu

   ?- sum_list(1, N).
      existence_error(matching_rule, sum_list(1, 0, _A)).
   ?- sum_list(L, N).
      existence_error(matching_rule, sum_list(_A, 0, _B)).

And then when adding the suggested catchall rule, both fail. Both. This throws us back by almost half a century. So now DEC10-style errors get into the 21st century. Prolog 0 and Prolog I were better.

   ?- sum_list(1, N).
      false.
   ?- sum_list(L, N).
      false, unexpected.

> "steadfastness", that you also mentioned, and that is a new term for me

Since 1987-09-29 in common use, see the Prolog Digest of that date, posting by Richard O'Keefe. Here is my definition: an argument Ai is steadfast w.r.t. a goal g(A1,..Ai,..AN), when executing this goal is equivalent (complete operationally) to, g(A1,..CAi,..AN), CAi = Ai with CAi a fresh new variable. And more generally an argument Ai is steadfast, if this property holds for all goals.

A good example of a steadfast argument is the second argument of append/3. You can replace any goal append(Xs, Ys, Zs) by append(Xs, CYs, Zs), CYs = Ys without observable effect (except efficiency, resource consumption and the like). But even non-termination is preserved! Another one is the 3rd argument of phrase/3.

Or take your run_length_encoding/2/3, where the second argument is steadfast. You (presumably on purpose) accumulated in the second argument all the elements only to reverse them finally in a highly lispeling manner. At least no destructive nreverse.

[YeGoblynQueenne]

>> That is not the worst that can happen. Nor is non-termination the worst. The worst is if an unexpected failure or success happens that looks like a legitimate answer. However, for many, non-termination is seen as being worse than such real errors. The reason is that there are only very few Prolog systems that reliably catch resource errors or have reliable timeouts. Well, to be true, there is one system. And the others most often abort.

Yes, those "right for the wrong reasons" results are a bad problem, too.

Which system are you referring to? I mainly use SWI-Prolog and in my experience it does exit with error when the stack size limit is exceeded, and it will even give you a message about possible unbounded recursion. Although on my Windows machine it then often crashes immediately after because it's left with no memory to recover from the error. I normally don't try-catch such errors because they're most of the time a sign of programming error (a well-written program should go into an infinite recursion without blowing the stack).

I've had more mixed experience with SWI-Prolog's resource-limited call/1 variants, like call_with_time_limit/2 and call_with_depth_limit/3 and call_with_inference_limit/3. The first one seems to work OK. The other two are a bit hit-and-miss, although there's warnings about that in the documentation if I remember correctly.

>> Or take your run_length_encoding/2/3, where the second argument is steadfast. You (presumably on purpose) accumulated in the second argument all the elements only to reverse them finally in a highly lispeling manner. At least no destructive nreverse.

Oh, is that lisp-y style? I had no idea. I use it because it makes tracing the program easier. There's a bit of overhead, but I feel it's made up for in spades by the ability to debug errors. If you put an accumulator variable in the head of the goal, it is only bound when the recursion terminates and the stack "unrolls", so you don't get to see the bindings of the accumulator and it's harder to verify the logic of the program. I normally trace with the "unify" port non-leashed but I don't think that changes anything.

>> A good example of a steadfast argument is the second argument of append/3. You can replace any goal append(Xs, Ys, Zs) by append(Xs, CYs, Zs), CYs = Ys without observable effect (except efficiency, resource consumption and the like). But even non-termination is preserved! Another one is the 3rd argument of phrase/3.

Thanks, this is a clear explanation. To be honest, I don't stay up-to-date with discussions about Prolog programming, as I did in the past. Even in the discussions that I've followed there's a lot that turns around principles that don't mean anything to me (as I said in another comment, I really can't be bothered about declarative, vs non-declarative programming, for example) and so I'm a bit suspicious of jargon that seems to be pointing to such principles. Maybe steadfastness is a more practical consideration. I'll have to think about it.

[falsissime]

> Which system are you referring to?

SICStus. The best for catching resource errors (that is with catch/3) and continues thereafter happily. SWI does catch many situations, but as you observe it is much less reliable. SICStus is also best for timeouts. I have not tested SWI recently, only noted that the compatibility library(timeout) has been updated. The call_with_inference_limit/3 might be interesting. The others you mention have just bad interfaces. But what both systems lack is a reproducible notion of time on a lower level. Walltime is not useful, as it is load dependent, CPU-time depends on TLBs, caches and all that. Think of number of (completed) CPU-instructions. PAPI-wise. While there is some JIT-tering in both systems this is still the best. Anyway, I start daydreaming...

> they're most of the time a sign of programming error

Often yes, but then you need to explain such errors which incurs a lot of attempts to execute fragments of the looping program. Like with a failure-slice https://stackoverflow.com/tags/failure-slice

And then, there is another kind of programs that effectively do not terminate but are still useful. Those that try to find counterexamples (posh expression for bugs).

> Oh, is that lisp-y style?

Recursive functions a la (cons x (recursive ...)) are not tail recursive (or at least have been not in the 1970s), whereas the corresponding code in Prolog is. So they were often transformed with an auxiliary argument that collected all conses in the wrong direction and were nreverse-d thereafter since (ideally) noone else referred to that list. I believe this has been solved in the meantime.

And yes, with a step-by-step tracer, this style makes things easier to watch. OTOH, you could use purer methods for debugging (for the pure part of your programs). At least, when you are into constraints, the tracers are of no use any longer and you have to switch.