4 ms·

> (From the POV of static types, dynamically typed languages are just untyped languages, that is, languages that have exactly one type) I believe the mistake y

> (From the POV of static types, dynamically typed languages are just untyped languages, that is, languages that have exactly one type) I believe the mistake you are making is assuming that TLA+ has some kind of runtime where things are "executed" (static vs. dynamic in the context of programming languages refers to, roughly, compile-time vs. execution-time). Model checking in TLA+ is actually analogous to "running a compiler", not running a REPL or whatever. It's definitely NOT a runtime. TLA+ has no "dynamic" aspect at all, it's all "static" from the POV of language theory. And since it DOES provide a way (read: syntax) to create and check types (as the GP shows) and those types are checked statically (again, remember: there is no execution happening here), TLA+ is formally (and practically) equivalent to a language that has a different syntactic way to specify types, i.e. one that programmers are more familiar with. What TLA+ is missing is any notion of executing code. That's what makes it a specification language and not an implementation language. Most programming languages are implementation languages, with a bit of "specification"—usually types—sprinkled on top. TLA+ is specification-only, no implementation stuff is written down. So to recap, from a language-theoretic POV, TLA+ absolutely supports "static typing" in the way that it is usually understood and used by programmers and language theoreticians, but with a syntax that is unfamiliar (because it is a specification language). For completeness, you can specify modern type inference algorithms like MLstruct/MLsub [0] in TLA+ and the model checker will happily apply them to your TLA+ specification (again, statically—at compile-time, which is really "model checking" time). [0] https://lptk.github.io/files/[v8.0]%20mlstruct.pdf https://lptk.github.io/files/[v8.0]%20mlstruct.pdf

abathologist 3y ago
I think you and Ron are confusing the relevant phase distinction the rest of us have in mind when we talk about type checking in this context. What we mean is that quint, the language, is expressed on top of a many-sorted logic which allows our tooling to find and diagnose typing errors prior to *running* any type checker. The TLA Toolbox doesn't support this, and TLA+, like TLA itself, is untyped. TLC can (and will) raise errors while in the process of checking your model due to the kinds of problems we usually call "typing errors". Quint and apalache shouldn't do this. I hope this helps clarify a bit. We understand that, viewed purely theoretically, TLA+ has no dynamic aspect at all, etc. However, our team is very focused on the fusion of theory and practice, and we think it's important to ground our understanding and our discourse in the way the specification languages are actually used by real people given the existing tools. While it may be true in theory that > What TLA+ is missing is any notion of executing code. The file `TLC.tla` in the `StandardModules` includes these lines: ``` Print(out, val) == val PrintT(out) == TRUE (* This expression equals TRUE, but evaluating it causes TLC to print *) (* the value of out. *) ``` https://github.com/tlaplus/tlaplus/blob/master/tlatools/org.lamport.tlatools/src/tla2sany/StandardModules/TLC.tla#L17-L22 https://github.com/tlaplus/tlaplus/blob/master/tlatools/org.... In any case, our whole team thinks TLA is great, and we're happy people like you and Ron find it so useful and insightful. We also think it is a very insightful tool for guiding understanding :)
- erichocean 3y ago
  > TLA+ is missing is any notion of executing code Which is why you had to talk about TLC, not TLA+. And yeah, while the presence of printf-style debugging for specifications being model checked by TLC is technically executable-ish, I think in this case, it kind of proves the point that TLA+ itself isn't. Regardless, I like the direction Quint is taking though I would quibble with this language in your reply: > our tooling to find and diagnose typing errors [runs] prior to [...] any type checker If you are diagnosing typing errors, then definitionally, you are running a type checker. It's great that it's fast and interactive, runs prior to the model checker, is easy for programmers to read/write, et cetera. But…it's still a type checker. :-) Anyway, keep up the great work! I've been on your newsletter for a long time and love the work you guys are doing.
  - bugarela 3y ago
    I think they meant "model checker" instead of "type checker" in that sentence. Otherwise, of course, we need to run the type checker to get the type diagnosis.
    - abathologist 3y ago
      That's what I meant! Thanks for the correction :)
- pron 3y ago
  > which allows our tooling to find and diagnose typing errors prior to running any model checker. Oh, I understand what you meant (and I think Quint is great!), but I'm trying to communicate to others how ineffective programming terminology is when it comes to things that aren't programming (I remember that many years ago, when I first learnt TLA+ I asked on the mailing list if TLA+ evaluates expressions eagerly or lazily because I was confused by the meaning of some expressions, and Leslie Lamport answered that there the question is meaningless because there is no program running and so no evaluation at all). Quint's tooling can find typing errors by running a tool called a type checker, while in TLA+ the tooling can find "typing" errors by running a tool called a model checker. Neither of these tools actually runs the program expressed in the text (although a model checker usually does evaluate some expressions), and so they're both "static" checks. This is important to point out because some people try to think of a model checker as something that runs a program many times, and that's why I gave an example showing how a model checker can check something in milliseconds without possibly having run "the program" even once, let alone all of its possible executions. But yes, the runtime efficiency of a type checker, i.e. a checker that performs its reasoning based on simple deductive rules is, is usually better than the efficiency of a model checker, which reasons in the semantic model, so even though in both cases you don't run "the program" to find type errors, the Quint tooling will report typing errors much more quickly (in wall-clock time) than the TLA+ tooling.

abathologist 3y ago

I think you and Ron are confusing the relevant phase distinction the rest of us have in mind when we talk about type checking in this context. What we mean is that quint, the language, is expressed on top of a many-sorted logic which allows our tooling to find and diagnose typing errors prior to *running* any type checker. The TLA Toolbox doesn't support this, and TLA+, like TLA itself, is untyped. TLC can (and will) raise errors while in the process of checking your model due to the kinds of problems we usually call "typing errors". Quint and apalache shouldn't do this. I hope this helps clarify a bit. We understand that, viewed purely theoretically, TLA+ has no dynamic aspect at all, etc. However, our team is very focused on the fusion of theory and practice, and we think it's important to ground our understanding and our discourse in the way the specification languages are actually used by real people given the existing tools. While it may be true in theory that > What TLA+ is missing is any notion of executing code. The file `TLC.tla` in the `StandardModules` includes these lines: ``` Print(out, val) == val PrintT(out) == TRUE (* This expression equals TRUE, but evaluating it causes TLC to print *) (* the value of out. *) ``` https://github.com/tlaplus/tlaplus/blob/master/tlatools/org.lamport.tlatools/src/tla2sany/StandardModules/TLC.tla#L17-L22 https://github.com/tlaplus/tlaplus/blob/master/tlatools/org.... In any case, our whole team thinks TLA is great, and we're happy people like you and Ron find it so useful and insightful. We also think it is a very insightful tool for guiding understanding :)

erichocean 3y ago
> TLA+ is missing is any notion of executing code Which is why you had to talk about TLC, not TLA+. And yeah, while the presence of printf-style debugging for specifications being model checked by TLC is technically executable-ish, I think in this case, it kind of proves the point that TLA+ itself isn't. Regardless, I like the direction Quint is taking though I would quibble with this language in your reply: > our tooling to find and diagnose typing errors [runs] prior to [...] any type checker If you are diagnosing typing errors, then definitionally, you are running a type checker. It's great that it's fast and interactive, runs prior to the model checker, is easy for programmers to read/write, et cetera. But…it's still a type checker. :-) Anyway, keep up the great work! I've been on your newsletter for a long time and love the work you guys are doing.
- bugarela 3y ago
  I think they meant "model checker" instead of "type checker" in that sentence. Otherwise, of course, we need to run the type checker to get the type diagnosis.
  - abathologist 3y ago
    That's what I meant! Thanks for the correction :)
pron 3y ago
> which allows our tooling to find and diagnose typing errors prior to running any model checker. Oh, I understand what you meant (and I think Quint is great!), but I'm trying to communicate to others how ineffective programming terminology is when it comes to things that aren't programming (I remember that many years ago, when I first learnt TLA+ I asked on the mailing list if TLA+ evaluates expressions eagerly or lazily because I was confused by the meaning of some expressions, and Leslie Lamport answered that there the question is meaningless because there is no program running and so no evaluation at all). Quint's tooling can find typing errors by running a tool called a type checker, while in TLA+ the tooling can find "typing" errors by running a tool called a model checker. Neither of these tools actually runs the program expressed in the text (although a model checker usually does evaluate some expressions), and so they're both "static" checks. This is important to point out because some people try to think of a model checker as something that runs a program many times, and that's why I gave an example showing how a model checker can check something in milliseconds without possibly having run "the program" even once, let alone all of its possible executions. But yes, the runtime efficiency of a type checker, i.e. a checker that performs its reasoning based on simple deductive rules is, is usually better than the efficiency of a model checker, which reasons in the semantic model, so even though in both cases you don't run "the program" to find type errors, the Quint tooling will report typing errors much more quickly (in wall-clock time) than the TLA+ tooling.