Hacker News Reader

Comment by jhbadger | Hacker News Reader

Have they ever fixed the problem with let not being implemented (which was the case when I looked at it a couple of years ago). Lisp-like languages without let lose a lot of the inherent power of lisps.

View on HN

It's been implemented, found to be subtly broken, and re-implemented a number of times. Ultimately trying to impose a different model of scoping on top of Python's scoping is living in a state of sin; thus the difficulties. The status quo is an apparently working `let` due to a truly heroic effort by Sunjay Cauligi ( https://github.com/hylang/hy/pull/2125 ). You're encouraged to try to find bugs in it.

In truth, I've never understood the fuss about `let`. It's taken a huge amount of Hy development effort, but it adds very little. Python's native scoping works well in the large majority of cases, and in tricky cases, I'd rather use `nonlocal` or `global` explicitly so I'm sure what's happening at the level of Python semantics. I suspect that people who heard that previous versions of Hy were missing `let` assumed that some other feature (such as nested lexical scopes) was also missing that has actually always been there.

> In truth, I've never understood the fuss about `let`. It's taken a huge amount of Hy development effort, but it adds very little

My use case is Jupyter notebooks. With normal Python scoping you quickly accumulate a lot of global variables and you might accidentally use the wrong one. With let-bindings you have variables that clean up after themselves.

I'm not familiar with Jupyter internals, but I think functions would work well for this. Hy allows multi-statement anonymous functions, and you can do stuff like `((fn [] … ))` to define and immediately call the function. One advantage of this over `let` is that it allows old locals to be garbage-collected. If memory isn't an issue, though, `let` should work well.

What problems did you encoutner implementing 'let'? Shouldn't it be equivalent to lambda expressions somehow?

In some other Lisps, yes, but Python has its own idea of how function scoping works that generally isn't what you'd want with `let`. For example, in

    (let [a 1]
      (defn foo []
        (setv a 2))
      (foo)
      (print a))

you probably `2` to be printed, but the seeming equivalent Python code

    a = 1
    def foo():
        a = 2
    foo()
    print(a)

prints `1`. Python interprets the assignment as declaring a new local variable. Frankly, Python got scoping wrong (in its quest to eliminate declarations), but here we are. So `let` needs to work around this somehow. The current implementation implicitly adds `global a` or `nonlocal a` as appropriate. (Actually, the name `a` is also mangled so that any references to `a` outside the `let` refer to a different variable.)

For me the example printing `1` makes sense because I think `let` as being something immutable, so I wouldn't expect `setv` to mutate that variable (would expect either a compiler error or the introduction of a new local variable).

But maybe it is because I am familiar with Clojure (where `setv` or equivalents doesn't exist) and not other Lisps.

This would be the equivalent Clojure code anyway:

  user=> (let [a 1] (defn foo [] (def a 2)) (foo) (println a))
  1
  nil

So yeah, makes sense for me. But can understand why some other people would expect the other way though.

Equivalent in Common Lisp:

    * (let ((a 1))
        (defun foo ()
          (setq a 2))
        (foo)
        (print a))
    
    2
    2

this made me wonder - I know I can use global variables inside a function definition:

    a = 1
    def foo():
      x = a
    foo()

so if this 'a' is a new variable, what happens when the name collides with a global variable?

    a = 1
    def foo():
      a = a
    foo()

    > UnboundLocalError: local variable 'a' referenced before assignment

yuck!

It works if you disambiguate your namespaces:

    >>> a = 1
    >>> def foo():
    ...     a = globals()['a']
    ...     print(a)
    ...
    >>> foo()
    1

Still yuck?

Here's another way to do it:

  >>> def foo():
  ...     import __main__
  ...     a = __main__.a
  ...     print(a)
  ...
  >>> foo()
  1

I'm not sure this is so much of a problem? Changing what you name the variable fixes the issue.

Python is generally quite good at having a verbose option to express things, so I'm surprised there's not an explicit way to reference the enclosing scope from an object.

Is there any non-horror piece of code where supporting this latter case would be useful?

It is not a particularly good example. A better one would be

    a = 1
    def foo():
      x = a
      a = 2
    foo()

    > UnboundLocalError: local variable 'a' referenced before assignment

Basically a name cannot be both global and local inside the same function. Otherwise it would be a total mess.

"a = a" just means assign the value of local variable "a" to a local variable "a", so the error makes sense.

it kinda sorta works when wrapped in a function

  def f():
      a = 1
      def g():
          nonlocal a
          a = 2
      g()
      print(a)

though nonlocal can't be used when referring to global variables... it's kinda bad.

You can still use `global` though.

It's sort of back, implemented as a macro. I've yet to go back to Hy (I left because of the removal of let, plus other things I found missing), but I try it every now and then and the last time I did the namespacing was still a tad weird.

Have you tried Hissp yet?

They are working on improving it. I wrote a short little book on Hy and I initially used the “let” macro. I stopped using it because in general Hy code can be automatically translated to very readable Python code, but using the “let” macro generates ugly Python code.

This might be interesting https://pypi.org/project/scoping/

Reply on news.ycombinator.com