[racket-users] Re: Help implementing an early return macro

[jackh...@gmail.com]

A brief update on this: I went with Ryan's approach and used an 
implementation of guarded-block throughout Rebellion's codebase. You can 
see the diff here: https://github.com/jackfirth/rebellion/pull/466. A 
couple of things to note:

- I added a define/guard form that's like define, but with the function 
body implicitly wrapped with guarded-block.
- I tweaked the syntax of guard to allow either (guard  else 
 ...) or (guard  then  ...). If else is used, the 
guard is taken if the condition is false. If then is used, the guard is 
taken if the condition is true. This made it easier to organize my code 
such that the short-circuiting precondition checks came before what I 
considered to be the main branch of the function. With just the else case, 
sometimes I had to awkwardly negate things.

There's a few problems related to robustness I'd like to figure out 
eventually, before documenting this macro and publicly exposing it:

- It allows shadowing. If a variable is defined both before and after a 
guard statement, the definition after shadows the one before it, instead of 
raising a duplicate definition error.
- If any definitions raise syntax errors, the error is reported in terms of 
the expanded define-values form instead of the definition the user actually 
wrote.
- The last body form of a guarded block shouldn't be allowed to be a guard 
statement, and a good error message should be raised if a user does that. 
The current implementation doesn't check for that.

The last one is pretty easy to fix, but the other two I'm not sure how to 
fix.

On Sunday, November 1, 2020 at 1:04:15 PM UTC-8 gneuner2 wrote:

> On Sat, 31 Oct 2020 03:25:32 -0700 (PDT),
> "jackh...@gmail.com"
>  wrote:
>
> >Wow, these are a lot of great responses. First of all, *awesome* job 
> Ryan. 
> >That implementation is exactly what I needed to figure out. I'm 
> definitely 
> >starting there first.
> >
> >> Are you looking for `let/ec`?
> >
> >I'd forgotten about that one. That has the *syntax* I want. However my 
> >issue with continuation-based approaches isn't the syntax, or even the 
> >performance. It's the semantics. What if someone writes code like this?
> >
> >(guarded-block
> > (define x (random 10))
> > (thread
> > (lambda ()
> > (guard (even? x) else #false)
> > ...)))
> >
> >If I implemented guarded-block in terms of let/ec, then what does this 
> code 
> >even *do*? I honestly don't know. It would probably run without error and 
> >do... something. 
>
> (let/ec return
> (thread
> (lambda ()
> (return -1)
> 42)))
>
> throws an error: "continuation application: attempt to jump into an
> escape continuation"
>
> There is a continuation barrier between the threads.
>
>
> However, let/cc works: e.g.,
>
> (let/cc return
> (thread
> (lambda ()
> (return -1)
> 42)))
>
> returns -1.
>
>
> >I am extremely sure that regardless of what it did, it 
> >would be confusing and it wouldn't solve any problem I had. 
>
> I am sure it would solve *some* problems.
>
>
> >I just flat out don't want to allow this or any related nonsense, 
> >such as:
> >
> >; Aliasing
> >(define return guard)
> >
> >; Higher-order usage
> >(map guard (some-list ...))
> >
> >; Capturing via closure
> >(guarded-block
> > (define (check-foo x) (guard (foo? x) else #false))
> > (check-foo ...)
> > ...)
> >
> >; Capturing via mutation
> >(set! previous-guard guard)
> >
> >; Oh great, now I have to think about even more continuation jumps
> >(dynamic-wind
> > (lambda () (guard ...))
> > (lambda () ...)
> > (lambda () ...))
> >
> >There might be valid use cases for some of these, but I certainly don't 
> >understand those use cases well enough to commit to a semantics for them.
>
> It always is safe to jump upwards OUT of a lower level computation.
> The sticky issues [and mental gyrations] with continuations all have
> to do with jumping downwards or sideways.
>
> I'm not sure what problems you might have with continuation barriers:
> the example of the thread shows that (upward-only) "escape"
> continuations don't work across threads ... but "full" continations do
> work, and would still work even if the threads were siblings rather
> than in a parent/child relationship.
>
> The issue for your purpose would be making sure the continuation is
> called from a position that is guaranteed to terminate the errant
> thread: you might need to recognize a thread as a special case, wrap
> it and (like an exception) catch/rethrow the continuation.
>
>
> Or, if you don't care about sibling threads, just use exceptions which
> always can be thrown upward out of child threads.
>
> George
>
>

-- 
You received this message because you are subscribed to the Google Groups 
"Racket Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to racket-users+unsubscr...@googlegroups.com.
To view this discussion on the web visit 
https://groups.google.com/d/msgid/racket-users/d7714f66-d791-4ec8-a1c1-da8a9c761caen%40googlegroups.com.

[racket-users] Re: Help implementing an early return macro

[George Neuner]

On Sat, 31 Oct 2020 03:25:32 -0700 (PDT),
"jackh...@gmail.com"
 wrote:

>Wow, these are a lot of great responses. First of all, *awesome* job Ryan. 
>That implementation is exactly what I needed to figure out. I'm definitely 
>starting there first.
>
>> Are you looking for `let/ec`?
>
>I'd forgotten about that one. That has the *syntax* I want. However my 
>issue with continuation-based approaches isn't the syntax, or even the 
>performance. It's the semantics. What if someone writes code like this?
>
>(guarded-block
>  (define x (random 10))
>  (thread
>(lambda ()
>  (guard (even? x) else #false)
>  ...)))
>
>If I implemented guarded-block in terms of let/ec, then what does this code 
>even *do*? I honestly don't know. It would probably run without error and 
>do... something. 

(let/ec return
  (thread
   (lambda ()
 (return -1)
 42)))

throws an error:  "continuation application: attempt to jump into an
escape continuation"

There is a continuation barrier between the threads.


However, let/cc works: e.g.,

(let/cc return
  (thread
   (lambda ()
 (return -1)
 42)))

returns -1.


>I am extremely sure that regardless of what it did, it 
>would be confusing and it wouldn't solve any problem I had. 

I am sure it would solve *some* problems.


>I just flat out don't want to allow this or any related nonsense, 
>such as:
>
>; Aliasing
>(define return guard)
>
>; Higher-order usage
>(map guard (some-list ...))
>
>; Capturing via closure
>(guarded-block
>  (define (check-foo x) (guard (foo? x) else #false))
>  (check-foo ...)
>  ...)
>
>; Capturing via mutation
>(set! previous-guard guard)
>
>; Oh great, now I have to think about even more continuation jumps
>(dynamic-wind
>  (lambda () (guard ...))
>  (lambda () ...)
>  (lambda () ...))
>
>There might be valid use cases for some of these, but I certainly don't 
>understand those use cases well enough to commit to a semantics for them.

It always is safe to jump upwards OUT of a lower level computation.
The sticky issues [and mental gyrations] with continuations all have
to do with jumping downwards or sideways.

I'm not sure what problems you might have with continuation barriers:
the example of the thread shows that (upward-only) "escape"
continuations don't work across threads ... but "full" continations do
work, and would still work even if the threads were siblings rather
than in a parent/child relationship.

The issue for your purpose would be making sure the continuation is
called from a position that is guaranteed to terminate the errant
thread: you might need to recognize a thread as a special case, wrap
it and (like an exception) catch/rethrow the continuation.


Or, if you don't care about sibling threads, just use exceptions which
always can be thrown upward out of child threads.

George

-- 
You received this message because you are subscribed to the Google Groups 
"Racket Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to racket-users+unsubscr...@googlegroups.com.
To view this discussion on the web visit 
https://groups.google.com/d/msgid/racket-users/2r6upftjom3aipt0g32nkcmikdc37je2t6%404ax.com.

Re: [racket-users] Re: Help implementing an early return macro

[jackh...@gmail.com]

Wow, these are a lot of great responses. First of all, *awesome* job Ryan. 
That implementation is exactly what I needed to figure out. I'm definitely 
starting there first.

> Are you looking for `let/ec`?

I'd forgotten about that one. That has the *syntax* I want. However my 
issue with continuation-based approaches isn't the syntax, or even the 
performance. It's the semantics. What if someone writes code like this?

(guarded-block
  (define x (random 10))
  (thread
(lambda ()
  (guard (even? x) else #false)
  ...)))

If I implemented guarded-block in terms of let/ec, then what does this code 
even *do*? I honestly don't know. It would probably run without error and 
do... something. I am extremely sure that regardless of what it did, it 
would be confusing and it wouldn't solve any problem I had. I just flat out 
don't want to allow this or any related nonsense, such as:

; Aliasing
(define return guard)

; Higher-order usage
(map guard (some-list ...))

; Capturing via closure
(guarded-block
  (define (check-foo x) (guard (foo? x) else #false))
  (check-foo ...)
  ...)

; Capturing via mutation
(set! previous-guard guard)

; Oh great, now I have to think about even more continuation jumps
(dynamic-wind
  (lambda () (guard ...))
  (lambda () ...)
  (lambda () ...))

There might be valid use cases for some of these, but I certainly don't 
understand those use cases well enough to commit to a semantics for them.

As for why not use condd, cond/else, or parendown: because they don't look 
right. Specifically, I often write code like this:

(define (f x y z)
  ... a few lines of checking preconditions ...
  ... a dozen or two lines of actual useful logic ...)

I don't want to indent the useful logic a bunch. That's the most important 
part of the function - the preconditions are minor things that you should 
be able to skim over and forget about. If my functions without 
preconditions look wildly different from my functions with preconditions, 
it becomes difficult to tell what the main focus of a function is. Not to 
mention that going from zero preconditions to one (and from one to zero) 
now introduces a bunch of busywork.

On naming: I like the symmetry between return-when / return-unless and when 
/ unless, but the problem is the word "return". If I call it a return 
statement, people will naturally expect this to work:

(define (first-owl animals)
  (for ([animal animals])
(return-when (owl? animal) animal)))

I don't want to have to explain why that doesn't work forty times.

The names guard-when and guard-unless are kind of ambiguous; does 
"guard-when" mean "when this condition is true, enter this guard block and 
escape" or does it mean "guard the function with this condition, if it 
fails enter the block and escape". Does "guard" mean "prevent this 
condition" or "ensure this condition"? Having two forms means you have to 
remember which one means which and figure out a way to keep them straight. 
I'd rather just have one form so there's only one way to do it, and you can 
teach yourself that "ah yes, guard always means 'make sure this is true'".

-- 
You received this message because you are subscribed to the Google Groups 
"Racket Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to racket-users+unsubscr...@googlegroups.com.
To view this discussion on the web visit 
https://groups.google.com/d/msgid/racket-users/3607a1a0-e303-4d9d-a62e-c958712aa2c8n%40googlegroups.com.

Re: [racket-users] Re: Help implementing an early return macro

[Dominik Pantůček]

Hi racketeers,

I would second this suggestion. Although it might look slightly
un-rackety at first sight, for certain types of code flow it does the
job really well.

And most importantly - I am using escape continuations in much wilder
setup (yes, futures) and it imposes no noticeable performance impact. It
is still on my TODO list to measure that empirically though. On CS that
is - on BC I didn't have time to test it yet.


Dominik

P.S.: Thanks go out to Matthew for pointing ECs out at RacketCon
informal discussions. Futures with generic CC come with absurdly huge
performance penalty even if they stay within the future thread ...

On 28. 10. 20 12:23, Alex Harsanyi wrote:
> Are you looking for `let/ec`?
> 
> (let/ec return
>   (define x (random 10))
>   (unless (even? x)
>     (log-info "x wasn't even, x = ~a" x)
>     (return -1))
>   (define y (random 10))
>   (unless (even? y)
>     (log-info "y wasn't even, y = ~a" y)
>     (return -1))
>   (+ x y))
> 
> Alex.
> 
> On Wednesday, October 28, 2020 at 6:54:44 PM UTC+8 jackh...@gmail.com wrote:
> 
> So I'm a little tired of writing code like this:
> 
> (define x ...)
> (cond
>   [(take-shortcut? x) (shortcut x)]
>   [else
>(define y (compute-y x))
>(cond
> [(take-other-shortcut? x y) (other-shortcut x y)]
> [else
>  (define z ...)
>  (cond ...)])])
> 
> That is, I have some logic and that logic occasionally checks for
> conditions that make the rest of the logic irrelevant, such as an
> empty or false input or something else that should trigger an early
> exit. Each check like this requires me to write a |cond| whose
> |else| clause wraps the remainder of the body, leading to an awkward
> nesting of |cond| forms. I don't have this issue when the early
> exits involve raising exceptions: in those cases I can just use
> |when| and |unless| like so:
> 
> (define x ...)
> (unless (passes-check? x) (raise ...))
> (define y ...)
> (unless (passes-other-check? x y) (raise ...))
> (define z ...)
> ...
> 
> I'm aware of a few macros in the racket ecosystem that try to solve
> this problem. For example, Jay wrote a blog post
>  that
> creates a |condd| form that's like |cond| but allows embedded
> definitions using a |#:do| keyword. I've also seen various
> approaches that use escape continuations to implement the early
> exit. There's drawbacks I'm not happy about however:
> 
>   *
> 
> For |cond|-like macros that allow embedded definitions, it looks
> too different from regular straight-line Racket code. I like my
> function bodies to be a sequence of definitions and expressions,
> with minimal nesting, just like the |when| and |unless| version
> above. I don't have to use a keyword or extra parentheses to
> signal whether a form is a definition or a |when| / |unless|
> check in error-raising code, why should I have to do that in
> code that uses early returns?
> 
>   *
> 
> Continuation-based solutions impose a nontrivial performance
> penalty and have complex semantics. I don't like that the
> generated code behaves differently from the |cond| tree I would
> normally write. What happens if I stick an early exit inside a
> lambda? Or a thread? What if I set up a continuation barrier?
> Does that matter? I don't know and I don't want to think about
> that just to write what would be a simple |if (condition) {
> return ... }| block in other languages.
> 
> So I wrote a basic macro for this and I have some questions about
> how to make it more robust. The macro is called |guarded-block| and
> it looks like this:
> 
> (guarded-block
>   (define x (random 10))
>   (guard (even? x) else
> (log-info "x wasn't even, x = ~a" x)
> -1)
>   (define y (random 10))
>   (guard (even? y) else
> (log-info "y wasn't even, y = ~a" y)
> -1)
>   (+ x y))
> 
> Each |guard| clause contains a condition that must be true for
> evaluation to proceed, and if it isn't true the block takes the else
> branch and finishes. So the above would expand into this:
> 
> (block
>   (define x (random 10))
>   (cond
> [(not (even? x))
>  (log-info "x wasn't even, x = ~a" x)
>  -1]
> [else
>  (define y (random 10))
>  (cond
>[(not (even? y))
> (log-info "y wasn't even, y = ~a" y)
> -1]
>[else (+ x y)])]))
> 
> This part I got working pretty easily. Where I hit problems, and
> where I'd like some help, is trying to extend this to support two
> important features:
> 
>   *
> 
> I should be able to define macros that /expand/ into |guard|
> 

[racket-users] Re: Help implementing an early return macro

[Alex Harsanyi]

Are you looking for `let/ec`?

(let/ec return
  (define x (random 10))
  (unless (even? x)
(log-info "x wasn't even, x = ~a" x)
(return -1))
  (define y (random 10))
  (unless (even? y)
(log-info "y wasn't even, y = ~a" y)
(return -1))
  (+ x y))

Alex.

On Wednesday, October 28, 2020 at 6:54:44 PM UTC+8 jackh...@gmail.com wrote:

> So I'm a little tired of writing code like this:
>
> (define x ...)
> (cond
>   [(take-shortcut? x) (shortcut x)]
>   [else
>(define y (compute-y x))
>(cond
> [(take-other-shortcut? x y) (other-shortcut x y)]
> [else
>  (define z ...)
>  (cond ...)])])
>
> That is, I have some logic and that logic occasionally checks for 
> conditions that make the rest of the logic irrelevant, such as an empty or 
> false input or something else that should trigger an early exit. Each check 
> like this requires me to write a cond whose else clause wraps the 
> remainder of the body, leading to an awkward nesting of cond forms. I 
> don't have this issue when the early exits involve raising exceptions: in 
> those cases I can just use when and unless like so:
>
> (define x ...)
> (unless (passes-check? x) (raise ...))
> (define y ...)
> (unless (passes-other-check? x y) (raise ...))
> (define z ...)
> ...
>
> I'm aware of a few macros in the racket ecosystem that try to solve this 
> problem. For example, Jay wrote a blog post 
>  that creates a 
> condd form that's like cond but allows embedded definitions using a #:do 
> keyword. I've also seen various approaches that use escape continuations to 
> implement the early exit. There's drawbacks I'm not happy about however:
>
>- 
>
>For cond-like macros that allow embedded definitions, it looks too 
>different from regular straight-line Racket code. I like my function 
> bodies 
>to be a sequence of definitions and expressions, with minimal nesting, 
> just 
>like the when and unless version above. I don't have to use a keyword 
>or extra parentheses to signal whether a form is a definition or a when 
>/ unless check in error-raising code, why should I have to do that in 
>code that uses early returns?
>- 
>
>Continuation-based solutions impose a nontrivial performance penalty 
>and have complex semantics. I don't like that the generated code behaves 
>differently from the cond tree I would normally write. What happens if 
>I stick an early exit inside a lambda? Or a thread? What if I set up a 
>continuation barrier? Does that matter? I don't know and I don't want to 
>think about that just to write what would be a simple if (condition) { 
>return ... } block in other languages.
>
> So I wrote a basic macro for this and I have some questions about how to 
> make it more robust. The macro is called guarded-block and it looks like 
> this:
>
> (guarded-block
>   (define x (random 10))
>   (guard (even? x) else
> (log-info "x wasn't even, x = ~a" x)
> -1)
>   (define y (random 10))
>   (guard (even? y) else
> (log-info "y wasn't even, y = ~a" y)
> -1)
>   (+ x y))
>
> Each guard clause contains a condition that must be true for evaluation 
> to proceed, and if it isn't true the block takes the else branch and 
> finishes. So the above would expand into this:
>
> (block
>   (define x (random 10))
>   (cond
> [(not (even? x))
>  (log-info "x wasn't even, x = ~a" x)
>  -1]
> [else
>  (define y (random 10))
>  (cond
>[(not (even? y))
> (log-info "y wasn't even, y = ~a" y)
> -1]
>[else (+ x y)])]))
>
> This part I got working pretty easily. Where I hit problems, and where I'd 
> like some help, is trying to extend this to support two important features:
>
>- 
>
>I should be able to define macros that *expand* into guard clauses. 
>This is important because I want to implement a (guard-match  
> else  ...) form that's like match-define 
>but with an early exit if the pattern match fails. I'd also really like to 
>add a simple (guard-define   else 
> ...) form that expects option-expression to produce an 
>option 
>
> 
>  
>(a value that is either (present v) or absent) and tries to unwrap it, 
>like the guard let construct in Swift 
>.
>- 
>
>Begin splicing. The begin form should splice guard statements into the 
>surrounding body. This is really an offshoot of the first requirement, 
>since implementing macros that expand to guard can involve expanding 
>into code like (begin (define some-temp-value ...) (guard ...) (define 
>some-result ...)).
>
> Having been around the Racket macro block before, I know I need to do some 
> kind of partial expansion here. But honestly I can't