GNU bug report logs - #24171
25.1; Bytecode returns nil instead of expected closure

Previous Next

Full log

Message #26 received at 24171 <at> debbugs.gnu.org (full text, mbox):

From: Alex Vong <alexvong1995 <at> gmail.com>
To: Stefan Monnier <monnier <at> iro.umontreal.ca>
Cc: Michael Heerdegen <michael_heerdegen <at> web.de>, 24171 <at> debbugs.gnu.org
Subject: Re: bug#24171: 25.1; Bytecode returns nil instead of expected closure
Date: Mon, 08 Aug 2016 18:38:53 +0800

Hi,

Stefan Monnier <monnier <at> iro.umontreal.ca> writes:

> You can test the problem with:
>
>    M-: (cconv-closure-convert '(let ((x 1)) (let ((x 2) (f (function (lambda (y) (+ y x))))) (funcall f x))))
>
> where you'll see that the lambda-lifting used by cconv.el is too naive
> and uses `x' to refer to the outer variable without noticing that that
> variable is shadowed by the inner `x'.
>
> The patch below should fix it and is the best I can come up with so far.
>
> Can you confirm that it fixes the original problem?
>
Yes, this fixes the original problem.
(https://lists.gnu.org/archive/html/help-gnu-emacs/2016-08/msg00038.html)
The test is performed on master branch with patch applied.

> The bug was filed against 25.1, so I have (very lightly) tested the
> patch against the emacs-25 branch, but since this bug dates back to
> Emacs-24.1, I think there's no hurry to fix it.
>
> IOW I intend to install it into master.  Please holler if you think it
> deserves to be on emacs-25.
>
I have no problem with this.

>
>         Stefan
>
>
>
> diff --git a/lisp/emacs-lisp/cconv.el b/lisp/emacs-lisp/cconv.el
> index 50b1fe3..2d68066 100644
> --- a/lisp/emacs-lisp/cconv.el
> +++ b/lisp/emacs-lisp/cconv.el
> @@ -253,6 +253,32 @@ Returns a form where all lambdas don't have any free variables."
>        `(internal-make-closure
>          ,args ,envector ,docstring . ,body-new)))))
>  
> +(defun cconv--remap-llv (new-env var closedsym)
> +  ;; In a case such as:
> +  ;;   (let* ((fun (lambda (x) (+ x y))) (y 1)) (funcall fun 1))
> +  ;; A naive lambda-lifting would return
> +  ;;   (let* ((fun (lambda (y x) (+ x y))) (y 1)) (funcall fun y 1))
> +  ;; Where the external `y' is mistakenly captured by the inner one.
> +  ;; So when we detect that case, we rewrite it to:
> +  ;;   (let* ((closed-y y) (fun (lambda (y x) (+ x y))) (y 1))
> +  ;;     (funcall fun closed-y 1))
> +  ;; We do that even if there's no `funcall' that uses `fun' in the scope
> +  ;; where `y' is shadowed by another variable because, to treat
> +  ;; this case better, we'd need to traverse the tree one more time to
> +  ;; collect this data, and I think that it's not worth it.
> +(mapcar (lambda (mapping)
> +            (if (not (eq (cadr mapping) 'apply-partially))
> +                mapping
> +              (cl-assert (eq (car mapping) (nth 2 mapping)))
> +              `(,(car mapping)
> +                apply-partially
> +                ,(car mapping)
> +                ,@(mapcar (lambda (arg)
> +                            (if (eq var arg)
> +                                closedsym arg))
> +                          (nthcdr 3 mapping)))))
> +          new-env))
> +
>  (defun cconv-convert (form env extend)
>    ;; This function actually rewrites the tree.
>    "Return FORM with all its lambdas changed so they are closed.
> @@ -350,34 +376,13 @@ places where they originally did not directly appear."
>                     (if (assq var new-env) (push `(,var) new-env))
>                     (cconv-convert value env extend)))))
>  
> -           ;; The piece of code below letbinds free variables of a λ-lifted
> -           ;; function if they are redefined in this let, example:
> -           ;;   (let* ((fun (lambda (x) (+ x y))) (y 1)) (funcall fun 1))
> -           ;; Here we can not pass y as parameter because it is redefined.
> -           ;; So we add a (closed-y y) declaration.  We do that even if the
> -           ;; function is not used inside this let(*).  The reason why we
> -           ;; ignore this case is that we can't "look forward" to see if the
> -           ;; function is called there or not.  To treat this case better we'd
> -           ;; need to traverse the tree one more time to collect this data, and
> -           ;; I think that it's not worth it.
> -           (when (memq var new-extend)
> -             (let ((closedsym
> -                    (make-symbol (concat "closed-" (symbol-name var)))))
> -               (setq new-env
> -                     (mapcar (lambda (mapping)
> -                               (if (not (eq (cadr mapping) 'apply-partially))
> -                                   mapping
> -                                 (cl-assert (eq (car mapping) (nth 2 mapping)))
> -                                 `(,(car mapping)
> -                                   apply-partially
> -                                   ,(car mapping)
> -                                   ,@(mapcar (lambda (arg)
> -                                               (if (eq var arg)
> -                                                   closedsym arg))
> -                                             (nthcdr 3 mapping)))))
> -                             new-env))
> -               (setq new-extend (remq var new-extend))
> -               (push closedsym new-extend)
> +           (when (and (eq letsym 'let*) (memq var new-extend))
> +             ;; One of the lambda-lifted vars is shadowed, so add
> +             ;; a reference to the outside binding and arrange to use
> +             ;; that reference.
> +             (let ((closedsym (make-symbol (format "closed-%s" var))))
> +               (setq new-env (cconv--remap-llv new-env var closedsym))
> +               (setq new-extend (cons closedsym (remq var new-extend)))
>                 (push `(,closedsym ,var) binders-new)))
>  
>             ;; We push the element after redefined free variables are
> @@ -390,6 +395,21 @@ places where they originally did not directly appear."
>               (setq extend new-extend))
>             ))                           ; end of dolist over binders
>  
> +       (when (not (eq letsym 'let*))
> +         ;; We can't do the cconv--remap-llv at the same place for let and
> +         ;; let* because in the case of `let', the shadowing may occur
> +         ;; before we know that the var will be in `new-extend' (bug#24171).
> +         (dolist (binder binders-new)
> +           (when (memq (car-safe binder) new-extend)
> +             ;; One of the lambda-lifted vars is shadowed, so add
> +             ;; a reference to the outside binding and arrange to use
> +             ;; that reference.
> +             (let* ((var (car-safe binder))
> +                    (closedsym (make-symbol (format "closed-%s" var))))
> +               (setq new-env (cconv--remap-llv new-env var closedsym))
> +               (setq new-extend (cons closedsym (remq var new-extend)))
> +               (push `(,closedsym ,var) binders-new)))))
> +
>         `(,letsym ,(nreverse binders-new)
>                   . ,(mapcar (lambda (form)
>                                (cconv-convert

Thanks,
Alex

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.