I'm open to what you're saying, and I think that optimizing away renaming when it doesn't need to be done would go a long way toward reconciling the tradeoff. I'll have more suggestions once we've actually started using the release with lexical scoping (can't do it yet because of that symbol-macro bug I reported).
Dan
On Thu, May 14, 2009 at 6:48 PM, Vladimir Sedach vsedach@gmail.com wrote:
The advantages of this approach outweigh the disadvantages by such a wide margin that there's no question in my mind. It's also not very hard to optimize away the renaming if it does not need to be done (which I guess puts the burden of proof on me). I just pushed a patch that does not do gensym renaming if the variable in question has not been bound in any enclosing lexical context, which deals with the majority of the use cases of let.
Another heuristic that could be applied is to see if the variable is referenced in any statements that execute after the binding and are not captured by any closures in their lexical context, but that requires a lot more hairy analysis because of things like scoping in 'for' loops. To be clear, this scenario most commonly occurs in code like:
(defun foo (x) (let ((x (do-something x))) body))
Given the amount of hairy code vs potential readability improvement ('x1' vs 'x') I'm not sure this case is common enough to be worth addressing.
Vladimir
On Wed, May 6, 2009 at 1:54 PM, Daniel Gackle danielgackle@gmail.com wrote:
Whoa. I've just had a chance to look at this lexical scoping business.
This
is a major change, and needs a vigorous discussion. I'm not convinced
that
the benefits outweigh the costs.
All local variables in our JS are now mangled by having a gensym-style suffix added to them. This considerably degrades the readability of the
JS.
One of the PS design goals has always been to balance nice abstractions
at
the source level with readability, ease of debugging, and efficiency in
the
generated code. This is a must-have because PS is not a compiler where
you
can just forget about the object code; those of us using PS on real
projects
have our hands immersed in the generated JS every day. I spend a great
deal
of time in Firebug, debugging and tracing and testing our code in the browser, and anything that degrades readability is a real drag.
Against that cost, what is the benefit? I agree that it is a cool hack
that
gets lexical scoping into PS without resorting to nasty things like try blocks or nested lambdas. But I don't think that's a sufficient argument
for
putting it into PS. What's the practical benefit? Has anyone expressed a need for this feature? We've written many thousands of lines of PS code
and
not run into a need for it that I can recall.
One option that won't work for us is to replace all usages of LET with
VAR.
A lot of our code is written in a subset of CL that is also emitted as
PS.
We can't throw out LET from that code. I suppose I could redefine PS's
LET
to just prepend a bunch of VAR statements to the block. It would be an unfortunate workaround, though.
I'm not saying I'm adamantly opposed to this change, just that it's a
major
one that has some drawbacks. I think it ought to be discussed and
debated.
The pros and cons are not all obvious, and need to be flushed out.
I'll post separately about a bug that seems to have been introduced. If
that
bug (and a few others) can be fixed, I'm willing to give this latest PS a try for a few weeks, as long as there's an alternative (short of
regressing
to an old version) to fall back on.
Dan
This is a major change. I think there needs to be a vigorous discussion about it. I'm not at all convinced that the benefits outweigh the costs,
and
mangling, debugging, readability obfuscation/minification
On Sun, May 3, 2009 at 3:40 PM, Vladimir Sedach vsedach@gmail.com
wrote:
Hi,
I just pushed a patch into the repository that may break existing code. Here is the patch description:
Implemented LET and LET* by variable renaming, which provides the correct scoping semantics, and simplifies both the producing and the produced code.
Removed the "." and "[]" name-mangling conventions from symbol-to-js-string. Any code that uses symbols such as
"foo.bar[baz]"
will now issue a warning, and needs to be rewritten to use standard Lisp accessors. This is needed for variable renaming to work, and is an extension of the patch that eliminated the ".method"
method-calling
convention.
Thanks to Daniel Gackle and Dough Hoyte for opening my eyes to this technique, which was right in front of me all along.
This is a win-win situation for everyone involved (yes, even if it breaks your code, think of it as an opportunity for refactoring).
With this patch Parenscript is finally rid of the evil symbol-abuse conventions, which IMO is the only thing from the original design of Parenscript that can in retrospect be called a mistake.
Let me know what you guys think.
Thank you, Vladimir
parenscript-devel mailing list parenscript-devel@common-lisp.net http://common-lisp.net/cgi-bin/mailman/listinfo/parenscript-devel
parenscript-devel mailing list parenscript-devel@common-lisp.net http://common-lisp.net/cgi-bin/mailman/listinfo/parenscript-devel
parenscript-devel mailing list parenscript-devel@common-lisp.net http://common-lisp.net/cgi-bin/mailman/listinfo/parenscript-devel