On May 12, 2011, at 7:42 AM, Claus Reinke wrote:
ECMAScript has a large set of problems. ..
- fixing ECMAScript's oddities would be worth a language revision,
even without adding anything new
This is a mistake. The Web does not permit "stop the world and fix all
known bugs". Neither users nor competing browser or server-side software
vendors will stop.
Ideals are like stars - out of reach, but good for orientation.
Bodies were heaped by the millions over the last few centuries in the name
of ideals. I detest _a priori_ systems that ignore empircal results. I
don't buy "ideal" without "real". Your message is woefully lacking in
"real".
The fixes alone would be worth a revision, and a competitive
advantage (for language implementers and programmers,
though the both would be limited by code having to work
with other browsers). That ideal does not mean that language
changes should not be introduced incrementally (refactoring
beats rewriting). It only means that fixes can be as valuable
as new features.
This is all vague sermonizing. Please stop.
There's a dream syntax Doug can glimpse (maybe I can too; working on it
still) that might be called a "fix". It relieves us of semicolons and ASI
without creating runtime migration traps.
Ok, show it. Or show a way to find it, a concrete method. The endless
ideal-mongering and vagueness don't cut it, and they are actually adding
noise and making trouble.
If we keep this up, we will have neither fixes (including the ones *in
hand*, e.g. lexical scoping) nor new features in ES.next.
We do not all even agree on what the problems are. The ones that we
agree on, we are successfully working on (modules, e.g.). Classes have
had a rough time because we don't even agree on premises or problems,
never mind conclusions or solutions.
Douglas will speak for himself, I'm sure, but my own interpretation of
his remarks was that most current work
seems focused on extensions to the house, while there
are some concerns about the foundation (not limited to
functions).
Vague words, and where testable, precisely wrong. Lexical scope all the
way up, modules, let, const, some of the object initialiser extensions,
and (I argue) paren-free and the for-in reform it enables -- these are all
fixes.
Another way to look at it: independent of the concrete
problems which you are working on, there seem to be
some meta-level problems that keep coming up and
get in the way of designing solutions. If that happens
often enough, it may be worth stepping back and solve
the meta-level problem first (no matter how horribly
difficult to handle that elephant looks;-).
For me to keep parlaying here, you need to be specific or else stop
injecting so many generalizations which are either useless, or actually
false when applied to specifics I've listed above.
I'm working on better syntax, to present at the next TC39 meeting. I
don't think "perfect" is an option in this life. I'm aiming at "better".
Better is the road to perfect.
No, there is no perfect. Get this right, or you are in dreamland (and I am
elsewhere).
More concise function syntax is important, as long as the deeper issues
get resolved, too.
What deeper issue beyond |this| binding (lexical, dynamic, soft, etc.)
do you mean, precisely?
I don't yet have a complete enough overview for detailed language
criticism, but I've been posting my concerns here as I encountered them.
It would be good to have a process for registering information and
suggestions with the committee and get feedback about the official
reactions to such suggestions (an es-discuss bug tracker perhaps,
separate from the spec bug tracker?).
At this stage of my exploration, the main items seem to be:
- expressions vs statements
this is well-known and would be hard to fix (too big a change for
javascript), beyond simplifying workarounds (such as parameterless
functions as block "values");
Ok, so you say this is not realistically solvable. Why bring it up?
Next:
- grammar
too many details are hard-coded here: that makes the language
complex and inflexible for the language design process (syntax
should not get in the way as often as it does here), and
limits language expressiveness for programmers (especially for
abstraction providers); more of the changes that occupy
TC39 should be possible at the level of libraries, not language;
You're arguing we can't improve the syntax of the language, even by
extension? False: let, const (already in browsers in a form that can be
fixed), lexical scope to get the global object off of the scope chain
(this is an important security fix), and modules.
Also: libraries are many and evolving, TC39 design-by-committee is about
the worst way to invent and evolve a library. We're de-facto standardizing
(ES5 did the Array extras, which still have some flaws we cannot fix).
- program equivalences
what equivalences hold for js code? Language implementers,
tool builders, and programmers depend on being able to reason about
code (optimizations, refactorings, code reading,
maintenance, bug fixing);
language-level equivalences are the most practical side of
semantic tools: no need to reason about abstract machine
details, or machine code, or mathematical constructs, just "this
high-level code fragment is equivalent to that one"/
"we can replace this with that and nothing will go wrong";
obviously, reflection breaks all equivalences (which is one reason
why mirrors and the like have been introduced: to delimit the code
affected by reflection); but even without
reflection interference, it seems difficult to find valid code
equivalences for javascript;
(It's not that bad in practice once you [the refactoring entity] buy into
having to lex and parse everything. This seems a formal objection.)
the more each instance of code has to be understood on its
own -rather than by relation/equivalence/simplification, the more
complex the language gets;
it is great to have a detailed language spec (really! not all
languages have one), but if we have to keep looking there, to
understand why obvious code equivalences fail in javascript, the
language becomes arcane (which is good for consultant experts, bad for
everyone else);
This is long-winded and vague still (also badly formatted). What is your
specific, actionable point?
We can't lose statements with control effects without a big compatibility
break, and you ruled out new syntax above. Are you arguing against
changing JS, and for inventing a wholly new language?
Just yesterday I re-raised the {|a,b| a+b} Ruby/Smalltalk-inspired
block-as-lambda idea recently. It has the advantage of being concrete. It
preserves certain program equivalences at the price of creating new
runtime error conditions (return from an outer function call that has
already returned, e.g.). Please address it specifically, or make a
competitive alternative proposal.
There are smaller items (some are instances of the above):
- objects vs blocks; grammar ambiguities;
This is a glitch JS hackers I've surveyed informally simply learn and get
past. They don't even complain about it much, compared to lack-of-ASI
(lack of significant newline) traps. I've talked to many over the years.
Anyway, it is not going to be fixed without a new grammar, which you seem
to have ruled out above.
- ASI coupling to combination of linebreak, error, restricted
productions serves neither full-ASI nor no-ASI proponents,
and complicates language extensions;
Yes, this is true and we've had proposals, including Mark's concrete
proposal for a better ASI and Harmony parsing both ways, declaring an
early error on disagreement. Unfortunately this requires backtracking and
error recovery in a "forked parser", so it looks too complicated for
implementors *and* users, but it was a concrete proposal in the right
direction. Please do likewise and stop rehashing the problem statement.
http://www.mail-archive.com/[email protected]/msg05606.html
- mixing numeric indices and string properties (every object
is an array, only Array instances have proper array methods);
dynamic typing is much less problematic without implicit
conversions muddying the waters;
I proposed tuples. They got a mixed reaction at the last tc39 meeting, and
they are not likely to get into ES.next at this point. Do you have
anything to add to this specific proposal, or another alternative specific
proposal to make?
http://wiki.ecmascript.org/doku.php?id=strawman:tuples
- var initializers shadow bindings in their rhs (similar to 'this'
and 'arguments' shadowing, but worse through 'undefined');
do we really need a different solution to each instance of
shadowing problems? differences make for language
complexity, extracting common solutions to common
problems simplifies understanding;
This is a topic for let and const in blocks, which is still being debated
on-and-off. IIRC most of TC39 agreed to make let and const start a new
implicit scope, so an outer name could be used in the initialiser.
Again, we can't change var unless you want us to make incompatible
changes, which you seemed to advise against above. Which is it?
- argument lists as arrays (instead of arrays as arguments)
[this seems to be worked on];
It *is* being worked on, you should know that -- rest and spread are in
harmony -- yet you rehash and inject noise again here. Why? To channel
Nancy Kerrigan: WHYYYYY???? :-|
http://wiki.ecmascript.org/doku.php?id=harmony:rest_parameters
http://wiki.ecmascript.org/doku.php?id=harmony:spread
- ephemeral nature of References; leads to syntactic junk in
the language (syntactically valid code that predictably leads
to runtime errors); also limits object methods as first-class
values (they lose their base reference), breaks a dozen
of "obvious" code equivalences (is this also behind direct
vs indirect eval? I haven't checked yet);
This is not only a minor issue, truly small potatoes -- it is required for
web compatibility due to VBScript-tainted JScript of this kind:
d.items(i) = j;
That kind of crap exists in IE-only branches of JS on the web, so it must
parse.
Let's focus on real problems affecting real users.
- hardcoding of binary ops/precedences/associativity in grammar; these
are much too variable to be written in stone, or be discussed by a
language committee; for javascript to be adapted quickly to new
domains, these should be in the hands of library authors and other
abstraction providers;
much too valuable a tool to ignore; if combined with
lightweight function definition and application syntax,
it helps in defining readable domain-specific languages;
commonalities in embedded domain-specific languages
feed back into general purpose control abstractions, which
again should be provided in libraries (moving more quickly
than language committees);
Sorry, we're not turning JS into Haskell or making the syntax,
particularly the operators, have adjustable precedence and associativity.
This is not something developers or implementors want. It has never been
on the agenda, even for value types/proxies. You are barking up the wrong
tree here.
Perhaps there were other items I don't recall right now (with a tracker,
I could just look up my tickets, you could assign priorities, milestones,
and committee members for those items you would want to adopt, and we'd
all know where our suggestions stand).
I'll keep it short by stopping here. Precision in stating problems would
be appreciated.
I hope this helps?
Not really.
I'm being rough on you here because we have a serious (serous enough, for
a discussion list I care about not going in ideal-mongering circles or
filling with useless noise and posturing) problem. We have real proposals
for ES.next on the table. They aren't all necessarily going to make it,
but trying to reset the discussion from them to vague ideals or non-issues
like programmable operator precedence does a disservice to our common
effort here.
We need to get back to discussing the live strawman proposals, and
refining the proposals in harmony but not quite spec'ed fully. Will you
help?
/be
