Lazy evaluation

Alex Kodat alexkodat at gmail.com
Tue Sep 12 23:38:43 UTC 2017


I think to the degree that almost anyone would care, the public and private key proposal would address most of these issues. The private keys would not be affected by object shape so would have rip-roaring good performance. You'd simply need to provide getter functions in the prototype that check for undefined (or whatever) in the private variable, if not undefined return value, otherwise calculate value, set private variable, and return the value. Unless you're trying to calculate pi to a million decimal places this should be pretty much good enough for anything.

And it gives you the encapsulation you get in many other programming languages such as Java and C++. Public JS APIs would end up looking a lot like these for many applications -- a bunch of private variables and public functions including getters and setters. 

Basically, most of our code already looks like this but we use closures rather than private variable to accomplish this. And yes, there is a dark side to closures, specifically the challenge in implementing closed function that don't gobble up a lot of memory. In V8 if I construct an object with say 20 function properties that's 20 JSFunction objects V8 has to create, each one being on the order of 72 bytes per function. It's not too horrible if we don't have a lot of these and they have relatively long lives but problematic for lots of instantiation of objects with lots of functions. The private key proposal would address this issue. 

----
Alex Kodat

From: es-discuss [mailto:es-discuss-bounces at mozilla.org] On Behalf Of Matthew Robb
Sent: Tuesday, September 12, 2017 4:49 PM
To: Steve Fink <sphink at gmail.com>
Cc: es-discuss at mozilla.org >> es-discuss <es-discuss at mozilla.org>
Subject: Re: Lazy evaluation

I think it would be nice to be able to define an own data property on an object that while defined delegated up the prototype chain. This would allow a getter in the proto to lazily assign to the own property without triggering the property setter. This is even more nice when combined with super as you could conceivably hit the setter which itself could assign to the own property of the current instance.

Is that all too complicated? It seems perfect if only a little complicated.

On Sep 12, 2017 5:39 PM, "Steve Fink" <mailto:sphink at gmail.com> wrote:
My intent was only to respond to the performance analysis, specifically the implication that the only performance cost is in building the new hidden class. That is not the case; everything that touches those objects is affected as well.

Whether or not it's still the right way to accomplish what you're after, I wasn't venturing an opinion. I could probably come up with a benchmark showing that your WeakMap approach can be faster -- eg by only accessing the property once, but feeding the old and new versions of the object into code that executes many many many times (doing something that never looks at that property, but is now slightly slower because it isn't monomorphic). But I suspect that for practical usage, redefining the property *is* faster than a WeakMap.

If I were to look beyond for other solutions for your problem, then I'm just speculating. Can decorators populate multiple properties once the expensive work is done?

I really want to tell the VM what's going on. I guess if it knew that accessing a getter property would convert it into a value property, and that it was doing something that would access the getter, then it could know to use the outgoing shape instead of the incoming shape. If only it knew that the getter was pure... but that way lies madness.

Given that most code that would slow down would also trigger the lazy defineProperty(), it's really not going to be that much of an issue. Any access after the first will see a single shape.

meh. Just take the perf hit, with awareness that you may be triggering slight slowdowns in all users of that object. Or you might not. I doubt it'll be that big, since you'll probably just end up with an inline cache for both shapes and there won't be all that much to optimize based on knowing a single shape.

Oh, and I think I was wrong about property enumeration order. The properties already existed, so defineProperty shouldn't modify the order IIUC. (I am awful with language semantics.)

On 9/11/17 2:48 PM, Andrea Giammarchi wrote:
Steve it's not solved in any other way. Even if you use a WeakMap with an object, you gonna lazy attach properties to that object. 

I honestly would like to see alternatives, if any, 'cause so far there is a benchmark and it proves already lazy property assignment is around 4x faster.

So, it's easy to say "it's not the best approach" but apparently hard to prove that's the case?

Looking forward to see better alternatives.


On Mon, Sep 11, 2017 at 8:15 PM, Steve Fink <mailto:sphink at gmail.com> wrote:
On 9/11/17 5:36 AM, Matthew Robb wrote:
> I think it's irrelevant if internally VMs are not too happy. VMs are there to solve our problems, not vice-versa ;-)
​
This ^​ is very important for everyone to get on board with. Regardless the cost should be negligible as the shape is only changing at the point of delayed init. This will cause, for example V8, to deop the object and have to build a new hidden class but only the one time. I guess it would potentially be interesting to support an own property that when undefined would delegate up the proto chain.

(I don't know, but) I would expect it to be worse than this. The shape is changing at the point of delayed init, which means that if an engine is associating the possible set of shapes with the constructor (or some other form of allocation site + mandatory initialization), then that site will produce multiple shapes. All code using such objects, if they ever see both shapes, will have to handle them both. Even worse, if you have several of these delayed init properties and you end up lazily initializing them in different orders (which seems relatively easy to do), then the internal slot offsets will vary.

You don't need to bend over backwards to make things easy for the VMs, but you don't want to be mean to them either. :-)

Not to mention that the observable property iteration order will vary.

On Mon, Sep 11, 2017 at 7:09 AM, Andrea Giammarchi <mailto:andrea.giammarchi at gmail.com> wrote:

Hi Peter. 

Unless you have a faster way to do lazy property assignment, I think it's irrelevant if internally VMs are not too happy. VMs are there to solve our problems, not vice-versa ;-)

Regards



On Mon, Sep 11, 2017 at 11:54 AM, peter miller <mailto:fuchsia.groan at virgin.net> wrote:
Hi Andrea,
```
class CaseLazy {
  get bar() {
    var value = Math.random();
    Object.defineProperty(this, 'bar', {value});
    return value;
  }
}
```

Doesn't this count as redefining the shape of the object? Or are the compilers fine with it?


_______________________________________________
es-discuss mailing list
mailto:es-discuss at mozilla.org
https://mail.mozilla.org/listinfo/es-discuss



_______________________________________________
es-discuss mailing list
mailto:es-discuss at mozilla.org
https://mail.mozilla.org/listinfo/es-discuss



More information about the es-discuss mailing list