Understanding V8
s3.mrale.ph
s3.mrale.ph
I'll add a note on the first slide as soon as I get home.
When using the same website both for the actual presentation and for disseminating the slides afterwards, it might make sense to have some kind of navigation visible by default. You could then have a "presentation mode" shortcut key that hides the nav element when you're showing the slides on the big screen. This way, the same site would serve both audiences.
In hot code
Avoid const
What's the relevance of the "GC cost" slide to the two code-change slides that follow it? Does manually converting tail-recursion have a significant impact on the number of long-lived objects?
Are you using cons-strings somewhere? (I was thinking of something like Erlang IO lists. But maybe you just mean the for (...) { x.push(s); } return x.join('') trick?)
Why is "indexing" relevant in "don't mix indexing and concatenation"? It looks to me like the problem is building up a big string by repeated concatenation, no? Not indexing.
What's ~ToNumber? Is that a destructor? Or is it just the conversion applied by Number()?
What did you use the typed arrays for?
Where do you get HiddenClass from? It's not in the global namespace of my copy of node.js by default.
Did you have an example where you were able to convert a polymorphic call site to monomorphic? Also, does V8 use PICs at all, or does it just have a monomorphic inline cache like in Deutsch & Schiffman?
What's the connection between dictionary mode and the code on the dictionary-mode slide?
What causes the arguments object to materialize?
On "Two types of variables", it appears that retaining g will result in v3 not being garbage collected, potentially resulting in retaining an arbitrarily large amount of garbage. Does that really happen?
2. I am not sure about this question. V8 uses cons-strings internally: so it is important to understand that c = a + b does not result (at least not immediately) in an allocation of a new sequential string and copying characters from a and b to it. That means that concatenating many strings does not cost much (especially if they are large) at the moment of concatenation. Actual concatenation happens later, when you start indexing into this string: V8 will flatten it (convert from cons-string form into sequential form).
3. See 2. If you intermix indexing and concatenation you basically force s to oscillate between cons and sequential forms: cons-string is created, then flattened on indexing, and result is concatenated with the string forming a cons-string which gets flattened... Resulting performance is bad because cons-strings actually become pure overhead (you waste more on their creation than you benefit from them) in this cycle of concatenation and flattening.
4. '*' as a prefix means optimized function and '~' as a prefix means non-optimized function.
5. Everything starting from 'Understanding Numbers' slide is a collection of facts unrelated to the first part of the talk about profiling. To get the idea of how typed arrays affect performance you can read my blog post: http://blog.mrale.ph/post/5436474765/dangers-of-cross-langua...
6. HiddenClass slides try to illustrate the idea of inline caching (http://en.wikipedia.org/wiki/Inline_caching) without using assembly. It's pseudo-code. HiddenClass is a structure that describes layout of an object. Slides #42-#50 illustrate how VM builds hidden classes hierarchy while executing the code. Slides #51-#52 try to illustrate how you can utilize hidden classes when you compile code. There are couple presentations that discuss hidden classes at the lower and more detailed level: for example Mads Ager's talk at Google IO 2009: http://youtu.be/FrufJFBSoQY
7. No, there were no example (it's impossible to fit that much into a single 40m talk). In V8 ICs have a special megamorphic state in addition to monomorphic state. Megamorphic stubs are not specialized for types IC had seen (unlike monomorphic stubs), instead they rely on a cache of monomorphic stubs (indexed by hidden class of the receiver).
8. Those are functions that force object from fast-mode to dictionary-mode.
9. In non optimized code it is materialized when used in the code. In optimized it is never materialized (but as the slide says: if you "misuse" arguments in a function it will not be optimized).
10. I have not seen any real world code that suffers from this problem. But people ask about scopes and closed-variables quite often that is why it is included.
Ohhh. I had no idea. That makes the next point a lot more understandable.
> '~' as a prefix means non-optimized function.
Where does ToNumber come from? It sounds like the kind of thing that ought to get optimized!
> Those are functions that force object from fast-mode to dictionary-mode
You mean, if you use delete, getters, setters, seal, or freeze, then the object will use dictionary-mode thereafter under he covers?
> if you "misuse" arguments in a function it will not be optimized).
What constitutes "misuse"? Anything other than .length, [i], and applying a function to it, I guess (e.g. passing it to some other function such as Array.prototype.slice)?
> I have not seen any real world code that suffers from this problem.
So it exists theoretically? The benefit of being implemented this way would be that creating closures is a lot faster and they use less space.
ToNumber comes from Number(l). I can't see how it can be optimized (l is a string so parsing is unavoidable).
> You mean, if you use delete, getters, setters, seal, or freeze, then the object will use dictionary-mode thereafter under he covers?
Yes.
> What constitutes "misuse"?
Your understanding of misuse is correct.
> So it exists theoretically?
Yes.
If people will have a lot of questions about slides (there is nothing novel there though) I can do a follow up post.
Might be a lot to ask, depending on the time needed to write a couple of lines - but I guess right now it's most helpful for the people that are already educated on v8 performance.
It's just that you could easily educate a broader audience (cough - like the author here) with a short write up. Some things are 'obvious' (GC pressure), some (everything about 'In hot code', some of the trace parameters) things are opaque and ~meaningless~.
I wondered too if they had more to say about the usefulness of some of the trace parameters. I am pretty sure they refer to what the profiler traces (and logs), but I don't know whether they got any usefulness out of them.
What I meant was: The whole 'in hot code' part of the slides didn't make sense to me, because it's something like a list of (unexplained) statements. I know what 'hot code' refers to, I just don't particularly like to accept things as facts without a minimum of explanation. Probably the author is dead on and I'm sure he knows a lot more about the technology than I do, but 'don't use X, use Y' or 'avoid Z' is something that should be qualified.
Probably the talk did that.
Trace parameters are there to give actual nodejs/v8 users a chance to experiment, to trace their app and see what optimizing compiler does to their code.