[youngthugger]

I think more languages should adopt Swift and Objective-C's automatic reference counting. The only downside to ARC is retain cycle's which rarely happen in my experiance. A deterministic object life cycle just feels right.

[rayiner]

Automatic reference counting has its own issues: 1) allocating short-lived objects is not cheap, because they call down into an underlying malloc/free; 2) as a consequence of (1), throughput is lower; 3) getting acceptable overhead for manipulating references in the heap/stack requires compiler optimizations to elide those reference counts, which adds a layer of complexity to your mental model of the program; 4) implementing the reference-count operating in a multi-core system requires pretty heavy-weight atomic primitives, and race conditions can result in incorrect reference counts.[1]

One of the more interesting avenues of research, especially in mobile devices where you don't want to pay the power cost of having a 2x larger heap your live size, are various hybrids of garbage collection and reference counting. A neat, easy to understand one is: http://users.cecs.anu.edu.au/~steveb/downloads/pdf/rcix-oops....

Also interesting is what you can do when you have more information about what pointers can point to (as in Rust). It's not so much that reference counting in Rust is cheap, but that the language offers a lot of tools to let you avoid reference counted pointers in the first place, in favor of references with static lifetime guarantees.

[1] What Swift or Obj-C do when you overwrite a pointer-valued field is to do a objc_release() for the old pointer, and a objc_retain() for the new pointer. If two threads write to a field at the same time, you can corrupt the reference count (even if the objc_release()/objc_retain() operations are themselves atomic!) As far as I know, Apple's obj-c runtime does not attempt to handle this situation. See: http://clang.llvm.org/docs/AutomaticReferenceCounting.html#o....

[ezy]

I prefer well known memory lifetimes, but I'm undecided about ARC (or somewhat equivalently: std::shared_ptr, etc.). It can be a little confusing and overwrought at times, and I don't think the lifetimes are as clear as they could be.

IMO, more languages shouldn't assume that there is a single memory allocator. That's one of the worst assumptions I see in systems languages -- even C++ (before C++11) got this entirely wrong. Swift gets this wrong, Go gets this wrong. Rust is probably a little better because it actually has a static idea of memory scope, but I haven't seen a way to swap out the memory allocator in various contexts.

Most projects I've been involved with have used region/arena allocators. Not only do you mostly avoid the non-determism of your average GC, but you avoid the hassle of fine grained reference counting (in most cases). This relies on you choosing the scopes for your regions appropriately, but there usually is a clear scope to attach things to (e.g. frames, iteration of an event loop, etc.).

[mamcx]

You are talking about a arena collector? Like this:

http://wiki.luajit.org/New-Garbage-Collector

Exist some pre-madethat can be used for a new project, for use with LLVM or luajit?

[kevinnk]

Maybe you're referring to something else, but std::allocator was in c++98

[ezy]

Yes, but (according to the spec) it was stateless -- which made it worthless for the purposes I'm describing. The original purpose was to support custom static strategies for allocation (e.g. i86 near/far pointers or one memory pool for one kind of object for the whole program), not dynamic memory pools and arenas. C++11 fixed that, which is why I mentioned it.

That said, by C++03, most STL implementations supported stateful allocators (and that's what I've used when I've had to use C++), but the standard took a while to catch up.

[Dewie]

> , but I haven't seen a way to swap out the memory allocator in various contexts.

You mean something like this?

https://github.com/rust-lang/rfcs/pull/244

I think part of it was inspired by something similar from C++. D-lang also seems to have something like this.

[ezy]

Yup. That's excellent. Rust changes so fast, I can hardly keep up. :-)

[Dewie]

Sadly though I haven't been able to find any mention of if and when it will be realized. But the sentiment - from what I've gleaned from discussions on it - seems to be that something like that should/need to happen.

[chrisseaton]

"The only downside to ARC is retain cycles"

Does ARC allow moving? If not, what happens when your heap becomes fragmented? Isn't that a downside?

Another downside is that many lock-free data-structures require a GC.

[ltta]

As far as I know ARC (like manual memory managenent) can lead to release cascades that can also cause application delays in games eg. Something you have to watch for. Malloc() and free() and equivalents do quite a bit of work under the hood (check today's linked tcmalloc article for example) that takes time.

[psykotic]

Release cascades can be a big problem even with plain old RAII in C++. I once diagnosed a performance issue in a C++ network server where the server listening thread would sometimes temporarily hang when clients disconnected. The culprit turned out to be the destructor of a large std::map that directly and indirectly accounted for tens of millions of heap-allocated objects that had to be individually destructed and freed. It's very rare for destructors to have intentional global side effects, so this kind of work could almost always have been done asynchronously, at least in principle.

An unfortunately common symptom of large C++ applications is that they take forever to shut down because they insist on calling destructors on everything in the known universe. In reality, most applications only have a tiny handful of resources that you truly have to release yourself at shut down, and memory certainly is not one of them.

[mamcx]

So, what do? Perhapsh mark a object as "kill me without mercy?"

[psykotic]

If the mass objects have trivial (ignorable) destructors and don't own any heap-allocated subobjects then it's not a problem. You can have top-level objects like levels or documents or sessions own everything directly and indirectly under them and allocate those things out of a private heap that can ideally be reused but otherwise be bulk freed in a few calls. Unfortunately, the standard C++ philosophy around RAII and value semantics works against this. Since there is something to be said for the elegance and upfront convenience of the value semantics approach, it comes down to a trade-off.

[mamcx]

So, if I design a language, how design it well? I suspect is easy to kill all values (ints, str, etc) and arrays/list of it. But how know what is a "don't own any heap-allocated subobject" (sorry to ask if this is obvious, I have not deep experience in this matters)?

Is not circular references the real problem? And what when the object reference a resource like a file/handle/database/etc?

So, could be good idea to mark objects like this (maybe in separated areas of memory?):

Instant kill: Ints, Strs, Bools, Array of all of this

Safe destructors: If it hold a resource (file, handle)

But don't know what to do for objects like Customer.Orders = List<Orders>[Order1.Customer]

[ltta]

This is not necessarily a language-level decision (except for GC, it helps in this case spread out the deallocations when using incremental GC).

You just have to be aware what happens in the destructor when you manually free an object. Libraries can make knowing this more difficult.

I guess you could build incremental alloc/release pools (even with reuse) or such things but it comes down to being aware of the problem as described and avoiding cascaded releases.

And to be honest this is not a super common problem but it can happen.

[Ded7xSEoPKYNsDd]

This is just an idea, but how about adding a function to your allocator that turns all future deallocations into a no-op?

[ltta]

And also check out this paper comparing reference counting and GC, and the accompanying discussion on LtU.

http://lambda-the-ultimate.org/node/2552

(Although I think the paper has problems and bias)

[nraynaud]

I had big troubles with cycles, the trick is that you don't know that you are leaking until it's too late.

Another trouble is the cache pollution.

[gkuan]

There is also this paper by Bacon et al that develops a framework that shows the relationship between refcount and tracing GC (duals of each other) plus the various optimizations of the two: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.143....

[Dewie]

> The only downside to ARC

Doesn't reference counting typically have a worse throughput than garbage collection? That's what I've read anyway.

[lumpypua]

GC has better throughput at the expense of significantly increased memory usage, and variable latency for any individual task.

On servers, that's fine.

On my phone, I'll take lower memory usage and predictable latency any day.

Yay refcounting.

[tolmasky]

There are certainly classes of mobile apps that need the guarantees of lower/deterministic memory usage. In my experience, that is not the common app being written however. These concerns seem largely like fantasies not backed up by any concrete evidence for your every-day twitter client/mail app/weather/whatever. Generic list-based apps simply do not need to be acting as if you need to squeeze every ounce out of the processor/ram anymore. They would instead benefit much more from not having to worry about whether self in the closure you're creating should be weak or not. Especially when you consider that many (most?) of the apps on the App Store that are actually pushing things to the limit -- games -- are running in a (old) GC environment (C# in Unity).

[mnem]

I would be quite surprised if most App Store games were created with Unity - it's expensive and (comparatively) cumbersome to develop in, especially if you're just using 2D graphics (even with the fancy new Unity2D stuff). I can't find any stats on mobile game engine use though, which is a shame because I think it would be interesting to know.

When creating Unity3D games, you have to be reasonably careful about garbage collection - lots of nasty performance dips result if you assume that it just works. Generally you have to cobble together a mixture of object pools and statements attempting to force collection at a convenient point in the game flow amongst other things. It would be quite nice if you could turn it off for specific portions of code :)

[mnem]

Found some interesting statistics on this site: http://www.iresearchchina.com/views/5604.html

I suspect to get anything more detailed you may have to purchase a paid report from one of the various analytics companies.

Sorry, I've wandered rather off topic...

[bad_user]

Unity uses an older Mono (2.10 I believe) which has a poor GC that's not even deterministic, so you can easily suffer memory leaks. It is also not generational. Newer versions of Mono come with a much better GC.

[bad_user]

Ref counting is not deterministic. First of all because while looping on an atomic reference is non blocking, it is not wait free. Malloc/free are not deterministic either, can be quite expensive and while real-time implementations exist, that's not the malloc/free you end up using. With a generational GC allocating new objects involves just incrementing a pointer, deallocating short lived objects happens in bulk, so the cost for short-lived objects is similar to stack allocation. With ref counting memory gets fragmented, apps like Firefox have been suffering for years from fragmentation and for Firefox there was a sustained and very significant effort to fix it.

This can only be solved by either doing stack allocation or by building object pools. And this is for people that know what they are doing.

Rust is the only language (I know) that tries solving this with the ownership concept in the language, but then Rust will have problems in implementing immutable data structures that can be shared amongst threads, data structures which are doing structural sharing, so people will expect reads to scale, except there will be non-obvious contention happening due to usage of reference counting.

The latency in state of the art mainstream GCs is also NOT variable. Good garbage collectors allow you to control the max latency and frequency of STW pauses. That is not the issue for real-time requirements - the issue with real-time being that STW is unacceptable. But so is reference counting.

On the memory requirements, I don't buy it. My Android phone has more memory and more CPU capacity than my computer did 7 years ago. And I'm not seeing a difference in behavior to my iPad. Surely for games it pretty bad to drop frames, but most apps are not games and games are using highly optimized engines built in C++anyway.

[pjmlp]

> Rust is the only language (I know) ...

Have a look at Cyclone, ATS and ParaSail.

[nraynaud]

it's not really "significant" if you don't optimise it that way. It's only the first generation scavenger that needs space for breathing.

[lumpypua]

I do agree with this critique. Object allocation patterns that suck down space in a GC'ed system equally abuse a refcounting system, just you're paying in CPU time rather than memory.

[adrianm]

All GC algorithms are not created equal.

What algorithm are you referring to when you say "better throughput at the expense of significantly increased memory usage"?

Also, why is unpredictable latency acceptable for you on servers but not phones? Wouldn't a latency spike on remote requests from an application degrade user experience just as much as if that latency were localized to the phone?

[lumpypua]

> What algorithm are you referring to when you say "better throughput at the expense of significantly increased memory usage"?

Here's a discussion of the particular paper behind that statement: [1]

I'd also like to recommend the paper "A Unified Theory of Garbage Collection" [2] that breaks down the divide between GC and refcounting. There is a lot of gray area between tracing GC and refcounting. You can make different tradeoff decisions in different parts of your collection algorithm. But fundamentally it breaks down to time-space tradeoffs—you want to save time and get throughput, you're gonna eat some extra space.

> Also, why is unpredictable latency acceptable for you on servers but not phones? Wouldn't a latency spike on remote requests from an application degrade user experience just as much as if that latency were localized to the phone?

We as developers make UI efforts to mitigate network unreliability (fallacy #1 of distributed computing: the network is reliable) so it's ok if a server is being temporarily shitty. It's a lot harder to keep responsive, smooth UI behavior in the face of dropped frames and long GC pauses.

[1] http://stackoverflow.com/questions/2982325/quantifying-the-p...

[2] http://www.cs.virginia.edu/~cs415/reading/bacon-garbage.pdf