- RAII-style freeing means a lot of reference counting. Reference counting and multiple threads don't mix well. You can circumvent that by having multiple separate heaps, but that's not always a practical solution.
- Reference counting is slow; if you want GC-like speed and RC-like pause time guarantees, you're probably better off with deferred reference counting, but that's not easy to engineer into C++; it's essentially a form of garbage collection.
- Lack of fully automatic memory management may affect modularity adversely; see Jones/Lins for the gory details. In brief, wholly or partly manual memory management creates leaky abstractions (this is if you eschew RC for some of your code/types for speed reasons or because you need to deal with cyclic references).
- It is not difficult to have scoped deallocation for resources (scope statement in D, using statement in C#, etc., destructor pragma in Nimrod). Resource lifetime, in any event, does not always coincide with the lifetime of a variable, so this is an incomplete solution. Using GC does not mean not using RAII where its downsides do not matter.
You can't say reference counting is bad for multi-threaded programs without mentioning that GC is also bad for multi-threaded programs, since the garbage collector has to pause all other threads, either for the entirety of the GC run or (if you're lucky and using a good GC) for parts of it.
The modularity argument is reasonable, and I've been annoyed by this aspect several times, though in the majority of cases it doesn't seem to be an issue.
The problem with reference counting in concurrent programs is one of correctness, not one of speed [1] because every change of a reference count is a potential race condition. That's quite different from the challenges garbage collection faces in a multi-threaded environment.
Whether a garbage collector has to pause all threads is an implementation issue. Modern garbage collectors can limit that pause effectively to the root scanning phase, then let the collector run concurrently with the mutator. You can also work around pausing threads entirely, though the tradeoffs are rarely worth it outside of hard realtime environments.
Note also that this is an issue of pause time, not performance. While HN can get obsessed with pause times, not everyone programs video games, embedded systems, or kernel drivers where that matters. High-performance computing, for example, is a huge field where pause times are all but irrelevant and where amortized cost is what matters. (HPC is also where correctness can become easily more important than squeezing out a percent or more of performance through low-level hacks; if a job that takes several days to run crashes halfway through, that's an enormous loss.)
[1] Technically, you can make reference count updates atomic, but the overhead is absurd, especially under contention. Hence why SNZIs [2] exist, which reduce the speed overhead at the expense of additional memory overhead. [2] http://dl.acm.org/citation.cfm?id=1281106
The (my) problem is that it also matters whenever a program holds lots of data in memory (databases, data analysis, caching, etc). As the cost of RAM decreases, the importance of this problem increases, i.e. fast.
Incremental garbage collection isn't a new technology [1] or one that's particularly difficult to implement by itself (a grad student could probably implement Baker's treadmill in a couple of days); what makes it hard is primarily compaction and multiple threads [1]. You can't effectively use (naive) RC with multiple threads and you don't get compaction with RC, either. Multi-threading is in practice a major driver for the need of GC, since (naive) RC isn't thread-safe and unpredictable object lifetimes don't go well with manual memory management.
Also, deferred RC strategies can under certain circumstances contain the cost for cycle collection. Trial deletion is already limited to nodes reachable from other nodes whose reference count has been decremented; type information can be leveraged further to exclude nodes that cannot possibly be parts of cycles (this is particularly easy in ML-style languages, which make mutually recursive types explicit in the source code, but is not limited to those [2]).
Finally, you can also use multiple disjoint heaps to simplify implementation and cap GC cost (one per thread and zero or more shared heaps). This can also bring naive RC back into play as a viable strategy, though you'd still lose compaction. Multiple heaps are particularly attractive for NUMA architectures.
[1] I note that hard realtime guarantees are difficult to make with a basic incremental collector due to the potential of pathological behavior, but we are not talking about hard realtime guarantees here.
[2] Obviously, JITs and dynamically or weakly typed languages have a harder time with this approach.
Not necessarily, but in practice they do go hand in hand. I didn't say the problem was impossible to solve, just that it is an important problem that needs solving. Azul solved the technology issues (or so they claim) but not the economics of it, and they solved it for a language that isn't a good fit for in-memory computing in the first place (to put it politely).
If I have to write software today that keeps a lot of data in-memory and requires reasonable latency (and I do) my only realistic option is C++.
I know all the drawbacks of naive reference counting. C++ shared pointers are a horrible kludge. Fortunately they are only needed in very few places, thanks to RAII and unique_ptr. The problem is that C++ has tons of other issues that will never be solved (antiquated modularity, header files, crazy compile times, excessive complexity and generally lots of baggage from the past).
I don't necessarily have a fundamental disagreement here, but I offer two caveats (one of which you may consider disagreement at a certain level).
One is that a lot of the discussion in this thread is not about what is possible today, but about where language implementations can realistically go.
The second is that there are language implementations that do allow you to keep lots of data in memory and still give you low latency; the catch is that most of them are functional programming languages and do not tie into the ecosystems of C++, Java, etc. which limits their applicability. But consider for example, that Jane Street Capital, which has pretty significant requirements for latency, is using OCaml. This is in part because OCaml has an incremental garbage collector for major collections (in addition to generational garbage collection). As I said, it's not rocket science.
The same goes for Erlang, which uses a model of thread-local heaps. Erlang heaps tend to be small (lots of threads [1] with little data each, even though the total memory footprint may be in the gigabytes), so Erlang can use a straightforward garbage collector that can run concurrently with other threads, does a complete collection fast (because the entire heap fits inside the L3 or even L2 cache) or can avoid collection entirely (if you specify the heap size explicitly with spawn_opt and no collection is needed before thread termination). As a result, Erlang can easily satisfy the low latency requirements for telecommunications (where its being used primarily).
Functional programming languages just had an unavoidable need for garbage collection for half a century now and so implementations of functional languages have seen a lot more GC tuning than imperative languages. Thus, you do at least in theory have other options, but I expect "realistic" in your case also implies being able to tap into certain existing software ecosystems.
Let me finally note that a big problem has arguably been too much focus on the JVM; not that there's anything wrong with supporting the JVM, but it has too often come at the cost of alternative execution models. JIT compilation, lack of value types, a very general threading model, loss of static type information, etc. all can very much get in the way of tuning automatic memory management. Luckily, some of the newer emergent languages target alternative backends, in part to avoid these problems.
[1] Technically, Erlang uses the term "process" in lieu of "thread"; I'm using "thread" to avoid confusion with OS-level processes.
[1]http://www.hpl.hp.com/personal/Pramod_Joisha/Publications/is...
The attraction of RAII is that it's a pretty simple mechanism that doesn't require compiler intervention outside of basic, already existing optimizations such as inlining.
That being said, D does not require the GC. It's easier if you use it for at least some of the features, but D makes it very easy to use RAII and manual memory management, which also helps a lot in making the GC work better, because it's much easier to avoid making a lot of garbage for it to collect when you don't need to. A lot of stuff in D ends up on the stack rather than the heap, and D's GC ends up having a lot less work to do than the GC does in languages like C# or Java. That being said, the current GC implementation needs some work (and effort is definitely put forth in that area), but if you don't want to use the GC, you can minimize its use or even outright avoid it (though _completely_ avoiding it can be a bit annoying since that means avoiding a few language features that require it; that list is short though).
So, while D definitely uses the GC and promotes its use where appropriate, you have full control over memory just like you would in C++. And the few features that are hampered by avoiding the GC don't even exist in C++ in the first place.
This is true, however there's a compromise involved. Because D also allows for manual memory management and unsafe memory access, it means that the GC is not free to move stuff in memory at will ... which really means that garbage collectors, like the ones available for Java (precise, generational, non-blocking, fairly predictable and compacting) are very hard to develop, probably next to impossible. This is the price you pay for using a lower level language and it's not a price that will go away easily.
I've been using Scala a lot for the past two years, using it to build high-traffic web services and while Scala is pretty wasteful in terms of allocating short-lived objects, I've been surprised at how well the JVM handles it.
In terms of throughput for example, Java GCs are much better than manual memory management. Allocating memory usually involves just incrementing a pointer, so it's basically as cheap as stack allocation. Deallocating short-lived objects is also fairly cheap, since it happens in bulk and so the amortized cost is pretty similar to dealocating stuff on the stack (!!!). The JVM can do some pretty neat things, like for example if it detects that certain references do not escape their local context, it can decide to allocate those objects straight on the stack.
What really sucks about garbage collection is the unpredictability. Java's CMS for example, awesome as it is, still blocks the world from time to time. And when it does, you have no real control over how much time it keeps the process hostage. The new G1 in JDK7 is much better and if you want the state of the art in terms of near-real-time GCs, you can buy into Azul's pauseless GC. But they still suck for certain apps. Allocating objects on the heap also means you have to pay the price of boxing/unboxing the references involved. This sucks too.
On the other hand, by having a good GC at disposal, it's much easier to build multi-threaded architectures. In C++ for example, it's so freaking painful to deal with non-blocking concurrent algorithms, or really, multi-threading of any kind.
Since I wrote a moving GC for Java years ago, and hence know how they work, I set the D semantics so it allows a moving GC. It's just that nobody has written one for D.
I'd just like to say there are sane solutions out there already to things like this. With the proper abstractions you get things like custodians[0] in Racket and several alternatives in terms of "call-with-..."[1] functions that do handling for you.
These might not be the languages that are being discussed here, but they offer things like these precisely because they are not the languages discussed here right now.
[0] http://docs.racket-lang.org/reference/eval-model.html?q=cust...
[1] http://docs.racket-lang.org/reference/file-ports.html?q=call...
It's not fundamentally different from having bound checks for array indices. You can either do that pretty easily at runtime or use a simple type system to do it at compile time (e.g., Boyapati's [1] or Flanagan's [2] work). It's been done, e.g. for Cyclone [3]. This is decades old stuff, not rocket science. Honestly, Monitors had the basic idea right in the 1970s, except as Per Brinch Hansen lamented [4], language designers keep screwing it up.
[1] http://www.pmg.lcs.mit.edu/~chandra/publications/oopsla02.pd...
[2] http://users.soe.ucsc.edu/~cormac/papers/esop99.pdf
[3] https://homes.cs.washington.edu/~djg/papers/cycthreads.pdf
Also, you can use shared memory in Rust. If you do, the type system ensures there can be no data races: either the data is immutable or you must take the lock before you mutate the data.
In Rust, this style is enforced by the language. In C++11 you can enforce it by the convention of using std::unique_ptr and passing those between threads via a shared data structure like a blocking queue.
Took awhile to get used to, but it is a damned fine way to avoid memory leaks of any kind!
But even that aside, if you are writing something like a render loop with hard timing guarantees, (or anything else with timing needs!) having non-deterministic allocation costs can easily blow your cycle budget.
Because "defer" is tied to function definitions, it means that pulling code out into a new function or inlining a function will silently change its behavior in subtle ways.
Like most complex language semantics, "defer" is less easily optimizable than the simple semantics of RAII—with RAII the compiler statically knows what it is supposed to call, which helps exceptions (panicking goroutines in Go) because it allows their cleanups to be optimized into table-driven unwinding. But because Go chose "defer", the compiler must insert code that dynamically keeps track of cleanups to handle the general case. This results in indirect function calls, and bookkeeping overhead, whereas RAII is purely static function calls. The bookkeeping overhead is significant: it allows RAII to be zero-cost in the no-panic path (very important in C++'s case!), while I don't see any way to implement "defer" in a zero-cost way.
Note that if your language supports RAII and lambdas (e.g. C++11), it's trivial to implement defer using a class that takes a function as an argument to its constructor and calls that function in its destructor. So you can still have the flexibility of defer in a language with RAII.
http://dlang.org/statement.html#ScopeGuardStatement
and an article about it:
Is it just that maintaining ownership semantics for heap memory tends to be more complicated than doing the same for file objects (much more rarely shared, I'd guess) or mutexes (not sure if "sharing" is even sensibly defined here).