Hacker News Reader

Comment by kouteiheika | Hacker News Reader

> the shared pointer implementation uses a POSIX mutex [...] C++’s shared pointer has the same problem

It doesn't. C++'s shared pointers use atomics, just like Rust's Arc does. There's no good reason (unless you have some very exotic requirements, into which I won't get into here) to implement shared pointers with mutexes. The implementation in the blog post here is just suboptimal.

(But it's true that C++ doesn't have Rust's equivalent of Rc, which means that if you just need a reference counted pointer then using std::shared_ptr is not a zero cost abstraction.)

View on HN

To be clear, the “same problem” is that it’s not a zero-cost abstraction, not that it uses the same specific suboptimal approach as this blog post.

I think that's an orthogonal issue. It's not that C++'s shared pointer is not a zero cost abstraction (it's as much a zero cost abstraction as in Rust), but that it only provides one type of a shared pointer.

But I suppose we're wasting time on useless nitpicking. So, fair enough.

I think they’re one and the same: C++ doesn’t have program-level thread safety by construction, so primitives like shared pointers need to be defensive by default instead of letting the user pick the right properties for their use case.

Edit: in other words C++ could provide an equivalent of Rc, but we’d see no end of people complaining when they shoot themselves in the foot with it.

(This is what “zero cost abstraction” means: it doesn’t mean no cost, just that the abstraction’s cost is no greater than the semantically equivalent version written by the user. So both Arc and shared_ptr are zero-cost in a MT setting, but only Rust has a zero-cost abstraction in a single-threaded setting.)

I can't say I agree with this? If C++ had an Rc equivalent (or if you'd write one yourself) it would be just as zero cost as it is in Rust, both in a single-threaded setting and in a multithreaded-setting. "Zero cost abstraction" doesn't mean that it cannot be misused or that it doesn't have any cognitive overhead to use correctly, just that it matches whatever you'd write without the abstraction in place. Plenty of "zero cost" features in C++ still need to you pay attention to not accidentally blow you leg off.

Simply put, just as a `unique_ptr` (`Box`) is an entirely different abstraction than `shared_ptr` (`Arc`), an `Rc` is also an entirely different abstraction than `Arc`, and C++ simply happens to completely lack `Rc` (at least in the standard; Boost of course has one). But if it had one you could use it with exactly the same cost as in Rust, you'd just have to manually make sure to not use it across threads (which indeed is easier said than done, which is why it's not in the standard), exactly the same as if you'd manually maintain the reference count without the nice(er) abstraction. Hence "zero cost abstraction".

Sorry, I realized I’m mixing two things in a confusing way: you’re right that C++ could easily have a standard zero-cost Rc equivalent; I’m saying that it can’t have a safe one. I think this is relevant given the weight OP gives to both performance and safety.

Isn't the point of using atomics that there is virtually no performance penalty in single threaded contexts?

IMO "zero cost abstraction" just means "I have a slightly less vague idea of what this will compile to."

No, atomics do have a performance penality compared to the equivalent single threaded code due to having to fetch/flush the impacted cache lines in the eventuality that another thread is trying to atomically read/write the same memory location at the same time.

Atomics have almost no impact when reading, which is what would happen in a shared pointer the vast majority of the time.

> which is what would happen in a shared pointer the vast majority of the time.

This seems workload dependent; I would expect a lot of workloads to be write-heavy or at least mixed, since copies imply writes to the shared_ptr's control block.

I think it's pretty rare to do a straight up atomic load of a refcount. (That would be the `use_count` method in C++ or the `strong_count` method in Rust.) More of the time you're doing either a fetch-add to copy the pointer or a fetch-sub to destroy your copy, both of which involve stores. Last I heard the fetch-add can use the "relaxed" atomic ordering, which should make it very cheap, but the fetch-sub needs to use the "release" ordering, which is where the cost comes in.

> C++ doesn’t have program-level thread safety by construction

It does. It’s called a process.

Everyone chose convenience and micro-benchmarks by choosing threads instead.

"Thread truther" is not one of the arguments I had on the bingo card for this conversation.

I guessed as much. I’m not alone - there is a whole chapter on this topic in “The art of UNIX programming”.

> very exotic requirements

I'd be interested to know what you are thinking.

The primary exotic thing I can imagine is an architecture lacking the ability to do atomic operations. But even in that case, C11 has atomic operations [1] built in. So worst case, the C library for the target architecture would likely boil down to mutex operations.

[1] https://en.cppreference.com/w/c/atomic.html

Well, basically, yeah, if your platform lacks support for atomics, or if you'd need some extra functionality around the shared pointer like e.g. logging the shared pointer refcounts while enforcing consistent ordering of logs (which can be useful if you're unfortunate enough to have to debug a race condition where you need to pay attention to refcounts, assuming the extra mutex won't make your heisenbug disappear), or synchronizing something else along with the refcount (basically a "fat", custom shared pointer that does more than just shared-pointering).

Does there exist any platform which has multithreading but not atomics? Such a platform would be quite impractical as you can't really implement locks or any other threading primitive without atomics.

> Does there exist any platform which has multithreading but not atomics?

Yes. Also, almost every platform I know that supports multi threading and atomics doesn’t support atomics between /all/ possible masters. Consider a microcontroller with, say, two Arm cores (multithreaded, atomic-supporting) and a DMA engine.

Yes but "atomic" operations with the DMA engine are accomplished through interrupts (atomic) or memory mapped IO configuration (atomic).

Certainly such systems can pretty readily exist. You merely need atomic reads/writes in order to implement locks.

You can't create userspace locks which is a bummer, but the OS has the capability of enforcing locks. That's basically how early locking worked.

The main thing needed to make a correct lock is interrupt protection. Something every OS has.

To go fast, you need atomic operations. It especially becomes important if you are dealing with multiple cores. However, for a single core system atomics aren't needed for the OS to create locks.

> You merely need atomic reads/writes in order to implement locks.

Nit: while it's possible to implement one with just atomic reads and writes, it's generally not trivial/efficient/ergonomic to do so without an atomic composite read-write operation, like a compare-and-swap.

I wrote "multithreaded" but I really meant "multicore". If two cores are contending for a lock I don't see how irq protection help. As long as there is only one core, I agree.

On most multicore systems you can pin the IRQ handling to a single core. Pinning locking interrupts to a single core would be how you handle this.

True, but locks are not only needed inside IRQ handler routines.

The boring answer is that standard atomics didn't exist until C++11, so any compiler older than that didn't support them. I think most platforms (certainly the popular desktop/server platforms) had ways to accomplish the same thing, but that was up to the vendor, and it might not've been well documented or stable. Infamously, `volatile` used to be (ab)used for this a lot before we had proper standards. (I think it still has some atomic-ish properties in MSVC?)

Which platforms might that be? Even MIPS has atomics (at least pointer sized last i checked).

AFIAK, and I'm not MIPS expert, but I believe it doesn't have the ability to add a value directly to a memory address. You have to do something like

    // Not real MIPS, just what I've gleaned from a brief look at some docs
    LOAD addr, register
    ADD 1, register
    STORE register, addr

The LOAD and STORE are atomic, but the `ADD` happens out of band.

That's a problem if any sort of interrupt happens (if you are multi-threading then a possibility). If it happens at the load, then a separate thread can update "addr" which mean the later STORE will stomp on what's there.

x86 and ARM can do

    ADD 1, addr

as well as other instructions like "compare and swap"

    LOAD addr, register
    MOV register, register2
    ADD 1, register2
    COMPARE_AND_SWAP addr, register, register2
    if (cas_failed) { try again }

On MIPS you can simulate atomics with a load-linked/store-conditional (LL/SC) loop. If another processor has changed the same address between the LL and SC instructions, the SC fails to store the result and you have to retry. The underlying idea is that the processors would have to communicate memory accesses to each other via the cache coherence protocol anyway, so they can easily detect conflicting writes between the LL and SC instructions. It gets more complicated with out-of-order execution...

    loop: LL r2, (r1)
          ADD r3, r2, 1
          SC r3, (r1)
          BEQ r3, 0, loop
          NOP

Reply on news.ycombinator.com