▲ 87 points
back
39 comments
> NOTE: Throughout our implementation we strictly use compare_exchange_strong, but the C++ standard suggests that for some systems like ARM it mighe be a better idea to run a loop with compare_exchange_weak for better performance. The problem with that is a compare_exchange_weak call may fail spuriously, which essentially it can randomly fail even if everything is correct, thus it makes code code a bit more complicated, so I avoided it for this implementation.
The reason using `compare_exchange_weak` is a better idea for lock-free algorithms is that, in most cases, you'll run it in a retry loop anyway. Since `compare_exchange_strong` is compiled to a retry loop, if you do a retry loop of `compare_exchange_strong` you basically have a loop in a loop. Using `compare_exchange_weak` makes things both simpler and more performant.
This is overstated in value; on arm systems with LSE it's faster to use that even for weak operations than to use ll/sc. Even if you are limited to ll/sc the compiler may not put your cas-loop body into the ll/sc region as there's limitations on how many and what type of instructions are permitted there.
On ARM systems with LSE both weak and strong compile to the same CAS instruction anyway. On LL/SC, using strong means CPU will attempt to retry the same exchange on spurious failure, which increases likelihood that another thread will update the value in the meantime, requiring an outer loop retry.
On any plaform which implements strong exchange with zero overhead over weak exchange, we can safely assume they compile to the same thing, and use weak. The only reason to use `strong` is if you have a one-shot CAS operation and you don't want to write a retry loop.
Sure, with LSE. But you have to target LSE and get a system that actually contains it, i.e. not the default compiler flags and not the original AWS Graviton.
Not trying to be critical, but there are a number of misspellings and grammatical issues and it was actually a breath of fresh air to be reminded while I was reading that a real human being wrote this. I feel a little inspired to turn off spell check for my own writing.
Strong "oll korrect" vibes.
Nice article. There a few issues with your code however from a cursory glance; your dtor seems to allow for spurious/double frees due to custom deleter support (you wanna check up on that), you also seem to use seq_cst far too much even if not needed (you want to avoid them is queues as much as possible), lastly class FastQueueNodeSlot.. isn't aligned (plus 64b alignment is only a thing for amd64 cpus, apple silicon is larger).
a C++ experienced programmer, I spoke recently to, told me that using new and delete is basically prohibited nowadays in C++ in favour of std::make_unique, std::make_shared etc.
For 95% of all code that is correct. std::make_* is easy to use and prevents a lot of mistakes, while having no loss of performance. That last 5% though you are doing weird things and so need to do something manually. (as the other poster said, make_unique is implemented with new) Of course the 5% is overall. Some projects never have anything that gets into that last 5%, while others it is more like 50% of the code can't use make_*. If at all possible your use of new/delete should be limited to a data structure/container than handles it, and not scatters all over.
"For 95% of all code that is correct. std::make_* is easy to use and prevents a lot of mistakes, while having no loss of performance."
For 95% of code std::unique_ptr has no loss of performance. Perhaps. Just remember that it is not a zero-cost abstraction, the compiler won't always be able to entirely eliminate the overhead:
I watched that about 7 years ago when it first came out, and at an hour long I'm not about to watch it again... From what I can recall those are things that rarely are important. There is a cost, but in most real world code they only add a few nanoseconds - I have more important things to worry about.
As a firmware developer commonly operating on devices with 32MHz of cpu, i am offended sir.
If your doing firmware you probably shouldn't be doing much memory allocation at runtime anyway to prevent memory fragmentation.
Even at 32Mhz you typically have more important things to worry about. These nanoseconds are unlikely to add up - the overhead from the underlying new/delete is going to be much worse.
The unique_ptr destructor is called on scope exit, and there are times you want this earlier before other critical code.
That's why you can define your own scope to control that.
you can call release or reset as needed if you want the delete to happen early - and if you forget an obscure code path it still gets deleted on scope exit which is probably good enough for that path.
In many application code bases no doubt. But how do you think make_unique and make_shared are implemented?
With what will most likely become [[unsafe]] profile in C++29, assuming WG21 actually gets their profiles story right.
Classic C++, any issue is at most three years away from going away! There will definitely not be any other issues that are also three years from getting fixed when you reach that year though
Actually no, because you are missing the time between standard being ratified and actuality being widely available in all major compilers.
By the way, it is also classical Web standards, C, Vulkan, OpenCL, and everything else that has a similar process.
Ideally we should have gotten rid of C and C++, but I haven't yet found a better Typescript for C.
Fair enough, three years might be more of an upper bound than a lower bound. And yeah, the issue of having to wait is common to standard-based processes, but I've never seen anything else based on a standard so frequently have people need to jump in and try to dissuade people that issues matter because of the rolling timeline so frequently, regularly, and for such a long time as C++. Obviously it would be unreasonable to expect it to become perfect and not need any more changes, but like, I was hearing people complain about various aspects of `unique_ptr` a decade ago, and it seems like we're not even past that yet.
I don't think adding unsafe will be possible - it will break too much existing code. We might be able to add something new that is only possible in an unsafe context, but there is too much existing code.
However I do expect a [[safe]] profile (or perhaps several, depending on which paper you read) that everyone is encourage to opt-in to. Likely combines with compiler warnings and static analysis to encourage that use. (Also syntax is still open for debate)
Check the WG21 mailing proposals.
I was for Safe C++ paper, based on Circle experience, but it was shot down due to politics.
So we're left with the profiles camp actually delivering, followed by the remaining compilers caring to actually implement them, otherwise it will be static and dynamic analysis as usual.
It wasn't shot down due to politics. It was shot down because it solved the wrong problem.
Once you use atomic cmpxchg you've lost a great deal of scalability because it implies a retry loop (internal or by the user)
The last thing you want is all of the threads failing to cmpxchg (spuriously or otherwise ) spinning on a shared cacheline
Real world alternatives show atomic xchg only solutions scale to hundreds of threads.
Agreed. But you make it sound like the worst case is necessarily fatal. It depends on the use-case. The workable cmpxchg algorithms will make progress on at least one core each round. In a push or pop operation one of the cmpxchg must have succeeded for another to fail. The atomic xchg algorithms that I know of have other undesirable pathologies (e.g. a suspended producer can stall the consumer).
cmpxchg can fail spuriously on some platforms
> Real world alternatives show atomic xchg only solutions scale to hundreds of threads
But notably only with certain workloads
Once you get to using custom lock free queues you should be picking something that matches your workload/broader design anyway.
Whether the allocator calls are "naive" or not depends entirely on the allocator in use. If you need thread/core locality and batching you can get that by replacing global new/delete functions with a decent allocator.
I would have thought that std::optional would have been a likely candidate for use with the Pop() method?
Is this similar to moodycamel’s?
How can I make a lock free queue without OS support, ESP32?
I'd ask your friendly neighborhood LLM for design advice, but off the top of my head, so long as `static_assert(std::atomic<whatever>::is_lock_free)` passes, you could build a lock-free SPSC queue with just plain ol' std::atomics and probably liberal use of `alignas()` calls
Most local LLMs I've used are not particularly great with the advice unfortunately, probably because I haven't spent thousands of dollars on new hardware recently
Edited my comment for clarity. I meant "friendly local" as a euphemism, like "neighborhood pub".
Anyways. Ask Claude.
Fair enough! It occurred to me after sending my comment that maybe you didn't mean it literally, but I'm easily confused, so clarity is helpful!