> So the standard library plays it safe: if your move constructor might throw (because you didn’t mark it noexcept), containers just copy everything instead. That “optimization” you thought you were getting? It’s not happening.
This is a bit of a footgun and clang-tidy has a check for it: performance-noexcept-move-constructor. However, I don't think it's enabled by default!
performance-noexcept-move-constructor is great but it also complains about move assignment operators, which are completely different beasts and are practically impossible to make noexcept if your destructors throw.
Most sensible Compiler flags aren't enabled by default... I keep a list of arguments for gcc to make things better, but even then you'll also wanna use a static analysis tool like clang-tidy
The reason performance-noexcept-move-constructor is not enabled by default is likely because blindly applying noexcept is dangerous if the underlying logic isn't actually exception-free. If you let clang-tidy slap noexcept on a move constructor that does end up throwing (perhaps because it calls into a legacy member or allocates memory internally), the runtime behavior changes from caught exception to std::terminate().
The documentations seems to say that option only causes the compiler to issue a warning when move constructors are not marked noexcept - it doesn't override anything.
Note that the way std::vector (and other STL containers) require noexcept move constructors for reallocation is by using template matching, and of course any other code might be doing this too, so having a compiler option that forced a constructor (or anything) to have a type signature different than the way it was declared would be a pretty dangerous thing to do since it'd be hard to know what the consequences would be.
I would argue performance-noexcept-move-constructor should always be on. Move constructors should almost always be noexcept since they typically just move pointers around and don't do allocations normally.
clang-tidy checks but doesn't change things for you.
Since you can also put noexcept(false) to indicate something throws exceptions and you didn't just forget to mark it noexcept, it's not a bad policy to say every move constructor should have a noexcept marker.
I think what he means is that on a 64-bit system you have a massive virtual address space (typically only 48-bit, but that's still 256TB), and since malloc allocates from virtual address space, not limited by physical memory, it is unlikely you will get a malloc failure (unless you are trying to allocate more than 256TB per process, maybe due to a memory leak).
Value categories and move semantics are great examples of programming concepts that can cause confusion, and it's a great example of how not having a bad documentation can still lead to confusion through bad mental models. ~
Intuitively you think you understand what is going on, and you think you can answer what is going on, and you can even use it due to understanding it on an operational level, but you can't explain it due to your confusion.
As a result, you most likely are going to create a lot of small bugs in your software and a lot of code that you don't really understand. So, I'm curious to know what others think.
What concept did you learn later than you thought you would? What knowledge did you struggle with the most? What finally helped you understand it?
The issue TFA is describing isn't really about not understanding move semantics, it's about not having read the documentation for the STL container classes, and not therefore realizing that anything requiring reallocation needs a noexcept move constructor (else will fall back to copy construction).
Note that a move constructor that is NOT declared with noexcept is perfectly valid, and will happily be used most of the time (other than where code, such as the STL, is explicitly looking for a noexcept one).
Value categories actually just are confusing in a language as complicated as C++. I'm not willing to bet that even senior C++ developers are always going to be able to deduce the correct value category.
And worse, in typical C++ fashion, there is still little guaranteed as far as when std::move will actually cause a move. The implementation is still given a lot of leeway. I've been surprised before and you basically have no choice but to check the assembly and hope it continues to be compiled that way as minor changes make their way into the code base.
Luckily, with C++17's if-constexpr and C++20's concepts, SFINAE has become mostly obsolete for new C++ code (unless you have/want to support older C++ standards).
About 28 years ago, I figured out that I’m just not smart enough to use C++. There are so many foot guns and so much rampant complexity that I can’t keep it all straight. I crave simplicity and it always felt like C++ craved the opposite.
I hear this a lot, but I don’t really understand how this manifests in language complexity like the stuff in TFA in practice.
Like, I can understand how systems programming requiring programmers to think about questions like “how can I proceed if allocation fails? How does this code work in an embedded context with no heap?” is hard and irreducible.
But I can’t understand why a language’s choice to impose complex rules like C++ move constructor hell is an inevitable outcome of irreducible complexity in systems programming. Put another way: C is also a systems programming language that works for many people, and it doesn’t have any of these Byzantine rules (unless you build them yourself). That’s not to say C is better/preferable, but it swims in the same “official Big Gun systems language” pond as C++, which seems to indicate that revalue semantics as complex as C++’s are a choice, not an inevitability.
I wouldn't say issues like this are dues to irreducible complexity, but more symptomatic of long-lived languages that continually get extended but don't give up on backwards compatibility. It's basically the 2nd law of thermodynamics applied to programming languages that they will eventually die due to increased entropy.
Maybe if move semantics, and noexcept, had been designed into C++ from the beginning then the designers might have chosen to insist that move constructors be noexcept, but since these were added later there is code out there with move constructors that do throw exceptions...
Note by the way that the issue being described isn't strictly about std::move or move semantics in general, but more about the STL and containers like std::vector that have chosen to define behavior that makes noexcept move constructors necessary to be used when reallocating.
The difference is that in C one is supposed to do allocations and deallocations oneself. Then move semantics is just pointer assignment with, of course, the catch that one should make sure one does not do a double-free because ownership is implicit. In C++ ownership is indicated by types so one has to write more stuff to indicate the ownership.
> The difference is that in C one is supposed to do allocations and deallocations oneself
No, you should only use the heap if necessary.
The bigger issue in C is there is no concept of references, so if you want to modify memory, the only recourse is return-by-value or a pointer. Usually you see the latter, before return value optimization it was considered a waste of cycles to copy structs.
In the embedded world, its often the case you won't see a single malloc/free anywhere. Because sizes of inputs were often fixed and known at compile time for a particular configuration.
I have no problem with systems programming issues. That complexity is essential complexity inherent in the problem itself, regardless of language. I have a problem with C++’s accidental complexity. I find C much more tractable. It certainly has a few of its own footguns, but it has much less accidental complexity.
As the author of the FQA noted (Yosef K-something), in C++ its more the combinations of features which causes so many issues.
And here we see this principle rear its ugly head yet again. In this case, its the combination of exceptions, manual memory allocation and the desire to make things work efficiently - of which the move constructor was developed as a "solution"
> [std::move silently copies const values, because] If something is const, you can’t move from it by definition.
Whoever wrote that definition should have a thing or two to learn from Rust. Different language I know, but it proves that it wasn't needed to cause so much confussion and collectively so much time and performance lost.
Also, who writes rules like that and ends the day satisfied with the result? It seems unlikely to feel content with leaving huge footguns and being happy to push the Publish button. I'd rather not ship the feature than doing a half-assed work at it. Comparing attitudes on language development and additions, it makes me appreciate more the way it's done for the Go lang, even though it also has its warts and all.
The point is not a comparison with Rust per-se, but the fact that a better implementation of the idea was mathematically and/or technically possible; and the personal opinion that such huge footguns that the language accumulates over the years are maybe signals of having needed more thought to them before they were considered ready.
e.g. if something as simple of a inconspicuous std::move in the wrong place can break the whole assumption about move semantics, then make that impossible to do, or at least do not make it the default happy path, before you consider it production ready. What the heck, at the very least ensure it will become a compiler warning?
Hence the mention to Go and how they follow exactly this path of extending discussion as long as needed, even if it takes 10 years, until a reasonable solution is found with maybe small gaps, but never huge ones such as those explained in this article (plus tens of others in any other text about the language)
A bit more if we consider the "bugfixing" release that was C++14 :)
But yeah it makes sense, given how that was the jumpstart of the whole modernization of the language. I believe it was a big undertake that required the time it took. Still years have passed and footguns keep accumulating... it wouldn't hurt to have a mechanism to optionally drop the old cruft from the language. Otherwise everything stacks on top in the name of backwards compatibility, but at this pace, how will C++36 look like?
While I share the sentiment, compare C#14 with C# 1.0, Java 25 with Java 1.0, Python 3.14 with Python 1.0.
While C++ might be worse, when you have 300+ doing proposals every three years, others aren't safer from similar churn, even if on smaller volume, and trying to keep backwards compatibility going.
And we all know what happened in Python.
Also Rust editions contrary to what many think, only cover a specific set of language evolution scenarios, it isn't anything goes, nor there is support for binary libraries.
As for a better C++, contrary to C, where it is business as usual, there are actually people trying to sort things out on WG21, even if isn't as well as we would like to.
"Making C++ Safe, Healthy, and Efficient - CppCon 2025"
This is a weird call-out because it's both completely incorrect and completely irrelevant to the larger point.
Rust absolutely supports binary libraries. The only way to use a rust library with the current rust compiler is to first compile it to a binary format and then link to it.
More so than C++ where header files (and thus generics via templates) are textual.
Cargo, the most common build system for rust, insists on compiling every library itself (with narrow exceptions - that include for instance the precompiled standard library that is used by just about everyone). That's just a design choice of cargo, not the language.
The story is that it must not matter which edition a library was compiled with - it's the boundary layer at which different editions interoperate with eachother.
a member of the c++ committee (herb sutter) is writing an compiler for an alternative c++ syntax [0], to c++, with the intent to restrict some semantics of the language for less UB, surprises, etc. i think less implementation-defined behavior is incredibly important; rvo vs std::move, dynamic function call optimization, i wish i didn't have to search asm to check for...
What’s the problem? It makes perfect sense to me that a const object cannot be moved from, since it violates the constness. Since constness goes hand in hand with thread safety you really don’t want that violation.
There are cases where you would not want to reject such code, though. For example, if std::move() is called inside a template function where the type in some instantiations resolves to const T, and the intent is indeed for the value to be copied. If move may in some cases cause a compiler error, then you would need to write specializations that don't call it.
It's weird that they made a mistake of allowing this after having so many years to learn from their mistake about copies already being non-obvious (by that I mean that references and copies look identical at the call sites)
Before move semantics the HeavyObject problem was solved in most cases by specializing std::swap for each container.
The design lesson I draw from this is that pursing a 100% general solution to a real problem is often worse than accepting a crude solution which covers the most important cases.
Regarding mistake 1: return std::move(local_var), it is worth clarifying why this is technically a pessimization beyond just breaking NRVO. It comes down to the change in C++17 regarding prvalues.
> Pre-C++17, a prvalue was a temporary object.
> Post-C++17, a prvalue is an initializer. It has no identity and occupies no storage until it is materialized.
The name predates the standardisation. The committee did not come with the whole thing themselves, rather they adopted and expanded already existing library implementations. You could move in C++, with this exact name, long before C++11.
thanks to the incredible advances in terms of developer tooling over the last 50 years (i.e. tab-autocompletion) there should be no difference in writing those two.
Sounds more like a contract thing. Of course std::move should be able to throw exceptions (like when it runs out of memory), but when it throws an exception it should still guarantee that memory is in a consistent state.
So the fault here is with std::vector who didn't write that contract.
> This code works. It compiles. It runs. But depending on how you’ve implemented your types, it might be performing thousands of expensive copy operations instead of cheap moves without you realizing it.
I've spent the last two decades in the .net platform. But for a decade or so before that I was a C++/Unix dev. I remember old style "C with classes" C++ as being fairly small and elegant, and approximately as easy to reason about as C# - albeit that you had the overhead of tracking object ownership and deallocation.
What the language has become now, boggles my mind. I get hints of elegance/power and innovation when I read about it, but the sheer number of footguns is astonishing. I'm very sure that I'm not clever enough to understand it.
But some very smart people have guided the language's evolution. So, what are the forces that have determined the current state of C++?
> So, what are the forces that have determined the current state of C++?
I'm very confident that the main driving factors are:
1. "performance" (not wanting to do more allocations than necessary)
2. abi compatibility
3. adding features without caring how well they integrate
Example for 1:
"emplace", you normally have "append" but emplace directly constructs the object in the container instead of having to be constructed first and then moved into the container. This nice and all but breaks when using pairs (for reasons you can google but I don't wanna explain here). So now you have these obscure classes like https://en.cppreference.com/w/cpp/utility/piecewise_construc... which solve this.
Example for 2:
Basically they never break the ABI and this leads to tons of old stuff hanging around and never being changed and just more stuff being added on top. std::iostream is famously slow and a big reason is because you can't fix it without breaking the abi which they don't wanna do.
Example for 3:
The whole template thing adds so much complexity it's bonkers, I think c++ without templates would be pretty manageable comparatively. For example because C++ has constructors and they don't quite mix well with templates you suddenly end up in the situation that you have 2 concepts: "normal" template argument deduction and constructor template argument deduction (CTAD). Because of this asymmetry you need a custom language feature called "deduction guides" to maneuver yourself out of the problems that come from this.
Or another short one: std::expected without something like the "!" that rust has. You end up with endless "if(result.has_value()) { return result; }" cascades and it's horribly unergonomic. So now we have a Result class but it's practically unusable that it will only fragment the ecosystem even more.
Note that C# 14 versus C# 1.0 isn't suffering from feature creap as well.
What has guided C++ are the 300+ volunteers that get to submit papers, travel around the world attending the meetings, and win the election rounds of what gets into the standard.
Unfortunately design by committee doesn't lead to a clear product roadmap.
std::move is definitely for there for optimizing application code and is often used there. another silly thing you often see is people allocating something with a big sizeof on the stack and then std::moving it to the heap, as if it saves the copying
> another silly thing you often see is people allocating something with a big sizeof on the stack and then std::moving it to the heap, as if it saves the copying
I'm convinced naming things is equivalent to choosing the right abstraction, and caching things is creating a correct "view" from given normalized data.
std::move itself doesn't move ownership, though. It allows the compiler to transfer ownership to the receiver of the value, but it doesn't force it in any way. This is important, because it means YOU may still be the owner of a value even after you called std::move on it.
Not to mention, ownership in C++ is not entirely lost with moves in the traditional sense. For example, your code still has to destruct the object even if you did move it to somewhere else.
Std move doesn’t move ownership. It simply casts into something that could have its ownership taken. Whether or not that actually happens is impossible to identify statically and the value after ownership is consumed is unspecified - sometimes it’s UB to access the value again, sometimes it’s not.
It needs to remain destructible, and if the type satisfies things like (move-)assignable/copyable, those still need to work as well.
For boxed types, it's likely to set them into some null state, in which case dereferencing them might be ill-formed, but it's a state that is valid for those types anyway.
Well it’s unspecified what empty/size return for collections after a move. Not a dereference, not UB but unspecified as I said. UB pops up in hand written code - I’ve seen it and the language doesn’t provide any protection here.
Thankfully clippy lints do exist here to help if you integrate that tooling
> it's intended for transferring of ownership versus copying data.
It's intended for transferring ownership, but what it actually does is mark the value as transferrable, whether or not the value is actually transferred is up to the callee.
"Validity" is an extremely low bar in C++, it just means operations with no preconditions are legal, which in the most general case may be limited to destruction (because non-destructive moves means destruction must always be possible).
So you're saying if you use the language to write UB, then you get UB?
Seems kinda circular. Ok, you're not the same user who said it can be UB. But what does it then mean to same "sometimes it's UB" if the code is all on the user side?
"Sometimes code is UB" goes for all user written code.
I mean the language doesn't dictate what post-condition your class has for move-ctor or move-assignment.
It could be
- "don't touch this object after move" (and it's UB if you do) or
- "after move the object is in valid but unspecified state" (and you can safely call only a method without precondition) or
- "after move the object is in certain state"
- or even crazy "make sure the object doesn't get destroyed after move" (it's UB if you call delete after move or the object was created on the stack and moved from).
But of course it's a good practice to mimic the standard library's contract, first of all for the sake of uniformity.
It is absolutely knowable statically if ownership will be taken. It's not necessarily very easy to do so, but the decision is 100% up to the compiler, as part of overload resolution and optimization choices (like the NRVO analysis that the article mentions). Since ownership is an inherently static concept, it doesn't even make sense to think about "runtime ownership".
My claim is that, if I call `foo(std::move(myObj))`, it is statically knowable if `foo` receives a copy of `myObj` or whether it is moved to it. Of course, `foo` can choose to further copy or move the data it receives, but it can't choose later on if it's copied or not.
Now, if I give `foo` a pointer to myObj, it could of course choose to copy or move from it later and based on runtime info - but this is not the discussion we are having, and `std::move` is not involved from my side at all.
If I call `foo(std::move(my_unique_ptr))`, I know for sure, statically, that my_unique_ptr was moved from, as part of the function call process, and I can no longer access it. Whether `foo` chooses to further move from it is irrelevant.
The only thing that is statically known here is that you’re wrong. The function I posted only moves its parameter half the time, at random. You may want to treat it as moved-from either way, but factually that’s just not what is happening.
> This is like trying to defend that you can't statically know the result of 1 + 2
It is completely unlike that. tsimionescu is asserting that they can always know statically whether `foo` will move its parameter. The function I provided is a counter-example to that assertion.
Of course the branch body always moves, that's what it's there for. That has no bearing on the argument.
aside from what others wrote, it’s also non local - whether std::move even does anything is dependent on the signature of foo - if foo takes it by const& you may think you’ve transferred ownership when it hasn’t actually happened.
In Rust if you pass say a Box<Goose> (not a reference, the actual object) into a function foo, it's gone, function foo might do something with that boxed goose or it might not, but it's gone anyway. If a Rust function foo wanted to give you it back they'd have to return the Box<Goose>
But C++ doesn't work that way, after calling foo my_unique_ptr is guaranteed to still exist, although for an actual unique_ptr it'll now be "disengaged" if foo moved from it. It has to still exist because C++ 98 (when C++ didn't have move semantics) says my_unique_ptr always gets destroyed at the end of its scope, so newer C++ versions also destroy my_unique_ptr for consistency, and so it must still exist or that can't work.
Creating that "hollowed out" state during a "move" operation is one of the many small leaks that cost C++ performance compared to Rust.
Look, the act of calling std::move and and calling a function taking an rvalue reference in no way invokes a move constructor or move assignment. It does not "move".
It's still just a reference, albeit an rvalue reference. std::move and the function shape is about the type system, not moving.
(Edit: amusingly, inside the callee it's an lvalue reference, even though the function signature is that it can only take rvalue references. Which is why you need std::move again to turn the lvalue into rvalue if you want to give it to another function taking rvalue reference)
I didn't reply to this thread until now because I thought you may simply be disagreeing about what "move" means (I would say move constructor or move assignment called), but the comment I replied to makes a more straightforward factually incorrect claim, that can easily be shown in godbolt.
If you mean something else, please sketch something up in godbolt to illustrate your point. But it does sound like you're confusing "moving" with rvalue references.
Static analysis is about proving whether the code emitted by a compiler is actually called at runtime. It's not simply about the presence of that code.
Well, no, because CAN take isn't the same as WILL take.
Changing something to an rvalue means it'll now match a move constructor, but there is no guarantee a move constructor will be used, even if defined, because you've got classes like std::vector that are picky and are explicitly looking for a noexcept move constructor.
In that sense, std::move() is no different than other passing semantics. Just because you wrote at the call site that you want to pass a copy of your object doesn't mean that the callee will actually make a copy of it.
> Rust’s borrow checker doesn’t actually borrow anything either
Why would it? It's called the borrow checker, not the borrower. So it checks that your borrows are valid.
std::move looks and feels like a function, but it doesn't do what it says, it makes objects movable but does never moves them (that's up to whatever is using the value afterwards). If you want something similar in Rust, Pin is a much better candidate.
Sure, but from the perspective of the code that has the move() its good to assume the value is moved at that call, which I guess was the intention of picking the name.
> So the standard library plays it safe: if your move constructor might throw (because you didn’t mark it noexcept), containers just copy everything instead. That “optimization” you thought you were getting? It’s not happening.
This is a bit of a footgun and clang-tidy has a check for it: performance-noexcept-move-constructor. However, I don't think it's enabled by default!
performance-noexcept-move-constructor is great but it also complains about move assignment operators, which are completely different beasts and are practically impossible to make noexcept if your destructors throw.
Most sensible Compiler flags aren't enabled by default... I keep a list of arguments for gcc to make things better, but even then you'll also wanna use a static analysis tool like clang-tidy
Would you mind sharing your list?
The reason performance-noexcept-move-constructor is not enabled by default is likely because blindly applying noexcept is dangerous if the underlying logic isn't actually exception-free. If you let clang-tidy slap noexcept on a move constructor that does end up throwing (perhaps because it calls into a legacy member or allocates memory internally), the runtime behavior changes from caught exception to std::terminate().
The documentations seems to say that option only causes the compiler to issue a warning when move constructors are not marked noexcept - it doesn't override anything.
https://clang.llvm.org/extra/clang-tidy/checks/performance/n... constructor.html
Note that the way std::vector (and other STL containers) require noexcept move constructors for reallocation is by using template matching, and of course any other code might be doing this too, so having a compiler option that forced a constructor (or anything) to have a type signature different than the way it was declared would be a pretty dangerous thing to do since it'd be hard to know what the consequences would be.
I would argue performance-noexcept-move-constructor should always be on. Move constructors should almost always be noexcept since they typically just move pointers around and don't do allocations normally.
clang-tidy checks but doesn't change things for you.
Since you can also put noexcept(false) to indicate something throws exceptions and you didn't just forget to mark it noexcept, it's not a bad policy to say every move constructor should have a noexcept marker.
Exceptions should never be enabled by default. We live in a 64bit world so allocations failing indicates some other problem.
What does processor but width have to do with the likelihood of allocation failures?
I think what he means is that on a 64-bit system you have a massive virtual address space (typically only 48-bit, but that's still 256TB), and since malloc allocates from virtual address space, not limited by physical memory, it is unlikely you will get a malloc failure (unless you are trying to allocate more than 256TB per process, maybe due to a memory leak).
640K ought to be enough for anybody!
Value categories and move semantics are great examples of programming concepts that can cause confusion, and it's a great example of how not having a bad documentation can still lead to confusion through bad mental models. ~
Intuitively you think you understand what is going on, and you think you can answer what is going on, and you can even use it due to understanding it on an operational level, but you can't explain it due to your confusion.
As a result, you most likely are going to create a lot of small bugs in your software and a lot of code that you don't really understand. So, I'm curious to know what others think.
What concept did you learn later than you thought you would? What knowledge did you struggle with the most? What finally helped you understand it?
The issue TFA is describing isn't really about not understanding move semantics, it's about not having read the documentation for the STL container classes, and not therefore realizing that anything requiring reallocation needs a noexcept move constructor (else will fall back to copy construction).
Note that a move constructor that is NOT declared with noexcept is perfectly valid, and will happily be used most of the time (other than where code, such as the STL, is explicitly looking for a noexcept one).
So, for example:
HeavyObject t;
HeavyObject s(std::move(t));
Will cause t to be moved to s.
Value categories actually just are confusing in a language as complicated as C++. I'm not willing to bet that even senior C++ developers are always going to be able to deduce the correct value category.
And worse, in typical C++ fashion, there is still little guaranteed as far as when std::move will actually cause a move. The implementation is still given a lot of leeway. I've been surprised before and you basically have no choice but to check the assembly and hope it continues to be compiled that way as minor changes make their way into the code base.
Coming from other languages with generics, it took a while for me to internalize SFINAE when writing templated code.
Luckily, with C++17's if-constexpr and C++20's concepts, SFINAE has become mostly obsolete for new C++ code (unless you have/want to support older C++ standards).
About 28 years ago, I figured out that I’m just not smart enough to use C++. There are so many foot guns and so much rampant complexity that I can’t keep it all straight. I crave simplicity and it always felt like C++ craved the opposite.
Systems programming in the large is hard, owning the category for decades harder still.
Even languages that have tried to fast-follow and disrupt C++ end up looking a lot like C++. There is an irreducible complexity.
I hear this a lot, but I don’t really understand how this manifests in language complexity like the stuff in TFA in practice.
Like, I can understand how systems programming requiring programmers to think about questions like “how can I proceed if allocation fails? How does this code work in an embedded context with no heap?” is hard and irreducible.
But I can’t understand why a language’s choice to impose complex rules like C++ move constructor hell is an inevitable outcome of irreducible complexity in systems programming. Put another way: C is also a systems programming language that works for many people, and it doesn’t have any of these Byzantine rules (unless you build them yourself). That’s not to say C is better/preferable, but it swims in the same “official Big Gun systems language” pond as C++, which seems to indicate that revalue semantics as complex as C++’s are a choice, not an inevitability.
I wouldn't say issues like this are dues to irreducible complexity, but more symptomatic of long-lived languages that continually get extended but don't give up on backwards compatibility. It's basically the 2nd law of thermodynamics applied to programming languages that they will eventually die due to increased entropy.
Maybe if move semantics, and noexcept, had been designed into C++ from the beginning then the designers might have chosen to insist that move constructors be noexcept, but since these were added later there is code out there with move constructors that do throw exceptions...
Note by the way that the issue being described isn't strictly about std::move or move semantics in general, but more about the STL and containers like std::vector that have chosen to define behavior that makes noexcept move constructors necessary to be used when reallocating.
The difference is that in C one is supposed to do allocations and deallocations oneself. Then move semantics is just pointer assignment with, of course, the catch that one should make sure one does not do a double-free because ownership is implicit. In C++ ownership is indicated by types so one has to write more stuff to indicate the ownership.
> The difference is that in C one is supposed to do allocations and deallocations oneself
No, you should only use the heap if necessary.
The bigger issue in C is there is no concept of references, so if you want to modify memory, the only recourse is return-by-value or a pointer. Usually you see the latter, before return value optimization it was considered a waste of cycles to copy structs.
In the embedded world, its often the case you won't see a single malloc/free anywhere. Because sizes of inputs were often fixed and known at compile time for a particular configuration.
I have no problem with systems programming issues. That complexity is essential complexity inherent in the problem itself, regardless of language. I have a problem with C++’s accidental complexity. I find C much more tractable. It certainly has a few of its own footguns, but it has much less accidental complexity.
As the author of the FQA noted (Yosef K-something), in C++ its more the combinations of features which causes so many issues.
And here we see this principle rear its ugly head yet again. In this case, its the combination of exceptions, manual memory allocation and the desire to make things work efficiently - of which the move constructor was developed as a "solution"
I understand the individual rationales of C++ things but I lost the faith on the whole thing.
Same. I’ve read all the books. Written all these things at least a few times. It’s just not doable post C++11.
C++ is the high rocky mountain pass between the fertile great plains of C and the weird but ultimately survivable California of Rust.
> [std::move silently copies const values, because] If something is const, you can’t move from it by definition.
Whoever wrote that definition should have a thing or two to learn from Rust. Different language I know, but it proves that it wasn't needed to cause so much confussion and collectively so much time and performance lost.
Also, who writes rules like that and ends the day satisfied with the result? It seems unlikely to feel content with leaving huge footguns and being happy to push the Publish button. I'd rather not ship the feature than doing a half-assed work at it. Comparing attitudes on language development and additions, it makes me appreciate more the way it's done for the Go lang, even though it also has its warts and all.
There was no Rust in 2011.
Rust did exist in some form in 2011. Source: I ate lunch with part of the Rust team in 2011.
Some form, meaning not market relevant.
The point is not a comparison with Rust per-se, but the fact that a better implementation of the idea was mathematically and/or technically possible; and the personal opinion that such huge footguns that the language accumulates over the years are maybe signals of having needed more thought to them before they were considered ready.
e.g. if something as simple of a inconspicuous std::move in the wrong place can break the whole assumption about move semantics, then make that impossible to do, or at least do not make it the default happy path, before you consider it production ready. What the heck, at the very least ensure it will become a compiler warning?
Hence the mention to Go and how they follow exactly this path of extending discussion as long as needed, even if it takes 10 years, until a reasonable solution is found with maybe small gaps, but never huge ones such as those explained in this article (plus tens of others in any other text about the language)
It took 13 years to get C++11, actually.
Go's discussion is interesting, given how much programming language design history, and flaws of existing languages, they ignore to this day.
A bit more if we consider the "bugfixing" release that was C++14 :)
But yeah it makes sense, given how that was the jumpstart of the whole modernization of the language. I believe it was a big undertake that required the time it took. Still years have passed and footguns keep accumulating... it wouldn't hurt to have a mechanism to optionally drop the old cruft from the language. Otherwise everything stacks on top in the name of backwards compatibility, but at this pace, how will C++36 look like?
While I share the sentiment, compare C#14 with C# 1.0, Java 25 with Java 1.0, Python 3.14 with Python 1.0.
While C++ might be worse, when you have 300+ doing proposals every three years, others aren't safer from similar churn, even if on smaller volume, and trying to keep backwards compatibility going.
And we all know what happened in Python.
Also Rust editions contrary to what many think, only cover a specific set of language evolution scenarios, it isn't anything goes, nor there is support for binary libraries.
As for a better C++, contrary to C, where it is business as usual, there are actually people trying to sort things out on WG21, even if isn't as well as we would like to.
"Making C++ Safe, Healthy, and Efficient - CppCon 2025"
https://youtu.be/p52mNWsh-qs?si=AEs1dtr6_iknzATv
> nor there is support for binary libraries.
This is a weird call-out because it's both completely incorrect and completely irrelevant to the larger point.
Rust absolutely supports binary libraries. The only way to use a rust library with the current rust compiler is to first compile it to a binary format and then link to it.
More so than C++ where header files (and thus generics via templates) are textual.
Cargo, the most common build system for rust, insists on compiling every library itself (with narrow exceptions - that include for instance the precompiled standard library that is used by just about everyone). That's just a design choice of cargo, not the language.
Rust does, Rust editions do not have a story for them.
The story is that it must not matter which edition a library was compiled with - it's the boundary layer at which different editions interoperate with eachother.
a member of the c++ committee (herb sutter) is writing an compiler for an alternative c++ syntax [0], to c++, with the intent to restrict some semantics of the language for less UB, surprises, etc. i think less implementation-defined behavior is incredibly important; rvo vs std::move, dynamic function call optimization, i wish i didn't have to search asm to check for...
[0]: https://github.com/hsutter/cppfront
What’s the problem? It makes perfect sense to me that a const object cannot be moved from, since it violates the constness. Since constness goes hand in hand with thread safety you really don’t want that violation.
Maybe a compiler error that a const object cannot be “moved”?
That would force the programmer to remove the std::move, making it clear that its a copy.
There are cases where you would not want to reject such code, though. For example, if std::move() is called inside a template function where the type in some instantiations resolves to const T, and the intent is indeed for the value to be copied. If move may in some cases cause a compiler error, then you would need to write specializations that don't call it.
I didn’t think of that, but you are right. At some point I thought I understood templates r-value references work but now I’ve forgotten.
It's weird that they made a mistake of allowing this after having so many years to learn from their mistake about copies already being non-obvious (by that I mean that references and copies look identical at the call sites)
clang-tidy has a check for this case
To be honest I agree that it makes sense, at least if we put our hats of puritanism on the conceptual and semantical way of seeing it.
But having std::move silently fall back to a copy constructor is not a good solution.
Before move semantics the HeavyObject problem was solved in most cases by specializing std::swap for each container.
The design lesson I draw from this is that pursing a 100% general solution to a real problem is often worse than accepting a crude solution which covers the most important cases.
Regarding mistake 1: return std::move(local_var), it is worth clarifying why this is technically a pessimization beyond just breaking NRVO. It comes down to the change in C++17 regarding prvalues.
> Pre-C++17, a prvalue was a temporary object.
> Post-C++17, a prvalue is an initializer. It has no identity and occupies no storage until it is materialized.
Should have be called give(). But naming things correctly is hard, and the C++ committee is known to do a lot of things incorrectly
There is no giving (or taking).
I think std::rvalue would be the least confusing name.
That has about the same issue: like std::move it doesn't really explain that the receiver decides.
std::offer
Maybe a std::take to pair with it?
∆ That's actually quite accurate. I like it!
std::abandon() ?!
The name predates the standardisation. The committee did not come with the whole thing themselves, rather they adopted and expanded already existing library implementations. You could move in C++, with this exact name, long before C++11.
See, for example, this implementation https://stlab.adobe.com/group__move__related.html
std::movable
Maybe std::make_movable would have been a slightly better name, but it's so much simpler to write std::move.
Split the difference with std::moveable().
Also signals it doesn't actually move, while remaining just as fast to type.
But that misses too much of the semantics. It also implies ownership transfer, even if copied.
thanks to the incredible advances in terms of developer tooling over the last 50 years (i.e. tab-autocompletion) there should be no difference in writing those two.
There is a difference, lots of stuff starts with make_, so lots of possible completions.
std::rvalue
Sounds more like a contract thing. Of course std::move should be able to throw exceptions (like when it runs out of memory), but when it throws an exception it should still guarantee that memory is in a consistent state.
So the fault here is with std::vector who didn't write that contract.
I found the previous discussion and article very helpful
https://news.ycombinator.com/item?id=45799157 (87 comments)
> This code works. It compiles. It runs. But depending on how you’ve implemented your types, it might be performing thousands of expensive copy operations instead of cheap moves without you realizing it.
I've spent the last two decades in the .net platform. But for a decade or so before that I was a C++/Unix dev. I remember old style "C with classes" C++ as being fairly small and elegant, and approximately as easy to reason about as C# - albeit that you had the overhead of tracking object ownership and deallocation.
What the language has become now, boggles my mind. I get hints of elegance/power and innovation when I read about it, but the sheer number of footguns is astonishing. I'm very sure that I'm not clever enough to understand it.
But some very smart people have guided the language's evolution. So, what are the forces that have determined the current state of C++?
> So, what are the forces that have determined the current state of C++?
I'm very confident that the main driving factors are:
1. "performance" (not wanting to do more allocations than necessary)
2. abi compatibility
3. adding features without caring how well they integrate
Example for 1:
"emplace", you normally have "append" but emplace directly constructs the object in the container instead of having to be constructed first and then moved into the container. This nice and all but breaks when using pairs (for reasons you can google but I don't wanna explain here). So now you have these obscure classes like https://en.cppreference.com/w/cpp/utility/piecewise_construc... which solve this.
Example for 2:
Basically they never break the ABI and this leads to tons of old stuff hanging around and never being changed and just more stuff being added on top. std::iostream is famously slow and a big reason is because you can't fix it without breaking the abi which they don't wanna do.
Example for 3:
The whole template thing adds so much complexity it's bonkers, I think c++ without templates would be pretty manageable comparatively. For example because C++ has constructors and they don't quite mix well with templates you suddenly end up in the situation that you have 2 concepts: "normal" template argument deduction and constructor template argument deduction (CTAD). Because of this asymmetry you need a custom language feature called "deduction guides" to maneuver yourself out of the problems that come from this.
Or another short one: std::expected without something like the "!" that rust has. You end up with endless "if(result.has_value()) { return result; }" cascades and it's horribly unergonomic. So now we have a Result class but it's practically unusable that it will only fragment the ecosystem even more.
>Example for 1: ...breaks when using pairs
No, it doesn't. But sometimes you want to construct pair's elements in-place too and that's what piecewise_construct is for.
Note that C# 14 versus C# 1.0 isn't suffering from feature creap as well.
What has guided C++ are the 300+ volunteers that get to submit papers, travel around the world attending the meetings, and win the election rounds of what gets into the standard.
Unfortunately design by committee doesn't lead to a clear product roadmap.
> So, what are the forces that have determined the current state of C++?
A subset of the language aimed at library writers. As a user of those libraries all these weirdo features are likely to be transparent.
TFA explains how std::move is tricky to use and this is not a feature reserved for library writers
Of course it is not reserved for library writers - nothing is. But it is not a feature that application writers should worry about overmuch.
std::move is definitely for there for optimizing application code and is often used there. another silly thing you often see is people allocating something with a big sizeof on the stack and then std::moving it to the heap, as if it saves the copying
> another silly thing you often see is people allocating something with a big sizeof on the stack and then std::moving it to the heap, as if it saves the copying
never seen this - an example?
Naming things is hard.
I'm convinced naming things is equivalent to choosing the right abstraction, and caching things is creating a correct "view" from given normalized data.
The best way to think about it is that std::move is a cast.
https://stackoverflow.com/a/42340735
The article explains this.
I always understood move as moving ownership, so it's not a misnomer.
> std::move is like putting a sign on your object “I’m done with this, you can take its stuff.”
Which exactly is moving ownership.
std::move itself doesn't move ownership, though. It allows the compiler to transfer ownership to the receiver of the value, but it doesn't force it in any way. This is important, because it means YOU may still be the owner of a value even after you called std::move on it.
Not to mention, ownership in C++ is not entirely lost with moves in the traditional sense. For example, your code still has to destruct the object even if you did move it to somewhere else.
Std move doesn’t move ownership. It simply casts into something that could have its ownership taken. Whether or not that actually happens is impossible to identify statically and the value after ownership is consumed is unspecified - sometimes it’s UB to access the value again, sometimes it’s not.
That's quite inaccurate.
It needs to remain destructible, and if the type satisfies things like (move-)assignable/copyable, those still need to work as well.
For boxed types, it's likely to set them into some null state, in which case dereferencing them might be ill-formed, but it's a state that is valid for those types anyway.
Well it’s unspecified what empty/size return for collections after a move. Not a dereference, not UB but unspecified as I said. UB pops up in hand written code - I’ve seen it and the language doesn’t provide any protection here.
Thankfully clippy lints do exist here to help if you integrate that tooling
May be disown would be more descriptive, but the point is that it's intended for transferring of ownership versus copying data.
> it's intended for transferring of ownership versus copying data.
It's intended for transferring ownership, but what it actually does is mark the value as transferrable, whether or not the value is actually transferred is up to the callee.
After moving a value, it needs to remain in a "valid but unspecified state".
How do you mean accessing a valid object is UB?
"Validity" is an extremely low bar in C++, it just means operations with no preconditions are legal, which in the most general case may be limited to destruction (because non-destructive moves means destruction must always be possible).
>After moving a value, it needs to remain in a "valid but unspecified state".
No, it doesn't.
The standard library requires that for its classes, but not the language.
"Unless otherwise specified, such moved-from objects shall be placed in a valid but unspecified state."[0]
[0] https://timsong-cpp.github.io/cppwp/n4950/lib.types.movedfro...
Ok, fair enough.
So you're saying if you use the language to write UB, then you get UB?
Seems kinda circular. Ok, you're not the same user who said it can be UB. But what does it then mean to same "sometimes it's UB" if the code is all on the user side?
"Sometimes code is UB" goes for all user written code.
I mean the language doesn't dictate what post-condition your class has for move-ctor or move-assignment.
It could be
- "don't touch this object after move" (and it's UB if you do) or
- "after move the object is in valid but unspecified state" (and you can safely call only a method without precondition) or
- "after move the object is in certain state"
- or even crazy "make sure the object doesn't get destroyed after move" (it's UB if you call delete after move or the object was created on the stack and moved from).
But of course it's a good practice to mimic the standard library's contract, first of all for the sake of uniformity.
It is absolutely knowable statically if ownership will be taken. It's not necessarily very easy to do so, but the decision is 100% up to the compiler, as part of overload resolution and optimization choices (like the NRVO analysis that the article mentions). Since ownership is an inherently static concept, it doesn't even make sense to think about "runtime ownership".
My function can choose to move or not to move from an object based on io input.
Can you show an example of what you mean?
My claim is that, if I call `foo(std::move(myObj))`, it is statically knowable if `foo` receives a copy of `myObj` or whether it is moved to it. Of course, `foo` can choose to further copy or move the data it receives, but it can't choose later on if it's copied or not.
Now, if I give `foo` a pointer to myObj, it could of course choose to copy or move from it later and based on runtime info - but this is not the discussion we are having, and `std::move` is not involved from my side at all.
No, it is not statically knowable if it is actually moved.
Does not move `arg`. It's fairly easy to write code that assumes `std::move` moves the value, but that can lead to bugs. For example: The expectation is that `some_other_function` is always called with `{ 2 }`, but this will only happen if `some_function` actually moves `a`.> Can you show an example of what you mean?
This is exactly what I meant as irrelevant.
If I call `foo(std::move(my_unique_ptr))`, I know for sure, statically, that my_unique_ptr was moved from, as part of the function call process, and I can no longer access it. Whether `foo` chooses to further move from it is irrelevant.
The only thing that is statically known here is that you’re wrong. The function I posted only moves its parameter half the time, at random. You may want to treat it as moved-from either way, but factually that’s just not what is happening.
This is like trying to defend that you can't statically know the result of 1 + 2 because:
I can tell you for sure that the result of 1 + 2 will be 3.> This is like trying to defend that you can't statically know the result of 1 + 2
It is completely unlike that. tsimionescu is asserting that they can always know statically whether `foo` will move its parameter. The function I provided is a counter-example to that assertion.
Of course the branch body always moves, that's what it's there for. That has no bearing on the argument.
aside from what others wrote, it’s also non local - whether std::move even does anything is dependent on the signature of foo - if foo takes it by const& you may think you’ve transferred ownership when it hasn’t actually happened.
Yeah no.
In Rust if you pass say a Box<Goose> (not a reference, the actual object) into a function foo, it's gone, function foo might do something with that boxed goose or it might not, but it's gone anyway. If a Rust function foo wanted to give you it back they'd have to return the Box<Goose>
But C++ doesn't work that way, after calling foo my_unique_ptr is guaranteed to still exist, although for an actual unique_ptr it'll now be "disengaged" if foo moved from it. It has to still exist because C++ 98 (when C++ didn't have move semantics) says my_unique_ptr always gets destroyed at the end of its scope, so newer C++ versions also destroy my_unique_ptr for consistency, and so it must still exist or that can't work.
Creating that "hollowed out" state during a "move" operation is one of the many small leaks that cost C++ performance compared to Rust.
No: https://godbolt.org/z/d7f6MWcb5
Look, the act of calling std::move and and calling a function taking an rvalue reference in no way invokes a move constructor or move assignment. It does not "move".
It's still just a reference, albeit an rvalue reference. std::move and the function shape is about the type system, not moving.
(Edit: amusingly, inside the callee it's an lvalue reference, even though the function signature is that it can only take rvalue references. Which is why you need std::move again to turn the lvalue into rvalue if you want to give it to another function taking rvalue reference)
I didn't reply to this thread until now because I thought you may simply be disagreeing about what "move" means (I would say move constructor or move assignment called), but the comment I replied to makes a more straightforward factually incorrect claim, that can easily be shown in godbolt.
If you mean something else, please sketch something up in godbolt to illustrate your point. But it does sound like you're confusing "moving" with rvalue references.
Edit: for the move to happen, you have to actually move. E.g. https://godbolt.org/z/b8M495Exq
I don't understand the downvoted here. Either the compiler emits the code to call a move constructor or it doesn't.
Static analysis is about proving whether the code emitted by a compiler is actually called at runtime. It's not simply about the presence of that code.
Code can be emitted but never executed.
Well, no, because CAN take isn't the same as WILL take.
Changing something to an rvalue means it'll now match a move constructor, but there is no guarantee a move constructor will be used, even if defined, because you've got classes like std::vector that are picky and are explicitly looking for a noexcept move constructor.
In that sense, std::move() is no different than other passing semantics. Just because you wrote at the call site that you want to pass a copy of your object doesn't mean that the callee will actually make a copy of it.
I'm not sure what you are saying.
If we have foo(std::string a, std string b), and then call it like this:
std::string x;
std::string y;
foo(std::move(x), y);
The x will be moved into a, and y will be copied into b.
The callee has no say in this - it's just the compiler implementing the semantics of the language.
Personally I see std::move more like removing ownership because it’s not explicit from its call where the ownership is transferred.
Even that is a bit suspect, because ownership may well remain with you even after the call, so it's not really removed.
For example, this is perfectly valid C++, and it is guaranteed to have no issue:
I don’t think this is particularly insightful, as move semantics and r-values are higher level language semantics, nothing more and nothing less.
Rust’s borrow checker doesn’t actually borrow anything either, it’s operating on a similar level of abstraction.
> Rust’s borrow checker doesn’t actually borrow anything either
Why would it? It's called the borrow checker, not the borrower. So it checks that your borrows are valid.
std::move looks and feels like a function, but it doesn't do what it says, it makes objects movable but does never moves them (that's up to whatever is using the value afterwards). If you want something similar in Rust, Pin is a much better candidate.
Sure, but from the perspective of the code that has the move() its good to assume the value is moved at that call, which I guess was the intention of picking the name.