What is that makes NVRO so much more difficult to implement? Why couldn't they mandate that just like RVO?
Do compilers literally just special case a simple return statement of a direct construction or something?
Is going to be much harder because you don't know are compile time which is returned and so cannot construct the one you return in the correct place. That is just off the top of my head, I'm not a compiler writer, I'm sure they have figured out the simple versions of the above, but you can start to see the complex versions that they can't.
N/RVO works by (at the machine language level, of course) rewriting the function signature to return void and take an extra pointer parameter, which is written to before returning. If you're returning a newly-constructed object, the compiler can rewrite that into calling the constructor on the pointer, but if you're returning a named object, the class may have a non-trivial destructor that needs to run after the move, such that it's not possible to rewrite uses of the local object into uses of the pointer.
I'm not too confident on that last part, because such an implementation would mess with semantics in case of an exception, so anyone feel free to correct me on that.
> N/RVO works by (at the machine language level, of course) rewriting the function signature to return void and take an extra pointer parameter
This sounds wrong, are you sure? Would you mind demonstrating with an example on godbolt? Whether NRVO applies or not, the ABI should be the same, AFAIK.
Yes, it works exactly like this, this is a demo on godbolt [0]. rdi stores the pointer in both cases, makeS1() uses RVO, makeS2() takes it explicitly and constructs with placement new.
I will say before testing this i didn't realize the RVO calling convention was to return the pointer you pass in, but apparently so. If makeS2() returned void, it's just a tail call to the constructor, but makeS1() has to spill rbx and use it to save the pointer.
Yes, of that I'm sure. This optimization is only possible if the compiler has control of both sides of a call. If the function may be callable from other translation units or modules I imagine it generates a thin wrapper that's externally callable.
The optimization is often possible even if the computer does not see the call, because most (all?) ABIs have always required hidden pointer parameters for class types with non-trivial destructors.
https://godbolt.org/z/9WvnEvEYh
Note how `std::unique_ptr<int>` effectively passed as a `int**`; and that the by-value unique_ptr is not destroyed at the end of the function -- destroying parameters is instead the caller's job (and commonly only happens at the end of the full expression containing the call -- though this choice is implementation-defined).
But that can only work if the caller can see the updated value of the parameter (to avoid double-free for `clear`) -> thus the need to pass the parameter by hidden pointer.
The point of (N)RVO is to directly construct the return value in-place at the calling frame. Which requires knowing what object will land there.
In RVO there is no problem because you know what object is the one you need to put there.
In NRVO there is a problem because you might have one of multiple objects being returned and you need to know which one to construct at the call site; it can't be all of them on top of each other. But you don't necessarily know at the time of construction whether that object will be the one that is actually returned. Doing so requires imperfect code analysis so the standard would need to define the complicated analyses to perform.
RVO is easy to detect since it happens only in expressions in return-statements.
NRVO requires the compiler to analyze the flow, like if 2 different variables/constructions can lead to the return (what one do we take, or can we do either later?).
Also, with RVO it's easy to detect and elide destruction calling for things going out of scope whilst NRVO would require more careful management of destruction order,etc.
Basically, NRVO touches a lot of things in "inconventient" places that can easily require reworking internal compiler structures to track destinations whilst RVO was probably far easier to just "hack in".
I figured any half decent compiler already do plenty of flow and liveness analysis on everything for register allocation, dead code elimination and what not.
Maybe it's the guaranteed elision that makes it a problem, like you can't fail the analysis, but then maybe you go the rust route - fail to compile and urge the programmer to rewrite their code so it accepts it.
Make it opt in with [[must_elide]] so old code still works I guess.
> What is that makes NVRO so much more difficult to implement?
I recall reading that at a high level RVO is implemented by treating the return value as an external object. In simple terms (simplistic terms) RVO then works by
- first instantiating the return variable,
- passing the var by reference to the function,
- and then use return value to actually initialize the variable passed by reference.
The moment there's some funny logic on what to write to that output value, the problem gets far more complex.
It did for sure, but the problem with C++ is its heritage, specifically that structures can be self-referential. For instance, the Rust's url::Url type has to use usize offsets for tracking the location of each of its components. Conversely, in C++, someone could have already created a similar Url type that would use std::string for the buffer and char pointers for the component locations. As such, you cannot simply memcpy from one struct into another and forget the former as std::string could have its own in-place storage and that would invalidate all pointers - you'll need to define a move constructor instead.
And the most important idea: destructive moves.
Since C++ doesn't track lifetimes it has to leave the object in a "valid state" after a move and the destructor still runs which has to have a check if it should do something or not.
std::move is a great tool when used correctly. However used incorrectly it makes code worse: more verbose and less performant. Since I have no idea how you are using it I can't comment on your experience. My experience is people (including me!) get it wrong fairly often. Fortunately tools can detect a lot of cases where you get it wrong.
I'm not too confident on that last part, because such an implementation would mess with semantics in case of an exception, so anyone feel free to correct me on that.
This sounds wrong, are you sure? Would you mind demonstrating with an example on godbolt? Whether NRVO applies or not, the ABI should be the same, AFAIK.
I will say before testing this i didn't realize the RVO calling convention was to return the pointer you pass in, but apparently so. If makeS2() returned void, it's just a tail call to the constructor, but makeS1() has to spill rbx and use it to save the pointer.
[0]: https://godbolt.org/z/ovd1n99P8
https://godbolt.org/z/9WvnEvEYh Note how `std::unique_ptr<int>` effectively passed as a `int**`; and that the by-value unique_ptr is not destroyed at the end of the function -- destroying parameters is instead the caller's job (and commonly only happens at the end of the full expression containing the call -- though this choice is implementation-defined). But that can only work if the caller can see the updated value of the parameter (to avoid double-free for `clear`) -> thus the need to pass the parameter by hidden pointer.
In RVO there is no problem because you know what object is the one you need to put there.
In NRVO there is a problem because you might have one of multiple objects being returned and you need to know which one to construct at the call site; it can't be all of them on top of each other. But you don't necessarily know at the time of construction whether that object will be the one that is actually returned. Doing so requires imperfect code analysis so the standard would need to define the complicated analyses to perform.
NRVO requires the compiler to analyze the flow, like if 2 different variables/constructions can lead to the return (what one do we take, or can we do either later?).
Also, with RVO it's easy to detect and elide destruction calling for things going out of scope whilst NRVO would require more careful management of destruction order,etc.
Basically, NRVO touches a lot of things in "inconventient" places that can easily require reworking internal compiler structures to track destinations whilst RVO was probably far easier to just "hack in".
Maybe it's the guaranteed elision that makes it a problem, like you can't fail the analysis, but then maybe you go the rust route - fail to compile and urge the programmer to rewrite their code so it accepts it.
Make it opt in with [[must_elide]] so old code still works I guess.
I recall reading that at a high level RVO is implemented by treating the return value as an external object. In simple terms (simplistic terms) RVO then works by
- first instantiating the return variable,
- passing the var by reference to the function,
- and then use return value to actually initialize the variable passed by reference.
The moment there's some funny logic on what to write to that output value, the problem gets far more complex.