2y ago

On harmful overuse of std::move - The Old New Thing

devblogs.microsoft.com

22 comments

To me it's not really excitement so much as "oh god how do i make this ridiculous language do the correct thing".
I'm way more scared of rogue implicit copy constructors. I wonder if cppfront has any plan to fix problems like implicit copy and pessimising move.
- In Rust, making something copyable is always explicit. I like that a lot.
  
  Cpp should have done ref by default and had & for copy, but here we are.
  
  Copy has a very different meaning between the two languages. In rust the equivalent of a c++ copy is a clone() call for anything non trivial
- To me it’s not really excitement so much as “oh god how do i make this ridiculous language do the correct thing”.
  I think this all comes down to having the right mental model.
  In this case, I think it helps to know that:
  returning by value by default implies making a copy,
  ...but if the output type supports move semantics and the function returns a temporary value (rvalue) then the language is on your side and tries to move it instead of doing a copy.
  and here's the kicker: now the language goes the extra mile and looks after you to avoid that copy, and if your return value allows the language to apply an optimization then the language is on your side and eliminates any potential copy without you having to do anything special.
  The zany behavior is that if you set up your function to push for a move (i.e., force a local variable to be treated as a temporary), the language ceases to be able to apply its optimization.
  That's basically it. No real mystery.
  
  The parts that seem likely to cause this confusion (which I shared when I first started using C++11) are:
  Moves in C++ are always a potentially-destructive operation on a reference, not just a memcopy.
  Consequently, "moving" a temporary still requires having two separate instances of the type, despite that generally not being what you want, hence RVO.
  ...but move-semantics are generally presented and understood as an "optimization", and conceptually "take the guts of this value and re-use them as a new value" is both what RVO is doing and what move-semantics are doing.
  std::move isn't a compiler intrinsic and doesn't force a move operation; it's just a function that returns an r-value reference. So it makes it harder, not easier, for the compiler to "see through" and optimize away, even in the case where "as if" rule should make that legal.
I wonder if the language could be updated so these extra std::move invocations actually become harmless? return std::move is something that I see used quite a bit.
- @QuadriLiteral @lysdexic We've been looking at a paper just recently in Kona, where the author proposed to not penalize "unfortunate" uses of std::move. I think this is user friendly and you might imagine what I've been voting for.
  
  That sounds great!
- I wonder if the language could be updated so these extra std::move invocations actually become harmless? return std::move is something that I see used quite a bit.
  These std::move invocations are harmless, as they only cast objects to their rvalue reference.
  The destructive bit takes place in the type they are assigned to, as it invokes either a move constructor or a move assignment operator, and calling those implies that the object you just passed to std::move will be invalidated after the call and should not be used subsequently.
  
  @lysdexic @QuadriLiteral Eh, no. Really. Changing the value category disables RVO
  
  I mean harmless in a way that using std::move on the return type doesn't prevent RVO?

22 comments