Memory management during copy

[malachib]

Firstly, let me say thank you for this awesome library @nlohmann !

OK. Question: When utilizing the copy (not move) assignment operator, what copy technique is employed? I can think of 3 possibilities:

• Deep copy (a complete reallocation of all keys, including subkeys)
• Shallow copy (just the one pointer)
• shared_ptr style - where there's a hidden ref counter at play

Any of these works for me, but I don't know which are at play! Thank you

Related: When I try to use the copy constructor in an initializer list, it tosses me into an array rabbit hole regardless of the const json& copy_from underlying type. Feels like a bug

[gregmarr]

The single assignment operator uses the by-value-parameter-and-swap idiom. It takes the RHS json object by value, and then swaps contents with that by-value parameter.

If the assignment operator is passed an rvalue, then the by-value parameter is constructed using the move constructor, otherwise it's constructed using the copy constructor. The copy constructor does a deep copy. The move constructor just steals the value.

The initializer list constructor can be a bit weird at times, and the behavior is actually different depending on which compiler you are using. #2311

[malachib]

If I understand you correctly, the single assignment operator behaves a lot like a move operation.

In a scenario of two json instances where a is empty and b has content, am I to understand a=b means a will have a shallow (pointer) copy of b and b will now be empty?

[gregmarr]

If I understand you correctly, the single assignment operator behaves a lot like a move operation.

Only when given an rvalue.

In a scenario of two json instances where a is empty and b has content, am I to understand a=b means a will have a shallow (pointer) copy of b and b will now be empty?

No, a = b; will be a copy. a = std::move(b); will behave as you described, and the same for a = makeJson() where makeJson() returns json.

[malachib]

That makes sense, makeJson() return would be an rvalue. As long as RVO is at play, I might be able to avoid some fuss.