A recursive parsing call stack has the overhead of pushing/popping return addresses and maybe a couple other registers. Maybe a bit more if there are security protections like stack cookies that didn't get optimized away, or if shadow stacks are enabled. It's not free, but is pretty cheap. A non-recursive use of std::stack dodges some of that, but instead has the overhead of dynamic allocation, which I'm generally going to consider more expensive - but does have the advantage of being more of a once-off cost - so perhaps it'd be cheaper for long but shallow JSON objects? The cost of the call stack vs std::stack is probably dwarfed by the cost of allocating arrays/objects/strings and decoding in either case.
If you want to over-optimize for that specific case anyways, you can take a hybrid approach with smaller stacks kept on the stack in a single linear root blob, with larger stacks falling back onto the heap, using a non-recursive style.
However, if you're capping JSON depths as an anti-DoS measure anyways (your parser isn't the only thing possibly vulnerable to blowing the stack via recursion - anything dealing with the parsed result is also in danger if it can be recursive), the heap fallback may be entirely dead code anyways... so why bother coding it?
If you want to over-optimize for that specific case anyways, you can take a hybrid approach with smaller stacks kept on the stack in a single linear root blob, with larger stacks falling back onto the heap, using a non-recursive style.
However, if you're capping JSON depths as an anti-DoS measure anyways (your parser isn't the only thing possibly vulnerable to blowing the stack via recursion - anything dealing with the parsed result is also in danger if it can be recursive), the heap fallback may be entirely dead code anyways... so why bother coding it?