[nlohmann]

The benchmark suite in tests/benchmarks covered only parsing, serialization and the binary formats. This adds 51 benchmarks for the rest of the everyday API, all on the existing test files:

benchmark	what it measures	inputs
SaxParse	sax_parse() with a handler that only counts events; the gap to ParseString is the cost of building the DOM	canada, citm_catalog, twitter
Copy, CopyAssign, CopyRing	copying a parsed document with the destruction of a copy timed as well: a temporary copy, an assignment over an earlier copy, and replacing the oldest of eight live copies	canada, citm_catalog, twitter
CopyOnly, DestroyOnly	only the copy or only the destruction, the other half with the timer paused; for telling the halves apart, not for comparing versions (see below)	canada, citm_catalog, twitter
Equal	== between two separate copies of a parsed document	canada, citm_catalog, twitter
Iterate	visiting every value with range-based for loops	canada, citm_catalog, twitter
Flatten, Unflatten	flatten() and unflatten()	canada, citm_catalog, twitter
FlattenOrdered	flatten() of an ordered_json (canada is left out: it takes minutes with 3.12.0)	citm_catalog, twitter
Diff, Patch	diff() against the document with every eighth scalar changed, and applying that patch	canada, citm_catalog, twitter
JsonPointerParse, JsonPointerToString, JsonPointerAt, JsonPointerContains	the JSON Pointer of every scalar: parsed from a string, turned back into a string, looked up with at() and contains()	citm_catalog, twitter
GetVectorInt, GetVectorDouble, GetMap	get<std::vector<std::int64_t>>(), get<std::vector<double>>(), get<std::map<std::string, json>>() (on the flattened document)	signed_ints, floats, citm_catalog
ToJsonStrings/copy, ToJsonStrings/move	creating a json array from a std::vector of 100,000 strings, copied or moved, and destroying it again (timed, for the reason given below)	generated

The benchmarks use only API that 3.12.0 also has, so releases can be compared on them. While preparing the 3.13.0 performance post, they found the unflatten(), operator==, get<std::map>() and destroy regressions against 3.12.0 that are now fixed in stack #5841.

Copies (and the arrays ToJsonStrings creates) are timed together with a destruction because how a document is freed shapes the heap the next copy allocates from. Since the faster destruction (#5762/#5767), part of the work it saves shows up in the next copy's allocations, so timing the copy alone made develop look slower than 3.12.0 (on glibc, citm_catalog 0.6× in a standalone loop) although copying and destroying together got faster. Medians against 3.12.0, develop header:

	macOS (Apple Clang)	Linux (GCC 13)
Copy	1.09–1.27×	1.09–1.52×
CopyAssign	1.05–1.26×	1.08–1.58×
CopyRing	1.06–1.27×	1.11–1.57×
CopyOnly	0.95–1.03×	0.83–0.99×
DestroyOnly	1.26–1.55×	1.28–2.12×

tests/benchmarks/README.md lists the new benchmarks.

There is no Accept benchmark here, because the json_view stack (#5627 and below) adds one. Merging this with that stack, #5828 or #5833 conflicts only in added #include lines, benchmark definitions and README rows; keep both sides.

Breaking changes

None. Only the benchmark suite changes; the library and its public API are untouched.

Testing

Built and run against the 3.12.0 header and against develop (Apple Clang, -O3 -flto, C++11), and as part of the 3.13.0 release candidate. The copy benchmarks were also run on Linux (GCC 13, x86-64) against 3.12.0, develop and the release candidate.

This PR was written by Claude Code on behalf of @nlohmann.

🤖 Generated with Claude Code