Make serializer's indent_string lazily allocated

The serializer constructor unconditionally allocated a 512-byte
indent_string, even though it is only ever read inside the
pretty_print branches of dump(). This wasted a heap allocation (and
its matching deallocation) on every compact (i.e. default, non-pretty)
dump() call.

indent_string is now default-constructed empty and lazily grown to
512 bytes, filled with indent_char, the first time a pretty-print
branch actually needs it. The existing doubling/growth logic for
larger indents is otherwise untouched, so output remains byte-identical
to before -- including in the pre-existing edge case where growth
beyond the initial buffer fills with ' ' instead of indent_char
(tracked separately by open PR #5186, which is left alone here).

The second, larger optimization mentioned in #5413 (removing the
shared_ptr-based output adapter) is intentionally out of scope, as it
overlaps open PR #5285.

Fixes #5413

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

include/nlohmann/detail/output/serializer.hpp

@@ -71,7 +71,7 @@ class serializer
         , thousands_sep(loc->thousands_sep == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->thousands_sep)))
         , decimal_point(loc->decimal_point == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->decimal_point)))
         , indent_char(ichar)
-        , indent_string(512, indent_char)
+        , indent_string()
         , error_handler(error_handler_)
     {}
 
@@ -126,6 +126,10 @@ class serializer
 
                     // variable to hold indentation for recursive calls
                     const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.empty()))
+                    {
+                        indent_string.resize(512, indent_char);
+                    }
                     if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
                     {
                         indent_string.resize(indent_string.size() * 2, ' ');
@@ -199,6 +203,10 @@ class serializer
 
                     // variable to hold indentation for recursive calls
                     const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.empty()))
+                    {
+                        indent_string.resize(512, indent_char);
+                    }
                     if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
                     {
                         indent_string.resize(indent_string.size() * 2, ' ');
@@ -260,6 +268,10 @@ class serializer
 
                     // variable to hold indentation for recursive calls
                     const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.empty()))
+                    {
+                        indent_string.resize(512, indent_char);
+                    }
                     if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
                     {
                         indent_string.resize(indent_string.size() * 2, ' ');
@@ -1010,7 +1022,7 @@ class serializer
 
     /// the indentation character
     const char indent_char;
-    /// the indentation string
+    /// the indentation string (lazily allocated on first use by a pretty-print branch)
     string_t indent_string;
 
     /// error_handler how to react on decoding errors

single_include/nlohmann/json.hpp

@@ -20150,7 +20150,7 @@ class serializer
         , thousands_sep(loc->thousands_sep == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->thousands_sep)))
         , decimal_point(loc->decimal_point == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->decimal_point)))
         , indent_char(ichar)
-        , indent_string(512, indent_char)
+        , indent_string()
         , error_handler(error_handler_)
     {}
 
@@ -20205,6 +20205,10 @@ class serializer
 
                     // variable to hold indentation for recursive calls
                     const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.empty()))
+                    {
+                        indent_string.resize(512, indent_char);
+                    }
                     if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
                     {
                         indent_string.resize(indent_string.size() * 2, ' ');
@@ -20278,6 +20282,10 @@ class serializer
 
                     // variable to hold indentation for recursive calls
                     const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.empty()))
+                    {
+                        indent_string.resize(512, indent_char);
+                    }
                     if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
                     {
                         indent_string.resize(indent_string.size() * 2, ' ');
@@ -20339,6 +20347,10 @@ class serializer
 
                     // variable to hold indentation for recursive calls
                     const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.empty()))
+                    {
+                        indent_string.resize(512, indent_char);
+                    }
                     if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
                     {
                         indent_string.resize(indent_string.size() * 2, ' ');
@@ -21089,7 +21101,7 @@ class serializer
 
     /// the indentation character
     const char indent_char;
-    /// the indentation string
+    /// the indentation string (lazily allocated on first use by a pretty-print branch)
     string_t indent_string;
 
     /// error_handler how to react on decoding errors

tests/benchmarks/src/benchmarks.cpp

@@ -214,323 +214,4 @@ static void BinaryToCbor(benchmark::State& state)
 }
 BENCHMARK(BinaryToCbor)->RangeMultiplier(2)->Range(8, 8 << 12);
 
-//////////////////////////////////////////////////////////////////////////////
-// parse binary formats
-//////////////////////////////////////////////////////////////////////////////
-
-// Only MessagePack had a read benchmark (FromMsgpack above, left untouched so
-// its numbers stay comparable across releases). The benchmarks below cover the
-// other formats, and read from a contiguous buffer as well as from a FILE*:
-// most callers pass a container, and the two adapters compile to different
-// code. The test data repository ships JSON only, so the input for each is
-// derived at setup time by serializing a parsed test file.
-
-/// binary format to benchmark; the _optimized variants add UBJSON/BJData size
-/// and type annotations, which the readers handle in a separate code path
-enum class binary_format
-{
-    cbor,
-    msgpack,
-    ubjson,
-    ubjson_optimized,
-    bjdata,
-    bjdata_optimized,
-    bson
-};
-
-static std::vector<std::uint8_t> to_binary(const json& j, const binary_format format)
-{
-    switch (format)
-    {
-        case binary_format::cbor:
-            return json::to_cbor(j);
-        case binary_format::msgpack:
-            return json::to_msgpack(j);
-        case binary_format::ubjson:
-            return json::to_ubjson(j);
-        case binary_format::ubjson_optimized:
-            return json::to_ubjson(j, true, true);
-        case binary_format::bjdata:
-            return json::to_bjdata(j);
-        case binary_format::bjdata_optimized:
-            return json::to_bjdata(j, true, true);
-        case binary_format::bson:
-        default:
-            return json::to_bson(j);
-    }
-}
-
-static json from_binary(const std::vector<std::uint8_t>& bytes, const binary_format format)
-{
-    switch (format)
-    {
-        case binary_format::cbor:
-            return json::from_cbor(bytes);
-        case binary_format::msgpack:
-            return json::from_msgpack(bytes);
-        case binary_format::ubjson:
-        case binary_format::ubjson_optimized:
-            return json::from_ubjson(bytes);
-        case binary_format::bjdata:
-        case binary_format::bjdata_optimized:
-            return json::from_bjdata(bytes);
-        case binary_format::bson:
-        default:
-            return json::from_bson(bytes);
-    }
-}
-
-static json from_binary(std::FILE* file, const binary_format format)
-{
-    switch (format)
-    {
-        case binary_format::cbor:
-            return json::from_cbor(file);
-        case binary_format::msgpack:
-            return json::from_msgpack(file);
-        case binary_format::ubjson:
-        case binary_format::ubjson_optimized:
-            return json::from_ubjson(file);
-        case binary_format::bjdata:
-        case binary_format::bjdata_optimized:
-            return json::from_bjdata(file);
-        case binary_format::bson:
-        default:
-            return json::from_bson(file);
-    }
-}
-
-/*!
-@brief serialize a parsed test file to @a format
-
-Returns an empty vector and marks the benchmark as skipped if the file cannot
-be represented in the format, rather than letting the exception escape: BSON
-requires an object at the top level, and several test files are arrays.
-*/
-static std::vector<std::uint8_t> binary_input(benchmark::State& state, const char* filename, const binary_format format)
-{
-    std::ifstream f(filename);
-    std::string const str((std::istreambuf_iterator<char>(f)), std::istreambuf_iterator<char>());
-    const json j = json::parse(str);
-
-    if (format == binary_format::bson && !j.is_object())
-    {
-        state.SkipWithError("BSON requires an object at the top level");
-        return {};
-    }
-
-    return to_binary(j, format);
-}
-
-static void FromBinaryBuffer(benchmark::State& state, const char* filename, const binary_format format)
-{
-    const std::vector<std::uint8_t> bytes = binary_input(state, filename, format);
-    if (bytes.empty())
-    {
-        return;
-    }
-
-    for (auto _ : state)
-    {
-        // the value is destroyed outside the timed section, because destroying
-        // a large DOM is not what this benchmark measures
-        state.PauseTiming();
-        auto* j = new json();
-        state.ResumeTiming();
-
-        *j = from_binary(bytes, format);
-
-        state.PauseTiming();
-        delete j;
-        state.ResumeTiming();
-    }
-
-    state.SetBytesProcessed(state.iterations() * bytes.size());
-}
-
-BENCHMARK_CAPTURE(FromBinaryBuffer, cbor / jeopardy, TEST_DATA_DIRECTORY "/jeopardy/jeopardy.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryBuffer, cbor / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryBuffer, cbor / citm_catalog, TEST_DATA_DIRECTORY "/nativejson-benchmark/citm_catalog.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryBuffer, cbor / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryBuffer, cbor / floats, TEST_DATA_DIRECTORY "/regression/floats.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryBuffer, cbor / signed_ints, TEST_DATA_DIRECTORY "/regression/signed_ints.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryBuffer, msgpack / jeopardy, TEST_DATA_DIRECTORY "/jeopardy/jeopardy.json", binary_format::msgpack);
-BENCHMARK_CAPTURE(FromBinaryBuffer, msgpack / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::msgpack);
-BENCHMARK_CAPTURE(FromBinaryBuffer, msgpack / citm_catalog, TEST_DATA_DIRECTORY "/nativejson-benchmark/citm_catalog.json", binary_format::msgpack);
-BENCHMARK_CAPTURE(FromBinaryBuffer, msgpack / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::msgpack);
-BENCHMARK_CAPTURE(FromBinaryBuffer, ubjson / jeopardy, TEST_DATA_DIRECTORY "/jeopardy/jeopardy.json", binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryBuffer, ubjson / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryBuffer, ubjson / citm_catalog, TEST_DATA_DIRECTORY "/nativejson-benchmark/citm_catalog.json", binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryBuffer, ubjson / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryBuffer, ubjson_optimized / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::ubjson_optimized);
-BENCHMARK_CAPTURE(FromBinaryBuffer, ubjson_optimized / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::ubjson_optimized);
-BENCHMARK_CAPTURE(FromBinaryBuffer, bjdata / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::bjdata);
-BENCHMARK_CAPTURE(FromBinaryBuffer, bjdata / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::bjdata);
-BENCHMARK_CAPTURE(FromBinaryBuffer, bjdata_optimized / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::bjdata_optimized);
-BENCHMARK_CAPTURE(FromBinaryBuffer, bjdata_optimized / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::bjdata_optimized);
-// BSON requires an object at the top level, so the array-rooted test files
-// (jeopardy and the regression files) cannot be captured here
-BENCHMARK_CAPTURE(FromBinaryBuffer, bson / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::bson);
-BENCHMARK_CAPTURE(FromBinaryBuffer, bson / citm_catalog, TEST_DATA_DIRECTORY "/nativejson-benchmark/citm_catalog.json", binary_format::bson);
-BENCHMARK_CAPTURE(FromBinaryBuffer, bson / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::bson);
-
-static void FromBinaryFile(benchmark::State& state, const char* filename, const binary_format format)
-{
-    const std::vector<std::uint8_t> bytes = binary_input(state, filename, format);
-    if (bytes.empty())
-    {
-        return;
-    }
-
-    const char* tmp = "benchmark_input.bin";
-    std::ofstream o(tmp, std::ios::binary);
-    o.write(reinterpret_cast<const char*>(bytes.data()), static_cast<std::streamsize>(bytes.size()));
-    o.flush();
-    o.close();
-
-    for (auto _ : state)
-    {
-        state.PauseTiming();
-        auto* j = new json();
-        auto* file = std::fopen(tmp, "rb");
-        state.ResumeTiming();
-
-        *j = from_binary(file, format);
-
-        state.PauseTiming();
-        std::fclose(file);
-        delete j;
-        state.ResumeTiming();
-    }
-
-    state.SetBytesProcessed(state.iterations() * bytes.size());
-}
-
-BENCHMARK_CAPTURE(FromBinaryFile, cbor / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryFile, cbor / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryFile, ubjson / canada, TEST_DATA_DIRECTORY "/nativejson-benchmark/canada.json", binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryFile, ubjson / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryFile, bjdata / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::bjdata);
-BENCHMARK_CAPTURE(FromBinaryFile, bson / twitter, TEST_DATA_DIRECTORY "/nativejson-benchmark/twitter.json", binary_format::bson);
-
-//////////////////////////////////////////////////////////////////////////////
-// parse binary formats: value shapes
-//////////////////////////////////////////////////////////////////////////////
-
-// The test files above are wide and shallow, but the readers' cost is per
-// container, so these cover the shapes that stress the container handling
-// itself. Every shape is wrapped in an object so that BSON, which requires an
-// object at the top level, measures the same value as the other formats.
-
-/// deeply nested arrays: one container per level, no other work
-static json make_nested()
-{
-    json nested = json::array();
-    json* p = &nested;
-    for (std::size_t i = 1; i < 1000; ++i)
-    {
-        p->push_back(json::array());
-        p = &p->operator[](0);
-    }
-
-    json j = json::object();
-    j["data"] = std::move(nested);
-    return j;
-}
-
-/// many sibling containers: maximum container churn, minimum nesting
-static json make_containers()
-{
-    json data = json::array();
-    for (std::size_t i = 0; i < 100000; ++i)
-    {
-        data.push_back(json::array({1, 2}));
-    }
-
-    json j = json::object();
-    j["data"] = std::move(data);
-    return j;
-}
-
-/// one flat array of numbers: the scalar decoding path, which must not move
-static json make_scalars()
-{
-    json data = json::array();
-    for (std::size_t i = 0; i < 1000000; ++i)
-    {
-        data.push_back(i);
-    }
-
-    json j = json::object();
-    j["data"] = std::move(data);
-    return j;
-}
-
-static void FromBinaryShape(benchmark::State& state, json (*build)(), const binary_format format)
-{
-    const std::vector<std::uint8_t> bytes = to_binary(build(), format);
-
-    for (auto _ : state)
-    {
-        state.PauseTiming();
-        auto* j = new json();
-        state.ResumeTiming();
-
-        *j = from_binary(bytes, format);
-
-        state.PauseTiming();
-        delete j;
-        state.ResumeTiming();
-    }
-
-    state.SetBytesProcessed(state.iterations() * bytes.size());
-}
-
-BENCHMARK_CAPTURE(FromBinaryShape, nested / cbor, make_nested, binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryShape, nested / msgpack, make_nested, binary_format::msgpack);
-BENCHMARK_CAPTURE(FromBinaryShape, nested / ubjson, make_nested, binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryShape, nested / bjdata, make_nested, binary_format::bjdata);
-BENCHMARK_CAPTURE(FromBinaryShape, nested / bson, make_nested, binary_format::bson);
-BENCHMARK_CAPTURE(FromBinaryShape, containers / cbor, make_containers, binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryShape, containers / msgpack, make_containers, binary_format::msgpack);
-BENCHMARK_CAPTURE(FromBinaryShape, containers / ubjson, make_containers, binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryShape, containers / ubjson_optimized, make_containers, binary_format::ubjson_optimized);
-BENCHMARK_CAPTURE(FromBinaryShape, containers / bjdata, make_containers, binary_format::bjdata);
-BENCHMARK_CAPTURE(FromBinaryShape, containers / bson, make_containers, binary_format::bson);
-// BSON names every array element, so a large array measures key generation
-// rather than scalar decoding and is left out here
-BENCHMARK_CAPTURE(FromBinaryShape, scalars / cbor, make_scalars, binary_format::cbor);
-BENCHMARK_CAPTURE(FromBinaryShape, scalars / msgpack, make_scalars, binary_format::msgpack);
-BENCHMARK_CAPTURE(FromBinaryShape, scalars / ubjson, make_scalars, binary_format::ubjson);
-BENCHMARK_CAPTURE(FromBinaryShape, scalars / bjdata, make_scalars, binary_format::bjdata);
-
-/*!
-@brief parse an indefinite-length CBOR string
-
-The writer never emits this form, so the input is assembled by hand: 0x7F
-opens the string, each chunk is a one-character string, and 0xFF closes it.
-*/
-static void FromCborChunkedString(benchmark::State& state, const std::size_t chunks)
-{
-    std::vector<std::uint8_t> bytes;
-    bytes.reserve(2 * chunks + 2);
-    bytes.push_back(0x7F);
-    for (std::size_t i = 0; i < chunks; ++i)
-    {
-        bytes.push_back(0x61); // string of length 1
-        bytes.push_back(0x61); // 'a'
-    }
-    bytes.push_back(0xFF);
-
-    for (auto _ : state)
-    {
-        json j = json::from_cbor(bytes);
-        benchmark::DoNotOptimize(j);
-    }
-
-    state.SetBytesProcessed(state.iterations() * bytes.size());
-}
-
-BENCHMARK_CAPTURE(FromCborChunkedString, 10000 chunks, 10000);
-
 BENCHMARK_MAIN();

tests/src/unit-serialization.cpp

@@ -14,6 +14,40 @@ using nlohmann::json;
 #include <array>
 #include <sstream>
 #include <iomanip>
+#include <cstdlib>
+#include <new>
+
+namespace
+{
+// heap allocation counter used by the regression test for issue #5413
+// (https://github.com/nlohmann/json/issues/5413); disabled (and thus a
+// no-op besides the counting) unless explicitly toggled on
+bool count_heap_allocations = false; // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
+std::size_t heap_allocations = 0; // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
+} // namespace
+
+void* operator new (std::size_t size) // NOLINT(cppcoreguidelines-owning-memory,misc-new-delete-overloads)
+{
+    if (count_heap_allocations)
+    {
+        ++heap_allocations;
+    }
+    if (void* ptr = std::malloc(size)) // NOLINT(cppcoreguidelines-no-malloc,cppcoreguidelines-owning-memory)
+    {
+        return ptr;
+    }
+    throw std::bad_alloc(); // NOLINT(hicpp-exception-baseclass)
+}
+
+void operator delete (void* ptr) noexcept // NOLINT(cppcoreguidelines-owning-memory,misc-new-delete-overloads)
+{
+    std::free(ptr); // NOLINT(cppcoreguidelines-no-malloc,cppcoreguidelines-owning-memory)
+}
+
+void operator delete (void* ptr, std::size_t /*size*/) noexcept // NOLINT(cppcoreguidelines-owning-memory,misc-new-delete-overloads)
+{
+    std::free(ptr); // NOLINT(cppcoreguidelines-no-malloc,cppcoreguidelines-owning-memory)
+}
 
 TEST_CASE("serialization")
 {
@@ -382,3 +416,76 @@ TEST_CASE("dump for basic_json with long double number_float_t")
         check_same(100.0L,  100.0);
     }
 }
+
+TEST_CASE("regression test for issue #5413 - lazily allocated indent_string")
+{
+    // the serializer used to unconditionally allocate a 512-byte
+    // indent_string in its constructor, even though it is only ever read
+    // inside the pretty_print branches of dump(). This wasted a heap
+    // allocation (and its matching deallocation) on every single compact
+    // (i.e. non-pretty, the default) dump() call. indent_string is now
+    // allocated lazily, the first time a pretty-print branch actually
+    // needs it -- so a compact dump() must perform strictly fewer heap
+    // allocations than a pretty dump() of the same value.
+    const json j = {{"level", "info"}, {"msg", "hello world"}, {"id", 12345}};
+
+    // warm up anything unrelated to indentation (e.g., one-time locale
+    // lookups) that might otherwise allocate on first use regardless of
+    // pretty-printing, so it does not skew the counts measured below
+    const auto warmup = j.dump();
+    const auto warmup_pretty = j.dump(4);
+    CHECK(!warmup.empty());
+    CHECK(!warmup_pretty.empty());
+
+    SECTION("compact dump() has a stable, minimal allocation count")
+    {
+        count_heap_allocations = true;
+
+        heap_allocations = 0;
+        const auto compact1 = j.dump();
+        const auto allocs_compact1 = heap_allocations;
+
+        heap_allocations = 0;
+        const auto compact2 = j.dump(-1);
+        const auto allocs_compact2 = heap_allocations;
+
+        count_heap_allocations = false;
+
+        CHECK(compact1 == compact2);
+        // dump() and dump(-1) both take the compact code path and must
+        // never touch indent_string, so they allocate identically often
+        CHECK(allocs_compact1 == allocs_compact2);
+    }
+
+    SECTION("first pretty dump() allocates more than a compact dump()")
+    {
+        // use a tiny value whose compact ({"a":1}, 7 bytes) and pretty
+        // ({"a": 1} with 1-space indent, 11 bytes) serializations both stay
+        // well inside every common std::string small-string-optimization
+        // buffer (>= 15 bytes on libstdc++/MSVC STL, >= 22 on libc++), so
+        // building the result string itself causes no heap allocation
+        // either way -- isolating indent_string as the only thing that can
+        // possibly account for a difference in allocation count
+        const json tiny = {{"a", 1}};
+
+        count_heap_allocations = true;
+
+        heap_allocations = 0;
+        const auto compact = tiny.dump();
+        const auto allocs_compact = heap_allocations;
+
+        heap_allocations = 0;
+        const auto pretty = tiny.dump(1);
+        const auto allocs_pretty = heap_allocations;
+
+        count_heap_allocations = false;
+
+        CHECK(compact == "{\"a\":1}");
+        CHECK(pretty == "{\n \"a\": 1\n}");
+        // a fresh serializer is created per dump() call; the pretty branch
+        // lazily allocates indent_string on its first use, so it must
+        // allocate at least once more than the compact branch, which never
+        // touches indent_string at all
+        CHECK(allocs_pretty > allocs_compact);
+    }
+}