Scan json_view strings with SIMD and index large objects (#5628)

* Scan json_view strings with SIMD and index large objects

Speed up json_view's parser with SIMD scanning and a hash table
for large objects.

Long runs of string bytes are scanned 16 bytes at a time with NEON
(AArch64, GCC and Clang) and SSE2 (x86-64), both baseline
instruction sets. Keys keep 16 table checks before the vector
loop, because their lengths repeat from record to record; string
values get 8, because their lengths vary more. Non-ASCII text is
validated 16 bytes at a time with simdjson's "lookup4" check
(Keiser and Lemire, 2021), with NEON on AArch64 and, on x86-64,
with SSSE3. SSSE3 is not part of baseline x86-64, so the check is
compiled for SSSE3 with a function attribute and used only where
CPUID reports it, which all x86-64 CPUs since about 2011 do; the
answer is cached in a statically initialized atomic, so there is
no guard of a local static and no global constructor. The same
input is accepted either way. JSON_VIEW_NO_SIMD selects the
portable code.

On x86-64, string runs are now checked vector-first: one SSE2
compare from the first byte finds the end of most keys and short
values, instead of a branch per byte for the first 8-16 bytes.
AArch64 keeps the byte-wise steps, where a NEON mask costs more and
the branches predict well. Entering an object or array no longer
stalls: open() stores the parent's frame field by field instead of
building it on the stack and reading it back with wider loads,
which waited for the narrower stores to retire.

Objects with 128 members or more get an open-addressing hash table
built when the object closes, so operator[], at(), find(),
contains(), count(), value(), and JSON pointers take constant time
on average in such objects; of duplicate keys, the first is kept,
as for the linear search. The idea comes from Boost.JSON.

simdjson is credited in simd.hpp's SPDX block, the README, and
license.md.

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

* Fix GCC module build and MSVC C4127 in the json_view SIMD code

GCC ignores the target attribute in modules, so the SSSE3 dispatch is
disabled for the module interface (the check stays portable, SSE2 is kept).
Make the 8-vs-16 byte unrolling condition in scan_string_run a
preprocessor/template split to avoid a constant condition (C4127).

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

* Remove an accidentally committed GCC module cache

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

* Bound the probe length of the large-object hash index

The key hash is not seeded, so keys chosen to collide made building the table quadratic (20,000 colliding keys took 470 ms to parse). A key may now sit at most 64 slots from its home slot; if a key would sit further away, the table is dropped and the object is searched linearly. Lookups stop after the same distance.

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

* Release the spare capacity of the hash index in shrink_to_fit

shrink_to_fit() now trims the tables of large objects like the node array and the decoded strings, and the list of large objects is released as soon as the tables are built.

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

* Use the portable string scan on non-x86 targets that define __SSE2__

WebAssembly with -msse2 defines __SSE2__, but has no <cpuid.h> or x86 intrinsics headers.

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

* Document the hash index number in node::extra

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

* Keep the last of duplicate keys in the object hash index

Lookups in objects with 128 members or more now return the last member of
a repeated key, like the linear search of smaller objects and like
materialize() and basic_json::parse(). build_object_index let the first
occurrence win, so the same text gave different results depending on the
size of the object.

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

* Keep the first of duplicate keys in the object hash index again

Lookups in objects with 128 members or more return the first member of a
repeated key again, like the linear search of smaller objects. This undoes
the code change of a69542046; its test now expects the first member from
lookups (with and without a table) and the last value from materialize()
and basic_json::parse().

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

* Fix clang-tidy 22 findings in the json_view headers

Silence bugprone-casting-through-void for the SSE loads (a reinterpret_cast
would trip -Wcast-align=strict), use auto for a cast initialiser, add
parentheses to a mixed expression, and name bugprone-std-namespace-modification
in the NOLINTs of the tuple_size/tuple_element specialisations.

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

* Fix clang-tidy 22 findings in unit-json_view.cpp and unit-json_view_builder.cpp

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

* Avoid GCC useless-cast and strict-overflow warnings in unit-json_view.cpp

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

* Index strings with size_t in the colliding-keys test (MSVC C4244 on Win32)

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

* Make json_view.hpp pass include-what-you-use

ci_single_binaries runs IWYU with --error on every header. For json_view.hpp
it suggested adding <array> (std::array is used), the headers that json.hpp
already provides (abi_config, abi_macros, input_adapters, json_pointer,
cpp_future, string_concat, value_t, json_fwd), and <version> for
std::nullptr_t, and removing <cstddef>.

- include <array>
- keep <cstddef> (nullptr_t, size_t; IWYU attributes them to <version> and <cstring>)
- tell IWYU not to suggest the headers that json.hpp provides: the amalgamated
  json_view.hpp only includes json.hpp, so including them here would duplicate
  their definitions

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

* Move the dump, comparison, and large-object tests to unit-json_view_dump.cpp

The dump and comparison tests moved to their own file in the dump pull
request; the hash index tests of this branch join them there.

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

* Regenerate the amalgamated headers

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

* Follow the renames from the review of the view builder

The review of #5620 renamed document_data's tape/tape_size/tape_cap to
nodes/node_count/node_capacity and made the decode arena a
std::vector<char>; adapt the code of this change to it.

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

* Rename slots in the object index for Qt compatibility

Qt defines slots as a macro, so a Qt header included before
json_view.hpp broke the declarations of the local variable (see #5865).
The test variable is renamed as well.

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

* Address the review of the json_view object index and SIMD scan

From gregmarr's review of #5628:

- Docs and comments say that the hash index is one of the object's keys.
- object_index.hpp: conditional_static_cast for the hash, "slot" for the
  table position, no flag for a degenerate table, and a static_assert
  that the smallest indexed object is larger than the probe window.
- scan.hpp unrolls with NLOHMANN_VIEW_REPEAT_FIRST_8/SECOND_8.

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

* Regenerate the amalgamated headers

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

* Encode JSON_VIEW_NO_SIMD and JSON_VIEW_USE_SSSE3 in the ABI namespace

Both macros select between definitions of the same inline functions, so
translation units that disagree on them violated the ODR. Like the other
ABI-affecting macros, they now add a tag to the inline namespace (_vns and
_vssse3), and the docs replace the "Define consistently" warning with the
usual ABI compatibility note. From gregmarr's review of #5628.

nlohmann_json.natvis covers twice as many tags now, so it lists four times
as many namespaces (14.7 MB instead of 3.4 MB).

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

* Regenerate the amalgamated headers

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

* Do not include type_traits.hpp directly in the object index

json.hpp, which object_index.hpp already includes, provides it; included
directly, the json_view amalgamation (which only treats nlohmann/json.hpp
as external) copied the json internals into json_view.hpp a second time,
so the fuzzers, which include both single headers, failed with
redefinitions (make_void, integer_sequence, priority_tag, ...).

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

* Compare keys with key_equals() in the object hash index

As the linear lookups do since the second review round of #5622.

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

* Regenerate the amalgamated headers

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

---------

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

include/nlohmann/detail/abi_macros.hpp

@@ -122,14 +122,26 @@
     #define NLOHMANN_JSON_ABI_TAG_DISABLE_UNLOCKED_FILE_IO
 #endif
 
+#ifdef JSON_VIEW_NO_SIMD
+    #define NLOHMANN_JSON_ABI_TAG_VIEW_NO_SIMD _vns
+#else
+    #define NLOHMANN_JSON_ABI_TAG_VIEW_NO_SIMD
+#endif
+
+#ifdef JSON_VIEW_USE_SSSE3
+    #define NLOHMANN_JSON_ABI_TAG_VIEW_USE_SSSE3 _vssse3
+#else
+    #define NLOHMANN_JSON_ABI_TAG_VIEW_USE_SSSE3
+#endif
+
 #ifndef NLOHMANN_JSON_NAMESPACE_NO_VERSION
     #define NLOHMANN_JSON_NAMESPACE_NO_VERSION 0
 #endif
 
 // Construct the namespace ABI tags component
-#define NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c, d, e, f, g, h, i, j) json_abi ## a ## b ## c ## d ## e ## f ## g ## h ## i ## j
-#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b, c, d, e, f, g, h, i, j) \
-    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c, d, e, f, g, h, i, j)
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c, d, e, f, g, h, i, j, k, l) json_abi ## a ## b ## c ## d ## e ## f ## g ## h ## i ## j ## k ## l
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b, c, d, e, f, g, h, i, j, k, l) \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c, d, e, f, g, h, i, j, k, l)
 
 #define NLOHMANN_JSON_ABI_TAGS                                       \
     NLOHMANN_JSON_ABI_TAGS_CONCAT(                                   \
@@ -142,7 +154,9 @@
             NLOHMANN_JSON_ABI_TAG_STRICT_BINARY_UTF8,                \
             NLOHMANN_JSON_ABI_TAG_OBJECTS_FOR_ENUM_KEYED_MAPS,       \
             NLOHMANN_JSON_ABI_TAG_DISABLE_STREAM_BULK_SCAN,          \
-            NLOHMANN_JSON_ABI_TAG_DISABLE_UNLOCKED_FILE_IO)
+            NLOHMANN_JSON_ABI_TAG_DISABLE_UNLOCKED_FILE_IO,          \
+            NLOHMANN_JSON_ABI_TAG_VIEW_NO_SIMD,                      \
+            NLOHMANN_JSON_ABI_TAG_VIEW_USE_SSSE3)
 
 // Construct the namespace version component
 #define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch) \

include/nlohmann/detail/view/builder.hpp

@@ -126,6 +126,13 @@ class builder
     frame_stack m_shallow;
     std::vector<frame> m_deep{};
 
+    /// remember an object to index after parsing (out of line, so that the
+    /// parse loop only has a call for it)
+    NLOHMANN_VIEW_NOINLINE void note_large_object(std::uint32_t idx)
+    {
+        m_doc.large_objects.push_back(idx);
+    }
+
     NLOHMANN_VIEW_NOINLINE bool fail(error_code c, const unsigned char* at) noexcept
     {
         m_failure.code = c;
@@ -539,7 +546,7 @@ class builder
     switch (cur())                                                              \
     {                                                                           \
         case '"':                                                               \
-            if (NLOHMANN_VIEW_UNLIKELY(!string())) { return false; }            \
+            if (NLOHMANN_VIEW_UNLIKELY(!string<true>())) { return false; }      \
             break;                                                              \
         case '{':                                                               \
             open(value_t::object);                                              \
@@ -623,7 +630,7 @@ obj_key:
             {
                 return fail(error_code::expected_key);
             }
-            if (NLOHMANN_VIEW_UNLIKELY(!string()))
+            if (NLOHMANN_VIEW_UNLIKELY(!string<false>()))
             {
                 return false;
             }
@@ -665,17 +672,25 @@ obj_next:
                 if (enabled(TrailingCommas) && cur() == '}')
                 {
                     ++p;
-                    goto close_container;
+                    goto close_object;
                 }
                 goto obj_key;
             }
             if (cur() == '}')
             {
                 ++p;
-                goto close_container;
+                goto close_object;
             }
             return fail(error_code::expected_object_end);
 
+close_object:
+            // a large object gets a hash index for its keys (objects only, so that closing
+            // an array pays nothing for this)
+            if (NLOHMANN_VIEW_UNLIKELY(cur_count >= document_data::index_min_members))
+            {
+                cold.note_large_object(cur_idx);
+            }
+
 close_container:
             close();
             if (NLOHMANN_VIEW_UNLIKELY(depth == 0))
@@ -728,7 +743,7 @@ root_done:
             switch (cur())
             {
                 case '"':
-                    return string();
+                    return string<true>();
                 case 't':
                     return literal_true();
                 case 'f':
@@ -867,14 +882,19 @@ indent_done:
             const auto idx = static_cast<std::uint32_t>(n - base);
             if (depth != 0)
             {
-                const frame f = {cur_idx, cur_count, cur_is_object};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    cold.m_shallow[depth - 1] = f;
+                    // field by field: a frame put together on the stack and
+                    // copied would be read back wider than it was written,
+                    // and that load waits until the stores are done
+                    frame& f = cold.m_shallow[depth - 1];
+                    f.idx = cur_idx;
+                    f.count = cur_count;
+                    f.is_object = cur_is_object;
                 }
                 else
                 {
-                    cold.m_deep.push_back(f);
+                    cold.m_deep.push_back(frame{cur_idx, cur_count, cur_is_object});
                 }
             }
             ++depth;
@@ -891,19 +911,21 @@ indent_done:
             n.next = static_cast<std::uint32_t>(out - base) - cur_idx;
             if (--depth != 0)
             {
-                frame f{};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    f = cold.m_shallow[depth - 1];
+                    const frame& f = cold.m_shallow[depth - 1];
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
                 else
                 {
-                    f = cold.m_deep.back();
+                    const frame f = cold.m_deep.back();
                     cold.m_deep.pop_back();
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
-                cur_idx = f.idx;
-                cur_count = f.count;
-                cur_is_object = f.is_object;
             }
         }
 
@@ -1028,12 +1050,13 @@ indent_done:
             return true;
         }
 
-        /// a string (value or key) at p
+        /// a string at p: a value (Value) or a key
+        template<bool Value>
         NLOHMANN_VIEW_ALWAYS_INLINE bool string()
         {
             ++p; // opening quote
             const unsigned char* const s = p;
-            p = scan_string_run(p, e);
+            p = scan_string_run<Value>(p, e);
             if (NLOHMANN_VIEW_LIKELY(p != e && *p == '"'))
             {
                 push_node(value_t::string, 0, 0, static_cast<std::size_t>(s - b), static_cast<std::uint64_t>(p - s));

include/nlohmann/detail/view/document_data.hpp

@@ -10,6 +10,7 @@
 
 #include <array> // array
 #include <cstddef> // size_t
+#include <cstdint> // uint32_t
 #include <cstring> // memcmp, memcpy
 #include <limits> // numeric_limits
 #include <new> // bad_alloc, operator new, placement new
@@ -40,6 +41,17 @@ struct document_data
     std::size_t max_depth = 0; ///< deepest nesting of arrays and objects (the root container is depth 1)
     std::vector<char> arena; ///< decoded strings that contained escapes (no terminator: they are handed out with their length)
     std::string owned; ///< owned copy of the input, if any (a std::string, so that its NUL serves as the parser's sentinel)
+
+    // hash indexes of the keys of large objects (see object_index.hpp)
+    static constexpr std::uint32_t index_min_members = 128;
+    struct object_index
+    {
+        std::size_t start;  ///< first slot in index_slots
+        std::uint32_t mask; ///< slot count - 1 (a power of two minus one)
+    };
+    std::vector<object_index> indexes;
+    std::vector<std::uint32_t> index_slots;
+    std::vector<std::uint32_t> large_objects; ///< positions of the objects to index (noted while parsing)
     std::array<const char*, 4> base = {{nullptr, nullptr, nullptr, nullptr}}; ///< string bases: source, arena (indexed by flags & node_flags::storage)
     bool discarded = true;
 
@@ -82,6 +94,9 @@ struct document_data
     document_data() noexcept
         : arena() // NOLINT(readability-redundant-member-init)
         , owned() // NOLINT(readability-redundant-member-init)
+        , indexes() // NOLINT(readability-redundant-member-init)
+        , index_slots() // NOLINT(readability-redundant-member-init)
+        , large_objects() // NOLINT(readability-redundant-member-init)
     {}
     document_data(const document_data&) = delete;
     document_data(document_data&&) = delete;

include/nlohmann/detail/view/lookup.hpp

@@ -18,6 +18,7 @@
 #include <nlohmann/detail/view/document_data.hpp>
 #include <nlohmann/detail/view/macro_scope.hpp>
 #include <nlohmann/detail/view/node.hpp>
+#include <nlohmann/detail/view/object_index.hpp>
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -89,6 +90,10 @@ class short_key
 /// their length, from the index alone
 inline const node* find_member(const document_data& d, const node* object, const char* key, std::size_t n) noexcept
 {
+    if (NLOHMANN_VIEW_UNLIKELY(object->extra != 0))
+    {
+        return find_indexed(d, object, key, n); // a large object
+    }
     const node* const end = document_data::child_end(object);
     const auto* const k = reinterpret_cast<const unsigned char*>(key); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
     if (NLOHMANN_VIEW_LIKELY(n <= 16))

include/nlohmann/detail/view/macro_scope.hpp

@@ -70,5 +70,11 @@
     #define NLOHMANN_VIEW_LITTLE_ENDIAN 0
 #endif
 
+/// eight checks at fixed offsets 0..7
+#define NLOHMANN_VIEW_REPEAT_FIRST_8(X) X(0) X(1) X(2) X(3) X(4) X(5) X(6) X(7)
+
+/// eight checks at fixed offsets 8..15
+#define NLOHMANN_VIEW_REPEAT_SECOND_8(X) X(8) X(9) X(10) X(11) X(12) X(13) X(14) X(15)
+
 /// sixteen checks at fixed offsets 0..15
-#define NLOHMANN_VIEW_REPEAT16(X) X(0) X(1) X(2) X(3) X(4) X(5) X(6) X(7) X(8) X(9) X(10) X(11) X(12) X(13) X(14) X(15)
+#define NLOHMANN_VIEW_REPEAT16(X) NLOHMANN_VIEW_REPEAT_FIRST_8(X) NLOHMANN_VIEW_REPEAT_SECOND_8(X)

include/nlohmann/detail/view/macro_unscope.hpp

@@ -19,3 +19,12 @@
 #undef NLOHMANN_VIEW_THROW
 #undef NLOHMANN_VIEW_LITTLE_ENDIAN
 #undef NLOHMANN_VIEW_REPEAT16
+#undef NLOHMANN_VIEW_REPEAT_SECOND_8
+#undef NLOHMANN_VIEW_REPEAT_FIRST_8
+#undef NLOHMANN_VIEW_NEON
+#undef NLOHMANN_VIEW_SSE2
+#undef NLOHMANN_VIEW_SSSE3
+#undef NLOHMANN_VIEW_SSSE3_DISPATCH
+#undef NLOHMANN_VIEW_SSSE3_TARGET
+#undef NLOHMANN_VIEW_VECTOR
+#undef NLOHMANN_VIEW_VECTOR_UTF8

include/nlohmann/detail/view/node.hpp

@@ -45,7 +45,7 @@ struct node
 {
     std::uint8_t kind;   ///< value_t
     std::uint8_t flags;  ///< node_flags
-    std::uint16_t extra; ///< numbers: integer digits (low byte) and fraction digits (high byte), 255 = "many"; otherwise 0
+    std::uint16_t extra; ///< numbers: integer digits (low byte) and fraction digits (high byte), 255 = "many"; objects: number of the object's key hash index (1-based, 0 = none); otherwise 0
     std::uint32_t off;   ///< source offset (string content, number token, literal, bracket); arena offset if node_flags::escaped
     std::uint32_t len;   ///< string: decoded bytes; float: token bytes; array/object: element count
     std::uint32_t next;  ///< array/object: number of nodes of the subtree (its extent in the enclosing sequence)

include/nlohmann/detail/view/object_index.hpp

@@ -0,0 +1,154 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2026 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <cstddef> // size_t
+#include <cstdint> // uint32_t, uint64_t
+
+#include <nlohmann/json.hpp>
+#include <nlohmann/detail/view/document_data.hpp>
+#include <nlohmann/detail/view/macro_scope.hpp>
+#include <nlohmann/detail/view/node.hpp>
+
+// Hash indexes of the keys of large objects, so that a lookup does not compare thousands
+// of keys (as Boost.JSON switches from a linear search to a hash table for
+// large objects). An object with document_data::index_min_members members or
+// more gets an open-addressing table after parsing; its node stores the
+// number of the table (1-based) in `extra`. A slot holds the offset of a key
+// node from its object node (0: empty). Of duplicate keys, the first is kept,
+// as for the linear search.
+//
+// The hash is not seeded, so keys chosen to collide could make the build
+// quadratic. A key therefore sits at most index_max_displacement slots away
+// from its home slot; if a key would sit further away, the table is dropped
+// and the object is searched linearly (like a small one). For the same
+// reason, a lookup visits at most index_max_displacement + 1 slots.
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+namespace view
+{
+
+/// the farthest a key may sit from its home slot (a table with at most half of
+/// its slots in use gives random keys a distance of about 50 for millions of
+/// members; and every member costs at most this many steps while building)
+constexpr std::size_t index_max_displacement = 64;
+
+// a table has at least 2 * len slots (len >= index_min_members), so the probe
+// window of index_max_displacement + 1 slots stays far smaller than the table
+// and a probe never wraps around onto its own start
+static_assert(document_data::index_min_members >= 2 * index_max_displacement,
+              "a table must be much larger than the probe window of a lookup");
+
+/// hash of a key: its bytes, eight at a time, in a fixed byte order
+inline std::uint64_t key_hash(const char* s, std::size_t n) noexcept
+{
+    std::uint64_t h = 0x9E3779B97F4A7C15u * (n + 1);
+    const auto* p = reinterpret_cast<const unsigned char*>(s); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+    while (n >= 8)
+    {
+        h = (h ^ read_eight_bytes(p)) * 0xBF58476D1CE4E5B9u;
+        h ^= h >> 29u;
+        p += 8;
+        n -= 8;
+    }
+    std::uint64_t w = 0;
+    for (std::size_t i = 0; i < n; ++i)
+    {
+        w |= static_cast<std::uint64_t>(p[i]) << (8u * i);
+    }
+    h = (h ^ w) * 0x94D049BB133111EBu;
+    return h ^ (h >> 31u);
+}
+
+/// build the table of a large object
+inline void build_object_index(document_data& d, node* obj)
+{
+    if (d.indexes.size() >= 0xFFFFu)
+    {
+        return; // LCOV_EXCL_LINE (the number must fit `extra`; more large objects are searched linearly)
+    }
+    std::size_t cap = 16;
+    while (cap < 2 * static_cast<std::size_t>(obj->len))
+    {
+        cap *= 2;
+    }
+    const std::size_t start = d.index_slots.size();
+    d.index_slots.resize(start + cap, 0);
+    std::uint32_t* const table = d.index_slots.data() + start;
+    const std::size_t mask = cap - 1;
+    for (const node* k = document_data::first_child(obj), *end = document_data::child_end(obj); k != end; k = document_data::after(k + 1))
+    {
+        const char* const key = d.str(*k);
+        std::size_t slot = conditional_static_cast<std::size_t>(key_hash(key, k->len)) & mask;
+        bool duplicate = false;
+        std::size_t distance = 0;
+        while (table[slot] != 0)
+        {
+            const node* const other = obj + table[slot];
+            if (d.key_equals(*other, key, k->len))
+            {
+                duplicate = true; // keep the first
+                break;
+            }
+            if (distance == index_max_displacement)
+            {
+                // too many keys share a home region
+                d.index_slots.resize(start); // no table: the object is searched linearly
+                return;
+            }
+            ++distance;
+            slot = (slot + 1) & mask;
+        }
+        if (!duplicate)
+        {
+            table[slot] = static_cast<std::uint32_t>(k - obj);
+        }
+    }
+    d.indexes.push_back(document_data::object_index{start, static_cast<std::uint32_t>(mask)});
+    obj->extra = static_cast<std::uint16_t>(d.indexes.size());
+}
+
+/// build the tables of the large objects the parser noted
+inline void build_object_indexes(document_data& d)
+{
+    for (const std::uint32_t i : d.large_objects)
+    {
+        build_object_index(d, d.nodes + i);
+    }
+}
+
+/// the key node of the first member with this key of an indexed object, or
+/// nullptr
+inline const node* find_indexed(const document_data& d, const node* obj, const char* key, std::size_t n) noexcept
+{
+    const document_data::object_index& ix = d.indexes[obj->extra - 1u];
+    const std::uint32_t* const table = d.index_slots.data() + ix.start;
+    std::size_t slot = conditional_static_cast<std::size_t>(key_hash(key, n)) & ix.mask;
+    for (std::size_t distance = 0; distance <= index_max_displacement; ++distance)
+    {
+        const std::uint32_t s = table[slot];
+        if (s == 0)
+        {
+            return nullptr;
+        }
+        const node* const k = obj + s;
+        if (d.key_equals(*k, key, n))
+        {
+            return k;
+        }
+        slot = (slot + 1) & ix.mask;
+    }
+    return nullptr; // (no key sits further from its home slot)
+}
+
+}  // namespace view
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END

include/nlohmann/detail/view/scan.hpp

@@ -16,6 +16,7 @@
 
 #include <nlohmann/json.hpp>
 #include <nlohmann/detail/view/macro_scope.hpp>
+#include <nlohmann/detail/view/simd.hpp>
 
 // Scanning primitives of the view's parser. The unrolled checks at fixed
 // offsets follow yyjson (https://github.com/ibireme/yyjson, MIT license): the
@@ -62,20 +63,43 @@ NLOHMANN_VIEW_ALWAYS_INLINE std::uint16_t load16(const unsigned char* p) noexcep
 
 /// Advance over plain string bytes and well-formed UTF-8. Stops at a quote,
 /// a backslash, a control character, ill-formed UTF-8, or the end. The first
-/// 16 bytes are checked one by one, so that the position advances by
-/// constants in predicted branches (most strings are short); longer runs
-/// continue eight bytes at a time.
+/// bytes are checked one by one, so that the position advances by constants
+/// in predicted branches: 16 for keys, whose lengths repeat from record to
+/// record, and 8 for string values (Value) where a vector loop follows, as
+/// their lengths vary more. Longer runs continue 16 bytes at a time with NEON
+/// or SSE2, else eight bytes at a time. With SSE2, the run is checked 16 bytes
+/// at a time from its first byte instead: on x86-64, one compare that finds
+/// the end of most keys and short values is faster than a branch per byte (on
+/// AArch64, where a NEON mask costs more and branches predict well, slower).
+template<bool Value = false>
 NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned char* p, const unsigned char* e) noexcept
 {
     const std::uint8_t* plain = string_plain();
     for (;;)
     {
+#if NLOHMANN_VIEW_SSE2
+        p = vector_plain_run(p, e);
+#else
         if (e - p >= 16)
         {
 #define NLOHMANN_VIEW_STEP(i) if (NLOHMANN_VIEW_LIKELY(plain[p[i]] != 0)) {} else { p += (i); goto stop; }
-            NLOHMANN_VIEW_REPEAT16(NLOHMANN_VIEW_STEP)
+            NLOHMANN_VIEW_REPEAT_FIRST_8(NLOHMANN_VIEW_STEP)
+#if NLOHMANN_VIEW_VECTOR
+            if (!Value) // (a plain `!Value || !NLOHMANN_VIEW_VECTOR` is a constant condition for MSVC, C4127)
+#endif
+            {
+                NLOHMANN_VIEW_REPEAT_SECOND_8(NLOHMANN_VIEW_STEP)
+                p += 8;
+            }
 #undef NLOHMANN_VIEW_STEP
-            p += 16;
+            p += 8;
+#if NLOHMANN_VIEW_VECTOR
+            p = vector_plain_run(p, e);
+            if (p != e && plain[*p] == 0)
+            {
+                goto stop;
+            }
+#else
             while (e - p >= 8)
             {
                 const std::uint64_t special = swar_string_special(read_eight_bytes(p));
@@ -86,8 +110,10 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
                 }
                 p += 8;
             }
+#endif
             continue;
         }
+#endif
         while (p != e && plain[*p] != 0)
         {
             ++p;
@@ -96,11 +122,23 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
         {
             return p;
         }
+#if !NLOHMANN_VIEW_SSE2
 stop:
+#endif
         if (*p < 0x80)
         {
             return p; // quote, backslash, or control character
         }
+#if NLOHMANN_VIEW_VECTOR_UTF8
+#if NLOHMANN_VIEW_SSSE3_DISPATCH
+        if (NLOHMANN_VIEW_LIKELY(cpu_has_ssse3()))
+#endif
+        {
+            // non-ASCII: the vector check, out of line
+            return scan_string_vector(p, e, plain);
+        }
+#endif
+#if !NLOHMANN_VIEW_VECTOR_UTF8 || NLOHMANN_VIEW_SSSE3_DISPATCH
         // non-ASCII: a run of well-formed sequences (the library's check, so
         // that exactly what json::parse accepts is accepted)
         do
@@ -113,6 +151,7 @@ stop:
             p += n;
         }
         while (p != e && *p >= 0x80);
+#endif
     }
 }
 

include/nlohmann/detail/view/simd.hpp

@@ -0,0 +1,339 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2026 Niels Lohmann <https://nlohmann.me>
+// SPDX-FileCopyrightText: 2018-2025 The simdjson authors <https://github.com/simdjson/simdjson>
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <array> // array
+#include <atomic> // atomic
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t, uint64_t
+
+#include <nlohmann/json.hpp>
+#include <nlohmann/detail/view/macro_scope.hpp>
+
+// Vector code for long runs of string bytes. NEON (AArch64) and SSE2 (x86-64)
+// belong to the baseline instruction sets and are used by default. The vector
+// UTF-8 check needs NEON or SSSE3. SSSE3 is not part of x86-64, and the code
+// must not depend on the flags of a translation unit (two translation units
+// with different flags would have different definitions of the same inline
+// functions): the check is compiled for SSSE3 with a function attribute and
+// used where the CPU has SSSE3 (all x86-64 CPUs since about 2011), else the
+// portable check. JSON_VIEW_USE_SSSE3 skips the CPU check (for code compiled
+// for SSSE3 anyway); JSON_VIEW_NO_SIMD selects the portable code.
+// NLOHMANN_VIEW_NO_TARGET_ATTRIBUTE (set by the GCC module interface, where GCC
+// ignores the target attribute and then rejects the SSSE3 intrinsics) keeps the
+// check portable without disabling the SSE2 code.
+#if !defined(JSON_VIEW_NO_SIMD) && defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) && NLOHMANN_VIEW_LITTLE_ENDIAN
+    #include <arm_neon.h>
+    #define NLOHMANN_VIEW_NEON 1
+#else
+    #define NLOHMANN_VIEW_NEON 0
+#endif
+// (x86 only: other targets can define __SSE2__ as well, e.g., WebAssembly with -msse2, but have no <cpuid.h>)
+#if !defined(JSON_VIEW_NO_SIMD) && !NLOHMANN_VIEW_NEON && (defined(__x86_64__) || defined(__i386__) || defined(_M_X64) || defined(_M_IX86)) && (defined(__SSE2__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 2))
+    #include <emmintrin.h>
+    #define NLOHMANN_VIEW_SSE2 1
+#else
+    #define NLOHMANN_VIEW_SSE2 0
+#endif
+#if NLOHMANN_VIEW_SSE2 && defined(JSON_VIEW_USE_SSSE3)
+    #include <tmmintrin.h>
+    #define NLOHMANN_VIEW_SSSE3 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+#else
+    #define NLOHMANN_VIEW_SSSE3 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+#endif
+#if NLOHMANN_VIEW_SSE2 && !NLOHMANN_VIEW_SSSE3 && !defined(NLOHMANN_VIEW_NO_TARGET_ATTRIBUTE) && ((defined(__clang__) && __clang_major__ >= 4) || (defined(__GNUC__) && !defined(__clang__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9))))
+    // (GCC before 4.9 has no SSSE3 intrinsics without -mssse3)
+    #include <cpuid.h>
+    #include <tmmintrin.h>
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET __attribute__((target("ssse3")))
+#elif NLOHMANN_VIEW_SSE2 && !NLOHMANN_VIEW_SSSE3 && defined(_MSC_VER)
+    // (MSVC compiles intrinsics of any instruction set)
+    #include <intrin.h>
+    #include <tmmintrin.h>
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET
+#else
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET
+#endif
+#define NLOHMANN_VIEW_VECTOR (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSE2)
+#define NLOHMANN_VIEW_VECTOR_UTF8 (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSSE3 || NLOHMANN_VIEW_SSSE3_DISPATCH)
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+namespace view
+{
+
+#if NLOHMANN_VIEW_VECTOR
+/*!
+@brief the first byte of a string run that is a quote, a backslash, a control
+character, or not ASCII, 16 bytes per step
+
+Stops at such a byte, or where fewer than 16 bytes are left (the caller tells
+the two apart). A signed compare with 0x20 finds control characters and
+non-ASCII bytes at once.
+*/
+NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned char* p, const unsigned char* e) noexcept
+{
+    while (e - p >= 16)
+    {
+#if NLOHMANN_VIEW_NEON
+        const uint8x16_t in = vld1q_u8(p);
+        const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('"')), vceqq_u8(in, vdupq_n_u8('\\'))),
+                                            vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)));
+        // one nibble per byte (the usual NEON replacement of x86's movemask, see
+        // D. Kutenin, "Porting x86 vector bitmask optimizations to Arm NEON", 2022)
+        const std::uint64_t bits = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(special), 4)), 0);
+        if (bits != 0)
+        {
+            return p + (count_trailing_zeros(bits) >> 2u);
+        }
+#else
+        const __m128i in = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(p))); // NOLINT(bugprone-casting-through-void)
+        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('"')), _mm_cmpeq_epi8(in, _mm_set1_epi8('\\'))),
+                                             _mm_cmplt_epi8(in, _mm_set1_epi8(0x20)));
+        const auto bits = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
+        if (bits != 0)
+        {
+            return p + count_trailing_zeros(bits);
+        }
+#endif
+        p += 16;
+    }
+    return p;
+}
+#endif
+
+#if NLOHMANN_VIEW_SSSE3_DISPATCH
+/// whether the CPU has SSSE3 (CPUID leaf 1, ECX bit 9)
+inline bool cpu_ssse3() noexcept
+{
+#if defined(_MSC_VER) && !defined(__clang__)
+    std::array<int, 4> regs {{}};
+    __cpuid(regs.data(), 1);
+    return (static_cast<unsigned>(regs[2]) & (1u << 9u)) != 0;
+#else
+    unsigned eax = 0;
+    unsigned ebx = 0;
+    unsigned ecx = 0;
+    unsigned edx = 0;
+    return __get_cpuid(1, &eax, &ebx, &ecx, &edx) != 0 && (ecx & (1u << 9u)) != 0;
+#endif
+}
+
+/// whether the CPU has SSSE3, asked once: the answer is kept in an atomic
+/// that is initialized at compile time, so that neither a guard of a local
+/// static nor a global constructor is needed (threads that ask at the same
+/// time all store the same answer)
+NLOHMANN_VIEW_ALWAYS_INLINE bool cpu_has_ssse3() noexcept
+{
+    static std::atomic<int> known{0}; // 0: not asked yet, 1: no, 2: yes
+    int state = known.load(std::memory_order_relaxed);
+    if (NLOHMANN_VIEW_UNLIKELY(state == 0))
+    {
+        state = cpu_ssse3() ? 2 : 1;
+        known.store(state, std::memory_order_relaxed);
+    }
+    return state == 2;
+}
+#endif
+
+#if NLOHMANN_VIEW_VECTOR_UTF8
+/// Tables of the UTF-8 check of J. Keiser and D. Lemire, "Validating UTF-8 In
+/// Less Than One Instruction Per Byte" (2021), as in simdjson ("lookup4"): each
+/// maps a nibble (high and low nibble of the previous byte, high nibble of the
+/// current byte) to the errors it allows; a byte pair is ill-formed if all
+/// three have an error bit in common.
+template<typename Dummy = void>
+struct utf8_lookup4
+{
+    static constexpr std::uint8_t too_short = 1u << 0u, too_long = 1u << 1u, overlong_3 = 1u << 2u, too_large = 1u << 3u;
+    static constexpr std::uint8_t surrogate = 1u << 4u, overlong_2 = 1u << 5u, too_large_1000 = 1u << 6u, overlong_4 = 1u << 6u;
+    static constexpr std::uint8_t two_conts = 1u << 7u, carry = too_short | too_long | two_conts;
+    static const std::array<std::uint8_t, 16> byte_1_high;
+    static const std::array<std::uint8_t, 16> byte_1_low;
+    static const std::array<std::uint8_t, 16> byte_2_high;
+};
+
+template<typename Dummy>
+const std::array<std::uint8_t, 16> utf8_lookup4<Dummy>::byte_1_high =
+{
+    {
+        too_long, too_long, too_long, too_long, too_long, too_long, too_long, too_long,
+        two_conts, two_conts, two_conts, two_conts,
+        too_short | overlong_2, too_short, too_short | overlong_3 | surrogate, too_short | too_large | too_large_1000 | overlong_4
+    }
+};
+
+template<typename Dummy>
+const std::array<std::uint8_t, 16> utf8_lookup4<Dummy>::byte_1_low =
+{
+    {
+        carry | overlong_3 | overlong_2 | overlong_4, carry | overlong_2, carry, carry,
+        carry | too_large, carry | too_large | too_large_1000, carry | too_large | too_large_1000, carry | too_large | too_large_1000,
+        carry | too_large | too_large_1000, carry | too_large | too_large_1000, carry | too_large | too_large_1000, carry | too_large | too_large_1000,
+        carry | too_large | too_large_1000, carry | too_large | too_large_1000 | surrogate, carry | too_large | too_large_1000, carry | too_large | too_large_1000
+    }
+};
+
+template<typename Dummy>
+const std::array<std::uint8_t, 16> utf8_lookup4<Dummy>::byte_2_high =
+{
+    {
+        too_short, too_short, too_short, too_short, too_short, too_short, too_short, too_short,
+        static_cast<std::uint8_t>(too_long | overlong_2 | two_conts | overlong_3 | too_large_1000 | overlong_4),
+        static_cast<std::uint8_t>(too_long | overlong_2 | two_conts | overlong_3 | too_large),
+        static_cast<std::uint8_t>(too_long | overlong_2 | two_conts | surrogate | too_large),
+        static_cast<std::uint8_t>(too_long | overlong_2 | two_conts | surrogate | too_large),
+        too_short, too_short, too_short, too_short
+    }
+};
+
+/// the end of scan_string_vector from block, where the vector loop stopped
+/// (ill-formed UTF-8, or fewer than 16 bytes left): one byte or sequence at a
+/// time, from the start of a sequence that crosses into the block
+inline const unsigned char* scan_string_finish(const unsigned char* p, const unsigned char* block, const unsigned char* e, const std::uint8_t* plain) noexcept
+{
+    for (int i = 1; i <= 3 && block - i >= p; ++i)
+    {
+        const unsigned char c = block[-i];
+        if (c < 0x80)
+        {
+            break;
+        }
+        if (c >= 0xC0)
+        {
+            const int len = 2 + static_cast<int>(c >= 0xE0) + static_cast<int>(c >= 0xF0);
+            if (len > i)
+            {
+                block -= i;
+            }
+            break;
+        }
+    }
+    for (p = block; p != e;)
+    {
+        if (*p < 0x80)
+        {
+            if (plain[*p] == 0)
+            {
+                return p;
+            }
+            ++p;
+            continue;
+        }
+        const std::size_t n = validate_one_utf8(p, static_cast<std::size_t>(e - p));
+        if (n == 0)
+        {
+            return p;
+        }
+        p += n;
+    }
+    return p;
+}
+
+/*!
+@brief the rest of a string from p (a character boundary), 16 bytes per step
+
+The first quote, backslash, or control character is found with vector
+compares, and the UTF-8 check covers the bytes up to it. Returns where the
+string scan stops, like scan_string_run: before ill-formed UTF-8 and for the
+last bytes of the input, the bytes are checked one sequence at a time. Out of
+line, so that no constants of the check occupy registers in the parse loop.
+On x86-64, it is compiled for SSSE3 (see cpu_has_ssse3()).
+*/
+NLOHMANN_VIEW_SSSE3_TARGET NLOHMANN_VIEW_NOINLINE inline const unsigned char* scan_string_vector(const unsigned char* p, const unsigned char* e, const std::uint8_t* plain) noexcept
+{
+    using lookup = utf8_lookup4<>;
+    const unsigned char* block = p;
+#if NLOHMANN_VIEW_NEON
+    const uint8x16_t t1h = vld1q_u8(lookup::byte_1_high.data());
+    const uint8x16_t t1l = vld1q_u8(lookup::byte_1_low.data());
+    const uint8x16_t t2h = vld1q_u8(lookup::byte_2_high.data());
+    uint8x16_t prev = vdupq_n_u8(0);
+    while (e - block >= 16)
+    {
+        const uint8x16_t in = vld1q_u8(block);
+        const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('"')), vceqq_u8(in, vdupq_n_u8('\\'))), vcltq_u8(in, vdupq_n_u8(0x20)));
+        const uint8x16_t prev1 = vextq_u8(prev, in, 15);
+        const uint8x16_t sc = vandq_u8(vandq_u8(vqtbl1q_u8(t1h, vshrq_n_u8(prev1, 4)), vqtbl1q_u8(t1l, vandq_u8(prev1, vdupq_n_u8(0x0F)))), vqtbl1q_u8(t2h, vshrq_n_u8(in, 4)));
+        const uint8x16_t must23 = vorrq_u8(vqsubq_u8(vextq_u8(prev, in, 14), vdupq_n_u8(0xE0 - 0x80)), vqsubq_u8(vextq_u8(prev, in, 13), vdupq_n_u8(0xF0 - 0x80)));
+        const uint8x16_t err = veorq_u8(vandq_u8(must23, vdupq_n_u8(0x80)), sc);
+        const std::uint64_t special_bits = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(special), 4)), 0);
+        const std::uint64_t err_bits = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(vtstq_u8(err, err)), 4)), 0);
+        if (special_bits != 0)
+        {
+            // errors up to the special byte count (an incomplete sequence
+            // before a quote shows at the quote); the bytes after it do not
+            const unsigned k = static_cast<unsigned>(count_trailing_zeros(special_bits)) >> 2u;
+            const std::uint64_t upto = k == 15 ? ~std::uint64_t{0} :
+                                       (std::uint64_t{1} << (4u * (k + 1u))) - 1u;
+            if ((err_bits & upto) == 0)
+            {
+                return block + k;
+            }
+            break;
+        }
+        if (err_bits != 0)
+        {
+            break;
+        }
+        prev = in;
+        block += 16;
+    }
+#else
+    // the same with SSSE3 (pshufb for the table lookups; nibbles from 16-bit
+    // shifts, as there are no byte shifts)
+    const __m128i t1h = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(lookup::byte_1_high.data()))); // NOLINT(bugprone-casting-through-void)
+    const __m128i t1l = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(lookup::byte_1_low.data()))); // NOLINT(bugprone-casting-through-void)
+    const __m128i t2h = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(lookup::byte_2_high.data()))); // NOLINT(bugprone-casting-through-void)
+    const __m128i nibble = _mm_set1_epi8(0x0F);
+    const __m128i zero = _mm_setzero_si128();
+    __m128i prev = zero;
+    while (e - block >= 16)
+    {
+        const __m128i in = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(block))); // NOLINT(bugprone-casting-through-void)
+        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('"')), _mm_cmpeq_epi8(in, _mm_set1_epi8('\\'))),
+                                             _mm_cmpeq_epi8(_mm_subs_epu8(in, _mm_set1_epi8(0x1F)), zero)); // in < 0x20
+        const __m128i prev1 = _mm_alignr_epi8(in, prev, 15);
+        const __m128i sc = _mm_and_si128(_mm_and_si128(_mm_shuffle_epi8(t1h, _mm_and_si128(_mm_srli_epi16(prev1, 4), nibble)),
+                                         _mm_shuffle_epi8(t1l, _mm_and_si128(prev1, nibble))),
+                                         _mm_shuffle_epi8(t2h, _mm_and_si128(_mm_srli_epi16(in, 4), nibble)));
+        const __m128i must23 = _mm_or_si128(_mm_subs_epu8(_mm_alignr_epi8(in, prev, 14), _mm_set1_epi8(0xE0 - 0x80)),
+                                            _mm_subs_epu8(_mm_alignr_epi8(in, prev, 13), _mm_set1_epi8(0xF0 - 0x80)));
+        const __m128i err = _mm_xor_si128(_mm_and_si128(must23, _mm_set1_epi8(static_cast<char>(-128))), sc);
+        const auto special_bits = static_cast<unsigned>(_mm_movemask_epi8(special));
+        const auto err_bits = ~static_cast<unsigned>(_mm_movemask_epi8(_mm_cmpeq_epi8(err, zero))) & 0xFFFFu;
+        if (special_bits != 0)
+        {
+            const auto k = static_cast<unsigned>(count_trailing_zeros(static_cast<std::uint64_t>(special_bits)));
+            if ((err_bits & ((2u << k) - 1u)) == 0)
+            {
+                return block + k;
+            }
+            break;
+        }
+        if (err_bits != 0)
+        {
+            break;
+        }
+        prev = in;
+        block += 16;
+    }
+#endif
+    return scan_string_finish(p, block, e, plain);
+}
+#endif
+
+}  // namespace view
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END

include/nlohmann/json_view.hpp

@@ -27,6 +27,7 @@
 #include <algorithm> // min
 #include <array> // array
 #include <cstddef> // nullptr_t, size_t // IWYU pragma: keep
+#include <cstdint> // uint32_t
 #include <cstring> // memcpy, strlen
 #include <istream> // istream // IWYU pragma: keep
 #include <iterator> // distance, input_iterator_tag, iterator_traits
@@ -73,6 +74,7 @@
 #include <nlohmann/detail/view/macro_scope.hpp>
 #include <nlohmann/detail/view/materialize.hpp>
 #include <nlohmann/detail/view/node.hpp>
+#include <nlohmann/detail/view/object_index.hpp>
 #include <nlohmann/detail/view/pointer.hpp>
 #include <nlohmann/detail/view/serializer.hpp>
 #include <nlohmann/detail/view/string_ref.hpp>
@@ -1009,7 +1011,9 @@ class basic_json_document
         }
         return sizeof(document_data) + (m_data->inline_capacity * sizeof(detail::view::node))
                + (m_data->nodes != m_data->inline_nodes ? m_data->node_capacity * sizeof(detail::view::node) : 0)
-               + m_data->arena.capacity() + m_data->owned.capacity();
+               + m_data->arena.capacity() + m_data->owned.capacity()
+               + (m_data->indexes.capacity() * sizeof(document_data::object_index)) + (m_data->index_slots.capacity() * sizeof(std::uint32_t))
+               + (m_data->large_objects.capacity() * sizeof(std::uint32_t));
     }
 
     /// release unused capacity of the index and the decoded strings; like
@@ -1030,6 +1034,8 @@ class basic_json_document
         std::vector<char> arena(shrink_arena ? d.arena : std::vector<char>());
         const bool shrink_nodes = d.nodes != d.inline_nodes && d.node_count != d.node_capacity;
         const bool into_header = d.node_count <= d.inline_capacity;
+        std::vector<document_data::object_index> indexes(d.indexes.capacity() > d.indexes.size() ? d.indexes : std::vector<document_data::object_index>());
+        std::vector<std::uint32_t> index_slots(d.index_slots.capacity() > d.index_slots.size() ? d.index_slots : std::vector<std::uint32_t>());
         node* fresh = (shrink_nodes && !into_header) ? static_cast<node*>(::operator new (d.node_count * sizeof(node))) : d.inline_nodes;
 
         if (shrink_nodes)
@@ -1047,6 +1053,14 @@ class basic_json_document
             d.arena.swap(arena);
             d.base[1] = d.arena.data();
         }
+        if (d.indexes.capacity() > d.indexes.size())
+        {
+            d.indexes.swap(indexes);
+        }
+        if (d.index_slots.capacity() > d.index_slots.size())
+        {
+            d.index_slots.swap(index_slots);
+        }
     }
 
   private:
@@ -1106,6 +1120,9 @@ class basic_json_document
         d.size = size;
         d.node_count = 0;
         d.arena.clear();
+        d.indexes.clear();
+        d.index_slots.clear();
+        d.large_objects.clear();
         d.discarded = true;
         detail::view::parse_failure failure;
         bool ok = false;
@@ -1121,9 +1138,12 @@ class basic_json_document
         {
             d.base[0] = d.src;
             d.base[1] = d.arena.data();
+            detail::view::build_object_indexes(d);
+            std::vector<std::uint32_t>().swap(d.large_objects); // (only needed while parsing)
             d.discarded = false;
             return;
         }
+        std::vector<std::uint32_t>().swap(d.large_objects);
         if (allow_exceptions)
         {
             detail::view::throw_parse_failure<BasicJsonType>(failure, src, size, comments, trailing_commas);