Scan strings with SSE2, SSSE3, or NEON (#5853)

* Scan strings with SSE2, SSSE3, or NEON

The string fast path found the end of a run of plain characters 8 bytes at
a time (SWAR) and validated non-ASCII text one sequence at a time, calling
validate_one_utf8() out of line for each.

The vector code of json_view's scanner, in the new shared header
detail/input/simd_scan.hpp, now does both 16 bytes at a time: SSE2 finds
the end of a run on x86-64, and the UTF-8 check of Keiser and Lemire runs
with NEON on AArch64 and with SSSE3 on x86-64. SSSE3 is not part of x86-64,
so the check is compiled for it with a function attribute and used when the
CPU has it (cached in a constant-initialized atomic), else the scalar check;
code compiled with -mssse3 skips the CPU check. The first word of a run is
still checked with SWAR (the first four with NEON, where a vector mask
costs more), as most keys and short values end there, and the scalar
helpers are always inlined. On anything else, the scalar code is unchanged.

The new macro JSON_NO_SIMD selects the portable code everywhere, including
the SSE2/NEON digit formatting of the floating-point writer. The lexer tests
now also run with JSON_NO_SIMD and, on x86-64, with -mssse3, and a new test
places every byte that ends a run at each offset across several 16-byte
blocks, after ASCII and after CJK text.

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

* Keep the string scan from slowing down short strings and dump() on Arm

Three benchmarks were 7-9% slower than develop on Apple M1 (Apple Clang,
-O3 -flto): Accept/jeopardy, Dump/citm_catalog, and Dump/jeopardy.

- dump(): the serializer passes a string without its closing quote, so
  plain_run_end() ran out of input and called the out-of-line NEON loop
  for every string, with fewer than 16 bytes left. The loop is now only
  called for 16 bytes or more.
- accept(): no string of jeopardy.json reaches the vector UTF-8 check,
  yet compiling it into the lexer's scan function made it 5% slower: the
  function runs once per token and now saved more registers and kept a
  stack slot. lexer_string_bulk_run() checks the first four words inline
  and continues out of line, and the serializer keeps string_bulk_run()
  inline, which is faster there.
- The byte that ends a run is read from the scanned buffer instead of
  through get(), which Apple Clang merged with the other get() calls that
  load the read position again right after it was stored (3%).

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

* Define string_bulk_run_from() only where lexer_string_bulk_run() calls it

With NEON and simdutf in C++17, lexer_string_bulk_run() takes the
string_bulk_run() path, so the out-of-line helper was unused.

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

* Do not dispatch the SSSE3 string scan with GCC's C++20 modules

GCC does not keep the target("ssse3") attribute on functions it reads
back from a module interface, so the module build (ci_module_cpp20,
gcc:latest) rejected the intrinsics in vector_utf8_run() with "needs isa
option -mssse3". With __cpp_modules, GCC now uses the portable UTF-8 check
like other compilers without the target attribute; the SSE2 scan stays,
and -mssse3 or more still compiles the SSSE3 check directly. Clang is not
affected.

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

* Include <cpuid.h> only once when json.hpp and json_view.hpp are both used

The SSSE3 dispatch of the string scan and the one of json_view each
include <cpuid.h>. Before GCC 11 the header has no include guard, so a
translation unit that includes both headers fails with GCC 10:
"redefinition of 'unsigned int __get_cpuid_max(unsigned int, unsigned int*)'".
Guard both includes with NLOHMANN_JSON_CPUID_INCLUDED.

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

* Add an ABI tag for JSON_NO_SIMD

JSON_NO_SIMD selects between two definitions of the internal scanning and
formatting functions. Like JSON_VIEW_NO_SIMD (_vns), it now appends _ns to
the inline namespace, so translation units built with and without it can
be linked into one program without ODR violations.

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

---------

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

single_include/nlohmann/json.hpp

@@ -186,6 +186,12 @@
     #define NLOHMANN_JSON_ABI_TAG_DISABLE_UNLOCKED_FILE_IO
 #endif
 
+#ifdef JSON_NO_SIMD
+    #define NLOHMANN_JSON_ABI_TAG_NO_SIMD _ns
+#else
+    #define NLOHMANN_JSON_ABI_TAG_NO_SIMD
+#endif
+
 #ifdef JSON_VIEW_NO_SIMD
     #define NLOHMANN_JSON_ABI_TAG_VIEW_NO_SIMD _vns
 #else
@@ -203,9 +209,9 @@
 #endif
 
 // Construct the namespace ABI tags component
-#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_CONCAT_EX(a, b, c, d, e, f, g, h, i, j, k, l, m) json_abi ## a ## b ## c ## d ## e ## f ## g ## h ## i ## j ## k ## l ## m
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c, d, e, f, g, h, i, j, k, l, m)
 
 #define NLOHMANN_JSON_ABI_TAGS                                       \
     NLOHMANN_JSON_ABI_TAGS_CONCAT(                                   \
@@ -219,6 +225,7 @@
             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_NO_SIMD,                           \
             NLOHMANN_JSON_ABI_TAG_VIEW_NO_SIMD,                      \
             NLOHMANN_JSON_ABI_TAG_VIEW_USE_SSSE3)
 
@@ -11299,6 +11306,331 @@ NLOHMANN_JSON_NAMESPACE_END
 
 // #include <nlohmann/detail/bit_ops.hpp>
 
+// #include <nlohmann/detail/input/simd_scan.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  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
+
+
+
+#include <array> // array
+#include <atomic> // atomic
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t, uint64_t
+
+// #include <nlohmann/detail/bit_ops.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+// Vector code for runs of string bytes, used by the lexer for contiguous input
+// (see string_bulk_run()). 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. Code compiled for SSSE3 anyway (-mssse3 or more) skips the
+// CPU check. JSON_NO_SIMD selects the portable code everywhere.
+// NLOHMANN_JSON_NO_TARGET_ATTRIBUTE (for compilers that ignore the target
+// attribute and then reject the SSSE3 intrinsics) keeps the check portable
+// without disabling the SSE2 code.
+#if !defined(JSON_NO_SIMD) && defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) && defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    #include <arm_neon.h>
+    #define JSON_SCAN_NEON 1
+#else
+    #define JSON_SCAN_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_NO_SIMD) && !JSON_SCAN_NEON && (defined(__x86_64__) || defined(__i386__) || defined(_M_X64) || defined(_M_IX86)) && !defined(_M_ARM64EC) && (defined(__SSE2__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 2))
+    #include <emmintrin.h>
+    #define JSON_SCAN_SSE2 1
+#else
+    #define JSON_SCAN_SSE2 0
+#endif
+#if JSON_SCAN_SSE2 && defined(__SSSE3__)
+    #include <tmmintrin.h>
+    #define JSON_SCAN_SSSE3 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+#else
+    #define JSON_SCAN_SSSE3 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+#endif
+#if JSON_SCAN_SSE2 && !JSON_SCAN_SSSE3 && !defined(NLOHMANN_JSON_NO_TARGET_ATTRIBUTE) && ((defined(__clang__) && __clang_major__ >= 4) || (defined(__GNUC__) && !defined(__clang__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9)) && !defined(__cpp_modules)))
+    // (GCC before 4.9 has no SSSE3 intrinsics without -mssse3; GCC with C++20
+    // modules drops the target attribute from functions it reads from a module
+    // interface and then rejects the intrinsics: "needs isa option -mssse3")
+    // <cpuid.h> of GCC before 11 has no include guard, and json.hpp and
+    // json_view.hpp both include it
+    #ifndef NLOHMANN_JSON_CPUID_INCLUDED
+        #define NLOHMANN_JSON_CPUID_INCLUDED
+        #include <cpuid.h>
+    #endif
+    #include <tmmintrin.h>
+    #define JSON_SCAN_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define JSON_SCAN_SSSE3_TARGET __attribute__((target("ssse3")))
+#elif JSON_SCAN_SSE2 && !JSON_SCAN_SSSE3 && defined(_MSC_VER)
+    // (MSVC compiles intrinsics of any instruction set)
+    #include <intrin.h>
+    #include <tmmintrin.h>
+    #define JSON_SCAN_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define JSON_SCAN_SSSE3_TARGET
+#else
+    #define JSON_SCAN_SSSE3_DISPATCH 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define JSON_SCAN_SSSE3_TARGET
+#endif
+#define JSON_SCAN_VECTOR (JSON_SCAN_NEON || JSON_SCAN_SSE2)
+#define JSON_SCAN_VECTOR_UTF8 (JSON_SCAN_NEON || JSON_SCAN_SSSE3 || JSON_SCAN_SSSE3_DISPATCH)
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+#if JSON_SCAN_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.
+*/
+JSON_HEDLEY_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned char* p, const unsigned char* e) noexcept
+{
+    while (e - p >= 16)
+    {
+#if JSON_SCAN_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 JSON_SCAN_VECTOR
+/// vector_plain_run() for runs that are already longer than a few words:
+/// inline with SSE2, out of line with NEON, where a larger string loop slows
+/// down the rest of the lexer
+#if JSON_SCAN_NEON
+    JSON_HEDLEY_NEVER_INLINE inline
+#else
+    JSON_HEDLEY_ALWAYS_INLINE
+#endif
+const unsigned char* vector_plain_run_long(const unsigned char* p, const unsigned char* e) noexcept
+{
+    return vector_plain_run(p, e);
+}
+#endif
+
+#if JSON_SCAN_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)
+JSON_HEDLEY_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 (JSON_HEDLEY_UNLIKELY(state == 0))
+    {
+        state = cpu_ssse3() ? 2 : 1;
+        known.store(state, std::memory_order_relaxed);
+    }
+    return state == 2;
+}
+#endif
+
+#if JSON_SCAN_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
+    }
+};
+
+/*!
+@brief the run of plain string bytes and well-formed UTF-8 at 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 of J. Keiser and D. Lemire covers the bytes up
+to it. Returns true with @a block at that byte; returns false with @a block at
+the start of the 16 bytes where the check found ill-formed UTF-8 or where
+fewer than 16 bytes are left, for the caller to finish one sequence at a time.
+Out of line, so that no constants of the check occupy registers in the
+caller's loop. On x86-64, it is compiled for SSSE3 (see cpu_has_ssse3()).
+*/
+JSON_SCAN_SSSE3_TARGET JSON_HEDLEY_NEVER_INLINE inline bool vector_utf8_run(const unsigned char* p, const unsigned char* e, const unsigned char*& block_out) noexcept
+{
+    using lookup = utf8_lookup4<>;
+    const unsigned char* block = p;
+#if JSON_SCAN_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)
+            {
+                block_out = block + k;
+                return true;
+            }
+            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)
+            {
+                block_out = block + k;
+                return true;
+            }
+            break;
+        }
+        if (err_bits != 0)
+        {
+            break;
+        }
+        prev = in;
+        block += 16;
+    }
+#endif
+    block_out = block;
+    return false;
+}
+#endif
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
 // #include <nlohmann/detail/macro_scope.hpp>
 
 
@@ -11320,7 +11652,7 @@ NLOHMANN_JSON_NAMESPACE_END
 // This file contains the byte-level string-scanning helpers used by the lexer's
 // contiguous fast path. They operate purely on raw bytes (no dependency on the
 // lexer's template parameters) so they are free functions, keeping the lexer
-// itself focused on the state machine; see lexer::scan_string_bulk().
+// itself focused on the state machine; see lexer::get_after_string_run().
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -11352,7 +11684,7 @@ inline std::uint64_t swar_string_special(std::uint64_t v) noexcept
 
 // return the index of the first is_string_special() byte in [data, data+n), or
 // n if every byte is ordinary; scans 8 bytes at a time
-inline std::size_t find_string_special(const unsigned char* data, std::size_t n) noexcept
+JSON_HEDLEY_ALWAYS_INLINE std::size_t find_string_special(const unsigned char* data, std::size_t n) noexcept
 {
     std::size_t i = 0;
     for (; i + 8 <= n; i += 8)
@@ -11521,35 +11853,143 @@ inline std::size_t valid_utf8_prefix(const unsigned char* data, std::size_t n) n
     return pos;
 }
 
-// Scalar (C++11) computation of the bulk run length: the number of leading
-// bytes in [data, data+n) that are ordinary ASCII or complete well-formed UTF-8
+#if JSON_SCAN_VECTOR_UTF8
+/// 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) 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 (is_string_special(*p))
+            {
+                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;
+}
+
+/// the rest of a string from p (a character boundary), with the vector UTF-8
+/// check and the scalar check for the end of the block where it stopped
+inline const unsigned char* scan_string_vector(const unsigned char* p, const unsigned char* e) noexcept
+{
+    const unsigned char* block = p;
+    if (vector_utf8_run(p, e, block))
+    {
+        return block;
+    }
+    return scan_string_finish(p, block, e);
+}
+#endif
+
+#if JSON_SCAN_VECTOR
+/// the end of the run of plain string characters at @a p: the first quote,
+/// backslash, control character, or non-ASCII byte, or @a e. The first words
+/// are checked with SWAR (one; four with NEON, where the vector mask costs
+/// more), as most keys and short values end there; longer runs continue 16
+/// bytes at a time.
+JSON_HEDLEY_ALWAYS_INLINE const unsigned char* plain_run_end(const unsigned char* p, const unsigned char* e) noexcept
+{
+    constexpr int prefix_words = JSON_SCAN_NEON ? 4 : 1;
+    for (int words = prefix_words; words > 0 && e - p >= 8; --words)
+    {
+        const std::uint64_t special = swar_string_special(read_eight_bytes(p));
+        if (special != 0)
+        {
+            return p + (static_cast<std::size_t>(count_trailing_zeros(special)) / 8);
+        }
+        p += 8;
+    }
+    // the vector loop only for 16 bytes or more: the serializer passes a
+    // string without its closing quote, so most of its runs end at e, and a
+    // call of the out-of-line NEON loop for a shorter rest made dump() up to
+    // 10% slower
+    if (e - p >= 16)
+    {
+        p = vector_plain_run_long(p, e);
+        if (e - p >= 16)
+        {
+            return p; // the vector loop stopped at a special byte
+        }
+    }
+    return p + find_string_special(p, static_cast<std::size_t>(e - p));
+}
+#endif
+
+// Computation of the bulk run length: the number of leading bytes in
+// [data, data+n) that are ordinary ASCII or complete well-formed UTF-8
 // sequences, stopping before the first byte that needs individual handling (the
 // closing quote, an escape, a control character, or an ill-formed/truncated
-// sequence). ASCII is skipped 8 bytes at a time.
-inline std::size_t scalar_string_bulk_run(const unsigned char* data, std::size_t n) noexcept
+// sequence). ASCII is skipped 16 bytes at a time with SSE2 and 8 bytes at a
+// time otherwise; runs of non-ASCII text use the vector UTF-8 check where
+// available.
+JSON_HEDLEY_ALWAYS_INLINE std::size_t scalar_string_bulk_run(const unsigned char* data, std::size_t n) noexcept
 {
-    std::size_t pos = 0;
-    while (pos < n)
+    const unsigned char* const e = data + n;
+    const unsigned char* p = data;
+    while (p != e)
     {
-        pos += find_string_special(data + pos, n - pos);
-        if (pos >= n || data[pos] < 0x80u)
+#if JSON_SCAN_VECTOR
+        p = plain_run_end(p, e);
+#else
+        p += find_string_special(p, static_cast<std::size_t>(e - p));
+#endif
+        if (p == e || *p < 0x80u)
         {
             break; // end of buffer, or a quote/escape/control byte
         }
+#if JSON_SCAN_VECTOR_UTF8
+#if JSON_SCAN_SSSE3_DISPATCH
+        if (JSON_HEDLEY_LIKELY(cpu_has_ssse3()))
+#endif
+        {
+            return static_cast<std::size_t>(scan_string_vector(p, e) - data);
+        }
+#endif
+#if !JSON_SCAN_VECTOR_UTF8 || JSON_SCAN_SSSE3_DISPATCH
         // a run of multi-byte sequences (e.g. CJK text) is validated sequence
         // by sequence without searching for the next special byte in between
         do
         {
-            const std::size_t seq = validate_one_utf8(data + pos, n - pos);
+            const std::size_t seq = validate_one_utf8(p, static_cast<std::size_t>(e - p));
             if (seq == 0)
             {
-                return pos; // ill-formed or truncated: let the byte path diagnose it
+                return static_cast<std::size_t>(p - data); // ill-formed or truncated: let the byte path diagnose it
             }
-            pos += seq;
+            p += seq;
         }
-        while (pos < n && data[pos] >= 0x80u);
+        while (p != e && *p >= 0x80u);
+#endif
     }
-    return pos;
+    return static_cast<std::size_t>(p - data);
 }
 
 #if defined(JSON_USE_SIMDUTF) && defined(JSON_HAS_CPP_17)
@@ -11590,7 +12030,7 @@ inline std::size_t find_string_delimiter(const unsigned char* data, std::size_t
 // next delimiter is validated in one shot by simdutf; on the rare failure the
 // scalar helper recomputes the exact valid prefix so the byte path still
 // produces the precise diagnostic. Without it, the pure scalar path is used.
-inline std::size_t string_bulk_run(const unsigned char* data, std::size_t n) noexcept
+JSON_HEDLEY_ALWAYS_INLINE std::size_t string_bulk_run(const unsigned char* data, std::size_t n) noexcept
 {
 #if defined(JSON_USE_SIMDUTF) && defined(JSON_HAS_CPP_17)
     const std::size_t run = find_string_delimiter(data, n);
@@ -11602,6 +12042,42 @@ inline std::size_t string_bulk_run(const unsigned char* data, std::size_t n) noe
     return scalar_string_bulk_run(data, n);
 }
 
+#if JSON_SCAN_NEON && !(defined(JSON_USE_SIMDUTF) && defined(JSON_HAS_CPP_17))
+/// scalar_string_bulk_run() from @a start on, out of line (see
+/// lexer_string_bulk_run())
+JSON_HEDLEY_NEVER_INLINE inline std::size_t string_bulk_run_from(const unsigned char* data, std::size_t start, std::size_t n) noexcept
+{
+    return start + scalar_string_bulk_run(data + start, n - start);
+}
+#endif
+
+// string_bulk_run() for the lexer. With NEON, only the first four words are
+// checked inline (most keys and short values end there), and longer runs and
+// non-ASCII text continue out of line. The lexer's scan function runs once per
+// token, and the vector code inlined into it needs more registers, which every
+// token pays for: 5% of json::accept() on jeopardy.json with Apple Clang on
+// Apple M1, even though no string of that file reaches the vector code. The
+// serializer keeps string_bulk_run() inline, which is faster there.
+JSON_HEDLEY_ALWAYS_INLINE std::size_t lexer_string_bulk_run(const unsigned char* data, std::size_t n) noexcept
+{
+#if JSON_SCAN_NEON && !(defined(JSON_USE_SIMDUTF) && defined(JSON_HAS_CPP_17))
+    std::size_t i = 0;
+    for (int words = 4; words > 0 && n - i >= 8; --words)
+    {
+        const std::uint64_t special = swar_string_special(read_eight_bytes(data + i));
+        if (special != 0)
+        {
+            const std::size_t pos = i + (static_cast<std::size_t>(count_trailing_zeros(special)) / 8);
+            return data[pos] < 0x80u ? pos : string_bulk_run_from(data, pos, n);
+        }
+        i += 8;
+    }
+    return string_bulk_run_from(data, i, n);
+#else
+    return string_bulk_run(data, n);
+#endif
+}
+
 // Decode the 4 hex digits at [data, data+4) - the digits following a `\u`
 // escape - into a codepoint 0x0000..0xFFFF via one table lookup per byte
 // (after yyjson's read_hex_u16), or return -1 if any of the 4 bytes is not a
@@ -12048,30 +12524,43 @@ class lexer : public lexer_base<BasicJsonType>
     }
 
     /// contiguous input: bulk-append the run of ordinary characters and complete
-    /// well-formed UTF-8 sequences starting at the current read position, leaving
-    /// the first byte that needs individual handling (the closing quote, an
-    /// escape, a control character, or an ill-formed UTF-8 byte) for get()
-    void scan_string_bulk(std::true_type /*bulk*/)
+    /// well-formed UTF-8 sequences starting at the current read position, then
+    /// read the first byte that needs individual handling (the closing quote, an
+    /// escape, a control character, or an ill-formed UTF-8 byte) like get().
+    /// That byte is taken from the buffer that was just scanned. Through get(),
+    /// Apple Clang merged this read with the other get() calls of the scan
+    /// function, which load the read position again right after it was stored
+    /// (3% of json::accept() on jeopardy.json on Apple M1).
+    char_int_type get_after_string_run(std::true_type /*bulk*/)
     {
         // a pending unget must be consumed through the normal path first
         if (next_unget)
         {
-            return;
+            return get();
         }
         const std::size_t remaining = ia.bulk_remaining();
         if (remaining == 0)
         {
-            return;
+            return get();
         }
-        const auto* const data = reinterpret_cast<const unsigned char*>(ia.bulk_data());
+        const auto* const raw = ia.bulk_data();
+        const auto* const data = reinterpret_cast<const unsigned char*>(raw);
 
-        const std::size_t pos = string_bulk_run(data, remaining);
-        if (pos == 0)
+        const std::size_t pos = lexer_string_bulk_run(data, remaining);
+        if (pos != 0)
         {
-            return;
+            token_buffer.append(reinterpret_cast<const typename string_t::value_type*>(data), pos);
+            bulk_consume(reinterpret_cast<const char*>(data), pos);
         }
-        token_buffer.append(reinterpret_cast<const typename string_t::value_type*>(data), pos);
-        bulk_consume(reinterpret_cast<const char*>(data), pos);
+        if (pos == remaining)
+        {
+            return get(); // end of the contiguous input
+        }
+        // what get() does, with the byte at hand
+        advance_position();
+        ia.bulk_skip(1);
+        current = char_traits<char_type>::to_int_type(raw[pos]);
+        return track_after_read();
     }
 
     /// consume @a n characters at @a data (the adapter's bulk_data()) and do
@@ -12104,7 +12593,10 @@ class lexer : public lexer_base<BasicJsonType>
     }
 
     /// streaming input: no bulk fast path
-    void scan_string_bulk(std::false_type /*bulk*/) const noexcept {}
+    char_int_type get_after_string_run(std::false_type /*bulk*/)
+    {
+        return get();
+    }
 
     /*!
     @brief scan a string literal
@@ -12132,11 +12624,8 @@ class lexer : public lexer_base<BasicJsonType>
         while (true)
         {
             // bulk-consume ordinary characters from contiguous input, then
-            // handle the next special byte through the switch below
-            scan_string_bulk(std::integral_constant<bool, bulk_scan> {});
-
-            // get the next character
-            switch (get())
+            // get the next character, the one that ends the run
+            switch (get_after_string_run(std::integral_constant<bool, bulk_scan>()))
             {
                 // end of file while parsing the string
                 case char_traits<char_type>::eof():
@@ -27610,12 +28099,12 @@ NLOHMANN_JSON_NAMESPACE_END
 #endif
 
 // SSE2 (every x86-64 CPU) and NEON (every 64-bit Arm CPU) convert the 16
-// digits of a double at once
-#if defined(__x86_64__) || (defined(_M_X64) && !defined(_M_ARM64EC))
+// digits of a double at once (unless JSON_NO_SIMD selects the portable code)
+#if !defined(JSON_NO_SIMD) && (defined(__x86_64__) || (defined(_M_X64) && !defined(_M_ARM64EC)))
     #include <emmintrin.h>
     #define JSON_DTOA_SSE2 1
     #define JSON_DTOA_NEON 0
-#elif (defined(__aarch64__) || defined(_M_ARM64)) && !defined(_M_ARM64EC) && !defined(__ARM_BIG_ENDIAN)
+#elif !defined(JSON_NO_SIMD) && (defined(__aarch64__) || defined(_M_ARM64)) && !defined(_M_ARM64EC) && !defined(__ARM_BIG_ENDIAN)
     #include <arm_neon.h>
     #define JSON_DTOA_SSE2 0
     #define JSON_DTOA_NEON 1
@@ -39627,6 +40116,13 @@ struct formatter<nlohmann::NLOHMANN_BASIC_JSON_TPL, char> // NOLINT(cert-dcl58-c
 #undef JSON_UNLOCKED_FILE_IO
 #undef JSON_DTOA_SSE2
 #undef JSON_DTOA_NEON
+#undef JSON_SCAN_NEON
+#undef JSON_SCAN_SSE2
+#undef JSON_SCAN_SSSE3
+#undef JSON_SCAN_SSSE3_DISPATCH
+#undef JSON_SCAN_SSSE3_TARGET
+#undef JSON_SCAN_VECTOR
+#undef JSON_SCAN_VECTOR_UTF8
 
 #ifndef JSON_TEST_KEEP_MACROS
     #undef JSON_CATCH

single_include/nlohmann/json_fwd.hpp

@@ -139,6 +139,12 @@
     #define NLOHMANN_JSON_ABI_TAG_DISABLE_UNLOCKED_FILE_IO
 #endif
 
+#ifdef JSON_NO_SIMD
+    #define NLOHMANN_JSON_ABI_TAG_NO_SIMD _ns
+#else
+    #define NLOHMANN_JSON_ABI_TAG_NO_SIMD
+#endif
+
 #ifdef JSON_VIEW_NO_SIMD
     #define NLOHMANN_JSON_ABI_TAG_VIEW_NO_SIMD _vns
 #else
@@ -156,9 +162,9 @@
 #endif
 
 // Construct the namespace ABI tags component
-#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_CONCAT_EX(a, b, c, d, e, f, g, h, i, j, k, l, m) json_abi ## a ## b ## c ## d ## e ## f ## g ## h ## i ## j ## k ## l ## m
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c, d, e, f, g, h, i, j, k, l, m)
 
 #define NLOHMANN_JSON_ABI_TAGS                                       \
     NLOHMANN_JSON_ABI_TAGS_CONCAT(                                   \
@@ -172,6 +178,7 @@
             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_NO_SIMD,                           \
             NLOHMANN_JSON_ABI_TAG_VIEW_NO_SIMD,                      \
             NLOHMANN_JSON_ABI_TAG_VIEW_USE_SSSE3)
 

single_include/nlohmann/json_view.hpp

@@ -690,7 +690,12 @@ NLOHMANN_JSON_NAMESPACE_END
 #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>
+    // <cpuid.h> of GCC before 11 has no include guard, and json.hpp and
+    // json_view.hpp both include it
+    #ifndef NLOHMANN_JSON_CPUID_INCLUDED
+        #define NLOHMANN_JSON_CPUID_INCLUDED
+        #include <cpuid.h>
+    #endif
     #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")))