Vector scans: one compare for the quote and the control characters

The SSE2/NEON tests for the bytes a string run stops at compared with the
quote, with the backslash and with 0x20 separately. Flipping bit 1 maps
0x00-0x1F onto itself and the quote onto 0x20, and every other ASCII byte
to 33 or more, so (c ^ 2) < 33 finds the quote and the control characters
with one compare: signed where the bytes from 0x80 stop the run too,
unsigned (or with a saturating subtraction / minimum on SSE2) where they
do not. This saves one vector operation per 16 bytes in all twelve vector
tests: the lexer's vector_plain_run() and vector_utf8_run(), dump()'s
copy_plain_halves() and copy_plain_vectors(), and json_view's scanner and
serializer. The DEL tests of the ensure_ascii paths stay.

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

include/nlohmann/detail/input/simd_scan.hpp

@@ -92,8 +92,12 @@ namespace detail
 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
+the two apart). One signed compare finds the quote, the control characters and
-non-ASCII bytes at once.
+the non-ASCII bytes at once: flipping bit 1 maps 0x00-0x1F onto itself and the
+quote (0x22) onto 0x20, keeps 0x80-0xFF negative, and maps every other byte to
+33 or more, so these are exactly the bytes with (int8_t)(c ^ 2) < 33 (see
+D. Lemire, "Detect control characters, quotes and backslashes efficiently
+using SWAR", 2025). The backslash needs its own compare.
 */
 JSON_HEDLEY_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned char* p, const unsigned char* e) noexcept
 {
@@ -101,8 +105,8 @@ JSON_HEDLEY_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned c
     {
 #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('\\'))),
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')),
-                                            vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)));
+                                            vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
         // 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);
@@ -112,8 +116,8 @@ JSON_HEDLEY_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned c
         }
 #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('\\'))),
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
-                                             _mm_cmplt_epi8(in, _mm_set1_epi8(0x20)));
+                                             _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
         const auto bits = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
         if (bits != 0)
         {
@@ -250,7 +254,8 @@ JSON_SCAN_SSSE3_TARGET JSON_HEDLEY_NEVER_INLINE inline bool vector_utf8_run(cons
     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)));
+        // the quote and the control characters: (c ^ 2) < 33 (see vector_plain_run())
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vcltq_u8(veorq_u8(in, vdupq_n_u8(2)), vdupq_n_u8(33)));
         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)));
@@ -290,8 +295,10 @@ JSON_SCAN_SSSE3_TARGET JSON_HEDLEY_NEVER_INLINE inline bool vector_utf8_run(cons
     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('\\'))),
+        // the quote and the control characters: (c ^ 2) < 33, that is,
-                                             _mm_cmpeq_epi8(_mm_subs_epu8(in, _mm_set1_epi8(0x1F)), zero)); // in < 0x20
+        // (c ^ 2) - 32 saturates to 0 (see vector_plain_run())
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
+                                             _mm_cmpeq_epi8(_mm_subs_epu8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(32)), zero));
         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))),

include/nlohmann/detail/input/string_scan.hpp

@@ -58,11 +58,10 @@ inline std::uint64_t swar_string_special(std::uint64_t v) noexcept
     constexpr std::uint64_t high = 0x8080808080808080ull;
     const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'  (0x22)
     const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\' (0x5C)
-    const std::uint64_t has_quote     = (q - ones) & ~q & high;
+    // The usual test for a zero byte is (x - ones) & ~x & high. The `& ~x`
-    const std::uint64_t has_backslash = (b - ones) & ~b & high;
+    // only removes bytes from 0x80, and `v & high` flags those anyway, so it
-    const std::uint64_t has_control   = (v - 0x2020202020202020ull) & ~v & high; // < 0x20
+    // can be dropped. Clang finds this reduction itself; GCC does not.
-    const std::uint64_t has_non_ascii = v & high;                                // >= 0x80
+    return ((q - ones) | (b - ones) | (v - 0x2020202020202020ull) | v) & high;
-    return has_quote | has_backslash | has_control | has_non_ascii;
 }
 
 // return the index of the first is_string_special() byte in [data, data+n), or
@@ -143,11 +142,8 @@ inline std::size_t find_ascii_copyable_run(const unsigned char* data, std::size_
         const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'  (0x22)
         const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\' (0x5C)
         const std::uint64_t d = v ^ 0x7F7F7F7F7F7F7F7Full; // DEL  (0x7F)
-        const std::uint64_t stop = ((q - ones) & ~q & high)   // == '"'
+        // (see swar_string_special(): the bytes from 0x80 are stops anyway)
-                                   | ((b - ones) & ~b & high)  // == '\\'
+        const std::uint64_t stop = ((q - ones) | (b - ones) | (d - ones) | (v - 0x2020202020202020ull) | v) & high;
-                                   | ((d - ones) & ~d & high)  // == 0x7F
-                                   | ((v - 0x2020202020202020ull) & ~v & high) // < 0x20
-                                   | (v & high);               // >= 0x80
         if (stop != 0)
         {
             // the lowest flagged byte is the first one to stop at (see

include/nlohmann/detail/output/serializer.hpp

@@ -724,10 +724,8 @@ class serializer
         constexpr std::uint64_t ones = 0x0101010101010101ull;
         const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'
         const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\'
-        return (((q - ones) & ~q) | ((b - ones) & ~b)         // == '"', == '\\'
+        // (see swar_string_special(): the bytes from 0x80 are flagged anyway)
-                | ((v - 0x2020202020202020ull) & ~v)          // < 0x20
+        return ((q - ones) | (b - ones) | (v - 0x2020202020202020ull) | v | (v + ones)) & 0x8080808080808080ull;
-                | v | (v + ones))                             // >= 0x7F
-               & 0x8080808080808080ull;
     }
 
     template<typename WordType>
@@ -836,8 +834,10 @@ class serializer
         const uint8x8_t lo = vld1_u8(reinterpret_cast<const std::uint8_t*>(p));
         const uint8x8_t hi = vld1_u8(reinterpret_cast<const std::uint8_t*>(p + n - 8));
         const uint8x16_t in = vcombine_u8(lo, hi);
-        const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('"')), vceqq_u8(in, vdupq_n_u8('\\'))),
+        // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
-                                            vorrq_u8(vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)), vceqq_u8(in, vdupq_n_u8(0x7F))));
+        // non-ASCII bytes (see vector_plain_run())
+        const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vceqq_u8(in, vdupq_n_u8(0x7F))),
+                                            vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
         if (vmaxvq_u8(special) != 0)
         {
             return false;
@@ -848,8 +848,8 @@ class serializer
         const __m128i lo = _mm_loadl_epi64(static_cast<const __m128i*>(static_cast<const void*>(p))); // NOLINT(bugprone-casting-through-void)
         const __m128i hi = _mm_loadl_epi64(static_cast<const __m128i*>(static_cast<const void*>(p + n - 8))); // NOLINT(bugprone-casting-through-void)
         const __m128i in = _mm_unpacklo_epi64(lo, hi);
-        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('"')), _mm_cmpeq_epi8(in, _mm_set1_epi8('\\'))),
+        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')), _mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F))),
-                                             _mm_or_si128(_mm_cmplt_epi8(in, _mm_set1_epi8(0x20)), _mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F))));
+                                             _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
         if (_mm_movemask_epi8(special) != 0)
         {
             return false;
@@ -868,10 +868,13 @@ class serializer
 #if JSON_SCAN_NEON
         const uint8x16_t a = vld1q_u8(reinterpret_cast<const std::uint8_t*>(p));
         const uint8x16_t b = vld1q_u8(reinterpret_cast<const std::uint8_t*>(p + n - 16));
-        const uint8x16_t special = vorrq_u8(vorrq_u8(vorrq_u8(vceqq_u8(a, vdupq_n_u8('"')), vceqq_u8(a, vdupq_n_u8('\\'))),
+        // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
-                                            vorrq_u8(vceqq_u8(b, vdupq_n_u8('"')), vceqq_u8(b, vdupq_n_u8('\\')))),
+        // non-ASCII bytes (see vector_plain_run())
-                                            vorrq_u8(vorrq_u8(vcltq_s8(vreinterpretq_s8_u8(a), vdupq_n_s8(0x20)), vcltq_s8(vreinterpretq_s8_u8(b), vdupq_n_s8(0x20))),
+        const uint8x16_t flip = vdupq_n_u8(2);
-                                                vorrq_u8(vceqq_u8(a, vdupq_n_u8(0x7F)), vceqq_u8(b, vdupq_n_u8(0x7F)))));
+        const int8x16_t limit = vdupq_n_s8(33);
+        const uint8x16_t special = vorrq_u8(vorrq_u8(vorrq_u8(vceqq_u8(a, vdupq_n_u8('\\')), vceqq_u8(b, vdupq_n_u8('\\'))),
+                                            vorrq_u8(vceqq_u8(a, vdupq_n_u8(0x7F)), vceqq_u8(b, vdupq_n_u8(0x7F)))),
+                                            vorrq_u8(vcltq_s8(vreinterpretq_s8_u8(veorq_u8(a, flip)), limit), vcltq_s8(vreinterpretq_s8_u8(veorq_u8(b, flip)), limit)));
         if (vmaxvq_u8(special) != 0)
         {
             return false;
@@ -881,14 +884,15 @@ class serializer
 #else
         const __m128i a = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(p))); // NOLINT(bugprone-casting-through-void)
         const __m128i b = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(p + n - 16))); // NOLINT(bugprone-casting-through-void)
-        const __m128i quote = _mm_set1_epi8('"');
+        // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
+        // non-ASCII bytes (see vector_plain_run())
         const __m128i backslash = _mm_set1_epi8('\\');
-        const __m128i space = _mm_set1_epi8(0x20);
         const __m128i del = _mm_set1_epi8(0x7F);
-        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(a, quote), _mm_cmpeq_epi8(a, backslash)),
+        const __m128i flip = _mm_set1_epi8(2);
-                                             _mm_or_si128(_mm_cmpeq_epi8(b, quote), _mm_cmpeq_epi8(b, backslash))),
+        const __m128i limit = _mm_set1_epi8(33);
-                                             _mm_or_si128(_mm_or_si128(_mm_cmplt_epi8(a, space), _mm_cmplt_epi8(b, space)),
+        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(a, backslash), _mm_cmpeq_epi8(b, backslash)),
-                                                 _mm_or_si128(_mm_cmpeq_epi8(a, del), _mm_cmpeq_epi8(b, del))));
+                                             _mm_or_si128(_mm_cmpeq_epi8(a, del), _mm_cmpeq_epi8(b, del))),
+                                             _mm_or_si128(_mm_cmplt_epi8(_mm_xor_si128(a, flip), limit), _mm_cmplt_epi8(_mm_xor_si128(b, flip), limit)));
         if (_mm_movemask_epi8(special) != 0)
         {
             return false;

include/nlohmann/detail/view/serializer.hpp

@@ -152,9 +152,12 @@ inline std::size_t plain_output_run(const unsigned char* s, std::size_t n) noexc
     for (; i + 16 <= n; i += 16)
     {
         const __m128i in = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(s + i))); // NOLINT(bugprone-casting-through-void)
-        // (a byte is below 0x20 if it is its own minimum with 0x1F)
+        // the quote and the control characters are the bytes with
-        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('"')), _mm_cmpeq_epi8(in, _mm_set1_epi8('\\'))),
+        // (c ^ 2) <= 32, that is, those that are their own minimum with 32
-                                             _mm_cmpeq_epi8(_mm_min_epu8(in, _mm_set1_epi8(0x1F)), in));
+        // (see vector_plain_run())
+        const __m128i flipped = _mm_xor_si128(in, _mm_set1_epi8(2));
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
+                                             _mm_cmpeq_epi8(_mm_min_epu8(flipped, _mm_set1_epi8(32)), flipped));
         const auto mask = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
         if (mask != 0)
         {
@@ -217,20 +220,21 @@ inline bool is_ascii_without_delete(const char* s, std::size_t n) noexcept
 
 #if NLOHMANN_VIEW_VECTOR
 /// the bytes of the 16 at p that ensure_ascii cannot copy (a quote, a
-/// backslash, a control character, DEL, or not ASCII; a signed compare with
+/// backslash, a control character, DEL, or not ASCII; the signed compare
-/// 0x20 finds control characters and non-ASCII bytes at once), one bit per
+/// (int8_t)(c ^ 2) < 33 finds the quote, the control characters and the
-/// byte (x86) or one nibble per byte (NEON)
+/// non-ASCII bytes at once, see vector_plain_run()), one bit per byte (x86) or
+/// one nibble per byte (NEON)
 NLOHMANN_VIEW_ALWAYS_INLINE std::uint64_t ascii_stop_mask(const unsigned char* p) noexcept
 {
 #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('\\'))),
+    const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vceqq_u8(in, vdupq_n_u8(0x7F))),
-                                        vorrq_u8(vceqq_u8(in, vdupq_n_u8(0x7F)), vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20))));
+                                        vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
     return vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(special), 4)), 0);
 #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('\\'))),
+    const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')), _mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F))),
-                                         _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F)), _mm_cmplt_epi8(in, _mm_set1_epi8(0x20))));
+                                         _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
     return static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
 #endif
 }

include/nlohmann/detail/view/simd.hpp

@@ -91,8 +91,10 @@ namespace view
 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
+the two apart). One signed compare finds the quote, the control characters and
-non-ASCII bytes at once.
+the non-ASCII bytes at once: these are exactly the bytes with
+(int8_t)(c ^ 2) < 33 (see detail::vector_plain_run()). The backslash needs its
+own compare.
 */
 NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned char* p, const unsigned char* e) noexcept
 {
@@ -100,8 +102,8 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned
     {
 #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('\\'))),
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')),
-                                            vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)));
+                                            vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
         // 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);
@@ -111,8 +113,8 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned
         }
 #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('\\'))),
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
-                                             _mm_cmplt_epi8(in, _mm_set1_epi8(0x20)));
+                                             _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
         const auto bits = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
         if (bits != 0)
         {
@@ -138,15 +140,17 @@ NLOHMANN_VIEW_ALWAYS_INLINE unsigned copy_plain_run16(const unsigned char* p, ch
 #if NLOHMANN_VIEW_NEON
     const uint8x16_t in = vld1q_u8(p);
     vst1q_u8(reinterpret_cast<std::uint8_t*>(w), in); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
-    const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('"')), vceqq_u8(in, vdupq_n_u8('\\'))),
+    // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
-                                        vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)));
+    // non-ASCII bytes (see vector_plain_run())
+    const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')),
+                                        vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
     const std::uint64_t bits = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(special), 4)), 0);
     return bits != 0 ? static_cast<unsigned>(count_trailing_zeros(bits) >> 2u) : 16u;
 #else
     const __m128i in = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(p))); // NOLINT(bugprone-casting-through-void)
     _mm_storeu_si128(static_cast<__m128i*>(static_cast<void*>(w)), in); // 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('\\'))),
+    const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
-                                         _mm_cmplt_epi8(in, _mm_set1_epi8(0x20)));
+                                         _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
     const auto bits = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
     return bits != 0 ? static_cast<unsigned>(count_trailing_zeros(bits)) : 16u;
 #endif
@@ -306,7 +310,8 @@ NLOHMANN_VIEW_SSSE3_TARGET NLOHMANN_VIEW_NOINLINE NLOHMANN_VIEW_NOCLONE inline c
     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)));
+        // the quote and the control characters: (c ^ 2) < 33 (see vector_plain_run())
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vcltq_u8(veorq_u8(in, vdupq_n_u8(2)), vdupq_n_u8(33)));
         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)));
@@ -345,8 +350,10 @@ NLOHMANN_VIEW_SSSE3_TARGET NLOHMANN_VIEW_NOINLINE NLOHMANN_VIEW_NOCLONE inline c
     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('\\'))),
+        // the quote and the control characters: (c ^ 2) < 33, that is,
-                                             _mm_cmpeq_epi8(_mm_subs_epu8(in, _mm_set1_epi8(0x1F)), zero)); // in < 0x20
+        // (c ^ 2) - 32 saturates to 0 (see vector_plain_run())
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
+                                             _mm_cmpeq_epi8(_mm_subs_epu8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(32)), zero));
         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))),

single_include/nlohmann/json.hpp

@@ -12186,8 +12186,12 @@ namespace detail
 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
+the two apart). One signed compare finds the quote, the control characters and
-non-ASCII bytes at once.
+the non-ASCII bytes at once: flipping bit 1 maps 0x00-0x1F onto itself and the
+quote (0x22) onto 0x20, keeps 0x80-0xFF negative, and maps every other byte to
+33 or more, so these are exactly the bytes with (int8_t)(c ^ 2) < 33 (see
+D. Lemire, "Detect control characters, quotes and backslashes efficiently
+using SWAR", 2025). The backslash needs its own compare.
 */
 JSON_HEDLEY_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned char* p, const unsigned char* e) noexcept
 {
@@ -12195,8 +12199,8 @@ JSON_HEDLEY_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned c
     {
 #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('\\'))),
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')),
-                                            vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)));
+                                            vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
         // 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);
@@ -12206,8 +12210,8 @@ JSON_HEDLEY_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned c
         }
 #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('\\'))),
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
-                                             _mm_cmplt_epi8(in, _mm_set1_epi8(0x20)));
+                                             _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
         const auto bits = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
         if (bits != 0)
         {
@@ -12344,7 +12348,8 @@ JSON_SCAN_SSSE3_TARGET JSON_HEDLEY_NEVER_INLINE inline bool vector_utf8_run(cons
     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)));
+        // the quote and the control characters: (c ^ 2) < 33 (see vector_plain_run())
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vcltq_u8(veorq_u8(in, vdupq_n_u8(2)), vdupq_n_u8(33)));
         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)));
@@ -12384,8 +12389,10 @@ JSON_SCAN_SSSE3_TARGET JSON_HEDLEY_NEVER_INLINE inline bool vector_utf8_run(cons
     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('\\'))),
+        // the quote and the control characters: (c ^ 2) < 33, that is,
-                                             _mm_cmpeq_epi8(_mm_subs_epu8(in, _mm_set1_epi8(0x1F)), zero)); // in < 0x20
+        // (c ^ 2) - 32 saturates to 0 (see vector_plain_run())
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
+                                             _mm_cmpeq_epi8(_mm_subs_epu8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(32)), zero));
         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))),
@@ -12537,11 +12544,10 @@ inline std::uint64_t swar_string_special(std::uint64_t v) noexcept
     constexpr std::uint64_t high = 0x8080808080808080ull;
     const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'  (0x22)
     const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\' (0x5C)
-    const std::uint64_t has_quote     = (q - ones) & ~q & high;
+    // The usual test for a zero byte is (x - ones) & ~x & high. The `& ~x`
-    const std::uint64_t has_backslash = (b - ones) & ~b & high;
+    // only removes bytes from 0x80, and `v & high` flags those anyway, so it
-    const std::uint64_t has_control   = (v - 0x2020202020202020ull) & ~v & high; // < 0x20
+    // can be dropped. Clang finds this reduction itself; GCC does not.
-    const std::uint64_t has_non_ascii = v & high;                                // >= 0x80
+    return ((q - ones) | (b - ones) | (v - 0x2020202020202020ull) | v) & high;
-    return has_quote | has_backslash | has_control | has_non_ascii;
 }
 
 // return the index of the first is_string_special() byte in [data, data+n), or
@@ -12622,11 +12628,8 @@ inline std::size_t find_ascii_copyable_run(const unsigned char* data, std::size_
         const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'  (0x22)
         const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\' (0x5C)
         const std::uint64_t d = v ^ 0x7F7F7F7F7F7F7F7Full; // DEL  (0x7F)
-        const std::uint64_t stop = ((q - ones) & ~q & high)   // == '"'
+        // (see swar_string_special(): the bytes from 0x80 are stops anyway)
-                                   | ((b - ones) & ~b & high)  // == '\\'
+        const std::uint64_t stop = ((q - ones) | (b - ones) | (d - ones) | (v - 0x2020202020202020ull) | v) & high;
-                                   | ((d - ones) & ~d & high)  // == 0x7F
-                                   | ((v - 0x2020202020202020ull) & ~v & high) // < 0x20
-                                   | (v & high);               // >= 0x80
         if (stop != 0)
         {
             // the lowest flagged byte is the first one to stop at (see
@@ -37015,10 +37018,8 @@ class serializer
         constexpr std::uint64_t ones = 0x0101010101010101ull;
         const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'
         const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\'
-        return (((q - ones) & ~q) | ((b - ones) & ~b)         // == '"', == '\\'
+        // (see swar_string_special(): the bytes from 0x80 are flagged anyway)
-                | ((v - 0x2020202020202020ull) & ~v)          // < 0x20
+        return ((q - ones) | (b - ones) | (v - 0x2020202020202020ull) | v | (v + ones)) & 0x8080808080808080ull;
-                | v | (v + ones))                             // >= 0x7F
-               & 0x8080808080808080ull;
     }
 
     template<typename WordType>
@@ -37127,8 +37128,10 @@ class serializer
         const uint8x8_t lo = vld1_u8(reinterpret_cast<const std::uint8_t*>(p));
         const uint8x8_t hi = vld1_u8(reinterpret_cast<const std::uint8_t*>(p + n - 8));
         const uint8x16_t in = vcombine_u8(lo, hi);
-        const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('"')), vceqq_u8(in, vdupq_n_u8('\\'))),
+        // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
-                                            vorrq_u8(vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)), vceqq_u8(in, vdupq_n_u8(0x7F))));
+        // non-ASCII bytes (see vector_plain_run())
+        const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vceqq_u8(in, vdupq_n_u8(0x7F))),
+                                            vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
         if (vmaxvq_u8(special) != 0)
         {
             return false;
@@ -37139,8 +37142,8 @@ class serializer
         const __m128i lo = _mm_loadl_epi64(static_cast<const __m128i*>(static_cast<const void*>(p))); // NOLINT(bugprone-casting-through-void)
         const __m128i hi = _mm_loadl_epi64(static_cast<const __m128i*>(static_cast<const void*>(p + n - 8))); // NOLINT(bugprone-casting-through-void)
         const __m128i in = _mm_unpacklo_epi64(lo, hi);
-        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('"')), _mm_cmpeq_epi8(in, _mm_set1_epi8('\\'))),
+        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')), _mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F))),
-                                             _mm_or_si128(_mm_cmplt_epi8(in, _mm_set1_epi8(0x20)), _mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F))));
+                                             _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
         if (_mm_movemask_epi8(special) != 0)
         {
             return false;
@@ -37159,10 +37162,13 @@ class serializer
 #if JSON_SCAN_NEON
         const uint8x16_t a = vld1q_u8(reinterpret_cast<const std::uint8_t*>(p));
         const uint8x16_t b = vld1q_u8(reinterpret_cast<const std::uint8_t*>(p + n - 16));
-        const uint8x16_t special = vorrq_u8(vorrq_u8(vorrq_u8(vceqq_u8(a, vdupq_n_u8('"')), vceqq_u8(a, vdupq_n_u8('\\'))),
+        // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
-                                            vorrq_u8(vceqq_u8(b, vdupq_n_u8('"')), vceqq_u8(b, vdupq_n_u8('\\')))),
+        // non-ASCII bytes (see vector_plain_run())
-                                            vorrq_u8(vorrq_u8(vcltq_s8(vreinterpretq_s8_u8(a), vdupq_n_s8(0x20)), vcltq_s8(vreinterpretq_s8_u8(b), vdupq_n_s8(0x20))),
+        const uint8x16_t flip = vdupq_n_u8(2);
-                                                vorrq_u8(vceqq_u8(a, vdupq_n_u8(0x7F)), vceqq_u8(b, vdupq_n_u8(0x7F)))));
+        const int8x16_t limit = vdupq_n_s8(33);
+        const uint8x16_t special = vorrq_u8(vorrq_u8(vorrq_u8(vceqq_u8(a, vdupq_n_u8('\\')), vceqq_u8(b, vdupq_n_u8('\\'))),
+                                            vorrq_u8(vceqq_u8(a, vdupq_n_u8(0x7F)), vceqq_u8(b, vdupq_n_u8(0x7F)))),
+                                            vorrq_u8(vcltq_s8(vreinterpretq_s8_u8(veorq_u8(a, flip)), limit), vcltq_s8(vreinterpretq_s8_u8(veorq_u8(b, flip)), limit)));
         if (vmaxvq_u8(special) != 0)
         {
             return false;
@@ -37172,14 +37178,15 @@ class serializer
 #else
         const __m128i a = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(p))); // NOLINT(bugprone-casting-through-void)
         const __m128i b = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(p + n - 16))); // NOLINT(bugprone-casting-through-void)
-        const __m128i quote = _mm_set1_epi8('"');
+        // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
+        // non-ASCII bytes (see vector_plain_run())
         const __m128i backslash = _mm_set1_epi8('\\');
-        const __m128i space = _mm_set1_epi8(0x20);
         const __m128i del = _mm_set1_epi8(0x7F);
-        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(a, quote), _mm_cmpeq_epi8(a, backslash)),
+        const __m128i flip = _mm_set1_epi8(2);
-                                             _mm_or_si128(_mm_cmpeq_epi8(b, quote), _mm_cmpeq_epi8(b, backslash))),
+        const __m128i limit = _mm_set1_epi8(33);
-                                             _mm_or_si128(_mm_or_si128(_mm_cmplt_epi8(a, space), _mm_cmplt_epi8(b, space)),
+        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(a, backslash), _mm_cmpeq_epi8(b, backslash)),
-                                                 _mm_or_si128(_mm_cmpeq_epi8(a, del), _mm_cmpeq_epi8(b, del))));
+                                             _mm_or_si128(_mm_cmpeq_epi8(a, del), _mm_cmpeq_epi8(b, del))),
+                                             _mm_or_si128(_mm_cmplt_epi8(_mm_xor_si128(a, flip), limit), _mm_cmplt_epi8(_mm_xor_si128(b, flip), limit)));
         if (_mm_movemask_epi8(special) != 0)
         {
             return false;

single_include/nlohmann/json_view.hpp

@@ -759,8 +759,10 @@ namespace view
 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
+the two apart). One signed compare finds the quote, the control characters and
-non-ASCII bytes at once.
+the non-ASCII bytes at once: these are exactly the bytes with
+(int8_t)(c ^ 2) < 33 (see detail::vector_plain_run()). The backslash needs its
+own compare.
 */
 NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned char* p, const unsigned char* e) noexcept
 {
@@ -768,8 +770,8 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned
     {
 #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('\\'))),
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')),
-                                            vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)));
+                                            vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
         // 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);
@@ -779,8 +781,8 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned
         }
 #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('\\'))),
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
-                                             _mm_cmplt_epi8(in, _mm_set1_epi8(0x20)));
+                                             _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
         const auto bits = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
         if (bits != 0)
         {
@@ -806,15 +808,17 @@ NLOHMANN_VIEW_ALWAYS_INLINE unsigned copy_plain_run16(const unsigned char* p, ch
 #if NLOHMANN_VIEW_NEON
     const uint8x16_t in = vld1q_u8(p);
     vst1q_u8(reinterpret_cast<std::uint8_t*>(w), in); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
-    const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('"')), vceqq_u8(in, vdupq_n_u8('\\'))),
+    // (int8_t)(c ^ 2) < 33: the quote, the control characters and the
-                                        vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20)));
+    // non-ASCII bytes (see vector_plain_run())
+    const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')),
+                                        vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
     const std::uint64_t bits = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(special), 4)), 0);
     return bits != 0 ? static_cast<unsigned>(count_trailing_zeros(bits) >> 2u) : 16u;
 #else
     const __m128i in = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(p))); // NOLINT(bugprone-casting-through-void)
     _mm_storeu_si128(static_cast<__m128i*>(static_cast<void*>(w)), in); // 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('\\'))),
+    const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
-                                         _mm_cmplt_epi8(in, _mm_set1_epi8(0x20)));
+                                         _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
     const auto bits = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
     return bits != 0 ? static_cast<unsigned>(count_trailing_zeros(bits)) : 16u;
 #endif
@@ -974,7 +978,8 @@ NLOHMANN_VIEW_SSSE3_TARGET NLOHMANN_VIEW_NOINLINE NLOHMANN_VIEW_NOCLONE inline c
     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)));
+        // the quote and the control characters: (c ^ 2) < 33 (see vector_plain_run())
+        const uint8x16_t special = vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vcltq_u8(veorq_u8(in, vdupq_n_u8(2)), vdupq_n_u8(33)));
         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)));
@@ -1013,8 +1018,10 @@ NLOHMANN_VIEW_SSSE3_TARGET NLOHMANN_VIEW_NOINLINE NLOHMANN_VIEW_NOCLONE inline c
     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('\\'))),
+        // the quote and the control characters: (c ^ 2) < 33, that is,
-                                             _mm_cmpeq_epi8(_mm_subs_epu8(in, _mm_set1_epi8(0x1F)), zero)); // in < 0x20
+        // (c ^ 2) - 32 saturates to 0 (see vector_plain_run())
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
+                                             _mm_cmpeq_epi8(_mm_subs_epu8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(32)), zero));
         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))),
@@ -8600,9 +8607,12 @@ inline std::size_t plain_output_run(const unsigned char* s, std::size_t n) noexc
     for (; i + 16 <= n; i += 16)
     {
         const __m128i in = _mm_loadu_si128(static_cast<const __m128i*>(static_cast<const void*>(s + i))); // NOLINT(bugprone-casting-through-void)
-        // (a byte is below 0x20 if it is its own minimum with 0x1F)
+        // the quote and the control characters are the bytes with
-        const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('"')), _mm_cmpeq_epi8(in, _mm_set1_epi8('\\'))),
+        // (c ^ 2) <= 32, that is, those that are their own minimum with 32
-                                             _mm_cmpeq_epi8(_mm_min_epu8(in, _mm_set1_epi8(0x1F)), in));
+        // (see vector_plain_run())
+        const __m128i flipped = _mm_xor_si128(in, _mm_set1_epi8(2));
+        const __m128i special = _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')),
+                                             _mm_cmpeq_epi8(_mm_min_epu8(flipped, _mm_set1_epi8(32)), flipped));
         const auto mask = static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
         if (mask != 0)
         {
@@ -8665,20 +8675,21 @@ inline bool is_ascii_without_delete(const char* s, std::size_t n) noexcept
 
 #if NLOHMANN_VIEW_VECTOR
 /// the bytes of the 16 at p that ensure_ascii cannot copy (a quote, a
-/// backslash, a control character, DEL, or not ASCII; a signed compare with
+/// backslash, a control character, DEL, or not ASCII; the signed compare
-/// 0x20 finds control characters and non-ASCII bytes at once), one bit per
+/// (int8_t)(c ^ 2) < 33 finds the quote, the control characters and the
-/// byte (x86) or one nibble per byte (NEON)
+/// non-ASCII bytes at once, see vector_plain_run()), one bit per byte (x86) or
+/// one nibble per byte (NEON)
 NLOHMANN_VIEW_ALWAYS_INLINE std::uint64_t ascii_stop_mask(const unsigned char* p) noexcept
 {
 #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('\\'))),
+    const uint8x16_t special = vorrq_u8(vorrq_u8(vceqq_u8(in, vdupq_n_u8('\\')), vceqq_u8(in, vdupq_n_u8(0x7F))),
-                                        vorrq_u8(vceqq_u8(in, vdupq_n_u8(0x7F)), vcltq_s8(vreinterpretq_s8_u8(in), vdupq_n_s8(0x20))));
+                                        vcltq_s8(vreinterpretq_s8_u8(veorq_u8(in, vdupq_n_u8(2))), vdupq_n_s8(33)));
     return vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(special), 4)), 0);
 #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('\\'))),
+    const __m128i special = _mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8('\\')), _mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F))),
-                                         _mm_or_si128(_mm_cmpeq_epi8(in, _mm_set1_epi8(0x7F)), _mm_cmplt_epi8(in, _mm_set1_epi8(0x20))));
+                                         _mm_cmplt_epi8(_mm_xor_si128(in, _mm_set1_epi8(2)), _mm_set1_epi8(33)));
     return static_cast<std::uint64_t>(static_cast<unsigned>(_mm_movemask_epi8(special)));
 #endif
 }