Performance

include/glaze/simd/sse.hpp

@@ -15,33 +15,38 @@ namespace glz::detail
    GLZ_ALWAYS_INLINE void sse2_string_escape(const char*& c, const char* e, Data*& data, size_t n,
                                              WriteEscape&& write_escape)
    {
-      // SSE2: 16 bytes at a time with direct comparison instructions
-      if (n > 15) {
-         const __m128i quote_vec = _mm_set1_epi8('"');
-         const __m128i bs_vec = _mm_set1_epi8('\\');
-         const __m128i ctrl_mask = _mm_set1_epi8(static_cast<int8_t>(0xE0));
-         const __m128i zero = _mm_setzero_si128();
-
-         for (const char* end_m15 = e - 15; c < end_m15;) {
-            __m128i v = _mm_loadu_si128(reinterpret_cast<const __m128i*>(c));
-            _mm_storeu_si128(reinterpret_cast<__m128i*>(data), v);
-
-            const uint32_t mask = static_cast<uint32_t>(
-               _mm_movemask_epi8(_mm_or_si128(_mm_or_si128(_mm_cmpeq_epi8(v, quote_vec), _mm_cmpeq_epi8(v, bs_vec)),
-                                              _mm_cmpeq_epi8(_mm_and_si128(v, ctrl_mask), zero))));
-
-            if (mask == 0) {
-               data += 16;
-               c += 16;
-               continue;
-            }
-
-            const uint32_t length = countr_zero(mask);
-            c += length;
-            data += length;
+      if (n < 16) return; // let SWAR → scalar handle it
+
+      const __m128i quote_vec = _mm_set1_epi8('"');
+      const __m128i bs_vec   = _mm_set1_epi8('\\');
+      const __m128i ctrl_mask = _mm_set1_epi8(static_cast<int8_t>(0xE0));
+      const __m128i zero     = _mm_setzero_si128();
+
+      while (c <= e - 16) {
+         const __m128i v = _mm_loadu_si128(reinterpret_cast<const __m128i*>(c));
+
+         // Speculative store
+         _mm_storeu_si128(reinterpret_cast<__m128i*>(data), v);
+
+         const __m128i is_ctrl  = _mm_cmpeq_epi8(_mm_and_si128(v, ctrl_mask), zero);
+         const __m128i is_quote = _mm_cmpeq_epi8(v, quote_vec);
+         const __m128i is_bs    = _mm_cmpeq_epi8(v, bs_vec);
+         const __m128i m = _mm_or_si128(_mm_or_si128(is_quote, is_bs), is_ctrl);
+
+         const uint32_t mask = static_cast<uint32_t>(_mm_movemask_epi8(m));
+
+         if (mask == 0) {
+            data += 16;
+            c += 16;
+         }
+         else {
+            const uint32_t len = countr_zero(mask);
+            data += len;
+            c += len;
             write_escape();
          }
       }
+      // Remaining < 16 bytes fall through to SWAR → scalar
    }
 }