Use barrent reduction

Use barrent reduction for CRC calculation.
Currently I use dogfood (sse2neon) for code prototype.

sse2neon.h

@@ -8498,6 +8498,28 @@ FORCE_INLINE uint32_t _mm_crc32_u8(uint32_t crc, uint8_t v)
 #elif ((__ARM_ARCH == 8) && defined(__ARM_FEATURE_CRC32)) || \
     (defined(_M_ARM64) && !defined(__clang__))
     crc = __crc32cb(crc, v);
+#elif defined(__ARM_FEATURE_PMULL)
+    // Barrent reduction
+    __m128i orig = _mm_set_epi64x(0, (uint64_t) (crc ^ v) << (32 - 8));
+    __m128i tmp = orig;
+
+    uint64_t p = 0x105EC76F1;
+    uint64_t mu =
+        0x1dea713f1;  // Barrett Reduction constant (u64`) = u` = (x**64 /
+                      // P(x))` = 0x11f91caf6 0b 1 0001 1111 1001 0001 1100 1010
+                      // 1111 0110 0x     1    f    3    1    7    a    e    d0
+
+    // Multiply by mu_{64}
+    tmp = _mm_clmulepi64_si128(tmp, _mm_set_epi64x(0, mu), 0x00);
+    // Divide by 2^64 (mask away the unnecessary bits)
+    tmp = _mm_and_si128(tmp, _mm_set_epi64x(0, 0xFFFFFFFF));
+    // Multiply by p (shifted left by 1 for alignment reasons)
+    tmp = _mm_clmulepi64_si128(tmp, _mm_set_epi64x(0, p), 0x00);
+    // Subtract original from result
+    tmp = _mm_xor_si128(tmp, orig);
+
+    // Extract the 'lower' (in bit-reflected sense) 32 bits
+    crc = (uint32_t) _mm_extract_epi32(tmp, 1);
 #else
     crc ^= v;