92   const __m128i v = *(const __m128i *)(const void *)(items + (i) * 8); \
93                     *(      __m128i *)(      void *)(items + (i) * 8) = \
105 SwapBytes2_128(CSwapUInt16 *items, const CSwapUInt16 *lim)  in SwapBytes2_128()  argument
110     SWAP2_128(0)  SWAP2_128(1)  items += 2 * 8;  in SwapBytes2_128()
111     SWAP2_128(0)  SWAP2_128(1)  items += 2 * 8;  in SwapBytes2_128()
113   while (items != lim);  in SwapBytes2_128()
119   __m128i v = *(const __m128i *)(const void *)(items + (i) * 4); \
126   *(__m128i *)(void *)(items + (i) * 4) = _mm_packus_epi16(v0, v1); }
134 SwapBytes4_128_pack(CSwapUInt32 *items, const CSwapUInt32 *lim)
141     SWAP4_128_pack(0); items += 1 * 4;
142     // SWAP4_128_pack(0); SWAP4_128_pack(1); items += 2 * 4;
144   while (items != lim);
149   __m128i v = *(const __m128i *)(const void *)(items + (i) * 4); \
157   *(__m128i *)(void *)(items + (i) * 4) = _mm_or_si128(v2, v); }
165 SwapBytes4_128_shift(CSwapUInt32 *items, const CSwapUInt32 *lim)
172     // SWAP4_128_shift(0)  SWAP4_128_shift(1)  items += 2 * 4;
173     // SWAP4_128_shift(0)  SWAP4_128_shift(1)  items += 2 * 4;
174     SWAP4_128_shift(0); items += 1 * 4;
176   while (items != lim);
236 #define SHUF_128(i)   *(items + (i)) = \
237      _mm_shuffle_epi8(*(items + (i)), mask); // SSSE3
249   __m128i *items = (__m128i *)items8;  in ShufBytes_128()  local
260     SHUF_128(0)  SHUF_128(1)  items += 2;  in ShufBytes_128()
261     SHUF_128(0)  SHUF_128(1)  items += 2;  in ShufBytes_128()
263   while (items != lim);  in ShufBytes_128()
278 #define SHUF_256(i)   *(items + (i)) = \
279   _mm256_shuffle_epi8(*(items + (i)), mask); // AVX2
291   __m256i *items = (__m256i *)items8;  in ShufBytes_256()  local
321     SHUF_256(0)  SHUF_256(1)  items += 2;  in ShufBytes_256()
322     SHUF_256(0)  SHUF_256(1)  items += 2;  in ShufBytes_256()
324   while (items != lim);  in ShufBytes_256()
399 #define SWAP2_128(i)   *(uint8x16_t *)      (void *)(items + (i) * 8) = \
400       vrev16q_u8(*(const uint8x16_t *)(const void *)(items + (i) * 8));
401 #define SWAP4_128(i)   *(uint8x16_t *)      (void *)(items + (i) * 4) = \
402       vrev32q_u8(*(const uint8x16_t *)(const void *)(items + (i) * 4));
412 SwapBytes2_128(CSwapUInt16 *items, const CSwapUInt16 *lim)  in SwapBytes2_128()  argument
417     SWAP2_128(0)  SWAP2_128(1)  items += 2 * 8;  in SwapBytes2_128()
418     SWAP2_128(0)  SWAP2_128(1)  items += 2 * 8;  in SwapBytes2_128()
420   while (items != lim);  in SwapBytes2_128()
431 SwapBytes4_128(CSwapUInt32 *items, const CSwapUInt32 *lim)  in SwapBytes4_128()  argument
436     SWAP4_128(0)  SWAP4_128(1)  items += 2 * 4;  in SwapBytes4_128()
437     SWAP4_128(0)  SWAP4_128(1)  items += 2 * 4;  in SwapBytes4_128()
439   while (items != lim);  in SwapBytes4_128()
457 …#define SWAP2_16(i)  { UInt32 v = items[i];  v += (v << 16);  v >>= 8;  items[i] = (CSwapUInt16)v;…
460     #define SWAP2_16(i)  { items[i] = _rotr16(items[i], 8); }
463   #define SWAP2_16(i)  { CSwapUInt16 v = items[i];  items[i] = Z7_BSWAP16(v); }
467   #define SWAP4_32(i)  { CSwapUInt32 v = items[i];  items[i] = Z7_BSWAP32(v); }
470     { UInt32 v = items[i]; \
473       items[i] = v; }
540     UInt64 v = *(const UInt64 *)(const void *)(items + (i) * 4); \
542     *(UInt64 *)(void *)(items + (i) * 4) = v; }
545 SwapBytes2_64(CSwapUInt16 *items, const CSwapUInt16 *lim)  in SwapBytes2_64()  argument
550     SWAP2_64(0)  SWAP2_64(1)  items += 2 * 4;  in SwapBytes2_64()
551     SWAP2_64(0)  SWAP2_64(1)  items += 2 * 4;  in SwapBytes2_64()
553   while (items != lim);  in SwapBytes2_64()
569     UInt64 v = *(const UInt64 *)(const void *)(items + (i) * 2); \
571     *(UInt64 *)(void *)(items + (i) * 2) = v; }
574 SwapBytes4_64(CSwapUInt32 *items, const CSwapUInt32 *lim)  in SwapBytes4_64()  argument
579     SWAP4_64(0)  SWAP4_64(1)  items += 2 * 2;  in SwapBytes4_64()
580     SWAP4_64(0)  SWAP4_64(1)  items += 2 * 2;  in SwapBytes4_64()
582   while (items != lim);  in SwapBytes4_64()
617     UInt32 v = *(const UInt32 *)(const void *)(items + (i) * 2); \
619     *(UInt32 *)(void *)(items + (i) * 2) = v; }
623 SwapBytes2_32(CSwapUInt16 *items, const CSwapUInt16 *lim)  in SwapBytes2_32()  argument
628     SWAP2_32(0)  SWAP2_32(1)  items += 2 * 2;  in SwapBytes2_32()
629     SWAP2_32(0)  SWAP2_32(1)  items += 2 * 2;  in SwapBytes2_32()
631   while (items != lim);  in SwapBytes2_32()
636 SwapBytes4_32(CSwapUInt32 *items, const CSwapUInt32 *lim)  in SwapBytes4_32()  argument
641     SWAP4_32(0)  SWAP4_32(1)  items += 2;  in SwapBytes4_32()
642     SWAP4_32(0)  SWAP4_32(1)  items += 2;  in SwapBytes4_32()
644   while (items != lim);  in SwapBytes4_32()
669 void z7_SwapBytes2(CSwapUInt16 *items, size_t numItems)  in z7_SwapBytes2()  argument
672 …for (; numItems != 0 && ((unsigned)(ptrdiff_t)items & (SWAP_VECTOR_ALIGN_SIZE - 1)) != 0; numItems…  in z7_SwapBytes2()
675     items++;  in z7_SwapBytes2()
683     lim = items + numItems2;  in z7_SwapBytes2()
690             ShufBytes_256((__m256i *)(void *)items,  in z7_SwapBytes2()
697             ShufBytes_128((__m128i *)(void *)items,  in z7_SwapBytes2()
705             SwapBytes2_128(items, lim);  in z7_SwapBytes2()
709             DEFAULT_Swap2(items, lim);  in z7_SwapBytes2()
711     items = lim;  in z7_SwapBytes2()
717     items++;  in z7_SwapBytes2()
723 void z7_SwapBytes4(CSwapUInt32 *items, size_t numItems)  in z7_SwapBytes4()  argument
726 …for (; numItems != 0 && ((unsigned)(ptrdiff_t)items & (SWAP_VECTOR_ALIGN_SIZE - 1)) != 0; numItems…  in z7_SwapBytes4()
729     items++;  in z7_SwapBytes4()
737     lim = items + numItems2;  in z7_SwapBytes4()
744             ShufBytes_256((__m256i *)(void *)items,  in z7_SwapBytes4()
751             ShufBytes_128((__m128i *)(void *)items,  in z7_SwapBytes4()
759             SwapBytes4_128(items, lim);  in z7_SwapBytes4()
763             DEFAULT_Swap4(items, lim);  in z7_SwapBytes4()
765     items = lim;  in z7_SwapBytes4()
771     items++;  in z7_SwapBytes4()