272   const __m128i m0 = LOADU( block + 00 );  in blake2b_compress()
273   const __m128i m1 = LOADU( block + 16 );  in blake2b_compress()
274   const __m128i m2 = LOADU( block + 32 );  in blake2b_compress()
275   const __m128i m3 = LOADU( block + 48 );  in blake2b_compress()
276   const __m128i m4 = LOADU( block + 64 );  in blake2b_compress()
277   const __m128i m5 = LOADU( block + 80 );  in blake2b_compress()
278   const __m128i m6 = LOADU( block + 96 );  in blake2b_compress()
279   const __m128i m7 = LOADU( block + 112 );  in blake2b_compress()
298   row1l = LOADU( &S->h[0] );  in blake2b_compress()
299   row1h = LOADU( &S->h[2] );  in blake2b_compress()
300   row2l = LOADU( &S->h[4] );  in blake2b_compress()
301   row2h = LOADU( &S->h[6] );  in blake2b_compress()
302   row3l = LOADU( &blake2b_IV[0] );  in blake2b_compress()
303   row3h = LOADU( &blake2b_IV[2] );  in blake2b_compress()
304   row4l = _mm_xor_si128( LOADU( &blake2b_IV[4] ), LOADU( &S->t[0] ) );  in blake2b_compress()
305   row4h = _mm_xor_si128( LOADU( &blake2b_IV[6] ), LOADU( &S->f[0] ) );  in blake2b_compress()
320   STOREU( &S->h[0], _mm_xor_si128( LOADU( &S->h[0] ), row1l ) );  in blake2b_compress()
321   STOREU( &S->h[2], _mm_xor_si128( LOADU( &S->h[2] ), row1h ) );  in blake2b_compress()
324   STOREU( &S->h[4], _mm_xor_si128( LOADU( &S->h[4] ), row2l ) );  in blake2b_compress()
325   STOREU( &S->h[6], _mm_xor_si128( LOADU( &S->h[6] ), row2h ) );  in blake2b_compress()