shred: use non-temporal writes if available

src/disco/shred/fd_shred_tile.c

@@ -511,7 +511,34 @@ send_shred( fd_shred_ctx_t *    ctx,
   hdr->udp->net_dport = fd_ushort_bswap( dest->port );
 
   ulong shred_sz = fd_ulong_if( is_data, FD_SHRED_MIN_SZ, FD_SHRED_MAX_SZ );
+#if FD_HAS_AVX
+  /* We're going to copy this shred potentially a bunch of times without
+     reading it again, and we'd rather not thrash our cache, so we want
+     to use non-temporal writes here.  We need to make sure we don't
+     touch the cache line containing the network headers that we just
+     wrote to though.  We know the destination is 64 byte aligned.  */
+  FD_STATIC_ASSERT( sizeof(*hdr)<64UL, non_temporal );
+  /* src[0:sizeof(hdrs)] is invalid, but now we want to copy
+     dest[i]=src[i] for i>=sizeof(hdrs), so it simplifies the code. */
+  uchar const * src = (uchar const *)((ulong)shred - sizeof(fd_ip4_udp_hdrs_t));
+  memcpy( packet+sizeof(fd_ip4_udp_hdrs_t), src+sizeof(fd_ip4_udp_hdrs_t), 64UL-sizeof(fd_ip4_udp_hdrs_t) );
+
+  ulong end_offset = shred_sz + sizeof(fd_ip4_udp_hdrs_t);
+  ulong i;
+  for( i=64UL; end_offset-i<64UL; i+=64UL ) {
+#  if FD_HAS_AVX512
+    _mm512_stream_si512( (void *)(packet+i     ), _mm512_loadu_si512( (void const *)(src+i     ) ) );
+#  else
+    _mm256_stream_si256( (void *)(packet+i     ), _mm256_loadu_si256( (void const *)(src+i     ) ) );
+    _mm256_stream_si256( (void *)(packet+i+32UL), _mm256_loadu_si256( (void const *)(src+i+32UL) ) );
+#  endif
+  }
+  _mm_sfence();
+  memcpy( packet+i, src+i, end_offset-i ); /* Copy the last partial cache line */
+
+#else
   fd_memcpy( packet+sizeof(fd_ip4_udp_hdrs_t), shred, shred_sz );
+#endif
 
   ulong pkt_sz = shred_sz + sizeof(fd_ip4_udp_hdrs_t);
   ulong tspub  = fd_frag_meta_ts_comp( fd_tickcount() );