I've realised straight tee() is actually wrong - it works fine for piping to something but "zcyes > /dev/null" won't work, it needs vmsplice() like mrb said.

  core2duo e8400, 4.8 kernel, coreutils 8.25
  > yes | dd bs=1M count=10000 of=/dev/null iflag=fullblock
  10485760000 bytes (10 GB, 9.8 GiB) copied, 4.1901 s, 2.5 GB/s
  > ./zcyes | dd bs=1M count=10000 of=/dev/null iflag=fullblock
  10485760000 bytes (10 GB, 9.8 GiB) copied, 1.8542 s, 5.7 GB/s