BPF or berkely packet filter was written to be a faster replacement of tcpdump. People saw that it was pretty neat and started using it for non-tcp dump like stuff and it became extended BPF/eBPF. I would guess that running eBPF on Windows would be a lot slower, but it would be interesting to see a performance comparison.
Point of order: BPF wasn't written simply for tcpdump; it's part of a line of research on using PL runtimes to configure and operate networking stacks; so, right after McCanne's BPF paper, you get MPF, which is proposed to do all of demux for Mach.
