A popcnt reduction loop before and after adding `xor eax, eax`

On an Intel Haswell or Skylake core, popcnt has 3-cycle latency and a throughput of one per cycle on port 1. This loop sums population counts over an array resident in L1D, and it measures about 3 cycles per iteration, roughly twice what its uop count on a 4-wide front end predicts. Inserting xor eax, eax immediately before the popcnt brings it well under 2 cycles per iteration, despite adding an instruction. What explains both observations?

.loop:
    popcnt rax, qword ptr [rdi + rcx*8]
    add    rdx, rax
    inc    rcx
    cmp    rcx, rsi
    jb     .loop

45 of 47 in Computer Architecture