The JCXZ instruction takes longer to run than a two-instruction sequence, which compares the count register to zero and jumps if the count is zero.
All branches are converted into 16-byte code fetches regardless of jump address or cacheability.