...and on the Haswell PEXT runs in one uop with a latency of 3! That is nothing short of amazing for an operation which, from its description, would seem to require a microcode loop or at least a few more cycles to collect an arbitrary number of bits with arbitrary gaps between them:

http://www.felixcloutier.com/x86/PEXT.html

The only unfortunate thing is that it was introduced quite recently in terms of x86 history (instead of being present early on but just microcoded), so earlier software can't take advantage of it nor any subsequent improvements, and uses a pretty complex encoding (VEX) only usable in protected mode. But if anything, this is another datapoint in the argument in favour of CISC --- try doing this on a MIPS, RISC-V or even ARM! With less powerful ISAs, even something as simple as the "shl ax, 4" (shift the lower 16 bits only) turns into a multi-instruction sequence of masking and combining.