Shifting a 64-bit value left by 37 bits sounds like it should take 37 cycles — one shift per position. Early CPUs actually worked that way, using a serial shifter that looped through iterations. Modern CPUs do it in one cycle using a barrel shifter: a fixed piece of combinational logic that routes every input bit to its correct output position in a single pass.

The trick is a logarithmic multiplexer tree. For a 64-bit shifter, you need only 6 stages (log₂ 64). Stage i conditionally shifts by 2ⁱ positions based on bit i of the shift amount. Want to shift by 37? Binary is 100101, so stages 0, 2, and 5 activate — shifting by 1, then 4, then 32. Total: 37, in one clock cycle. Each stage is just a row of 64 two-input muxes.

Real-world example: On modern x86, SHL rax, cl executes in a single cycle on Intel Skylake's port 0 or 6. Compare that to bit-parallel emulation: multiplying by 2ⁿ takes 3 cycles on the FMA unit. The barrel shifter isn't just faster — it's why constant-time cryptographic code can rely on shifts without timing leaks. AES key expansion, ChaCha20 rotations, and SHA-256 all lean heavily on single-cycle rotates that reuse the barrel shifter's wiring in a rotational configuration.

ARM took this further: nearly every ARM data-processing instruction has a free barrel shift on the second operand. ADD r0, r1, r2, LSL #3 shifts r2 left by 3 and adds to r1 in one cycle — the shifter sits in series with the ALU's input mux. This is why ARM code often has fewer instructions than x86 for address arithmetic.

The cost: Wiring. A 64-bit barrel shifter has 64 × 6 = 384 muxes and a wire fan-out that grows with each stage. In a 32-bit CPU it's cheap; in 64-bit, it's a noticeable chunk of the ALU's area and power. RISC-V's base ISA only has fixed shifts for this reason — the "B" extension adds funnel shifts and rotates when the silicon budget allows.

Rule of thumb: Any shift, rotate, or funnel shift by a variable amount on a modern desktop CPU is one cycle, one µop. If your algorithm can be expressed as shifts + masks + adds instead of multiplies, it'll beat the multiplier (3-5 cycles) every time — this is why bit-manipulation tricks in libraries like Hacker's Delight still matter.