3 ms·

A couple responses: 1. If the values the source registers aren't computed yet (due to prediction, renaming, etc.), then CMOV will need to stall until they're c

• gpderetta 4y ago
  1. When quoting the latency of an instruction, it is customary to quote the latency that the instruction itself adds to a dependency chain. CMOV itself had indeed higher latency on P4 (Although Agner only claims 6 cycles, not 10 like Linus). Still, the advantage of conditional branches is that they do indeed break dependency chains and allow OoO execution, whether that's a win or not depend on the prediction rate of the specific branch and the length of the chain. 2. Well I only checked Agner for Ice (which BTW came out in 2019). But CMOV has had a 1 clock latency since Broadwell (2014). It had been 2 clock cycles for a few generation before. The number of generated uops has also significantly decreased.
  • jabl 4y ago
    Wrt #2, there is the additional argument that when selecting compiler optimizations for a -mtune=generic (or equivalent) target, it makes sense to be a bit "forward leaning", as it'll be a while before your work ends up in a released version, and even further time goes by until that compiler release ends up in the hands of users.