[lkml]   [1998]   [Nov]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [patch] entry.S asm improvement (removed some ugly jmp)

    The processor has an internal stack of return addresses from CALL
    instructions. This stack holds upto 4 addresses, and is used to predict
    the destination address of a RET so that instruction prefetch can happen
    earlier. If the prediction misses, then the penalty is about a dozen
    cycles on a Pentium Pro, Pentium II, Celeron, Xeon, or any Intel x86
    processor with "dynamic execution". The penalty on a Pentium is about
    five cycles. Thus, the penalty is about the same as a pipeline flush, and
    on a dynamic execution processor it is also about the same as a cache
    miss. The first mis-predicted RET could cause all RETs then in the
    internal stack to be mis-predicted (the easy implementation). So, be
    careful in high-frequency areas. But if there will be a context switch,
    or if the depth is now only 1, or will exceed 4 before the RET, then
    definitely do use the PUSHL+JMP instead of CALL+JMP.

    ----------------------------------------------------------- (John Reiser)

    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to
    Please read the FAQ at

     \ /
      Last update: 2005-03-22 13:45    [W:0.017 / U:0.076 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site