• Downsizing Qupls5 instruction set

    From Robert Finch@robfi680@gmail.com to comp.arch on Sat Nov 14 21:39:50 2026
    From Newsgroup: comp.arch

    The insanity continues.

    Downsizing the size of instructions for Qupls5 from 48-bit to 32-bit.
    This was made possible by finding a way to use prefixes (and suffixes)
    with vector instructions. To perform a vector operation a prefix is
    needed now which effectively makes the instruction 64-bits instead of
    48. However, most instruction should be 32-bit now.

    Previously I thought vector instruction could not be done without
    feeding multiple micro-ops into the pipeline. The issue was it could not
    be done at the reservation station time because the reservation station
    only has access to physical register specs. The vector chunks needed to
    be renamed meaning that had to be done at the rename stage. Just
    incrementing the register spec at reservation would not work.

    However, I noticed that for vector registers only the first chunk of the vector needs to be renamed. Due to the way the renamer works, physical registers following the renamed one for a vector register are not
    renamed up to the length of the vector. So, it should be safe to just increment the physical register number. As long as the physical register
    is not marked valid until the vector instruction is finished, I think it should work.

    Previously things were set at 48-bits in part to support vector
    instructions which needed additional instruction bits. But was very
    wasteful for other instructions. Vector instructions were detected by
    the micro-op translator which fed a bunch of duplicate instructions with
    just the register spec changing into the instruction stream.

    The alternate means of handling vector instruction now is to hold the instruction in the reservation station and increment the register spec
    while decrementing the length every time the instruction is issued to
    execute. That way there is only one instruction being processed.
    However, it may be not quite as fast as it takes time to read the
    register file from the reservation station. But other instructions may
    now fly past the vector instruction; the pipeline is not polluted with a
    bunch of vector micro-ops.
    --- Synchronet 3.22a-Linux NewsLink 1.2