• Concertina IV Again?

    From quadibloc@quadibloc@invalid.com (John Savard) to comp.arch on Thu Nov 12 21:54:51 2026
    From Newsgroup: comp.arch

    Concertina II, in its present form, finally had me satisfied with it.
    But it was based on block structure, which many people here, whose
    opinions I respect, say is a bad thing.
    Since to finally be satisfied even _with_ block structure, I had
    needed to improve my skills at squeezing things into the instruction
    set, I was willing to hope that I could manage without block
    structure.
    I arrived at Concertina III. Since I had only 1/4 of the opcode space
    left, after the 32-bit instructions, I felt that 16-bit short
    instructions, all of which must begin with 11, would be useless, so I
    would have to go with paired short instructions instead. Those I had
    used before with success.
    However, I wanted to avoid the version of the register specification
    where the destination register could be any of the 32 registers, but
    the source register could only be one of the eight registers in the
    same group of eight as the destination registers.
    This had made sense for me for code that assisted efficient pipelining
    by interweaving four independent threads. But it had been strongly
    criticized.
    Having 15 bits free per instruction instead of l7 bits - and needing
    opcode space for instructions longer than 32 bits - I had to really
    squeeze the short instructions.

    The technique I chose could be taken further, though. Maybe it was
    possible to get useful short instructions even with only 14 bits
    available.
    In this case, only the start of the instruction would have a length
    prefix, it would be a conventional variable-length instruction set
    like that of the System/360 or the 680x0 and so on.
    --- Synchronet 3.22a-Linux NewsLink 1.2
  • From MitchAlsup@user5857@newsgrouper.org.invalid to comp.arch on Fri Nov 13 01:22:50 2026
    From Newsgroup: comp.arch

    quadibloc@invalid.com (John Savard) posted:

    Concertina II, in its present form, finally had me satisfied with it.
    But it was based on block structure, which many people here, whose
    opinions I respect, say is a bad thing.

    We argue that it is "at best" different and urge caution.
    We further argue that similar ISA from the past had "problems".

    Since to finally be satisfied even _with_ block structure, I had
    needed to improve my skills at squeezing things into the instruction
    set, I was willing to hope that I could manage without block
    structure.

    I arrived at Concertina III. Since I had only 1/4 of the opcode space
    left, after the 32-bit instructions, I felt that 16-bit short
    instructions, all of which must begin with 11, would be useless, so I
    would have to go with paired short instructions instead. Those I had
    used before with success.
    However, I wanted to avoid the version of the register specification
    where the destination register could be any of the 32 registers, but
    the source register could only be one of the eight registers in the
    same group of eight as the destination registers.
    This had made sense for me for code that assisted efficient pipelining
    by interweaving four independent threads. But it had been strongly criticized.

    Compilers hat that sort of thing.

    Having 15 bits free per instruction instead of l7 bits - and needing
    opcode space for instructions longer than 32 bits - I had to really
    squeeze the short instructions.

    s/squeeze/eliminate/

    The technique I chose could be taken further, though. Maybe it was
    possible to get useful short instructions even with only 14 bits
    available.
    In this case, only the start of the instruction would have a length
    prefix, it would be a conventional variable-length instruction set
    like that of the System/360 or the 680x0 and so on.
    --- Synchronet 3.22a-Linux NewsLink 1.2
  • From quadibloc@quadibloc@invalid.com (John Savard) to comp.arch on Fri Nov 13 04:45:35 2026
    From Newsgroup: comp.arch

    I stopped partway through in the initial post of this thread; I am
    finishing what I intended to write now.

    What I thought of for allowing 16-bit instructions was this:

    110
    (op code) 3 bits
    (destination register) 5 bits
    (source register) 5 bits

    111
    (op code) 7 bits
    (destination register) 3 bits
    (source register) 3 bits

    The first format would serve only the double-precision floating-point
    type. Thus, it shaves off one bit, compared to the Concertina III,
    which also allowed the Medium floating-point type to use all the
    registers.

    The second format serves all types. In the case of the repeated
    occurrence of the double-precision floating-point type, the source
    register would be one of the first eight registers in the 128-register
    extended register bank.

    11111 would not be used, and would begin all instructions longer than
    32 bits.

    For Concertina III, I thought of also including the more useful 17-bit
    short instructions by embedding four of them in an 80-bit instruction.
    Branches to the four bytes of the first 32 bits of the instruction
    would be the way to allow branches to any of the four embedded
    instructions, because in this way, the first 32 bits would be fetched
    and decoded.

    This would not work for Concertina IV. The basic unit at which the
    decoding of an instruction may begin is only 16 bits. But a 48-bit
    instruction, in which two 17-bit short instructions are placed, would
    still use only 24 bits for each instruction, yielding a savings of
    space.

    Branching to the second byte of the first two bytes of an instruction
    of this type would indicate that execution is to continue with the
    second embedded instruction.

    This, now, is my vision of an ISA which offers what RISC
    architectures, like the PowerPC, offered while at the same time
    offering much of what older architectures, like 680x0 and System/360
    offered.

    John Savard
    --- Synchronet 3.22a-Linux NewsLink 1.2