80386 Barrel Shifter

(nand2mario.github.io)

79 points | by jamesbowman 3 days ago

2 comments

  • rep_lodsb 21 hours ago
    Implementing rotate through carry like that was a really bad decision IMO - it's almost never by more than one bit left or right at a time, and this could be done much more efficiently than with the constant-time code which is only faster when the count is > 6.

    Is the full microcode available anywhere?

    • ajenner 21 hours ago
      I haven't published it yet as there are still some rough edges to clear up, but if you email me (andrew@reenigne.org) I'll send you the current work-in-progress (the same one that nand2mario is working from).
    • kjs3 21 hours ago
      Since the shifter is also used for bit tests, the 'most things are a 1-bit shift' might not be the case. Perhaps they did the analysis and it made sense.
      • rep_lodsb 20 hours ago
        There are separate opcodes for shift/rotate by 1, by CL, or by an immediate operand. Those are decoded to separate microcode entry points, so they could have at least optimized the "RCL/RCR x,1" case.

        And the microcode for bit test has to be different anyway.

  • cmovq 18 hours ago
    > For memory operands, there's an additional twist: the bit index is a signed offset that can address bits outside the nominal operand. A bit index of 35 on a dword accesses bit 3 of the next dword in memory.

    I wonder what is the use case for testing a bit outside of the memory address given.

    • rep_lodsb 18 hours ago
      So you can have bit arrays of any length in memory, rather than just 32 bits in a register.
      • cmovq 18 hours ago
        That makes sense. LLVM could probably do better here by using the memory operand version:

        https://godbolt.org/z/jeqbaPsMz

        • ack_complete 9 hours ago
          Don't think the memory operand version would work here. If I understand the x86 architectural manual description, the 32-bit operand form interprets the bit offset as signed. A 64-bit operand could work around that but then run into issues with over-read due to fetching 64 bits of data.
        • jxors 14 hours ago
          The memory operand version tends to be as slow or slower than the manual implementation, so LLVM is right to avoid it.
    • juancn 18 hours ago
      It was probably easier to just implement it that way, given that the barrel shifter is 64 bits wide.