Login
You're viewing the discuss.systems public feed.

Replies

  • I Can't Believe It's Not Zero!steve
    Oct 13, 2026, 12:47 AM

    One weird trick to go faster by breaking the loop-carried dependency chain. Validation suites hate it!

    💬 1🔄 0⭐ 7
  • Fabian Giesenrygorous@mastodon.gamedev.place
    Oct 13, 2026, 1:24 AM

    @steve the best part about 64-bit carry shenanigans is that if you're not specifically testing for it, you can have bugs for an outrageous amount of time until somebody notices

    (more of an issue with x86-64 style carry/borrow, with ARM-style true carry you will probably notice on the first subtraction)

    💬 1🔄 0⭐ 1
  • 💬 1🔄 0⭐ 1
  • I Can't Believe It's Not Zero!steve
    Oct 13, 2026, 1:29 AM

    @rygorous it's been long enough since I've thought about bignums on x86 that I'm actually not sure what the preferred idiom is these days, but it sure as heck isn't whatever this is.

    💬 2🔄 0⭐ 1
  • 💬 1🔄 0⭐ 1
  • 💬 1🔄 0⭐ 0
  • I Can't Believe It's Not Zero!steve
    Oct 13, 2026, 1:40 AM

    @rygorous If you only care about Apple Silicon, this will be not-terrible, since the cores are so absurdly wide now. If you just roll the increments into the LDR/STR, me pulling numbers out of my ass (which is generally more accurate than any public simulator) says this will asymptotically get you within about 30% of what you can get by unrolling, with lower fixed cost, so it's actually probably a good choice for most workloads.

    If you care about narrower cores, at all, yeah it should be 2x unrolled so you can benefit from LDP/STP.

    💬 0🔄 0⭐ 2
  • Fabian Giesenrygorous@mastodon.gamedev.place
    Oct 13, 2026, 1:37 AM

    @steve something something dual carry chains with ADCX/ADOX.

    Not relevant for actual adds but does matter for muls.

    No idea how you handle loops when you have to.

    Presumably unroll and spill/restore carry/overflow into a designated landing pad reg around the test

    💬 1🔄 0⭐ 1
  • Fabian Giesenrygorous@mastodon.gamedev.place
    Oct 13, 2026, 1:44 AM

    @steve idly wondering whether you can do something like

    mov r8d, 0
    adcx r8d, r8d
    adox r8d, r8d
    shl r8d, 30

    to save and then

    add r8d, r8d

    to restore. That will give you carry back, but I always forget what exactly overflow was. Carry-in into MSB != carry-out, right?

    in this case carry-out from MSB is saved_CF and carry-in is saved_OF, so it would set OF to saved_OF^saved_CF.

    Almost. Pity. No actual use for it but it felt cute. :)

    💬 1🔄 0⭐ 1
  • Fabian Giesenrygorous@mastodon.gamedev.place
    Oct 13, 2026, 1:57 AM

    @steve anyway, with low-ish trip count, probably something like

    sub rdx, rsi
    sub rdi, rsi
    clc

    lp:
    mov rax, [rsi]
    adc rax, [rdx + rsi]
    mov [rdi + rsi], rax
    lea rsi, [rsi + 8]
    dec rcx
    jnz lp

    or maybe (probably not)

    lea rsi, [rsi + rcx*8]
    lea rdx, [rdx + rcx*8]
    lea rdi, [rdi + rcx*8]
    neg rcx
    jz skip

    clc
    lp:
    mov rax, [rsi + rcx*8]
    adc rax, [rdx + rcx*8]
    mov [rdi + rcx*8], rax
    inc rcx
    jnz lp

    💬 1🔄 0⭐ 0
  • I Can't Believe It's Not Zero!steve
    Oct 13, 2026, 1:58 AM

    @rygorous yeah, that feels about right. `neg rcx` was pretty standard when I cared about this, FWIW (but x86 is wider now than it was then).

    💬 0🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 1