5 ms·

Hi, author here. My superpower is spending unreasonable amounts of time researching things with no practical purpose. Occasionally I blog about it - as a warnin

Hi, author here. My superpower is spending unreasonable amounts of time researching things with no practical purpose. Occasionally I blog about it - as a warning to others.

Ethee 1y ago
It's these kinds of posts that I appreciate reading the most, so thank you for sharing!
owls-on-wires 1y ago
“…no practical purpose” Nonsense, I learned something about compilation today. Thank you for sharing.
trws 1y ago
I liked the article. I saw your PS that we added it to the working draft for c++26, we also made it part of OpenMP as of 5.0 I think. It’s sometimes a hardware atomic like on arm, but what made the case was that it’s common to implement it sub-optimally even on x86 or LL-SC architectures. Often the generic cas loop gets used, like in your lambda example, but it lacks an early cutout since you can ignore any input value that’s on the wrong side of the op by doing a cheap atomic read or just cutting out of the loop after the first failed CAS if the read back shows it can’t matter. Also can benefit from using slightly different memory orders than the default on architectures like ppc64. It’s a surprisingly useful op to support that way. If this kind of thing floats your boat, you might be interested in the non-reading variants of these as well. Mostly for things like add, max, etc but some recent architectures actually offer alternate operations to skip the read-back. The paper calls them “atomic reduction operations” https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p3111r6.html https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p31...
- anematode 1y ago
  Curious: even with hardware atomics, wouldn't it be a good idea to first perform a non-atomic load to check for whether the store might be necessary (which would require the cache line to be locked), then only run the atomic max if it might change the value?
  - adgjlsfhk1 1y ago
    This depends heavily on what concurrency optimizations your processor implements (and unfortunately this is the sort of thing that doesn't get doccumented and is somewhat hard to test).
    - anematode 1y ago
      I did a little unscientific test here on an Apple M4 Pro with n threads spamming atomic operations with pseudorandom values on one memory location (the worst case). Used inline asm to make sure there was no funny business going on. atomic adds n = 1 -> 333e6 adds/second n = 2 -> 174e6 n = 4 -> 95e6 n = 8 -> 63e6 atomic maxs n = 1 -> 161e6 maxs/second n = 2 -> 59e6 n = 4 -> 39e6 n = 8 -> 27e6 atomic maxs with preceding check n = 1 -> 929e6 maxs/second n = 2 -> 1541e6 n = 4 -> 3494e6 n = 8 -> 5985e6 So evidently the M4 doesn't do this optimization. Of course if your distribution is different you'd get different results, and this level of contention is unrealistic, but I don't see why you'd EVER not do a check before running atomic max. I also find it interesting that atomic max is significantly slower than atomic add
      - thequux 1y ago
        I think that this can change the semantics though; with the preceding check you can miss the shared variable being decremented from another thread. In some cases, such as if the shared value is monotonic, this is done, but not in the general case.
        anematode 1y ago
        With a relaxed ordering I'm not sure if that's right, since the ldumax would have no imposed ordering relation with the (atomic) decrement on another thread and so could very well have operated on the old value obtained by the non-atomic load
        ibraheemdev 1y ago
- SkiFire13 1y ago
  > but it lacks an early cutout since you can ignore any input value that’s on the wrong side of the op by doing a cheap atomic read or just cutting out of the loop after the first failed CAS if the read back shows it can’t matter. I believe this is a bit trickier than that, you would also need at least some kind of atomic barrier to preserve the ordering semantics of the successful update case.
ajayka 1y ago
Great article! Did you end up hiring the candidate?!
xarope 1y ago
looks around room, heads nodding. Ah, a magician. welcome.
michalsustr 1y ago
Thank you for sharing, loved the article!

Hi, author here. My superpower is spending unreasonable amounts of time researching things with no practical purpose. Occasionally I blog about it - as a warning to others.

Ethee 1y ago

It's these kinds of posts that I appreciate reading the most, so thank you for sharing!

owls-on-wires 1y ago

“…no practical purpose” Nonsense, I learned something about compilation today. Thank you for sharing.

trws 1y ago

I liked the article. I saw your PS that we added it to the working draft for c++26, we also made it part of OpenMP as of 5.0 I think. It’s sometimes a hardware atomic like on arm, but what made the case was that it’s common to implement it sub-optimally even on x86 or LL-SC architectures. Often the generic cas loop gets used, like in your lambda example, but it lacks an early cutout since you can ignore any input value that’s on the wrong side of the op by doing a cheap atomic read or just cutting out of the loop after the first failed CAS if the read back shows it can’t matter. Also can benefit from using slightly different memory orders than the default on architectures like ppc64. It’s a surprisingly useful op to support that way. If this kind of thing floats your boat, you might be interested in the non-reading variants of these as well. Mostly for things like add, max, etc but some recent architectures actually offer alternate operations to skip the read-back. The paper calls them “atomic reduction operations” https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p3111r6.html https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p31...

anematode 1y ago
Curious: even with hardware atomics, wouldn't it be a good idea to first perform a non-atomic load to check for whether the store might be necessary (which would require the cache line to be locked), then only run the atomic max if it might change the value?
- adgjlsfhk1 1y ago
  This depends heavily on what concurrency optimizations your processor implements (and unfortunately this is the sort of thing that doesn't get doccumented and is somewhat hard to test).
  - anematode 1y ago
    I did a little unscientific test here on an Apple M4 Pro with n threads spamming atomic operations with pseudorandom values on one memory location (the worst case). Used inline asm to make sure there was no funny business going on. atomic adds n = 1 -> 333e6 adds/second n = 2 -> 174e6 n = 4 -> 95e6 n = 8 -> 63e6 atomic maxs n = 1 -> 161e6 maxs/second n = 2 -> 59e6 n = 4 -> 39e6 n = 8 -> 27e6 atomic maxs with preceding check n = 1 -> 929e6 maxs/second n = 2 -> 1541e6 n = 4 -> 3494e6 n = 8 -> 5985e6 So evidently the M4 doesn't do this optimization. Of course if your distribution is different you'd get different results, and this level of contention is unrealistic, but I don't see why you'd EVER not do a check before running atomic max. I also find it interesting that atomic max is significantly slower than atomic add
    - thequux 1y ago
      I think that this can change the semantics though; with the preceding check you can miss the shared variable being decremented from another thread. In some cases, such as if the shared value is monotonic, this is done, but not in the general case.
      - anematode 1y ago
        With a relaxed ordering I'm not sure if that's right, since the ldumax would have no imposed ordering relation with the (atomic) decrement on another thread and so could very well have operated on the old value obtained by the non-atomic load
        ibraheemdev 1y ago
SkiFire13 1y ago
> but it lacks an early cutout since you can ignore any input value that’s on the wrong side of the op by doing a cheap atomic read or just cutting out of the loop after the first failed CAS if the read back shows it can’t matter. I believe this is a bit trickier than that, you would also need at least some kind of atomic barrier to preserve the ordering semantics of the successful update case.

ajayka 1y ago

Great article! Did you end up hiring the candidate?!

xarope 1y ago

looks around room, heads nodding. Ah, a magician. welcome.

michalsustr 1y ago

Thank you for sharing, loved the article!