32-bit version

[void2012]

How portable is your implementation for 32-bit system? It uses a hack for a 48-bit pointer, however, is it possible to implement it on 32-bit? Thanks!

[vtyulb]

I think it depends on hardware. If you don't have a lot of ram (256mb) and you don't use a lot of threads (8), then you can pack into 32-bit variable. That will work for any hardware fast enough I think.

Situation is more painful if you talking about something like windows xp on an old i686 pc, where full address space is used. My wild guess is that my implementation will work without any changes. Compiler should replace lock xadd qword ptr [rdi], rax to lock cmpxchg8b QWORD PTR [ebp]: https://godbolt.org/z/rf4jn6j9d
Compare exchange cycle is always slower than fetch_add, however.

What kind of software/harware are you talking about?

[void2012]

Thank you for the input!
I'm targeting Windows XP 32-bit with Pentium 4

[vtyulb]

Isn't Pentium 4 pure i386 with only 1 core? That will not work, because there is no atomic instruction available. At least i686 required for windows. You can use spinlock + std::shared_ptr, there will be no contention for obvious reasons.

Is it about homm3 modding, or just some educational project? I thought no one targets old cpus now. ArchLinux dropped even i686 around 10 years ago, i386 much earlier.

[void2012]

The pentium 4 instruction set is the minimum requirement for my games. The game should work fine in any modern multicore x64 CPU under 32-bit similation mode. I do not run the multithreaded code if there is only one core, i.e. I check the number of cores in the runtime and switch the appropriate implementation(single or multi threaded)

Pentium 4 supports atomic instructions, it isn't pure i386, it's i686 with SSE2

[void2012]

Any updates?

[vtyulb]

Sorry for the delayed response. I don't have plans to add specific support for Pentium 4, as it would be simpler to rely on spinlocks or mutexes in scenarios involving low core counts and older architectures. But you can submit pull request if you want to - I can help with some logic and review.

[void2012]

Thanks for the response! Pentium4 is just an example. The question is rather about the IA-32 architecture itself (as modern CPUs handle everything just fine for backwards compatibility reasons). I'll be just fine if something requires even AVX instructions(because if the underlying x64 CPU supports a certain set of instructions, the x86 mode must too), for example, the entire code should work on 32-bit x86.

[EyeAndTea]

I did not read the code, but judging by the conversation, such solutions should be simpler on 32bit X86 and more reliable because of the availability of the CMPXCHG8B instruction. Only very old, pre K5, AMD processors do not have it if I am not wrong.