Hardware supported popcount breaks on Windows with CPUs that don't support SSE4 about dynamic_bitset

[jeking3]

cc: @Izaron

[jeking3]

cc: @Izaron

[glenfe]

What is the plan if the user being CC'd doesn't respond by 1.70 deadlines? Is the change that was committed going to be reverted for 1.70?

[jeking3]

That's the easiest thing to do, unless someone else steps up to resolve it.

[pdimov]

James, since you committed the PR, it's your responsibility to deal with the regression/crash, one way or another. (The gcc __builtin_popcount path is fine, but the msvc-specific section isn't, so there are options, but something must be done.)

[jeking3]

Okay, thanks.

[jeking3]

@pdimov what would your preference be to resolve this:

1. Disable the MSVC code (easiest).
2. Add a definition such as BOOST_DYNAMIC_BITSET_HAVE_SSE4 that folks must define in order to enable the code (also easy, requires documentation). Of course, this means they know their code will always run on a modern CPU so it's their choice.
3. Add static cpuid inspection (ideally C++11 only, so it is threadsafe) and fall back when not present (more difficult, may not have time to do...).

Right now I would opt for one of the first two, which makes the library no worse than it was in 1.68.0 for MSVC. It's unfortunate the original author @Izaron hasn't responded or taken action. For now we could do (1) and then open an issue for (3).

[glenfe]

I suggest #2 and document it accordingly: i.e. The opt-in macro would only affect MSVC; for GCC and Clang using __builtin_popcount unconditionally is fine. (#1 is also acceptable if you don't want to do #2).

[pdimov]

(1) is fine if we don't want to invest too much time into it.

On the surface silently doing the right thing, as in (3), is better than (2), but I'm not sure if the two tests - on the static guard variable and on the actual static variable - would not defeat the gains from using the popcnt instruction.

There is also the option of using an alternate popcount algorithm - the one used by gcc/clang for implementing the builtin when SSE4 isn't available, https://godbolt.org/z/VGP6Z_, https://en.wikipedia.org/wiki/Hamming_weight. This will need to be benchmarked if we choose to use it (on MSVC or in general).

[jeking3]

I was looking at https://github.com/kimwalisch/libpopcnt/ as there is an implementation in there as well, but I haven't done anything else but look. Benchmarks exist there for the decisions that were made to use different algorithms depending on the size of the bitset.

[pdimov]

FWIW, std::bitset on MSVC uses the same 8 bit table-based implementation as dynamic_bitset: https://godbolt.org/z/YuQqIj

This probably means that it was judged optimal there in the absence of hardware popcnt.