<bitset>: optimize streaming operator >>

[AlexGuteniev]

📝 What it does

• Instead of calling push_back in a loop for temporary string, use temporary buffer with integer index
  • Have it on stack if small enough
  • Lazy allocate to have exceptions handled as they were before
  • Index is also reused as _Changed, there's no need for explicit variable for that
  • We also don't need to check for string::max_size() then
  • Custom RAII for that buffer if it is large to avoid dragging in <memory>
• Also restore not calling Traits::length in operator <<. The previous optimization avoided allocation, but introduced implicit length, which is good in total effect, but could have been better

More conservative alternative

string::_Resize_and_overwrite could have been used.
It is slightly worse due to not avoiding the allocation for mid-sized bitset.
The benchmark results are listed too.

🔍 Where's coverage

Not a vectorization PR, so no need to stress. Existing coverage ought to be enough:

> python tests\utils\stl-lit\stl-lit.py ..\..\llvm-project\libcxx\test ..\..\tests\std --filter=bitset -v
-- Testing: 207 of 53614 tests, 12 workers --

⏱️ How much of improvement

Benchmark	main	resize_and_overwrite	custom buffer (current)
bitset_from_stream<15, char>	22858 ns	24288 ns	20708 ns
bitset_from_stream<16, char>	26394 ns	26224 ns	19523 ns
bitset_from_stream<36, char>	17921 ns	14053 ns	10970 ns
bitset_from_stream<64, char>	14559 ns	9911 ns	8045 ns
bitset_from_stream<512, char>	8799 ns	5241 ns	10640 ns
bitset_from_stream<2048, char>	9396 ns	4801 ns	4619 ns
bitset_from_stream<15, wchar_t>	24327 ns	26221 ns	21511 ns
bitset_from_stream<16, wchar_t>	27978 ns	25502 ns	20432 ns
bitset_from_stream<36, wchar_t>	18754 ns	14578 ns	11391 ns
bitset_from_stream<64, wchar_t>	15853 ns	9798 ns	8462 ns
bitset_from_stream<512, wchar_t>	9609 ns	5244 ns	5014 ns
bitset_from_stream<2048, wchar_t>	8348 ns	5109 ns	4334 ns

🔜 What's up next

operator>> non-performance

The loop structure can be improved.
It can avoid unnecessary _Istr.rdbuf()->snextc() on the final counter increment.

I'm calling that back< I realized that it is an intention.

operator>> performance

Sigh... locales and streams.
Part of it is is in CRT which means it cannot be fixed.

So overall performance cannot reach the performance of from string construction.

[AlexGuteniev]

Applied changes, verified that benchmark results are about the same (although there is seemingly random variation)

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[StephanTLavavej]

Thanks for optimizing this operator! 🚀 ⏱️ 🎉