<bit>: Improve has_single_bit() codegen

[pps83]

code avoids branches, provides 20% speed up on x64 targets.

x64 results

---------------------------------------
Benchmark                          Time
---------------------------------------
Rand16_has_single_bit            284 ns
Rand16_has_single_bit_new        232 ns
Rand32_has_single_bit            306 ns
Rand32_has_single_bit_new        246 ns
Rand64_has_single_bit            289 ns
Rand64_has_single_bit_new        233 ns

partially addresses #5359

[pps83]

seems like code is better for x86/x64/arm/arm64 : https://godbolt.org/z/fjshqMf4d

the only potential drawback: old code would check for _Val = 0 and jump. New code has no jumps (except on arm)

[AlexGuteniev]

Looks good to me, some nitpicks...

[pps83]

Can you please try this local change also:

_NODISCARD constexpr bool has_single_bit(const _Ty _Val) noexcept {
#if 1 // _POPCNT_INTRINSICS_ALWAYS_AVAILABLE
#if _HAS_CXX20
    if (_STD is_constant_evaluated())
        return 1 == _Popcount_fallback(_Val);
#endif // _HAS_CXX20
    return 1 == _Unchecked_popcount(_Val);
#else
    return (_Val ^ (_Val - 1)) > _Val - 1;
#endif
}

[AlexGuteniev]

Turns out this particular benchmark is very much affected by #4496 and popcnt version is slower.

That's why I've hidden the previous results, and obtained the updated results by changing also this line to say Release instead of RelWithDebInfo

STL/benchmarks/CMakeLists.txt

Line 62 in f2a2933

set(CMAKE_BUILD_TYPE RelWithDebInfo)

i5-1235U

x64

P-core

Benchmark	Before	After	popcnt
has_single_bit_if<uint16_t>	11.5 ns	10.7 ns	6.68 ns
has_single_bit_if<uint32_t>	11.3 ns	11.3 ns	10.6 ns
has_single_bit_if<uint64_t>	6.98 ns	11.7 ns	6.57 ns
has_single_bit<uint16_t>	6.36 ns	4.51 ns	3.51 ns
has_single_bit<uint32_t>	6.08 ns	4.71 ns	3.52 ns
has_single_bit<uint64_t>	5.27 ns	4.71 ns	3.53 ns

E-core

Benchmark	Before	After	popcnt
has_single_bit_if<uint16_t>	16.2 ns	14.1 ns	14.1 ns
has_single_bit_if<uint32_t>	16.7 ns	14.0 ns	14.0 ns
has_single_bit_if<uint64_t>	16.2 ns	16.2 ns	14.0 ns
has_single_bit<uint16_t>	12.2 ns	9.07 ns	7.24 ns
has_single_bit<uint32_t>	11.9 ns	9.07 ns	7.24 ns
has_single_bit<uint64_t>	11.6 ns	9.07 ns	7.25 ns

Apparently popcnt is a good idea. Even for the "if" case.

---

#4496 wasn't a severe obstacle for most previous optimization, but at such micro level it is really disrupting, we'll need to get into that.

[AlexGuteniev]

Can you please try this local change also:

Before you go with this in the PR:

• We don't compare like 1 == function() in new code, although there are some relic precedents. In new code, we use natural order function() == 1
• Always control flow braces {} even around single statement in if
• You don't need to check #if _HAS_CXX20. It is C++20 header, and the check is already there at the very beginning for the whole content

[pps83]

Before you go with this in the PR...

This PR doesn't have popcnt change. The popcnt change I suggest above doesn't use the dispatcher, it's the unchecked version that does the popcnt opcode directly and it should only be used when building for AVX+. The default (version with this PR) is better than using dispatcher imo. _Checked_popcount will use the dispatcher.

I can add popcnt it as a separate PR, i think it's better that way.

[StephanTLavavej]

For the record, ARM32 performance is indeed utterly irrelevant. Windows ripped out their ARM32 build, and will even be dropping support for ARM32 binaries running on ARM64 machines, see https://learn.microsoft.com/en-us/windows/arm/arm32-to-arm64 .

ARM32 still needs to behave correctly (until DevDiv/MSVC drops support for targeting it, which I am constantly asking about), but we don't want to waste any more time on it.

[StephanTLavavej]

Thanks! I pushed some changes to the benchmark for consistency with repo conventions.