<numeric>: Can we improve throughput or reduce repetition with is_constant_evaluated()?

[StephanTLavavej]

After Daniil Goncharov (@Neargye) implemented constexpr <numeric> in #399, Michael Schellenberger Costa (@miscco) observed in #399 (comment) that there is a potential throughput concern here (and another occurrence below):

STL/stl/inc/numeric

Lines 78 to 97 in 7ad0d63

	#ifdef __cpp_lib_is_constant_evaluated
	// TRANSITION, DevCom-878972
	if (_STD is_constant_evaluated()) {
	for (; _UFirst != _ULast; ++_UFirst) {
	_Val = _Reduce_op(_STD move(_Val), *_UFirst); // Requirement missing from N4713
	}
	return _Val;
	} else
	#endif // __cpp_lib_is_constant_evaluated
	{
	if constexpr (_Plus_on_arithmetic_ranges_reduction_v<_Unwrapped_t<const _InIt&>, _Ty, _BinOp>) {
	(void) _Reduce_op; // TRANSITION, VSO-486357
	return _Reduce_plus_arithmetic_ranges(_UFirst, _ULast, _Val);
	} else {
	for (; _UFirst != _ULast; ++_UFirst) {
	_Val = _Reduce_op(_STD move(_Val), *_UFirst); // Requirement missing from N4713
	}
	return _Val;
	}
	}

On a related note, I observe that there is code repetition between the is_constant_evaluated and !_Plus_on_arithmetic_ranges_reduction_v paths.

I suspect that we can't get both optimal throughput and zero repetition (because of the non-if constexpr nature of is_constant_evaluated). So perhaps the best pattern is:

if constexpr (_Plus_on_arithmetic_ranges_reduction_v) {
    if (!_STD is_constant_evaluated()) {
        return _Reduce_plus_arithmetic_ranges(ARGS);
    }
}

return CLASSIC_IMPLEMENTATION;