<random>: Floating-point random number generation suboptimal due to generate_canonical()

[statementreply]

Describe the bug

The function template generate_canonical() makes a call to std::log2() to determine the number of calls to the random number generator given (in order to acquire enough entropy):

STL/stl/inc/random

Lines 268 to 273 in 5be7d49

	const _Real _Gxmin = static_cast<_Real>((_Gx.min)());
	const _Real _Gxmax = static_cast<_Real>((_Gx.max)());
	const _Real _Rx = (_Gxmax - _Gxmin) + _Real{1};
	const int _Ceil = static_cast<int>(_STD ceil(static_cast<_Real>(_Minbits) / _STD log2(_Rx)));
	const int _Kx = _Ceil < 1 ? 1 : _Ceil;

Although the number of calls required (_Kx) is a compile-time constant, the call to std::log2() cannot be optimized away even if /fp:fast is specified so that every time we generate a floating-point random number through std::generate_canonical(), which is the case for all the std::*_distribution's instantiated for a floating-point type, we must pay for a cost of a function call to std::log2().

Also, there is another cost for (inlined) std::ceil() followed by a floating-point to int cast.

The incurred cost of totally unnecessary function calls and a cast can be significant for such an application that makes heavy use of, say, a randomized algorithm.

STL version

https://github.com/microsoft/STL/commit/5be7d49c243979231dff35919d3a3d813d75d714

Additional context

Also tracked by DevCom-717452 and Microsoft-internal VSO-976509 / AB#976509.

[MattStephanson]

I wasn't aware of this issue at the time, but it seems that #1964 is a duplicate of this and is fixed by #2498.

[CaseyCarter]

Closing as duplicate of #1964. Thanks, Matt Stephanson (@MattStephanson)!