<cmath>: Use ceil/floor/round/etc. intrinsics

[StephanTLavavej]

In VS 2019 16.8 Preview 3, the following intrinsics will be available (MS-internal MSVC-PR-265316):

double __ceil(double);
float __ceilf(float);
double __floor(double);
float __floorf(float);
double __round(double);
float __roundf(float);
double __trunc(double);
float __truncf(float);
double __copysign(double, double);
float __copysignf(float, float);
unsigned int __signbitvalue(double);
unsigned int __signbitvaluef(float);

These are available for all architectures, declared in <intrin0.h>. We should take advantage of them. (We may need to report this to the UCRT too.)

Also tracked by DevCom-987080 (MS-internal VSO-1101527 / AB#1101527) and DevCom-1038270 (MS-internal VSO-1129605 / AB#1129605).

[statementreply]

Precaution: maybe we should verify that __round/__roundf uses the rounding mode of std::round, not FE_NEAREST, on all architectures/ISAs//archs, before using it.

From OP of DevCom-1038270:

round = _MM_FROUND_TO_NEAREST_INT

If __round implements this rounding mode, it couldn't be used when the rounding behavior of std::round is required.

[sylveon]

After investigating, the only place where this can be applied from the STL's side is the handwritten float overloads. Functions with a f suffix and the double versions are from the UCRT.

I believe a best solution would be to give intrinsic value to those functions to the compiler (for example like how memcpy is).