Explicit rounding control per operation

[baryluk]

Hi,

add, sub, div, sqrt, mul, and float->integer conversions, should all have a third optional argument, to allow changing a rounding mode (nearest, toward plus infinity, towards minus infinity, to zero, out of zero, exact - last one excepting when result is not exact).

Default would be nearest and would generally create no overhead in code at all.

If the rounding is non standard / non default, it can translate to explicit rounding in instructions (AVX512 does that), or manipulate SSE control registers (preferably in optimized way to not switch there and back too much), to get proper behavior.

If not implemented in the first draft, at least make sure it can be added later without changing semantic or adding overheads.

[binji]

This is mentioned briefly in FutureFeatures.md for scalar floats.

[arunetm]

Floating point vector rounding instructions are considered here: #7
Efficient implementations across platforms is a concern in exposing these.

[baryluk]

@arunetm #7 only deals with explicit rounding of floating point conversions to integers. It doesn't deal with rounding control of all other arithmetic.