Add tests against MPFR for scalbn{f} and ldexp{f}

[tgross35]

No description provided.

[tgross35]

This discards the rounding of 2 ^ exp which should be okay since powers of two are representable if within range. Somebody should double check this though.

[tgross35]

Ah this doesn't work because scalbn(INF, -HUGE) works out to INF * 0 = NaN, but the result should be infinity.

For reference, from the C spec:

• scalbn(±0, n) returns ±0.
• scalbn(x, 0) returns x.
• scalbn(±∞, n) returns ±∞.

If the calculation does not overflow or underflow, the returned value is exact and independent of the current rounding direction mode.

[programmerjake]

after checking the docs, what you want is Float::shl, if you're not immediately converting back to a primitive float you'll probably also want to call subnormalize

[programmerjake]

you can look at subnormalize_ieee_round's source for an idea of how to use subnormalize, i'm assuming rust may gain support for non-ieee floats such as bf16 or f80.

[tgross35]

Thanks, that does seem to work simpler than what I had. The call to prep_retval does call subnormalize_ieee_round before casting to the float type (link), I don't think bf16 uses libm (unlike f16, all bf16 ops may be done at f32 precision right?) but this will probably need to be adjusted once we add the x87 float type. Easiest to just cross that bridge later though, I'm just getting started with f16/f128 here.