More efficient fmod()

[Wilco Dijkstra <Wilco.Dijkstra@arm.com>]

Hi Oliver,

> I've benchmarkey my code against e_fmod.c and
> the difference in the performance is still the same:

That looks like a random test over the full range - this tests 2 cases:
x < y and x > y with a large average exponent difference.
The first case is common and should have an early exit returning x.
The other is rare, and while it is worth reducing the worst-case timing,
this requires an algorithm which processes many bits at a time.
What is missing is the common case where the exponent difference is
small - here you can also have an early exit after doing a division.

For benchmarking it is best to test each case separately. When possible
we extract a representative trace from a benchmark and replay that so
we optimize for real-world inputs.

To add to Adhemerval's comments, the code needs to work efficiently on
32-bit targets as well as targets that don't have a count leading zeroes
instruction. Since there are many uses of both, the overhead of emulating
them is likely going to be expensive, so that might require an alternative
approach to make sure the new version is always faster.

Cheers,
Wilco