Fix #7420

[ivarne]

When indexing a FloatRange with a OrdinalRange, you have to be careful to use the actual r.step, instead of step(r) = r.step/r.divisor.

[ivarne]

This should definitely have some tests, but I don't have time to write them now.

[Jutho]

A quick test learns that you should also not use r[first(s)] but probably r[first(s)]*r.divisor, although I would need to look better into range.jl to be sure.

[ivarne]

Seems like you are right range.jl#197

I would probably rather use r.start to not have the rounding error from a multiply and divide.

[ivarne]

How about?

FloatRange(r.start + r.step*(first(s)-1), r.step*step(s),length(s),r.divisor)

Or maybe everything gets inlined and the extra /r.divisor * r.divisor gets translated into nothing.

[simonster]

The compiler cannot safely optimize out /r.divisor * r.divisor since, as you note above, this may change the result due to rounding (or if r.divisor is 0. or Inf).

[StefanKarpinski]

It was slightly trickier than this because the start value is also lifted by the divisor; there was also some fussiness with bounds checking.