RFC: add BitsFloat abstract type, leverage dispatch for rounding modes

[simonbyrne]

This adds an intermediate abstract type below FloatingPoint for floating point bitstypes (which I've termed BitsFloat, however it might be better to change this to make it more visually distinct from BigFloat). It also changes the behaviour of the rounding mode functions to leverage dispatch, for example:

# for ordinary floating point types
set_rounding(BitsFloat,RoundUp)
# for BigFloats
set_rounding(BigFloat,RoundUp)

It still requires some documentation, there are options for further behaviour:

• set_rounding(RoundUp) to change both modes
• set_rounding(Float64,RoundUp) == set_rounding(BitsFloat,RoundUp)

[StefanKarpinski]

I don't think you meant for this to sneak in here.

[ViralBShah]

I generally like this API cleanup.

[StefanKarpinski]

If we're going to do this we may want to also have BitsInteger.

[StefanKarpinski]

Which is relevant to @vtjnash's pr that had generic definitions of many integer operations.

[vtjnash]

Or allow specifying the type of the type in the function definitions

[JeffBezanson]

I'm not sure about this type. There is not really any connection between being a bitstype and setting rounding modes. It probably makes the most sense for the fenv rounding modes to apply to Union(Float32,Float64).

[simonbyrne]

@JeffBezanson The rationale was that the function changes the rounding mode of both types (or at least that's my understanding). But dispatching on concrete types is probably more useful (in fact that was my motivation for this).

If we do decide to keep it, perhaps IEEE754 might be a better name, since there might be other types for which we can't set rounding (such as a Float80 type?)

[kmsquire]

You could also make BitsFloat or IEEE754 an alias for
Union(Float32,Float64).

Kevin

On Wed, Dec 4, 2013 at 2:29 PM, Simon Byrne notifications@github.comwrote:

@JeffBezanson https://github.com/JeffBezanson The rationale was that
the function changes the rounding mode of both types (or at least that's my
understanding). But dispatching on concrete types is probably more useful
(in fact that was my motivation for this).

If we do decide to keep it, perhaps IEEE754 might be a better name, since
there might be other types for which we can't set rounding (such as a
Float80 type?)

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/5007#issuecomment-29852571
.

[simonbyrne]

I've removed the abstract type, and replaced it by a union. Before I update the documentation, we need to decide on the behaviour of set_rounding(r::RoundingMode). The possible options that I can think of are:
a. leave undefined
b. set rounding of Float32/Float64 only
c. set rounding of Float32/Float64 and BigFloat

[simonbyrne]

Okay, I've added docs and deprecation warnings (I've taken option b above, which corresponds to old usage).

[ivarne]

NEWS.md

[simonbyrne]

Okay, I've added the news and rebased.