cleanups

[oscardssmith]

This removes a bunch of compats for unsupported Julia versions, and slightly simplifies the annoying hacks we're making to handle the ABI issues.

[DilumAluthge]

This seems like a breaking change?

[simonbyrne]

Yeah, maybe bump the version number?

[DilumAluthge]

Yeah. If we're going to make a breaking release, might as well take the opportunity to do 1.0.0 now.

[oscardssmith]

Is this actually breaking? I would be very surprised if anyone was using the layout of this type. That said, no objection to making a 1.0 tag...

[simonbyrne]

That's a much better idea, I don't know why I didn't do it this way originally.

[oscardssmith]

Also, while I'm cleaning things up, Is the current approach of calling C functions for the basic conversions good? I see ~2x faster performance using the pure Julia UInt128 conversion than the ccalled UInt64 conversion (zen5 linux). If one of you has an aarc64 system (which has native Float128), can you run this benchmark?

julia> @btime Float128(x) setup=x=rand(UInt64)
  10.249 ns (0 allocations: 0 bytes)

@benchmark Float128(UInt128(x)) setup=x=rand(UInt64)
  3.727 ns (0 allocations: 0 bytes)

[simonbyrne]

On my M2 Mac

julia> @btime Float128(x) setup=x=rand(UInt64)
  2.750 ns (0 allocations: 0 bytes)
5.31167148155784243000000000000000000e+18

julia> @btime Float128(UInt128(x)) setup=x=rand(UInt64)
  2.208 ns (0 allocations: 0 bytes)
6.82926167534499090300000000000000000e+18

Yeah, so that looks faster...

[simonbyrne]

You can also clean up these bits as well:

Quadmath.jl/src/Quadmath.jl

Lines 191 to 197 in f6ea87e

	if Sys.iswindows()
	return reinterpret(Float128,d)
	else
	y1 = reinterpret(Float64,UInt64(d >> 64))
	y2 = reinterpret(Float64,(d % UInt64))
	return Float128((VecElement(y2),VecElement(y1)))
	end

Quadmath.jl/src/Quadmath.jl

Lines 215 to 221 in f6ea87e

	if Sys.iswindows()
	return reinterpret(Float128,d)
	else
	y1 = reinterpret(Float64,UInt64(d >> 64))
	y2 = reinterpret(Float64,(d % UInt64))
	Float128((VecElement(y2),VecElement(y1)))
	end

[oscardssmith]

@simonbyrne I might have gone slightly overboard and moved a lot of the simple implementations into Julia with pretty decent perf gains (usually ~2x sometimes as big as 8x)