Float16: convert to various number types; round, trunc, floor, ceil.

These additional operations are needed to get the random and linalg
tests closer to passing for the FloatRange change [#5636]. We really
need to make a decision about how first-class we want Float16 to be.
I'm starting to suspect that we should just bite the bullet and make
all the things that work for other floating-point types work for the
Float16 type also. It's just easier than having them be half-broken
and try to explain that they're "just for storage".

base/float16.jl

@@ -96,6 +96,16 @@ function convert(::Type{Float16}, val::Float32)
     reinterpret(Float16, h)
 end
 
+for T in (Bool, Int128, Uint128, Complex, Rational)
+    @eval convert(::Type{$T}, x::Float16) = convert($T,float32(x))
+end
+convert{T<:Number}(::Type{T}, x::Float16) = convert(T,float32(x))
+
+round(x::Float16) = float16(round(float32(x)))
+trunc(x::Float16) = float16(trunc(float32(x)))
+floor(x::Float16) = float16(floor(float32(x)))
+ ceil(x::Float16) = float16( ceil(float32(x)))
+
 isnan(x::Float16)    = reinterpret(Uint16,x)&0x7fff  > 0x7c00
 isinf(x::Float16)    = reinterpret(Uint16,x)&0x7fff == 0x7c00
 isfinite(x::Float16) = reinterpret(Uint16,x)&0x7c00 != 0x7c00

base/mpfr.jl

@@ -75,11 +75,12 @@ BigFloat(x::Integer) = BigFloat(BigInt(x))
 BigFloat(x::Union(Bool,Int8,Int16,Int32)) = BigFloat(convert(Clong,x))
 BigFloat(x::Union(Uint8,Uint16,Uint32)) = BigFloat(convert(Culong,x))
 
-BigFloat(x::Union(Float16,Float32)) = BigFloat(float64(x))
+BigFloat(x::Union(Float16,Float32,Float16)) = BigFloat(float64(x))
 BigFloat(x::Rational) = BigFloat(num(x)) / BigFloat(den(x))
 
 convert(::Type{Rational}, x::BigFloat) = convert(Rational{BigInt}, x)
 convert(::Type{BigFloat}, x::Rational) = BigFloat(x) # to resolve ambiguity
+convert(::Type{BigFloat}, x::Float16) = BigFloat(x) # to resolve ambiguity
 convert(::Type{BigFloat}, x::Real) = BigFloat(x)
 convert(::Type{FloatingPoint}, x::BigInt) = BigFloat(x)