Make floattype depend on the total number of bits, not just fractional bits

This ensures that conversion is lossless even for the largest representable values.

Author: timholy

src/FixedPointNumbers.jl

@@ -84,15 +84,10 @@ widen1(::Type{Int64})  = Int128
 widen1(::Type{UInt64}) = UInt128
 widen1(x::Integer) = x % widen1(typeof(x))
 
-if VERSION <= v"0.5.0-dev+755"
-    @generated function floattype{T,f}(::Type{FixedPoint{T,f}})
-        f>22 ? :(Float64) : :(Float32)
-    end
-else
-    @pure function floattype{T,f}(::Type{FixedPoint{T,f}})
-        f>22 ? Float64 : Float32
-    end
-end
+typealias ShortInts Union{Int8,UInt8,Int16,UInt16}
+
+floattype{T<:ShortInts,f}(::Type{FixedPoint{T,f}}) = Float32
+floattype{T,f}(::Type{FixedPoint{T,f}}) = Float64
 floattype(x::FixedPoint) = floattype(supertype(typeof(x)))
 
 

test/fixed.jl

@@ -103,7 +103,7 @@ end
 
 # Floating-point conversions
 @test isa(float(one(Fixed{Int8,6})),   Float32)
-@test isa(float(one(Fixed{Int32,18})), Float32)
+@test isa(float(one(Fixed{Int32,18})), Float64)
 @test isa(float(one(Fixed{Int32,25})), Float64)
 
 # Show

test/ufixed.jl

@@ -112,7 +112,7 @@ x = UFixed8(0b01010001, 0)
 @test -x == 0xafuf8
 
 @test isa(float(one(UFixed{UInt8,7})),   Float32)
-@test isa(float(one(UFixed{UInt32,18})), Float32)
+@test isa(float(one(UFixed{UInt32,18})), Float64)
 @test isa(float(one(UFixed{UInt32,25})), Float64)
 
 for T in (FixedPointNumbers.UF..., UF2...)