Missing methods for Float16 - Integer comparison

[RalphAS]

Currently comparisons between Float16 and integer types fall back to using promote_type (instead of specific methods as for other floating point types). The promotion goes to Float16. This leads to the following sort of errors and inconsistencies:

julia-1.1> Inf16 == typemax(Int)
true

julia-1.1> Inf16 == 10^5
true

Compare:

julia-1.1> Inf16 == 1.0e5  # promotion works for float-float comparisons
false

julia-1.1> Inf32 <= typemax(UInt128)
false

julia-1.1> Float32(typemax(UInt128))
Inf32  # overflow is ok when it doesn't confound comparisons

Verified on a variety of Julia versions from 0.6.4 through a recent nightly.

[eric-perkerson]

Why is Float16 not simply included in line 476 in float.jl:

for Ti in (Int64,UInt64,Int128,UInt128)
for Tf in (Float32,Float64) # Why not include Float16 here?
# Defines comparison methods
end
end

[StefanKarpinski]

Because LLVM (and hardware generally) does not support 16-bit float operations.

[sam0410]

Hi @StefanKarpinski, I added comparison by changing it to Float64. Please review my PR #29870

[cossio]

@StefanKarpinski Close this?