Incorrect conversion to unisigned integer

[LilithHafner]

julia> convert(UInt32, 4.2949673f9)
0xffffffff

4.2949673f9 is one greater than the maximum value of a UInt32 (4.2949673f9 == typemax(UInt32)+1) so an inexact error should be thrown here:

function (::Type{$Ti})(x::$Tf)
    if ($(Tf(typemin(Ti))) <= x <= $(Tf(typemax(Ti)))) && isinteger(x)
        return unsafe_trunc($Ti,x)
    else
        throw(InexactError($(Expr(:quote,Ti.name.name)), $Ti, x))
    end
end

However, the comparison that happens is 4.2949673f9 <= Float32(typemax(UInt32)) which is true even though 4.2949673f9 > typemax(UInt32).

[LilithHafner]

unsafe_trunc calls llvm's fptoui under the hood so this result is a poison value

[gbaraldi]

The issue is that at this value the eps of Float32 is 512 so

julia> Float32(typemax(UInt32))
4.2949673f9

julia> ans + 1 
4.2949673f9

and float32 can't represent typemax(UInt32) exactly (It's 4.294967295e9)

[LilithHafner]

Here's a full list of failures:

julia> UInt128(3.402823669209385e38)
0x00000000000000000000000000000000

julia> UInt64(1.8446744073709552e19)
0xffffffffffffffff

julia> UInt64(1.8446744f19)
0xffffffffffffffff

julia> UInt32(4.2949673f9)
0xffffffff

julia> UInt128(1.8446744073709552e19)
0x00000000000000010000000000000000

for reference about what those float values are actually:

julia> UInt128.((1.8446744073709552e19, 1.8446744f19, 4.2949673f9))
(0x00000000000000010000000000000000, 0x00000000000000010000000000000000, 0x00000000000000000000000100000000)