1e50 == 1?

[Matthew Paul Del Buono <delbu9c1@...>]

Hello Lua-L,

I was messing around today and may have found some kind of bug in Lua's addition method; I was posting here to confirm.

I was actually testing for stack overflows, etc., due to too complex statements, but ran across this instead which didn't seem accurate. This was tested in both 5.1.2 and 5.1.3rc5:

loadstring("return "..("1+"):rep(1e50).."1")()
Returns: 1

Surely this is incorrect. Different interpreters tend to have different limits (I have seen that the 5.1.3rc5 interpreter reports "1" when doing ("1+"):rep(1e10) while 5.1.2 takes a while to do the calculation and ends up with the correct result). 

Can anyone else confirm that they are having this problem as well? I have only tested in Windows and Cygwin, but see the same problem in both places.

If the problem does exist, can anyone see how to fix it? My belief (without any looking into the code whatsoever) is that Lua is performing some kind of optimization incorrectly. The compiled code doesn't seem to appear out of the ordinary:


function <lua.lua:1,3> (13 instructions, 52 bytes at 0x12d1128)
0 params, 6 slots, 0 upvalues, 0 locals, 7 constants, 0 functions
        1       [2]     GETGLOBAL       0 -1    ; assert
        2       [2]     GETGLOBAL       1 -2    ; loadstring
        3       [2]     LOADK           2 -3    ; "return "
        4       [2]     LOADK           3 -4    ; "1+"
        5       [2]     SELF            3 3 -5  ; "rep"
        6       [2]     LOADK           5 -6    ; 1e+50
        7       [2]     CALL            3 3 2
        8       [2]     LOADK           4 -7    ; "1"
        9       [2]     CONCAT          2 2 4
        10      [2]     CALL            1 2 0
        11      [2]     CALL            0 0 2
        12      [2]     CALL            0 1 1
        13      [3]     RETURN          0 1


(Note this was actually an older version of this code, so you see the "assert" in there. My older version was:

assert(loadstring("return "..("1+"):rep(1e50).."1"))()

No big difference, I just have a habit of using assert with loadstring to avoid "attempt to call a nil value")

Anyway, I'll keep looking through the C code, but I was just curious if anyone's picked this up before? I know I'm being picky, but personally I'd prefer this to be fixed, though I don't necessarily agree with taking all that time to get the correct value. Even just a "arithmetic expression too complex" error would be better than the incorrect result "1." I can understand incorrect results for things like floating point errors, etc., but even 1+1 is not 1... preferably this would not turn out this way :)


Thanks for any response,
-- Matthew P. Del Buono a.k.a Shirik