Weirdness

[Rafael Garcia-Suarez]

On 25/02/2008, Dan Rowles <d.rowles at outcometechnologies.com> wrote:
> Hello,
>
>  I have encountered some weirdness with floating point numbers, and I was
>  hoping for your collective wisdom on what to do next.
>
>  I have two values that are being generated by some test code, and are
>  being compared with the "is" method in Test::More. One is calculated
>  from a database, the other is calculated via Statistics::Descriptive, in
>  order to confirm that an SQL query is working properly.
>
>  My code says:-
>  is(sprintf('%.5f', $this), sprintf('%.5f', $that));
>
>  The test fails, stating:-
>  #   Failed test 'max score is ok'
>  #   in t/reports/13health_change_per_procedure.t at line 241.
>  #          got: '4.74812'
>  #     expected: '4.74811'
>
>  Fair enough - they are different. Printing out their values shows:-
>  # THIS: '4.748115'
>  # THAT: '4.748115'
>
>  The numbers are identical when printed, but are rounded differently. So
>  I used the "Dump" method in Devel::Peek to have a closer look at these
>  numbers:-
>
>  THIS: SV = PVNV(0xa2ac528) at 0xa1c3a6c
>    REFCNT = 1
>    FLAGS = (PADBUSY,PADMY,NOK,POK,pNOK,pPOK)
>    IV = 0
>    NV = 4.748115
>    PV = 0xa2b36e8 "4.748115"\0
>    CUR = 8
>    LEN = 20
>
>  THAT: SV = PVNV(0xa2ac4f8) at 0xa1c39e8
>    REFCNT = 1
>    FLAGS = (PADBUSY,PADMY,NOK,POK,pIOK,pNOK,pPOK)
>    IV = 4
>    NV = 4.748115
>    PV = 0xa2f2da8 "4.748115"\0
>    CUR = 8
>    LEN = 36
>
>  Finally, I attempted to dump out their raw representation using the
>  following code:-
>      unpack("B*", pack("d",$this));
>
>  This produced:
>  THIS: 1001001010010110110010101101101100010001111111100001001001000000
>  THAT: 1001000010010110110010101101101100010001111111100001001001000000
>  DIFF:       ^
>
>  These are two different numbers, and perl correctly says that they are
>  different when I compare them, but it prints them out as exactly the
>  same number.
>
>  Now, I know very little about machine representations of floating point
>  numbers, but I *guess* that this means that the numbers differ by too
>  small an amount to print out, even though they do differ.
>
>  So my question to you is, should I report this as a bug somewhere, or
>  should I just ignore it as "floating point weirdness"?

perl's printf (and Devel::Peek) don't print every available digit for
floating point numbers, but use a default precision of whatever your
platform defines the C macro DBL_DIG to be (that's often 6 on the
average Linux system. In your case, that's 6 indeed)

If you want more decimals, you need to tell perl to do so, eg. with a
format "%.10f".

I guess that could be documented more accurately...