Re: [HACKERS] 123.45 - 123 = 0.45

Tom Lane wrote:

> Sooner or later we ought to try to implement true fixed-point
> arbitrary-precision numeric data types per the SQL spec.  That'll be
> a lot slower than hardware float math, but its roundoff properties will
> be less surprising to novices.

    I  think  the string math done in bc(1) could be a good point
    to start from. In the old version (1.3 if I  remember  right)
    of  the minix sources (yepp - still have them), there are the
    algorithm's to get  sine,  logarithm  and  the  like  in  any
    precision  as  bc  functions.  Internally, only the four base
    operations, (think think) pow() and sqrt() are implemented in
    C.  Anything else is done with them on the higher level.

    We  need  to define what the precision of a result should be,
    if it is not assigned to a column (where the precision can be
    the  atttypmod).  Is  there any standard defined for? If not,
    what about this:

    Internal representation holds different precision for DISPLAY
    and CALC.

    On  any operation, the DISPLAY precision is set to the higher
    of the two operands.

    On add/subtract, the CALC precision becomes the higher of the
    two.

    On multiply, the CALC precision is adjusted to hold the exact
    result up to a (variable settable?) maximum.

    On divide, the CALC precision is set to max and after  it  to
    the number of used digits.

    If  the  result get's assigned to an attribute, it is rounded
    to it's atttypmod and both precisions set to that.

    The types output function rounds it to the DISPLAY precision.

    The  input  function  sets  both  precisions to the number of
    digits present after decimal point.

    Needless to say that there will be special functions to round
    explicitly and set the precisions.


Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#======================================== jwieck@debis.com (Jan Wieck) #