Re: [HACKERS] numeric data type on 6.5

Thomas Lockhart <lockhart@alumni.caltech.edu> writes:
> Seems to do OK with numeric tokens of unspecified type which will
> become int8 and numeric in the parser. There may be some edge-effect
> cases (e.g. decimal data with 17 characters) which aren't quite right.
> Comments?

I'd suggest backing off one more place on the length of string you will
try to convert to a float8.  Since the test is strlen() <= 17, you
actually can have at most 16 digits (there must be a decimal point in
there too).  But IEEE float is only good to 16-and-change digits; I'm
not sure I'd want to assume that the 16th digit will always be
reproduced exactly.  15 digits would be safer.

It could still break if the C library's float<=>string conversion
routines are sloppy :-(.  I suppose you're interested in preserving
the info that "this constant looks numeric-ish" to assist in type
resolution heuristics?  Otherwise the value could be left in string
form till later.

Is there any value in marking the constant as a numeric token, yet
leaving its specific value as a string until after type resolution
is done?
        regards, tom lane