Re: Zero-padding and zero-masking fixes for to_char(float)

On Tue, Mar 24, 2015 at 03:08:26PM -0400, Bruce Momjian wrote:
> On Tue, Mar 24, 2015 at 10:05:12AM -0400, Bruce Momjian wrote:
> > On Tue, Mar 24, 2015 at 09:47:56AM -0400, Noah Misch wrote:
> > > On Sun, Mar 22, 2015 at 10:53:12PM -0400, Bruce Momjian wrote:
> > > > On Sun, Mar 22, 2015 at 04:41:19PM -0400, Noah Misch wrote:
> > > > > On Wed, Mar 18, 2015 at 05:52:44PM -0400, Bruce Momjian wrote:
> > > > > > This "junk" digit zeroing matches the Oracle behavior:
> > > > > >
> > > > > > SELECT to_char(1.123456789123456789123456789d, '9.9999999999999999999999999999999999999') as x from dual;
> > > > > > ------
> > > > > > 1.1234567891234568000000000000000000000
> > > > > >
> > > > > > Our output with the patch would be:
> > > > > >
> > > > > > SELECT to_char(float8 '1.123456789123456789123456789', '9.9999999999999999999999999999999999999');
> > > > > > ------
> > > > > > 1.1234567891234500000000000000000000000
> > >
> > > > > These outputs show Oracle treating 17 digits as significant while PostgreSQL
> > > > > treats 15 digits as significant. Should we match Oracle in this respect while
> > > > > we're breaking compatibility anyway? I tend to think yes.
> > > >
> > > > Uh, I am hesistant to adjust our precision to match Oracle as I don't
> > > > know what they are using internally.
> > >
> > > http://sqlfiddle.com/#!4/8b4cf/5 strongly implies 17 significant digits for
> > > float8 and 9 digits for float4.
> >
> > OK, I am fine in using those values if you can find them as compiler
> > defines, but I don't see how we can grab those values from a user test
> > on Oracle.

We encounter no authority higher than the test results, so it would be wrong
to seek out and use a define that just happens to match a test result. Adding
"#define TO_CHAR_DBL_DIG 17" and "#define TO_CHAR_FLT_DIG 9" is good.

> > There are some "invisible" float digits that don't appear in %f but can
> > be shown if desired --- I think we used to do that in the regression
> > tests, but found they added too much platform-specific randomness. Do
> > we want to go in that direction?

Bare %f simply prints all digits before the decimal point and exactly six
digits after the decimal point. Whether implementation-defined digits appear
in that output depends on the number's magnitude. However, float8out and
float4out do behave along the lines of your description.

I do recommend pushing TO_CHAR in that direction, to make it more like Oracle
while we're already breaking compatibility with PostgreSQL 9.4.

> How about if we have to_char() honor our extra_float_digits GUC, so
> users who want those digits can get them?

It's not my first choice; for one thing, no value of extra_float_digits would
yield 17 digits for float8 and 9 digits for float4.