Here are two patches.
The 0001-*.patch simply adds a regression test to numeric.sql that bits
aren't lost casting from float[48] to numeric. Naturally, it fails at this
stage.
The 0002-*.patch is a proof-of-concept patching float4_numeric and
float8_numeric in the trivial way (just using FLT_DECIMAL_DIG and
DBL_DECIMAL_DIG in place of FLT_DIG and DBL_DIG). It makes the new
regression test pass. (It will only work under a compiler that has
__FLT_DECIMAL_DIG__ and __DBL_DECIMAL_DIG__ available, and I used
those internal versions to avoid mucking with build tooling to change
the target C standard, which I assume wouldn't be welcome anyway.
This patch is just POC and not proposed as anything else.)
It changes the output of an existing numeric test and a few numeric
aggregate tests.
I have adjusted the numeric test: its purpose is to check the direction
of numeric rounding of ties, but the value to be rounded was computed
in float8 and then cast to numeric, and because negative powers of ten
aren't tidy in binary, it can turn out that the float8 computation
produces a correctly-rounded-53-bit-result that is on the nearer-to-zero
side of a tie, and now that that result is correctly cast, the resulting
numeric doesn't round in the away-from-zero direction.
I changed that test because I concluded it wasn't meant to test
float8-to-numeric casting, but only the rounding of tied numerics,
so I just made the inner expression be typed numeric, and the expected
output is unchanged.
The three aggregate tests with changed output are working from a table
of float4 values, and my assumption is they are now simply producing
the correct aggregate results given the full precision of the input
values, but I haven't confirmed that yet, so this patch leaves those
three failing for now.
-Chap