json: insufficient precision for float64 serialization #15397
Description
Nim's standard json module serializes float numbers with 16 significant decimal digits
Nim/lib/system/formatfloat.nim
Line 32 in fe3211f
which is insufficient to represent all of 64 bit IEEE-754 floats. One needs at least 17 significant decimal digits, see https://en.wikipedia.org/wiki/Double-precision_floating-point_format for details.
The following MWE shows the problem and a possible solution that changes precision from 16 to 17 digits.
MWE
import json
import fenv
import strformat
let
x1eps = 1.0+epsilon(float64) # smallest float64 larger than 1
jsnStr1 = $(%*{"fnum": x1eps }) # json uses 16 significant digits, not enough!
jsnStr2 = fmt("""{{ "fnum": {x1eps:.17g} }}""") # need at least 17 significant digits
jsnNd1 = parseJson(jsnStr1)
jsnNd2 = parseJson(jsnStr2)
x1 = jsnNd1{"fnum"}.getFloat(NaN)
x2 = jsnNd2{"fnum"}.getFloat(NaN)
echo "jsbStr1 = ", jsnStr1
echo "jsbStr2 = ", jsnStr2
echo "jsnNd1 = ", jsnNd1
echo "jsnNd2 = ", jsnNd2
echo "x1 = ", x1
echo "x2 = ", x2
# Results:
doAssert (x1 != x1eps)
doAssert (x1 == 1.0)
doAssert (x2 == x1eps)Additional Information
$ nim -v
Nim Compiler Version 1.2.6 [MacOSX: amd64]
Compiled at 2020-09-02
Copyright (c) 2006-2020 by Andreas Rumpf
git hash: 211868b85cf8b11baabf06d75aaf7e2015d82e2b
active boot switches: -d:release