make dragonbox opt-in via -d:nimFpRoundtrips

[Araq]

No description provided.

[timotheecour]

@Araq this is a bad PR, for many reasons:

• it was merged without letting time for discussion which would've shown the flaws in this appproach
• users now have to choose between dropping roundtrip floats (which was a big feature of 1.6) or having broken semantics where you introduce an inconsistency between RT and VM or const folded expressions:

example 1

proc main(): string =
  let a = 0.1+0.2
  $a
static: echo main()
echo main()

nim r -d:nimFpRoundtrips main

0.3
0.30000000000000004

example 2

const a = 0.1+0.2
echo ($a, a)

nim r -d:nimFpRoundtrips main
("0.3", 0.30000000000000004)

example 3

const a1 = 0.1+0.2
let a2 = a1
doAssert a1 != 0.3 # just to be clear
echo [$a1, $a2]

nim r -d:nimFpRoundtrips main
["0.3", "0.30000000000000004"]

All this does is break roundtrip float and add complexity to the support matrix in nim libraries, creating more incompatibilities.

It also prevents users of 1.6 from having roundtrip floats at CT.

So, no, this doesn't make dragonbox opt-in, it introduces an opt-in flag for a now-broken feature.

And when that flag becomes opt-out, it still won't be compatible with a program that was compiled with -d:nimFpRoundtrips because of the above inconsistencies!

[kaushalmodi]

I was as well surprised seeing this get merged so soon too. Was there a bug report that inspired this PR?

[timotheecour]

I'm fixing those issues in #18531 and #18566