Test failure when compiled with -O0

[nh2]

If I compile with -O0, I get the following test failure:

*** Failed! Falsifiable (after 1 test):
0.0
0.0
(used seed -8074436915145524066)
  t_shortest: [Failed]
*** Failed! Falsifiable (after 1 test):
0.0
0.0
(used seed 5824167789149766384)
Regressions:
  toShortest_overflow: [OK]

         Properties  Test Cases  Total      
 Passed  0           1           1          
 Failed  2           0           2          
 Total   2           1           3          

The test code in question is https://github.com/bos/double-conversion/blob/e4b7956d51fbb79ada95412519bd2a6a680def23/tests/Properties.hs#L9-L13

where a = 0.0, b = 0.0.

The problem seems to be the use of realToFrac (0.0 / 0.0). Note that 0.0/0.0 is NaN.

• With -O, realToFrac NaN returns NaN
• With -O0, realToFrac NaN returns -Infinity

You would expect realToFrac :: Double -> Double to always be id, but apparently that is not the case in GHC right now. It is so only if compile with optimisations.

[nh2]

This is due to what's explained on:

• https://gitlab.haskell.org/ghc/ghc/issues/16519
• https://gitlab.haskell.org/ghc/ghc/issues/3676#note_39486

As mentioned in the latter issue by Simon Marlow 9 years ago:

`realToFrac cannot convert accurately between floating-point types

I think this means that the code in https://github.com/bos/double-conversion/blob/e4b7956d51fbb79ada95412519bd2a6a680def23/Data/Double/Conversion/ByteString.hs#L68-L72 is wrong, as it uses realToFrac to convert between floating point types (Double -> CDouble).

Of course it's somewhat unreasonable to blame double-conversion for this. The fact that Haskell does not provide a function to reasonably convert Double -> CDouble is insane (I posted here).

[Kleidukos]

Can still reproduce with GHC 9.6