[GHC] #10387: toRational should error out on NaN and Infinity values for Double/Floats

5 May 2015

      #10387: toRational should error out on NaN and Infinity values for Double/Floats
-------------------------------------+-------------------------------------
              Reporter:  lerkok      |             Owner:
                  Type:  bug         |            Status:  new
              Priority:  normal      |         Milestone:
             Component:  Compiler    |           Version:  7.10.1
              Keywords:              |  Operating System:  Unknown/Multiple
          Architecture:              |   Type of failure:  None/Unknown
  Unknown/Multiple                   |        Blocked By:
             Test Case:              |   Related Tickets:
              Blocking:              |
Differential Revisions:              |
-------------------------------------+-------------------------------------
 The current implementation of `toRational` for `Float`/`Double` types
 accept `NaN` and `Infinity` values and produce results. For `Infinity` the
 result nicely maps back to `Infinity`:

 {{{#!hs
 Prelude> fromRational (toRational (1/0)) :: Double
 Infinity
 Prelude> fromRational (toRational (-1/0)) :: Double
 -Infinity
 Prelude>
 }}}

 But the same isn't true for NaNs:

 {{{#!hs
 Prelude> fromRational (toRational (0/0)) :: Double
 -Infinity
 }}}

 Also, these results are probably architecture specific.

 In any case, I think `toRational` should simply error out if it receives
 `NaN` or `Infinity`, as there are really no legitimate values we can map
 them to, even though we might be able to complete the round-trip correctly
 in case if infinities. The following just doesn't make sense:

 {{{#!hs
 Prelude> toRational (1/0)
 179769313486231590772930519078902473361797697894230657273430081157732675805500963132708477322407536021120113879871393357658789768814416622492847430639474124377767893424865485276302219601246094119453082952085005768838150682342462881473913110540827237163350510684586298239947245938479716304835356329624224137216
 % 1
 }}}

 So, I propose the `toRational` instance for Float/Double's should simply
 recognize `NaN` and `Infinite` values and throw an error.

 [Note that this is also how the HW implementations for the corresponding
 conversions work, raising a floating-point-exception (in the unmasked
 case). See `CVTPS2PI`/`CVTPS2DQ` etc., for instance, in x86.]

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/10387
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler