[GHC] #10378: min/max for Double/Float instances are incorrect

[GHC]

#10378: min/max for Double/Float instances are incorrect
-------------------------------------+-------------------------------------
        Reporter:  lerkok            |                   Owner:
            Type:  bug               |                  Status:  new
        Priority:  high              |               Milestone:
       Component:  Compiler          |                 Version:  7.10.1
      Resolution:                    |                Keywords:
Operating System:  Unknown/Multiple  |            Architecture:
 Type of failure:  None/Unknown      |  Unknown/Multiple
      Blocked By:                    |               Test Case:
 Related Tickets:                    |                Blocking:
                                     |  Differential Revisions:
-------------------------------------+-------------------------------------
Description changed by lerkok:

Old description:

> This is similar to many other numeric issues around `Double`s and
> `Float`s.
>
> The IEEE754 requires `min` and `max` on floats to return the "other"
> number, if one of the arguments is `NaN`. The default definitions used in
> Haskell do not satisfy this property.
>
> Furthermore, the current definitions are not commutative, when given
> `NaNs` and `-0` arguments.
>
> The following cases demonstrate the issue with `max`. Note that `min` has
> the exact same problems.
>
> {{{#!hs
> Prelude> (0/0) `max` 5
> NaN
> Prelude> 5 `max` (0/0)
> 5.0
> Prelude> isNegativeZero ((-0) `max` 0)
> False
> Prelude> isNegativeZero (0 `max` (-0))
> True
> }}}
>
> It wouldn't be hard to fix the definitions appropriately; here are
> reference implementations that are IEEE754 compliant:
>
> {{{#!hs
> max x y
>     | isNaN x                               = y
>     | isNaN y                               = x
>     | x > y || (x == y && isNegativeZero y) = x
>     | True                                  = y
>
> min x y
>     | isNaN x                               = y
>     | isNaN y                               = x
>     | x < y || (x == y && isNegativeZero x) = x
>     | True                                  = y
> }}}
>
> Note that these reference implementations would be quite expensive if GHC
> were to use them directly. Luckily, IEEE754 compliant min/max operations
> are supported by all modern CPUs, so doing the "right" thing here is also
> the cheaper thing as well. (i.e., it'll even be faster than the current
> incorrect implementation.)
>
> On platforms that do not have hardware implementations, the above
> definitions can serve as the correct implementations, albeit they'll be
> slower. (Embedded devices, perhaps.)

New description:

 This is similar to many other numeric issues around `Double`s and
 `Float`s.

 The IEEE754 requires `min` and `max` on floats to return the "other"
 number, if one of the arguments is `NaN`. The default definitions used in
 Haskell do not satisfy this property.

 Unfortunately, the IEEE-spec isn't exactly clear on how min/max should
 work around +0/-0. It's deliberately underspecified. Most Intel
 implementations simply return the second argument in this case. So, the
 following would be the "reference" implementation, assuming we map to the
 Intel CPU instructions for floating-point min/max:

 {{{#!hs
 max x y
     | isNaN x                               = y
     | isNaN y                               = x
     | (x == 0) && (y == 0)                  = y
     | x > y                                 = x
     | True                                  = y

 min x y
     | isNaN x                               = y
     | isNaN y                               = x
     | (x == 0) &&& (y == 0)                 = y
     | x < y                                 = x
     | True                                  = y
 }}}

--

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10378#comment:6>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler