[Haskell-cafe] Trouble with asinh (c calls with Doubles) in Windows

[David James]

Hello all – I’m trying to debug a problem with asinh in Windows.

On Linux (Ubuntu 18.04.5 LTS) running GHC 8.10.7, I get (I think correctly):

GHCi, version 8.10.7: https://www.haskell.org/ghc/  :? for help
Prelude> asinh 1.7976931348623157e308
710.4758600739439

But on Windows I get:

GHCi, version 8.10.7: https://www.haskell.org/ghc/  :? for help
Prelude> asinh 1.7976931348623157e308
NaN

My understanding is that, since commit c6f4eb4f8<https://gitlab.haskell.org/ghc/ghc/-/commit/c6f4eb4f8bc5e00024c74198ab9126bf1750db40> (released in GHC 8.8.1, I think), asinh is defined as a primop that just calls the c asinh function, so I tried the following code:

In Main.hs:

{-# LANGUAGE ForeignFunctionInterface #-}
{-# LANGUAGE CApiFFI #-}

module Main (main) where

import Foreign
import Foreign.C.Types

foreign import ccall unsafe "math.h   asinh"   c_asinh      :: Double -> Double
foreign import ccall unsafe "CAsinh.c testFn"  c_testFn     :: IO ()

main :: IO ()
main = do
  putStrLn $ "asinh   1.7976931348623157e308 = " ++ show (  asinh 1.7976931348623157e308)
  putStrLn $ "c_asinh 1.7976931348623157e308 = " ++ show (c_asinh 1.7976931348623157e308)

  putStrLn "Calling c_testFn..."
  c_testFn

In CAsinh.c:

#include <stdio.h>
#include <math.h>

void testFn ()
{
  printf ("in testFn\n");

  printf ("asinh(1.7976931348623157e308) = %f\n", asinh(1.7976931348623157e308));

  double x = 1.7976931348623157e308;
  printf ("asinh(x) = %f\n", asinh(x));
}

In Linux, this all works fine:

asinh   1.7976931348623157e308 = 710.4758600739439
c_asinh 1.7976931348623157e308 = 710.4758600739439
Calling c_testFn...
in testFn
asinh(1.7976931348623157e308) = 710.475860
asinh(x) = 710.475860

But on Windows I get weird results:

asinh   1.7976931348623157e308 = NaN
c_asinh 1.7976931348623157e308 = NaN
Calling c_testFn...
in testFn
asinh(1.7976931348623157e308) = 710.475860
asinh(x) = -1.#IND00

The primop call from Haskell to asinh and the FFI call to c_asinh give the same (incorrect) result, as I expected. But the first call to asinh from the c testFn gives the correct result. So why is that different to the FFI call from Haskell?

And why doesn’t the second call from testFn return the right result?

I also tested on 9.0.1 on Windows, with a slightly different result for the final asinh:

...
asinh(x) = nan

I’ve investigated the things I can think of (different foreign declarations, size of double being different, checking the code in cpp.sh<http://cpp.sh/>), but can’t find an explanation. Am I doing something silly? Is this a (known) bug?

Note that on Windows, values up to asinh 1.3407807929942596e154 work correctly.

(As a little background: I was trying to implement the Kahan<https://people.freebsd.org/~das/kahan86branch.pdf> functions to give the correct branch cuts for complex trig functions, something I think the current Haskell implementation doesn’t do correctly. These depend on asinh, etc, on RealFloat numbers).

Thanks very much for any help,
David.


  *

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/haskell-cafe/attachments/20210903/b3d7eefa/attachment.html>