R devel - Mar 2017 - Hyperbolic tangent different results on Windows and Mac

Dear all,

We seem to have found a "strange" behaviour in the hyperbolic tangent
function tanh on Windows.
When running tanh(356 + 0i) the Windows result is NaN + 0.i while on Mac
the result is 1 + 0i. It doesn't seem to be a floating point error because
on Mac it is possible to run arbitrarily large numbers (say tanh(
999999677873648767519238192348124812341234182374817239847812738481234871823+0i)
) and still get 1 + 0i as result. This seems to be related to the imaginary
part as tanh(356) returns 1 in both Windows and Mac.

We have obtained those results in:
1) Mac with El Capitan v 10.11.6 *processor: 2.7 GHz Intel Core i5*
- 2) Mac with Sierra v 10.12.3 *processor: 3.2 GHz Intel Core i5*
- 3) Windows 10 Home v 1607 *processor: Intel Core m3-SY30 CPU@ 0.90 GHz
1.51 GHz*
- 4) Windows 7 Home Premium Service Pack 1 *processor: Intel Core i5-2410M
CPU @2.30 GHz 2.30GHz.*

?In all cases we are using R version 3.3.3 (64 bits)?

- *Does anybody have a clue on why is this happening?*
-
?PS: We have previously posted this issue in Stack Overflow (
http://stackoverflow.com/questions/42847414/hyperbolic-tangent-in-r-throws-nan-in-windows-but-not-in-mac).
A comment suggests it is related to a glibc bug.

?
?Thanks,?
Rod

	[[alternative HTML version deleted]]

>>>>> Rodrigo Zepeda <rzepeda17 at gmail.com>
>>>>>     on Fri, 17 Mar 2017 12:56:06 -0600 writes:
    > Dear all,
    > We seem to have found a "strange" behaviour in the hyperbolic
tangent
    > function tanh on Windows.
    > When running tanh(356 + 0i) the Windows result is NaN + 0.i while on
Mac
    > the result is 1 + 0i. It doesn't seem to be a floating point error
because
    > on Mac it is possible to run arbitrarily large numbers (say tanh(
    >
999999677873648767519238192348124812341234182374817239847812738481234871823+0i)
    > ) and still get 1 + 0i as result. This seems to be related to the
imaginary
    > part as tanh(356) returns 1 in both Windows and Mac.

    > We have obtained those results in:
    > 1) Mac with El Capitan v 10.11.6 *processor: 2.7 GHz Intel Core i5*
    > - 2) Mac with Sierra v 10.12.3 *processor: 3.2 GHz Intel Core i5*
    > - 3) Windows 10 Home v 1607 *processor: Intel Core m3-SY30 CPU@ 0.90
GHz
    > 1.51 GHz*
    > - 4) Windows 7 Home Premium Service Pack 1 *processor: Intel Core
i5-2410M
    > CPU @2.30 GHz 2.30GHz.*

(The hardware should not matter).

Yes, there is a bug here on Windows only, (several Linux
versions work correctly too).

    > ?In all cases we are using R version 3.3.3 (64 bits)?


    > - *Does anybody have a clue on why is this happening?*

    > ?PS: We have previously posted this issue in Stack Overflow (
    >
http://stackoverflow.com/questions/42847414/hyperbolic-tangent-in-r-throws-nan-in-windows-but-not-in-mac).
    > A comment suggests it is related to a glibc bug.

Yes, that would have been my guess too... as indeed, R on
Windows which should work for quite old versions of Windows has
been using a relatively old (gcc / libc) toolchain.

The upcoming version of R 3.4.0 uses a considerably newer
toolchain *BUT* I've just checked the latest "R-devel" binary
and the bug is still present there.

Here's a slight extension of the answer I wrote to the
above SO question here:  http://stackoverflow.com/a/42923289/161921

... Windows uses somewhat old C libraries, and here it is the
"mathlib" part of glibc. 

More specifically, according to the CRAN download page for R-devel for Windows
https://cran.r-project.org/bin/windows/base/rdevel.html ,
the R 3.3.z series uses the gcc 4.6.3 (March 2012) toolchain, whereas
"R-devel", the upcoming (not yet released!) R 3.4.z series uses
the gcc 4.9.3 (June 2015) toolchain.

According to Ben Bolker's comment on SO, the bug in glibc should have
been fixed in 2012 -- and so the change from 4.6.3 to 4.9.3
should have helped,
**however* I've just checked (installed the R-devel binary from CRAN on our
Windows server virtual machine) and I see that the problem is still present
there: In yesterday's version of R-devel, tanh(500+0i) still returns NaN+0i.

I now think a better solution would be to use R's internal
substitute (in R's src/main/complex.c): There, we have
------------------------------------------------
#ifndef HAVE_CTANH
#define ctanh R_ctanh
static double complex ctanh(double complex z)
{
    return -I * ctan(z * I); /* A&S 4.5.9 */
}
#endif
------------------------------------------------
and we should use it (by "#undef HAVE_CTAN" (or better by a
       	      	      configure check, using  ctanh("500 + 0i"),
as I see that on Windows,
   R> -1i * tan((500+0i)*1i)
gives
   [1] 1+0i
as it should for tanh(500+0i) --- but does not on Windows.

Martin Maechler
ETH Zurich and R Core

On 21.03.2017 10:54, Martin Maechler wrote:
>>>>>> Rodrigo Zepeda <rzepeda17 at gmail.com>
>>>>>>     on Fri, 17 Mar 2017 12:56:06 -0600 writes:
>
>     > Dear all,
>     > We seem to have found a "strange" behaviour in the
hyperbolic tangent
>     > function tanh on Windows.
>     > When running tanh(356 + 0i) the Windows result is NaN + 0.i while
on Mac
>     > the result is 1 + 0i. It doesn't seem to be a floating point
error because
>     > on Mac it is possible to run arbitrarily large numbers (say tanh(
>     >
999999677873648767519238192348124812341234182374817239847812738481234871823+0i)
>     > ) and still get 1 + 0i as result. This seems to be related to the
imaginary
>     > part as tanh(356) returns 1 in both Windows and Mac.
>
>     > We have obtained those results in:
>     > 1) Mac with El Capitan v 10.11.6 *processor: 2.7 GHz Intel Core
i5*
>     > - 2) Mac with Sierra v 10.12.3 *processor: 3.2 GHz Intel Core i5*
>     > - 3) Windows 10 Home v 1607 *processor: Intel Core m3-SY30 CPU@
0.90 GHz
>     > 1.51 GHz*
>     > - 4) Windows 7 Home Premium Service Pack 1 *processor: Intel Core
i5-2410M
>     > CPU @2.30 GHz 2.30GHz.*
>
> (The hardware should not matter).
>
> Yes, there is a bug here on Windows only, (several Linux
> versions work correctly too).
>
>     > ?In all cases we are using R version 3.3.3 (64 bits)?
>
>
>     > - *Does anybody have a clue on why is this happening?*
>
>     > ?PS: We have previously posted this issue in Stack Overflow (
>     >
http://stackoverflow.com/questions/42847414/hyperbolic-tangent-in-r-throws-nan-in-windows-but-not-in-mac).
>     > A comment suggests it is related to a glibc bug.
>
> Yes, that would have been my guess too... as indeed, R on
> Windows which should work for quite old versions of Windows has
> been using a relatively old (gcc / libc) toolchain.
>
> The upcoming version of R 3.4.0 uses a considerably newer
> toolchain *BUT* I've just checked the latest "R-devel" binary
> and the bug is still present there.
>
> Here's a slight extension of the answer I wrote to the
> above SO question here:  http://stackoverflow.com/a/42923289/161921
>
> ... Windows uses somewhat old C libraries, and here it is the
> "mathlib" part of glibc.
>
> More specifically, according to the CRAN download page for R-devel for
Windows
> https://cran.r-project.org/bin/windows/base/rdevel.html ,
> the R 3.3.z series uses the gcc 4.6.3 (March 2012) toolchain, whereas
> "R-devel", the upcoming (not yet released!) R 3.4.z series uses
> the gcc 4.9.3 (June 2015) toolchain.

Actually the R-3.3.z series already uses gcc-4.9.3.


Best,
Uwe Ligges
>
> According to Ben Bolker's comment on SO, the bug in glibc should have
> been fixed in 2012 -- and so the change from 4.6.3 to 4.9.3
> should have helped,
> **however* I've just checked (installed the R-devel binary from CRAN on
our Windows server virtual machine) and I see that the problem is still present
there: In yesterday's version of R-devel, tanh(500+0i) still returns NaN+0i.
>
> I now think a better solution would be to use R's internal
> substitute (in R's src/main/complex.c): There, we have
> ------------------------------------------------
> #ifndef HAVE_CTANH
> #define ctanh R_ctanh
> static double complex ctanh(double complex z)
> {
>     return -I * ctan(z * I); /* A&S 4.5.9 */
> }
> #endif
> ------------------------------------------------
> and we should use it (by "#undef HAVE_CTAN" (or better by a
>        	      	      configure check, using  ctanh("500 + 0i"),
> as I see that on Windows,
>    R> -1i * tan((500+0i)*1i)
> gives
>    [1] 1+0i
> as it should for tanh(500+0i) --- but does not on Windows.
>
> Martin Maechler
> ETH Zurich and R Core
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>

This looks like a bug in mingw-w64 CRT. The problem can be produced
with C++ without R:

  #include <iostream>
  #include <cmath>
  #include <complex>

  int main(){
    std::cout << std::fixed;
    std::complex<double> z(356, 0);
    std::cout << "tanh" << z << " = "
<< std::tanh(z)
         << " (tanh(356) = " << std::tanh(356) <<
")\n";
  }

On OS-X we get:

  tanh(356.000000,0.000000) = (1.000000,-0.000000) (tanh(356) = 1.000000)

But on Windows we get:

  tanh(356.000000,0.000000) = (nan,0.000000) (tanh(356) = 1.000000)

I was also able to reproduce the problem with gcc 6.3 in msys2 so it
has not been fixed upstream. You should file a bug report for
mingw-w64.

FWIF, we have run into NaN edge-case bugs before with mingw-w64.

 -
https://sourceforge.net/p/mingw-w64/mingw-w64/ci/6617ebd5fc6b790c80071d5b1d950e737fc670e1/
 -
https://github.com/wch/r-source/commit/e9aaf8fdeddf27c2a9078cd214a41475c8ff6f40

I am cc'ing Ray Donnelly who is an expert on mingw-w64.

For future reference:
https://sourceforge.net/p/mingw-w64/mailman/message/35747206/

On Wed, Mar 22, 2017 at 2:12 PM, Jeroen Ooms <jeroenooms at gmail.com>
wrote:
> This looks like a bug in mingw-w64 CRT. The problem can be produced
> with C++ without R:
>
>   #include <iostream>
>   #include <cmath>
>   #include <complex>
>
>   int main(){
>     std::cout << std::fixed;
>     std::complex<double> z(356, 0);
>     std::cout << "tanh" << z << " = "
<< std::tanh(z)
>          << " (tanh(356) = " << std::tanh(356)
<< ")\n";
>   }
>
> On OS-X we get:
>
>   tanh(356.000000,0.000000) = (1.000000,-0.000000) (tanh(356) = 1.000000)
>
> But on Windows we get:
>
>   tanh(356.000000,0.000000) = (nan,0.000000) (tanh(356) = 1.000000)
>
> I was also able to reproduce the problem with gcc 6.3 in msys2 so it
> has not been fixed upstream. You should file a bug report for
> mingw-w64.
>
> FWIF, we have run into NaN edge-case bugs before with mingw-w64.
>
>  -
https://sourceforge.net/p/mingw-w64/mingw-w64/ci/6617ebd5fc6b790c80071d5b1d950e737fc670e1/
>  -
https://github.com/wch/r-source/commit/e9aaf8fdeddf27c2a9078cd214a41475c8ff6f40
>
> I am cc'ing Ray Donnelly who is an expert on mingw-w64.