[Bug c++/64497] New: std::scalbln does not round correctly for long doubles

[Bug c++/64497] New: std::scalbln does not round correctly for long doubles

[walter.mascarenhas at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497

            Bug ID: 64497
           Summary: std::scalbln does not round correctly for long doubles
           Product: gcc
           Version: 4.8.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: walter.mascarenhas at gmail dot com

The overload std::scalbln(long double, long) may not round the result correctly
when the exponent is very small. For instance, with round mode = near,

    std::scalbln(1.1L, -16446) 

returns 0, whereas  std::scalbn(1.1L, -16446) and std::ldexp(1.1L, -16446)
return std::numeric_limits<long double>::denorm_min(), which I believe
is the correct result.

[Bug c++/64497] std::scalbln does not round correctly for long doubles

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497

--- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
GCC just calls the scalnlnl() function in libm, so it's not a GCC bug, and is
not specific to C++ either. I suggest you report it to your libc vendor.

Complete testcase in C:

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

int main()
{
  long double di = scalbnl(1.1L, -16446);
  assert( di != 0.0L );
  long double dl = scalblnl(1.1L, -16446L);
  assert( dl != 0.0L );
}

[Bug c++/64497] std::scalbln does not round correctly for long doubles

[walter.mascarenhas at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497

--- Comment #2 from Walter Mascarenhas <walter.mascarenhas at gmail dot com> ---
What if there is a difference in the expected behavior
for this function in  C and C++11? Is it not up to g++
for implementing what is mandated in C++11? (This
is not a rhetorical question, I really do not know the answer.)

In http://man7.org/linux/man-pages/man3/scalbn.3.html it
is written that scalbln should return 0 in case of underflow:

"If the result underflows, a range error occurs, and the functions
return zero, with a sign the same as *x*."

On the other hand,
http://en.cppreference.com/w/cpp/numeric/math/scalbn states that

"If a range error due to underflow occurs, the correct result (after
rounding) is returned."

I looked at the standard (N3797.pdf) but did not find anything
specific about std::scalbln.
If there is indeed a discrepancy in the definitions of scalbln in C
and C++11 then there
may be no bug in libm, and my vendor will not change it.

I do not have a copy of the ISO 60599 standard, and I do not know whether
the content of the pages http://man7.org/linux/man-pages/man3/scalbn.3.html and
http://en.cppreference.com/w/cpp/numeric/math/scalbn are compatible with
any standards. Therefore I am in no position to argue,
but maybe you could think a bit longer about this..














On Mon, Jan 5, 2015 at 10:29 AM, redi at gcc dot gnu.org <
gcc-bugzilla@gcc.gnu.org> wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497
>
> --- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
> GCC just calls the scalnlnl() function in libm, so it's not a GCC bug, and
> is
> not specific to C++ either. I suggest you report it to your libc vendor.
>
> Complete testcase in C:
>
> #include <stdio.h>
> #include <math.h>
> #include <assert.h>
>
> int main()
> {
>   long double di = scalbnl(1.1L, -16446);
>   assert( di != 0.0L );
>   long double dl = scalblnl(1.1L, -16446L);
>   assert( dl != 0.0L );
> }
>
> --
> You are receiving this mail because:
> You are on the CC list for the bug.
> You reported the bug.
>

[Bug c++/64497] std::scalbln does not round correctly for long doubles

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497

--- Comment #3 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Walter Mascarenhas from comment #2)
> What if there is a difference in the expected behavior
> for this function in  C and C++11?

There isn't any difference, so it doesn't matter.

> Is it not up to g++
> for implementing what is mandated in C++11? (This
> is not a rhetorical question, I really do not know the answer.)

Yes.

> On the other hand,
> http://en.cppreference.com/w/cpp/numeric/math/scalbn states that
> 
> "If a range error due to underflow occurs, the correct result (after
> rounding) is returned."

The C++ standard doesn't say that.

> I looked at the standard (N3797.pdf) but did not find anything
> specific about std::scalbln.

It says it is the same as C, so cppreference.com is wrong.

[Bug c++/64497] std::scalbln does not round correctly for long doubles

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497

--- Comment #4 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Looking at the C standard, it seems that the result is implementation-defined
on underflow, and zero is a valid result. C++ doesn't change that.

[Bug c++/64497] std::scalbln does not round correctly for long doubles

[cubbi at cubbi dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497

Sergey Zubkov <cubbi at cubbi dot org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |cubbi at cubbi dot org

--- Comment #5 from Sergey Zubkov <cubbi at cubbi dot org> ---
FYI, cppreference got that phrasing from POSIX's "If the correct value would
cause underflow, and is representable, a range error may occur and the correct
value shall be returned.", which is a part of its optional IEC 60559
Floating-Point extension to the C standard:
http://pubs.opengroup.org/onlinepubs/9699919799/functions/scalbln.html

[Bug c++/64497] std::scalbln does not round correctly for long doubles

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64497

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
                 CC|                            |jakub at gcc dot gnu.org
         Resolution|---                         |INVALID

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Closing as this is definitely not a GCC issue.
Please file it against glibc instead.  Note, on x86_64, scalbnl is implemented
in assembly, while scalblnl is not.