multiprecision acos() sometimes does not return

[steve@…]

Thibaut Paumard reported the following bug to Debian as ​https://bugs.debian.org/844495

Some inputs cause multiprecision acos() to hang. The following example hangs on s390x:

#include <boost/multiprecision/cpp_dec_float.hpp>

int main(int argc, char** argv) {

  boost::multiprecision::cpp_dec_float_100 alpha100, a, b;
  alpha100=5.9839860068377014046259e-02;
  b=cos(alpha100);
  a=acos(b);

  return 0;
}

Known affected architectures include arm64 armel armhf mips64el mipsel ppc64el s390x powerpc ppc64. It's not clear to me whether the value of alpha100 needs to be adjusted to trigger the bug on different architectures.

Thibaut has proposed the attached patch. Would appreciate knowing the maintainer's thoughts about the patch.

[steve@…]

patch for bug

[Thibaut Paumard <thibaut@…>]

Hi,

Although I haven't yet been able to demonstrate that acos() can hang on amd64, I confirm that ilogb(zero) is also 0 for a multiprecision zero on this architecture, therefore comparing it to FP_ILPOGB0 does not make sense:

#include <iostream>
#include <boost/multiprecision/cpp_dec_float.hpp>

int main(int argc, char** argv) {

  boost::multiprecision::number<boost::multiprecision::cpp_dec_float<100> >
    zero("0.");

  std::cout << ilogb(zero) << std::endl;
  std::cout << FP_ILOGB0 << std::endl;

  return 0;
}

Kind regards, Thibaut.

[John Maddock]

The attached patch fixes things for cpp_dec_float, but completely breaks other types in the library (as you would quickly discover running the regression tests). This is now fixed in develop in a series of commits:

​https://github.com/boostorg/multiprecision/commit/138a9055e2edc68dca2e78ff97ffe6276c17fa0a ​https://github.com/boostorg/multiprecision/commit/26ceb6ef386f5fb879ddccca59cc3bdc57c124f3 ​https://github.com/boostorg/multiprecision/commit/c0df5b6747d09268256b59501c3cab746112268e

These mostly relate to issue #12581 but fix this one as well.