Boost C++ Libraries: Ticket #10993: frexp with multiprecision cpp_dec_float_50 returns random values with VS 2013

attachment set

Paul A. Bristow — Wed, 04 Feb 2015 10:04:15 GMT

attachment → test.cpp

John Maddock — Thu, 05 Feb 2015 19:16:52 GMT

With expression templates on, then the return value is an expression template - no calculations have been performed at that point - you have to assign the returned value to something to force the function to be evaluated.

But.... this is probably a big trap for the unwary, so I'll change.

Paul A. Bristow — Fri, 06 Feb 2015 09:46:56 GMT

I was indeed unwary, and fell straight into the trap :-(

The original problem was within this templated function:

template <int N, class T = double>
T nth_2deriv(T x)
{ // return nth root of x using 1st and 2nd derivatives and Halley.
  //typedef double guess_type; // Works OK with this (and may be accurate enough anyway).
  typedef T guess_type; // Failed with this with cpp_dec_float_50 only.
  int exponent;
  frexp(static_cast<guess_type>(x), &exponent); // Get exponent of z (ignore mantissa).
  T guess = ldexp(static_cast<guess_type>(1.), exponent / N); // Rough guess is to divide the exponent by n.
  T min = ldexp(static_cast<guess_type>(1.) / 2, exponent / N); // Minimum possible value is half our guess.
  T max = ldexp(static_cast<guess_type>(2.), exponent / N); // Maximum possible value is twice our guess.
...

Others may appreciate this trap not existing! Thanks.

anonymous — Fri, 06 Feb 2015 20:25:46 GMT

But the sample code shows frexp obtaining faulty results for the integer exponent argument. This is not a multiprecision type. I don't understand what the exponent obtained in frexp has to do with et_on or et_off. I'm confused here.

Chris

anonymous — Sat, 07 Feb 2015 08:50:17 GMT

But the sample code shows frexp obtaining faulty results for the integer exponent argument. This is not a multiprecision type. I don't understand what the exponent obtained in frexp has to do with et_on or et_off. I'm confused here.

The function is never evaluated in the et_on case unless the returned value is assigned to something - so the integer argument remains uninitialized. It's a bug in the expression template code really, will fix shortly. John.

Paul A. Bristow — Sun, 08 Feb 2015 16:35:03 GMT

I confirm that providing something to take the result of frexp works as predicted.

T mantisssa = frexp(static_cast<guess_type>(x), &exponent);

This provides an easy workaround for the current release 1.56.

(But it is curious the GCC works OK without this when MSVC fails?).

status changed; resolution set

John Maddock — Sun, 22 Feb 2015 13:23:27 GMT

status new → closed
resolution → fixed

Fixed in https://github.com/boostorg/multiprecision/commit/f970a181414751b1de54c66ebec0e40543830450