_TLOSS in _matherr

_TLOSS in _matherr

• sepehr.Beigi

  Hi,
  
  I have to ask you for more specific list of concerns. Is there a specific code scenario when VC _matherr is different from Borland's one
  
  Thanks,
  Nikola
  VC++
  
  

• ginganinja

  Hi,
  
  In this sample the 1.0e300 is a valid double value (<DBL_MAX) and sin() is computed, exception is not thrown. My guess, that Borland folks assume that sin of such a big value is equal to sin(infinite) which is undefined. I guess if VC compiler was letting us enter 1e309 > DLB_MAX, then perhaps in this case CRT could throw TLOSS exception. But this literal is tranlated to a variable of double type and compiler checks for valid range. I do not think something should be change in VC CRT for this case. What kind of issues does this difference in CRTs create in your application
  
  Thanks,
  Nikola
  
  

• cjarvis

  I have created issue# 536175 for tracking down the bug. As Nikola suggested, it might be hard to get it fixed in the coming release but it will sure be considered for the next one.
  
  Thanks for reporting the issue!
  
  Ayman Shoukry
  VC++
  
  

• GerryT

  I will follow up and get the issue logged in our bug database.
  
  Thanks,
    Ayman.
  
  

• Sunil S

  If you can get your code to work on vc++ 6 it is easy to get it to work on UNIXES. Usually it will just compile without any major changes with the gcc compiler.
  Unforentally I never tried these fuctions so they may now work easly.
  If you cant get the gcc compiler to work you can find help in its usage all over the place. Most open source communities can help you.
  
  I never used the borland compiler so I cant help you with that.
  
  I hope that I was of some assistance.
  

• Rich_Cov

  If I compile the code fpe.cpp with Visual C++ I get
  
  _matherr (Domain): log(-2)
  log(-2.0) = -1.#IND00e+000

  _matherr (Domain): sqrt(-1)
  sqrt(-1.0) = -1.#IND00e+000

  sin(1.0e300) = 9.804836e-001
  
  but with Borland C++ I get the more sensible
  
  _matherr (Domain): log(-2)
  log(-2.0) = -NAN

  _matherr (Domain): sqrt(-1)
  sqrt(-1.0) = +NAN

  _matherr (Total loss of significance): sin(1e+300)
  sin(1.0e300) = +NAN
  
  ---------------------------------------------------------------------------
  -------------------------------------- fpe.cpp ---------------------------
  

  #ifdef __BORLANDC__

  #include <windows.h>

  #include <stdio.h>

  #else

  #include "stdafx.h"

  #include <fpieee.h>

  #endif

  #include <math.h>

  #include <excpt.h>

  #include <float.h>

  char *errs[] = { "Domain", "Singularity", "Overflow",

  "Partial loss of significance", "Total loss of significance",

  "Underflow" };

  #ifdef __BORLANDC__

  #define _EXC_MASK EM_UNDERFLOW + EM_OVERFLOW + EM_ZERODIVIDE + EM_INEXACT

  #define TEST(st,op) { _controlfp(_EXC_MASK, MCW_EM); \

  __try { \

  printf( st " = %e\n\n", op); \

  } \

  __except (0) {\

  printf("Exception raised: \n");\

  }}

  #else

  #define _EXC_MASK _EM_UNDERFLOW + _EM_OVERFLOW + _EM_ZERODIVIDE + _EM_INEXACT

  #define TEST(st,op) { _controlfp(_EXC_MASK, _MCW_EM); \

  __try { \

  printf( st " = %e\n\n", op); \

  } \

  __except (_fpieee_flt(GetExceptionCode(),GetExceptionInformation(),fpieee_handler)) {\

  printf("Exception raised: \n");\

  }}

  #endif

   

  int _matherr( struct _exception *e )

  {

  printf("_matherr (%s): %s(%lg)\n",errs[e->type-1],e->name,e->arg1);

  return 1;

  }

  #ifndef __BORLANDC__

  int fpieee_handler( _FPIEEE_RECORD *pieee )

  {

  return EXCEPTION_CONTINUE_EXECUTION;

  }

  #endif

   

  int main(int argc, char* argv[])

  {

  TEST("log(-2.0)",log(-2.0));

  TEST("sqrt(-1.0)",sqrt(-1.0));

  TEST("sin(1.0e300)",sin(1.0e300));

  return 0;

  }

  
  
  
  
   

  
  

• Shlomi Meoded

  The trouble is the answer given by VC is completely wrong: the right value is
  

  -0.9857504251603769966...

  as can be checked by means of any program with variable precision. Not even the sign is right in VC result!
  
  BC does not assume sin(1.0e300) is undefined, it only warns you that it cannot be computed with double precision. I think this is the only reasonable answer, because to compute sin(1.0e300)  you have to use a lot of decimal places to reduce the argument to the right interval: maybe [0,2 pi].
  

• Rohit Tela

  Agreed. But it is a very old code, I doubt we are going to change it in VS2005. But it may be a good thing to do in the next release.  Our FP-"experts" can take a look on this bug when it is in the database. Go ahead and open a bug on lab.msdn.microsoft.com, it will be re-assigned to Dev-owners accordingly. Thanks for reporting this.
  
  Nikola
  VC++
  
  
• _TLOSS in _matherr