IZ88082: FP MULTIPLY SUBTRACT ROUNDING ISSUE

January 2011 PTF for XL C/C++ for AIX, V11.1
May 2011 PTF for XL C for AIX, V11.1
May 2011 PTF for XL C/C++ for AIX, V11.1
August 2011 PTF for XL C/C++ for AIX, V11.1
November 2011 PTF for XL C/C++ for AIX, V11.1
January 2012 PTF for XL C for AIX, V11.1
January 2012 PTF for XL C/C++ for AIX, V11.1
April 2012 PTF for XL C for AIX, V11.1
April 2012 PTF for XL C/C++ for AIX, V11.1
July 2012 PTF for XL C/C++ for AIX, V11.1
October 2012 PTF for XL C for AIX, V11.1
October 2012 PTF for XL C/C++ for AIX, V11.1
February 2013 PTF for XL C for AIX, V11.1
February 2013 PTF for XL C/C++ for AIX, V11.1
XL C/C++ for AIX Fix Pack 14 (May 2013 PTF) for 11.1
XL C for AIX Fix Pack 14 (May 2013 PTF) for 11.1
XL C/C++ for AIX Fix Pack 15 (August 2013 PTF) for 11.1
XL C for AIX Fix Pack 15 (August 2013 PTF) for 11.1
XL C/C++ for AIX Fix Pack 16 (November 2013 PTF) for 11.1
XL C for AIX Fix Pack 17 (February 2014 PTF) for 11.1
XL C/C++ for AIX Fix Pack 17 (February 2014 PTF) for 11.1
XL C/C++ for AIX Fix Pack 18 (May 2014 PTF) for 11.1
XL C/C++ for AIX Fix Pack 19 (August 2014 PTF) for 11.1
XL C/C++ for AIX Fix Pack 20 (November 2014 PTF) for 11.1
XL C for AIX Fix Pack 21 (April 2015 PTF) for 11.1
XL C/C++ for AIX Fix Pack 21 (April 2015 PTF) for 11.1
XL C/C++ for AIX Fix Pack 22 (September 2015 PTF) for 11.1
XL C for AIX Fix Pack 23 (March 2016 PTF) for 11.1
XL C/C++ for AIX Fix Pack 23 (March 2016 PTF) for 11.1
XL C for AIX Fix Pack 24 (September 2016 PTF) for 11.1
XL C/C++ for AIX Fix Pack 24 (September 2016 PTF) for 11.1
January 2011 PTF for XL C for AIX, V11.1
August 2011 PTF for XL C for AIX, V11.1
November 2011 PTF for XL C for AIX, V11.1
July 2012 PTF for XL C for AIX, V11.1
XL C for AIX Fix Pack 16 (November 2013 PTF) for 11.1
XL C for AIX Fix Pack 18 (May 2014 PTF) for 11.1
XL C for AIX Fix Pack 19 (August 2014 PTF) for 11.1
XL C for AIX Fix Pack 20 (November 2014 PTF) for 11.1
XL C for AIX Fix Pack 22 (September 2015 PTF) for 11.1

APAR status

• Closed as program error.

Error description

• ****************************************************************
  * Floating point expressions of the form:                      *
  *                                                              *
  *     a - b * c                                                *
  *     -(a - b * c)                                             *
  *                                                              *
  * are returning results that differ by a small margin in the   *
  * least significant digit, as compared to results from other   *
  * versions and releases of the XL compilers.                   *
  ****************************************************************
  

Local fix

• ****************************************************************
  * Applications compiled with affected compiler versions and a  *
  * minimum optimization level of -O2 can workaround the issue   *
  * by adding the compile option -qstrict=noieeefp.              *
  ****************************************************************
  

Problem summary

• ****************************************************************
  *USERS AFFECTED:                                               *
  * Those compiling applications that use floating point         *
  * arithmetic in the form outlined in the error description.    *
  * Affected compiler versions are XL C/C++ for AIX, V11.1.0.0 - *
  * V11.1.0.3 and XL Fortran V13.1.0.0 - V13.1.0.3.              *
  ****************************************************************
  *PROBLEM DESCRIPTION:                                          *
  * The result difference was caused by a change to the          *
  * instructions generated by the compiler for calculations of   *
  * the form ' a - b * c '.  Generated code utilizes fmul and    *
  * fsub instructions instead of the single fused                *
  * multiply/subtract fnmsub instruction, which was used in      *
  * other compiler versions.                                     *
  *                                                              *
  * Depending on the input values, the rounding difference       *
  * between these two instruction sequences can sometimes lead   *
  * to small differences in the least significant digit of the   *
  * result.  The difference is within standard rounding          *
  * tolerances, however code that has a strict reliance on the   *
  * presence of the fnmsub instruction and it's rounding         *
  * properties may be affected in more significant ways.         *
  ****************************************************************
  *RECOMMENDATION:                                               *
  * The issue was corrected via APAR in the XLC V11.1.0.4 PTF    *
  * (IZ88082) and XLF V13.1.0.4 PTF (IZ91502).  It is            *
  * recommended to use the latest PTF release of the compiler.   *
  ****************************************************************
  

Problem conclusion

• ****************************************************************
  * The compiler code generation was reverted back to use the    *
  * fnmsub instruction by default, effectively restoring the     *
  * previous rounding behaviour.                                 *
  ****************************************************************
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

  IZ91502 LI76141

Fix information

• Fixed component name

  XL C FOR AIX

• Fixed component ID

  5724X1200

Applicable component levels

• RB10 PSY

     UP