IV23305: INACCURATE/INVALID RESULTS RETURNED BY SOME MATH LIBRARY APIS APPLIES TO AIX 7100-00

Obtain the fix for this APAR.

APAR status

• Closed as program error.

Error description

• Some APIs in the math library are showing incorrect
  values.
  For example lrint() API returns wrong value.
  

Local fix

• Use a copy of 6.1 version of libm.a instead.
  

Problem summary

• ****************************************************************
  * USERS AFFECTED:
  * Systems running the 7100-00 Technology Level with
  * bos.adt.prof below the 7.1.0.20 level and/or bos.adt.libm
  * below the 7.1.0.17 level.
  ****************************************************************
  * PROBLEM DESCRIPTION:
  * This APAR addresses inaccuracies/invalid results returned by
  * some of the functions in the AIX math library in early AIX 7.1
  * levels that do not have this APAR installed. It also addresses
  * a linker error that may be seen when using the sqrt()
  * function.
  *
  * Inaccuracies may be seen in math functions such as pow(),
  * exp() and assorted trigonometric functions. The symptoms range
  * from valid but slightly different values in the least
  * significant digits, to significant inaccuracies, depending on
  * the function and the arguments it is called with.
  *
  * Examples of the data discrepancies for known impacted APIs
  * includes small errors in acos, asin, and atan (but still
  * under 1 ulp), and larger errors in cosh, exp, expm1, pow,
  * sinh, up to 500 ulps. The trigonometric functions cos, cosf,
  * sin, sinf, tan, tanf can be accurate for small-magnitude
  * arguments but can have complete loss of significance for very
  * large magnitude arguments. The most serious impact to users of
  * the problem library is likely to occur with applications
  * calling the double-precision trigonometric functions with
  * large arguments (>16), or the single-precision trigonometric
  * functions with very large arguments (>2ª18).
  *
  * Invalid results can be seen with llrint(), llrintl() and
  * lrint() functions in 64 bit mode.  The issue is caused by an
  * internal max value check using an invalid literal value,
  * resulting in a return value of -1, and errno set to 33. The
  * issue was fixed, and correct max value literals are now used.
  *
  * Linker error of the form .ld:  0711-768 WARNING: Object
  * /usr/lib/libm.a sqrt_raise_xcp.o , section 1, function
  * .strcmp: The branch at address 0x40 is not  followed by a
  * recognized no-op or TOC-reload instruction. may be seen when
  * using the sqrt() function. The issue was resolved and the
  * __sqrt_raise_xcp now contains the necessary no-op instruction.
  *
  * For any programs using libm.a functions that were linked on
  * AIX 7.1,  it is recommended that this APAR be installed and
  * the program relinked.  The program must be relinked since
  * libm.a is a static library.
  *
  * The inaccuracy in certain floating-point calculations is a
  * result of a change that had been made in the initial releases
  * of the XL C/C++ for AIX, V11.1 and XL Fortran for AIX, V13.1
  * compilers.  Affected versions are XLC/C++ V11.1.0.0 through
  * V11.1.0.4 and XLF V13.1.0.0 through V13.1.0.4.  Code generated
  * for floating point calculations of the form ' a - b * c '
  * utilizes fmul and fsub instructions with these versions,
  * instead of the fnmsub instruction used in other compiler
  * versions.  The rounding difference between these two
  * instruction sequences can lead to small differences in the
  * computed result.  The issue was corrected in
  * XLC V11.1.0.5 (IZ96351), XLC/C++ V11.1.0.5 (IZ96352), and
  * XLF V13.1.0.5 (IZ96353).  If using the affected compiler
  * versions noted above,  it is also recommended to move to the
  * latest PTF release of the compiler.
  ****************************************************************
  * RECOMMENDATION:
  * Install APAR IV23305.
  ****************************************************************
  

Problem conclusion

• Makefile changes to pick new compiler for libm
  

Temporary fix

• *********
  * HIPER *
  *********
  

Comments

• 7100-00 - use AIX APAR IV23305
  

APAR Information

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

  IV22812

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

  U857205

Fix information

• Fixed component name

  AIX V7.1

• Fixed component ID

  5765H4000

Applicable component levels

• R710 PSY U857205

     UP12/12/11 I 1000