LI77037: NAN FROM VEC_EXTRACT FROM DOUBLE COMPLEX CAST

APAR status

• Closed as suggestion for future release.

Error description

• When using optimization level -O3 and above, a -NaN value is
  being returned from the vec_extract vector built-in function.
  In this scenario, a double complex value is loaded into a
  vector4double datatype using a vec_ld2 vector built-in function.
  When the values are extracted from the vector4double, the result
  is an incorrect -NaN value.
  
  
      =====  CODE EXAMPLE  =====
  
    double complex test;
    test= 1+2*I;
    vector4double test_ex= vec_ld2(0, (double *)&test);
    double result = vec_extract(test_ex,0);
  

Local fix

• - Use a lower optimization level
  

Problem summary

Problem conclusion

Temporary fix

Comments

• The problem is due to an ansi aliasing violation in the usage
  of the 'vec_ld2' built-in function:
  
   vector4double test_ex= vec_ld2(0, (double *)&test);
  
  When optimization is disabled, the trivial case works because no
  assumptions are made about the data types being used.  Casting
  the second parameter to (double *) is a violation of ansi
  aliasing rules, and will affect how the code gets optimized at
  higher opt levels, such as -O3.  When this happens, bad code can
  be generated.
  
  The vec_ld2 built-in was not designed for use with complex data
  types.  For complex datatypes, the vec_ld built-in should be
  used instead.  The issue may also be worked around by compiling
  with the -qalias=noansi option, which tells the compiler that
  aliasing rules may be violated.  Usage of this option limits the
  optimization potential of the program, so it should be used
  carefully.
  
  
  A related APAR, LI76958, may also affect aliasing cases using
  the vec_ld2 and vec_ld built-in functions.
  

APAR Information

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

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

Fix information

Applicable component levels