IJ24655: DIVIDE VECTORIZATION CAUSING UNEXPECTED DIVIDE BY ZERO FP EXCEPTION

APAR status

• Closed as program error.

Error description

• When compiling with -O3 and HOT optimization, multiple double
  floating point divide operations are combined into a single
  vector instruction.  In the example code, to avoid an exception,
  divide-by-zero cases are protected with an IF condition:
  
      for (i = 0; i < len; i++) {
  
          if (value[i] != 0.0)
              result[i] /= value[i];
          else
              result[i] = 0.0;
      }
  
  The code generated respects the zero-case scenario logically,
  however the bad vectorized divide operation still occurs, thus
  triggering a runtime floating point exception.
  

Local fix

• Lower optimization level
  

Problem summary

• PROBLEM DESCRIPTION:
  The IPA component of the compiler is optimizing the looping
  divide using vector instructions unsafely.
  
  USERS AFFECTED:
  Users compiling code at O3 and up.
  

Problem conclusion

• The fix was to identify the unsafe operations in the control
  flow in the loop, and mark the unsafe operations un-simdizable.
  

Temporary fix

Comments

APAR Information

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

  LI81271

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

Fix information

• Fixed component name

  XL C/C++ FOR AI

• Fixed component ID

  5725C7200

Applicable component levels