Resolve "IVAS DEC : segmentation fault errors with "-funsafe-math-optimizations" option"

• Related issues: #1560 (closed)
• Requested reviewers: @weckbecker

Reason why this change is needed

• The issue is in the SVD, more precisely in biDiagonalReductionLeft()

  for ( iCh = currChannel + 1; iCh < nChannelsC; iCh++ ) /* nChannelsC */
  {
      norm_x = 0.0f;
      for ( jCh = currChannel; jCh < nChannelsL; jCh++ ) /* nChannelsL */
      {
          norm_x += ( singularVectors[jCh][currChannel] * singularVectors[jCh][iCh] );
      }
  
      f = norm_x / maxWithSign( r );
  
  
      for ( jCh = currChannel; jCh < nChannelsL; jCh++ ) /* nChannelsL */
      {
          singularVectors[jCh][iCh] += ( f * singularVectors[jCh][currChannel] );
      }
  }

  The expression singularVectors[jCh][iCh] += ( f * singularVectors[jCh][currChannel] ); becomes -inf, which results in nan later on.

  This can only happen in f = norm_x / maxWithSign( r );, but the result of this expression is not available in the optimized build. As soon as I start adding e.g., a printf(), the issue disappears.

  Unless we now start changing the precision of the complete SVD, I would propose addressing the issue by making the variables which store the results of expressions like X / maxWithSign( Y ) "volatile". At least for me, this seems to resolve the issue.

Description of the change

• Add volatile keyword, to avoid over-optimization

Affected operating points

• Without -funsafe-math-optimizations, everything remain bit-exact

Related to #1560 (closed)