LI75703: BUSS ERROR WITH VA ARGS

September 2010 Update for XL C/C++ Advanced Edition for Blue Gene/L, V9.0

APAR status

• Closed as program error.

Error description

• The following test case leads to a bus error at runtime:
  
  ===== COMPILE COMMAND:
  /bgsys/drivers/ppcfloor/comm/bin/mpixlcxx_r -o bustest test.cpp
  
  ===== TESTCASE:
  =========Start of test.cpp===============
  #define MPICH_IGNORE_CXX_SEEK 1
  
  #include <stdio.h>
  #include <stdlib.h>
  #include <stdarg.h>
  #ifdef USEMPI
  #include <mpi.h>
  #endif
  
  // constants and enumerated types
  
  const int  Output_Text_File = 1 << 0 ;
  const int  Output_Screen    = 1 << 1 ;
  const int  Output_Error     = 1 << 2 ;
  
  // Class Prototypes
  
  class autumn_class
  {
    public:
  
      int Print(int output_files,
                const char *format, ...);
  
  };
  
  // Global Vars
  
  autumn_class autumn;
  void gyllen_test(void);
  
  // Classes
  
  int autumn_class::Print(int output_files,
                         const char *format, ...)
  {
      int num_chars;
      va_list args;
  
      va_start( args, format );
  
  // Eureka! This Line Generates Bus Error
  // Comment this line, and un-comment the next call to
  // vfprintf to see code run without a bus error.
      num_chars = vfprintf( stdout, format, args );
  
  // This Line Does Not!
  // Only difference is storing the return from the call
  // to vfprintf into the num_chars local variable.
      // vfprintf( stdout, format, args );
  
      va_end( args );
  
      return 0;
  }
  
  // Small routine to loop over problem code many times
  
  void gyllen_test(void)
  {
      for (int i = 0; i < 500; i ++ )
      {
              autumn.Print(Output_Screen,
                          "\t%03d: X-Coordinate Offset to Move
  Surfaces :\t       %13.6e %s\n",
                          i,
                          99.9,
                          "abcdefghijklmnop");
      }
  }
  
  // Main routine
  // Open file pointer
  // Call gylle_test routine
  
  int main(int argc, char **argv)
  {
  #ifdef USEMPI
      if (MPI_Init(&argc, &argv) != 0) {
          fprintf(stderr,"MPI_Init failed\n");
          abort();
      }
  #endif
  
      gyllen_test();
  
  #ifdef USEMPI
      if (MPI_Finalize() != 0) {
          fprintf(stderr,"MPI_Init failed\n");
      }
  #endif
  }
  ==============End of test.cpp=============
  

Local fix

• n/a
  

Problem summary

• USER AFFECTED:
  This affects Blue Gene users who are using variable arguments.
  
  PROBLEM DESCRIPTION:
  There is a runtime failure with 'signal 7'.
  var args creates a buffer in the automatic area of the callee
  to store var args passed in registers.
  The register is only 4 byte aligned, not 8.
  As a result, when saving floating point values, an alignment
  error occurs and the program traps to the OS to fix up the load
  After the default value of 1000 of these misaligned stores, the
  OS throws a signal 7.
  

Problem conclusion

• The compiler has been fixed so that the callee buffer is now
  8-byte aligned.
  

Temporary fix

Comments

APAR Information

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

  LI74903

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

Fix information

• Fixed component name

  XL C/C++ AE BG/

• Fixed component ID

  5799HJH00

Applicable component levels

• R900 PSN LI74903

     UP06/09/13