BUG: Incorrect Calculations with Both /Og and /G5

• Microsoft C/C++ Compiler (CL.EXE) included with Microsoft Visual C++, 32-bit Edition, versions 1.0, 2.0, and 2.1

SYMPTOMS

Code compiled with both /G5 and the /Og options generate incorrect floating point results.

RESOLUTION

To work around the problem, do not use both the /G5 and the /Og compiler switches together; use either /G5 or /Og. Alternatively, you can disable the global optimization (provided by /Og) for the area of code generated incorrectly with the #pragma optimize.

STATUS

Microsoft has confirmed this to be a bug in the Microsoft products listed at the beginning of this article. We are researching this bug and will post new information here in the Microsoft Knowledge Base as it becomes available.

MORE INFORMATION

The /G5 compiler switch tells the compiler to place the assembly instructions generated into an order that will be more beneficial to the pipelined architecture of the Pentium processor. The /Og switch enables global optimizations.

The following sample code illustrates the problem. Remove the comment slashes (//) from the #pragma line to optimze the line and illustrate a workaround to the problem.

Sample Code

/* Compile options needed: /G5 /Og
*/

#include <stdio.h>

//#pragma optimize("g", off)    //uncomment to fix the problem
void main()

   double   polylgd[5] = {0.0, 0.0, 0.0, 1.0, 1.0};
   double   valpol,an,an1,temp,dtemp;
   double   *polyptr;
   int      i;

   polyptr = &polylgd[4];
   an  = *polyptr--;
   an1 = *polyptr--;

   for (i = 4; i >= 2; i--)
   {
      dtemp = (double) (i - 1.0) / (double) i ;
      temp  =  dtemp * an;
      temp  = *polyptr - temp;
      *polyptr--;
      an    = an1 + (dtemp + 1.0) * (-0.5) * an;
      an1   = temp;
   }

   valpol = an1 + (-0.5) * an;
   printf("value should be:  0.14843750\n");
   printf("value this run is %.8lf\n", valpol);