On Intel 32-bit platforms such as Windows and Linux, floating point unit (FPU) registers offer greater precision than floating point memory locations. It is not uncommon in C code for the same expression appearing in different parts of a program to produce slightly different results, because the compiler keeps intermediate results for a computation in the FPU registers. This permits a higher degree of precision, since the floating point register has 12 extra bits for the mantissa and 4 extra bits for the exponent. For example:

   data _null_;
     p=1;
     p=p/100;
     q=1;
     q=q/100;
     r=p-q;
     put p= q= r=;
   run;

produces:

   p=0.01  q=0.01  r=2.083701E-19

The value of R is not exactly 0. Internally, P is loaded in an FPU register and divided by 100. The result is written to memory, which causes the processor to convert the value from an 80-bit to a 64-bit floating point number. Q is then loaded and divided by 100. The result is again written to memory but is also retained in the FPU. The subtraction is then performed, using the stored 64-bit value for P and the 80-bit retained value for Q, which results in the slight difference in the generated result. The difference would be zero if Q was reloaded to the FPU prior to the subtraction.

Operating System and Release Information