AMD Typewriter x86 사용자 설명서

22007E/0—November 1999

lead to unexpected results. Fortunately, in the vast majority of
cases, the final result will differ only in the least significant
bits.

double a[100],sum;
int i;

sum = 0.0f;

for (i=0; i<100; i++) {

sum += a[i];

}

double a[100],sum1,sum2,sum3,sum4,sum;
int i;

sum1 = 0.0;
sum2 = 0.0;
sum3 = 0.0;
sum4 = 0.0;
for (i=0; i<100; i+4) {

sum1 += a[i];
sum2 += a[i+1];
sum3 += a[i+2];
sum4 += a[i+3];
}

sum = (sum4+sum3)+(sum1+sum2);

Notice that the 4-way unrolling was chosen to exploit the 4-stage
fully pipelined floating-point adder. Each stage of the floating-
point adder is occupied on every clock cycle, ensuring maximal
sustained utilization.

In certain situations, C compilers are unable to extract common
subexpressions from floating-point expressions due to the
guarantee against reordering of such expressions in the ANSI
standard. Specifically, the compiler can not re-arrange the
computation according to algebraic equivalencies before
ex tracting  com mo n subexpre ss ions . I n such ca se s,  the
p r o g r a m m e r   s h o u l d   m a n u a l l y   e x t r a c t   t h e   c o m m o n
subexpression. It should be noted that re-arranging the
expression may result in different computational results due to
the lack of associativity of floating-point operations, but the
results usually differ in only the least significant bits.