"Keith Thompson" <ks***@mib.or g> wrote in message
news:ln******** ****@nuthaus.mi b.org...
ag************* @gmail.com writes: result=(x-y) //subtract y from x
//if x>y result is +ve else -ve
result=result >>31 //get the sign bit to last bit
result=result&0 x01 //mask the last bit
return x*result + y*(~result & 0x1)
if result =1 x will be returned as ~result&0x1 would be 0
That has so many portability problems I'm not going to bother to
decide whether it's correct.
The correct would be (on a 2's complemented target where the right shift
operator performed on a signed value keeps the MSBit/sign bit, aka Shift
Right Arithmetic):
#include <stdio.h>
/* BITS_IN_LONG says how many bits are in long, can be computed (at run time
but then it'll be a variable not macro) or if exists such a define somewhere
in the std libc headers, can be taken right from there: */
#define BITS_IN_LONG 32
long Max(long x, long y)
{
long res;
res = (x-y) >> BITS_IN_LONG; /* res=0 (all zeroes) or -1 (all ones) */
res = (x & ~res) | (y & res); /* depending on prev value of res, we just
select/multiplex x or y by means of bitwise masking */
return res;
}
int main()
{
printf ("%ld\n", Max(-1,-1));
printf ("%ld\n", Max(-1, 0));
printf ("%ld\n", Max(-1, 1));
printf ("%ld\n", Max( 0,-1));
printf ("%ld\n", Max( 0, 0));
printf ("%ld\n", Max( 0, 1));
printf ("%ld\n", Max( 1,-1));
printf ("%ld\n", Max( 1, 0));
printf ("%ld\n", Max( 1, 1));
return 0;
}
Gives as expected:
-1
0
1
0
0
1
1
1
1
I'd suggest the newbies learn boolean stuff and digital logic better (or at
all if never studied before). I know that the above thing has a limited use,
but the limit is many millions of CPUs where it would work just fine.
And honestly, fixed point math (where people often use similar tricks) is
much more portable than floating point. It will yield the same result (if
coded properly, of course) on all targets and compilers, whereas floating
point will almost always give different LSBit (due to optimization,
operation reordering or because of simply being not IE754 compliant or using
shorter floating types) and since the errors tend to accumulate, the
difference can be bigger.
There're a few great docs on the subjects:
Sun's Numerical Computation Guide / What Every Computer Scientist Should
Know About Floating-Point Arithmetic by Goldberg
and
Algorithms For Programmers ideas and source code by Arndt
HTH
Alex
