https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79151
--- Comment #5 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #3)
> The question is of course whether vector division has comparable latency /
> throughput as the scalar one.
Here's a test case on a rather old CPU, a Core 2 Q8200:
$ cat foo.c
#include <stdio.h>
double foo(double a, double b)
#ifdef SCALAR
{
return 1/a + 1/b;
}
#else
{
typedef double v2do __attribute__((vector_size (16)));
v2do x, y;
x[0] = a;
x[1] = b;
y = 1/x;
return y[0] + y[1];
}
#endif
#define NMAX 1000000000
int main()
{
double a, b, s;
s = 0.0;
for (a=1.0; a<NMAX+0.5; a++)
{
b = a+0.5;
s += foo(a,b);
}
printf("%f\n", s);
}
$ gcc -DSCALAR -O3 foo.c && time ./a.out
41.987257
real 0m19,508s
user 0m19,500s
sys 0m0,000s
$ gcc -O3 foo.c && time ./a.out
41.987257
real 0m9,146s
user 0m9,140s
sys 0m0,000s
