optimize `pow(x, i1) / pow(x, i2)` with fast-math

[k-arrows]

Consider the following:

#include <cmath>

double f1(double x)
{
    return pow(x, 2) / pow(x, 1);
}

double f2(double x)
{
    return pow(x, 3) / pow(x, 1);
}

https://godbolt.org/z/jjKE5n8o1
If -Ofast is specified, GCC just calculates x * x for f2, but Clang performs floating-point division.

When I lower the version to Clang 14.0.0, it behaves as expected, has anything changed?