optimize `exp(exp(x)) / exp(x)` with fast-math

[k-arrows]

Consider the following two functions:

#include <cmath>

double f1(double x)
{
    return exp(exp(x)) / exp(x);
}

double f2(double x)
{
    return exp(exp(x) - x);
}

When these functions are compiled with -Ofast, Clang emits code for f1 that performs floating-point division. GCC does not emit code that performs floating-point division for either function.

[zjaffal]

taking a look