Why aren't cmath's pow and sqrt templates?

They are overloaded for types float, double, and long double.

There are no other floating-point types, so there would be no advantage in a more general solution using templates.

The C++ <cmath> header is almost, but not quite, a copy of C's <math.h> header.

C doesn't have overloading, so it provides three different square root functions:

float sqrtf(float arg);
double sqrt(double arg);
long double sqrtl(long double arg);

and similarly for the other floating-point functions.

C++'s <cmath> provides those functions and overloaded versions of sqrt for all three floating-point types.

Here's an example that demonstrates this:

#include <iostream>
#include <iomanip>
#include <cmath>

int main() {
    auto f = std::sqrt(2.0F);
    auto d = std::sqrt(2.0);
    auto ld = std::sqrt(2.0L);
    std::cout << std::setprecision(64);
    std::cout << ' ' << sizeof f << ' '  << f  << '\n';
    std::cout << ' ' << sizeof d << ' '  << d  << '\n';
    std::cout        << sizeof ld << ' ' << ld << '\n';
}

The output on my system, demonstrating different size and precision for each call:

 4 1.41421353816986083984375
 8 1.4142135623730951454746218587388284504413604736328125
16 1.4142135623730950487637880730318329369765706360340118408203125

The 'c' in <cmath> literally means that it is a collection of C library functions. By analogy, <stdio.h> can be included in C++ as <cstdio> etc. and since C doesn't have templates, then neither does anything in <cmath>.

Aside from that, what would parameterizing the type get you? These functions take and return doubles, which you can always covert to a plain int or a float if that's what you need.

I stand corrected on the second point: as M.A. and Francois Andrieux point out, there are situations where this would be useful. However, I believe the answer still stands as far as <cmath> is concerned.