#include <metal_stdlib>

using namespace metal;

template<typename T, int N, int M>
inline auto operator*(matrix<T, N, M> lhs, packed_vec<T, N> rhs) {
  return lhs * vec<T, N>(rhs);
}

template<typename T, int N, int M>
inline auto operator*(packed_vec<T, M> lhs, matrix<T, N, M> rhs) {
  return vec<T, M>(lhs) * rhs;
}

struct S {
  /* 0x0000 */ packed_float3 v;
  /* 0x000c */ int8_t tint_pad[4];
};

void f(device S& U) {
  U.v = float3(1.0f, 2.0f, 3.0f);
  U.v[0] = 1.0f;
  U.v[1] = 2.0f;
  U.v[2] = 3.0f;
}