#include <metal_stdlib>
using namespace metal;

template<typename T, size_t N>
struct tint_array {
  const constant T& operator[](size_t i) const constant { return elements[i]; }
  device T& operator[](size_t i) device { return elements[i]; }
  const device T& operator[](size_t i) const device { return elements[i]; }
  thread T& operator[](size_t i) thread { return elements[i]; }
  const thread T& operator[](size_t i) const thread { return elements[i]; }
  threadgroup T& operator[](size_t i) threadgroup { return elements[i]; }
  const threadgroup T& operator[](size_t i) const threadgroup { return elements[i]; }
  T elements[N];
};

struct tint_module_vars_struct {
  device float* s;
};

tint_array<float, 4> f1() {
  return tint_array<float, 4>{};
}

tint_array<tint_array<float, 4>, 3> f2() {
  tint_array<float, 4> const v = f1();
  tint_array<float, 4> const v_1 = f1();
  return tint_array<tint_array<float, 4>, 3>{v, v_1, f1()};
}

tint_array<tint_array<tint_array<float, 4>, 3>, 2> f3() {
  tint_array<tint_array<float, 4>, 3> const v_2 = f2();
  return tint_array<tint_array<tint_array<float, 4>, 3>, 2>{v_2, f2()};
}

kernel void tint_symbol(device float* s [[buffer(0)]]) {
  tint_module_vars_struct const tint_module_vars = tint_module_vars_struct{.s=s};
  tint_array<float, 4> const a1 = f1();
  tint_array<tint_array<float, 4>, 3> const a2 = f2();
  tint_array<tint_array<tint_array<float, 4>, 3>, 2> const a3 = f3();
  (*tint_module_vars.s) = ((a1[0] + a2[0][0]) + a3[0][0][0]);
}