CGLM_INLINE
void
glm_rotate_make_sse(mat4 m, float angle, vec3 axis) {
  __m128 x0, x1, x2, x3, x4, x5, x6, x7, x8;
  float  c, s;

  c  = cosf(angle);
  s  = sinf(angle);

  x0 = glm_simd_load_v3(axis);
  x0 = _mm_div_ps(x0, glm_simd_norm(x0));
  x1 = _mm_set_ps(s, c, 1.0f, 0.0f);

  x2 = _mm_mul_ps(x0, _mm_sub_ps(_mm_shuffle1_ps1(x1, 1),
                                 _mm_shuffle1_ps1(x1, 2)));

  x3 = _mm_mul_ps(x0, _mm_shuffle1_ps1(x2, 0));
  x4 = _mm_mul_ps(x0, _mm_shuffle1_ps1(x2, 1));
  x5 = _mm_mul_ps(x0, _mm_shuffle1_ps1(x2, 2));

  x2 = _mm_mul_ps(x0, _mm_shuffle1_ps1(x1, 3));

  x6 = _mm_shuffle2_ps(x1, x2, 1, 2, 2, 0, 0, 3, 2, 1);
  x7 = _mm_shuffle2_ps(x1, x2, 0, 2, 2, 0, 0, 3, 1, 2);
  x8 = _mm_shuffle2_ps(x1, x2, 0, 1, 2, 0, 0, 1, 3, 2);

  x6 = _mm_xor_ps(x6, _mm_set_ps(0.f, -0.f,  0.f,  0.f));
  x7 = _mm_xor_ps(x7, _mm_set_ps(0.f,  0.f,  0.f, -0.f));
  x8 = _mm_xor_ps(x8, _mm_set_ps(0.f,  0.f, -0.f,  0.f));

  _mm_store_ps(m[0], _mm_add_ps(x3, x6));
  _mm_store_ps(m[1], _mm_add_ps(x4, x7));
  _mm_store_ps(m[2], _mm_add_ps(x5, x8));
  _mm_store_ps(m[3], _mm_shuffle1_ps(x1, 1, 0, 0, 0));
}