From IRC:

<roconnor> The documentation for secp256k1_fe_equal_var suggest both inputs have to have magnitude 1; however only the first input needs to have magnitude 1.
<roconnor> thus this call to fe_normalize_weak could be eliminated.
<roconnor> But it is so very not critical, I'm not sure it is even worth PRing.
<real_or_random> roconnor: same thing happens in secp256k1_gej_eq_x_var which is used in ECDSA verification. but ok, the speedup is still negligible
<real_or_random> but I believe it's worth a PR (or an issue) because it would improve the internal docs