re-use D(fake) for optimizing both, G and D when training GANs

Question

When training GANs, I can do this:

pseudo code

opt_g = Optimizer(G.params)
opt_d = Optimizer(D.params)

fake1 = G(z1)
l = loss(D(fake1)
l.backward()
opt_g.step()

fake2 = G(z2).detach()
l = loss(-D(fake2)) + loss(D(real))
l.backward()
opt_d.step()

However, I am wondering, if I can reuse G(z1) or even D(G(z1))

fake = G(z1)
fake_pred = D(fake)
l = loss(fake_pred)
l.backward()
opt_g.step()

# variant 1 (shared fake)
l = D(fake.detach()) + loss(D(real))

# variant 2 (shared prediction)
l = loss(-fake_pred) + loss(D(real))
l.backward()


opt_d.step()

If this was possible, I wonder why it was implemented differently in StyleMapGAN?