Deep learning (DL)-based attenuation correction (AC) for dedicated myocardial perfusion (MP) SPECT has been proposed recently, based on attenuation map (µ-map) generation or direct AC. This study aims to provide a direct comparison of the effectiveness of these two AC methods using simulation. We used a population of 100 XCAT phantoms modelling various body and organ sizes, 99mTc-sestamibi distributions, defect sizes and locations. An analytical projector of a LEHR collimator with attenuation, scatter and collimator-detector response modeling was used to simulate 64 noisy projections for 180° based on a standard clinical count level. Projections were then reconstructed by OS-EM method with and without AC (NAC) using 12 iterations and 6 subsets. A 3D conditional generative adversarial network was implemented and optimized for the two DL-based AC methods respectively with training based on: (i) NAC SPECT paired with the corresponding µ-map. The projections were then reconstructed with the DL-generated µ-map for AC (DL-ACµ); (ii) NAC SPECT paired with the corresponding AC SPECT to perform direct AC (DL-AC). We randomly used 70, 10 and 20 phantoms for training, validation and testing respectively. The relative defect size difference (RSD) on polar maps, normalized mean square error (NMSE) and structural similarity index measure (SSIM) on a 3D cardiac VOI (36×36×36) were compared for DL-ACµ and DL-AC, using AC as the gold standard. The NMSE and SSIM were significantly lower for DL-ACµ as compared to DL-AC (p<0.0001). The RSD is also significantly lower for DL-ACµ (p<0.05). We conclude that DL-ACµ is superior to DL-AC for MP SPECT.