Well, rerendering the scene is quite costly, especially in terms of CPU overhead in Gamestudio, but even then the cost is usually worth it if it adds nice shadows. And just rendering the depth is a lot faster than calculating complex shading for each fragment.
From my experience VSM suffers a lot from light bleeding and there is no good solution for it, except using exponential shadow mapping which mostly fixes that issue, but only when using two 32bit floating point channels, which from my experience is still quite costly on laptop GPUs and often slower than a small filter kernel for percentage closer filtering, even without hardware filtering (with vendor specific extensions or newer DirectX versions it is possible to get 2*2 percentage closer filtering with bilinear blending for the performance of a hard shadow). Also the possibility to use a slope scaled depth bias without having to implement it in your own shaders is quite an advantage of single depth texture based approaches.

Okay, I am not sure what I wanted to say
Other than that it depends a lot on the hardware how good things work and that DirectX 9 sucks .