As far as I know, the SH test in RAxML is implemented for testing different topologies based on the SAME underlying sequence data, therefore both the unconstrained and the constrained tree must be derived from the same alignment. The only difference is that you compute the constrained tree by adding a fixed topology to the analysis.The test statistics is based on the comparative likelihood values, which only works if based on the same alignment. Basically, you want to know if the same data would still allow a different topology with a similar likelhood value.

There are several alternatives to test instead for topological conflict between different data sets, but I do not know whether one is implemented in RAxML. One commonly used is Kauff and Lutzoni's compat.py. Basically, the idea is to compare supported clades between different trees and test whether this differ. [http://lutzonilab.org/downloadable-programs/]