I still am confused as to what you refer to as normalization. What do you mean by "expression stages"? If it is during library preparation, then its what Huang already mentioned. If its regarding gene expression, then normalization tries to correct for biases in gene expression, expressed as a measure of read counts, that are introduced due to inherent GC bias, amplification bias, poly-dT/random priming, gene length etc..

For SNP detection, normally people remove duplicate reads and then call for SNPs. I guess this is what you are calling as normalization? If so, yes, I believe that SNP calling should be done on reads after removing duplicates.