Current models for the generation of new gametophytic self-incompatibility specificities require that neutral variability segregates within specificity classes. Furthermore, one of the models predicts greater ratios of nonsynonymous to synonymous substitutions in pollen than in pistil specificity genes. All models assume that new specificities arise by mutation only. To test these models, 21 SFB (the pollen S-locus) alleles from a wild Prunus spinosa (Rosaceae) population were obtained. For seven of these, the corresponding S-haplotype was also characterized. The SFB data set was also used to identify positively selected sites. Those sites are likely to be the ones responsible for defining pollen specificities. Of the 23 sites identified as being positively selected, 21 are located in the variable (including a new region described here) and hypervariable regions. Little variability is found within specificity classes. There is no evidence for selective sweeps being more frequent in pollen than in pistil specificity genes. The S-RNase and the SFB genes have only partially correlated evolutionary histories. None of the models is compatible with the variability patterns found in the SFB and the S-haplotype data.