Currently observed means are used, and especially for ANCOVA designs one would prefer that the LS means which are corrected for the covariate(s) be used. See my case nr 01525480 for more details and examples. SAS, SPSS and R provide post hoc tests based on the LS means.