1.-Do not use Tophat2. Nowadays you have many much better options, just look a the bechmarks.
2.- STAR it is pretty good, I would totally recommend it. However the issue with STAR is the high memory requirement. If you are working with human and you have less than 28 GB of RAM memory, you should use HISAT2 instead. Otherwise, both aligners programs should perform very similar.
I would say it depends on what you want to do with your data.
I've sometimes found that the TopHat alignments work better than STAR alignments with some splicing analysis programs, possibly due to the format of the alignment.
I would consider the run-time for TopHat to be sufficiently quick that you could run comparisons and see what works best with your data (while the benchmark papers can be a useful starting point, the optimal strategy is not necessarily the same for every dataset). So, if the combination of latest aligner and downstream algorithm gives results that don't make sense, it may be a good idea to try other aligners / algorithms.
There are also options for gene expression quantification without the alignment step (Salmon, kallisto, Sailfish, etc.). If you have a two-group comparison with triplicates and clear expression differences, then that should work fine. However, I've found the accuracy for gene assignments for a given sample may be less accurate, and having replicates can give you some sense in the robustness of the read / expression assignments for transcripts (or the sum of transcripts for genes).