Interesting blog post on single precision arithmetic, compiler flags, and FastTree via Jonathan Eisen on Facebook (no, really) http://darlinglab.org/blog/2015/03/23/not-so-fast-fasttree.html
I doubt whether wide alignments of closely-related sequences should be analyzed using standard ML, as evolution on these time scales is often dominated by recombination. But in case I’m wrong, FastTree now supports very short branch lengths if compiled with -DUSE_DOUBLE. And the standard executable now issues a warning if given an alignment with many positions and very closely-related sequences. (These changes are as of version 2.1.8; if you are running FastTree on wide alignments of nearly-identical sequences then please update.)
Has anyone noticed any speed issues when bumping up precision? I know the Darling Lab blog post was actually faster, but the tree itself is not that large.