Regression tests succeed or fail depending on which version of gcc was
used for compiling moses. This "fix" (it's a pretty awful hack, actually)
rounds the scores reported by Moses to one less digit of precision in the
Moses output for
tests/phrase.source-word-deletion.nbest and
tests/phrase.sparse-features
To update the "truth" file, the following was done
./round.pl truth/results.txt > foo && mv foo truth/results.txt