Tested using sma - various configurations & compilation times for the
default set of fst's.
The tests were run quite some time ago (early 2017 or late 2016), but the relative speed diffs should still be valid.
./configure
time make -j
real 0m23.681s
user 0m53.518s
sys 0m3.117s
Using Hfst to compile twolc rules.
./configure --with-foma --without-xfst
time make -j
real 1m20.850s
user 2m40.157s
sys 0m8.248s
NB! This version fails massively in the tests (using make check), and can
not be used. This configuration is presently blocked (foma can only be used when
an .xfscript file is used for the morphophonology.
./configure --with-hfst --with-backend-format=foma --without-xfst
time make -j
real 1m31.903s
user 3m20.402s
sys 0m11.716s
./configure --with-hfst --with-backend-format=sfst --without-xfst
time make -j
real 1m43.750s
user 3m45.265s
sys 0m10.695s
./configure --with-hfst --without-xfst
time make -j
real 3m30.414s
user 9m23.566s
sys 0m15.767s
Except for the plain Foma configurations, all configurations give the same
results when running make check. That is, the fastest Hfst compilation
we get is the one using Hfst with the Foma backend.
There are further optimisations that can be done when compiling with Hfst, but the selection of backend is the most important one.