Lazy evaluation has many advantages, but it can cause bad
performance. Consequently, Haskell allows users to
force eager evaluation at certain program points by inserting
strictness annotations, known and written as bangs (!).
Unfortunately, manual bang placement
is difficult. Autobahn 1.0
uses a genetic algorithm to infer bang annotations
that improve performance. However, Autobahn 1.0 often generates
large numbers of superfluous bangs,
which is problematic because users must inspect each such bang to
determine whether it is safe.
We introduce Autobahn 2.0, which uses GHC
profiling information to reduce the number of superfluous bangs.
When evaluated on the NoFib benchmark suite,
Autobahn 2.0 reduced the number of inferred bangs by 90.2% on average,
while only degrading program performance by 15.7% compared with the
performance produced by Autobahn 1.0.
In a case study on a garbage collection simulator,
Autobahn 2.0 eliminated 81.8% of the recommended bangs, with the
same 15.7% optimization degradation.