I finally got around to investigating why the Simplifier was sometimes
iterating so often.  There's a nice example in Text.ParserCombinators.ReadPrec,
which produced:

NOTE: Simplifier still going after 3 iterations; bailing out.  Size = 339
NOTE: Simplifier still going after 3 iterations; bailing out.  Size = 339
NOTE: Simplifier still going after 3 iterations; bailing out.  Size = 339

No progress is being made.  It turned out that an interaction between
eta-expansion, casts, and eta reduction was responsible. The change is
small and simple, in SimplUtils.mkLam: do not require the body to be
a Lam when floating the cast outwards.  

I also discovered a missing side condition in the same equation, so fixing
that is good too.  Now there is no loop when compiling ReadPrec.

Should do a full nofib run though.