Thanks for the review and the fixes.  I found a couple more things. 
First, it's not portable to cast wint_t * to wchar_t *, since the 
pointed-to types might be different sizes or representations. Second, we 
can put the cache directly in the struct dfa, saving the overhead of 
doing a separate malloc.

The attached further patch should address these problems.  I pushed 
this, along with the earlier two patches in this sequence, and am 
marking this as done.