I think personally that knowing a bit of XS is very useful (especially if at some point you want to do core hacking). Still for easy prototyping Inline::C is just great; for packaging into a a module InlineX::C2XS is useful (or at least the ideas are useful to know what and where you eventually need to patch). You can also try InlineX::XS

If you want to interface for speed, I think it is important to know that going in-and-out of perl takes time, and that you probably want to minimize that. There are some tricks out there: one of the most useful was mentioned on the p5p list and is used by Imager. Still the most serious problem is using a contract of the type malloc/free, while wanting automatic (end of scope) "freeing" on the perl side; for this it is probably better to use ideas like that of the Scope::Guard module (and the documentation mentioned therein). Arena-like programming can be ok too: you "malloc" (for each new object) on the C-side and "free" only in bunchs on the perl side (in the end it is probably better to do this via special objects -- thus part of the perl API).