In fixing Trac #9390 I discovered that I *still* didn't really understand
what the can_fail and has_side_effects properties of a PrimOp mean, precisely.

The big new things I learned are

* has_side_effects needs to be true only of *write* effects,
  Reads (which are, strictly speaking, effects) don't matter here.

* has_side_effects must be true of primops that can throw a synchronous
  Haskell exception (eg raiseIO#)

* can_fail is true only of primops that can cause an *unchecked* (not
  Haskell) system exception, like divide by zero, or accessing memory
  out of range through an array read or write.

I've documented all this now.  The changes in this patch are only
in comments.