If you're writing a linear algebra library, then the bulk of executing code should be floating point vector operations in tight loops with known bounds. A NULL pointer check in the prolog where you set up the loop is going to be negligible.
This is very different from kernel code, which by its nature isn't very computational but doing resource management all day long and pretty much all it does is stuff that looks like walking linked lists.
That was a very specific linear algebra, we toyed around with huge but sparse matrices. Not anything graphics-related, rather number theory-related. But there was a lot of integer comparisons in there.
This was around 2001-2002 anyway. Things might have changed.
This is very different from kernel code, which by its nature isn't very computational but doing resource management all day long and pretty much all it does is stuff that looks like walking linked lists.