	The 2.9 bsd distribution long int math uses the floating point
processor for machines with an FP processor.  Of course, the nonfpcrt
directory still contained code to do it with integer arithmetic.
	There were two problems.  "ldiv", and friends would produce
a floating exception (signal SIGFPE) when faced with a divide by zero.
"lmul" and "almul" would return zero if there was a floating overflow.
rand(3) expected to see the LS 32 bits of the result.
	It is undesirable to change the integer arithmetic behaviour,
except in speed.  It is doubly undesirable to have a non-fp machine
behave differantly than a machine with fp.
	When csv/cret is avioded (in "lmul"), the integer arithmetic
becomes faster than the floating point code, except on 70's, where
they are the same speed.
	Thus, "ldiv", and friends still use floating point, if available,
with a divide by zero check.  This is faster.  "lmul", "almul" use
integer math.
