🚀 go-pugleaf

All Sparc implementations do not have any hardware support for integer
divide and limited support for integer multiply in the form of the MULS
multiply step instruction. There should never be any support beyond this,
as this partitioning is part of the Sparc ARCHITECTURE. For multiply, the
support is really sufficient, as most multiplications can be catched at 
compile time (read a paper about HP Precision Architecture and their measure-
ments) and the speed of a software routine for the other cases with
support from the MULS instruction is quite good. Quoted from the Cypress
CY7C601 Sparc Architecture manual:
 
  ... Code for general mul subroutine ...
 
  This code has an optimization built in for short (less than 13-bit)
  multiplies. Short multiplies require 26 or 27 instruction cycles, long
  ones require 47 to 51 instruction cycles. For two positive numbers (the
  most common case), the cycle count is 47.
 
They go on and claim 25 and 46 to 48 cycles respectively for unsigned
multiplication. The integer divide code is much harder to understand and
the only give an estimate for the most common case, worst case being up to
4 times slower (under strange circumstances). The formula is
 
	CEIL(log2(dividend/divisor)/3) x ( 21.5 )  + some setup cycles
 
which translates into something from 30 to 260 cycles  if I understand
the formula correctly (which I probably don't).
 
	Burkhard Neidecker-Lutz, Digital CEC Karlsruhe, Project NESTOR
	neideck@nestvx.dec.com
 
PS: How comes that John Mashey always answers my questions in my "notes read
    delay slot" so that I'm looking plain stupid when my question appears
    AFTER his answer... :-)