The multiple machine support relies on redirecting all functions which will
need to be machine specific through a table of function pointers, the
machvec. These functions fall into a number of categories:

 - I/O functions to IO memory (inb etc) and PCI/main memory (readb etc).
 - I/O remapping functions (ioremap etc)
 - some initialisation functions
 - a 'heartbeat' function
 - some miscellaneous flags

The tree can be built in two ways:
 - as a fully generic build. All drivers are linked in, and all functions
   go through the machvec
 - as a machine specific build. In this case only the required drivers
   will be linked in, and some macros may be redefined to not go through
   the machvec where performance is important (in particular IO functions).

There are three ways in which IO can be performed:
 - none at all. This is really only useful for the 'unknown' machine type,
   which us designed to run on a machine about which we know nothing, and
   so all all IO instructions do nothing.
 - fully custom. In this case all IO functions go to a machine specific
   set of functions which can do what they like
 - a generic set of functions. These will cope with most situations,
   and rely on a single function, mv_port2addr, which is called through the
   machine vector, and converts an IO address into a memory address, which
   can be read from/written to directly.

Thus adding a new machine involves the following steps (I will assume I am
adding a machine called fred):

 - add a new file include/asm-sh/io_fred.h which contains prototypes for
   any machine specific IO functions prefixed with the machine name, for
   example fred_inb. These will be needed when filling out the machine
   vector. In addition, a section is required which defines what to do when
   building a machine specific version. For example:

     #ifdef __WANT_IO_DEF
     #define inb  fred_inb
     ...
     #endif

   This is the minimum that is required, however there are ample
   opportunities to optimise this. In particular, by making the prototypes
   inline function definitions, it is possible to inline the function when
   building machine specific versions. Note that the machine vector
   functions will still be needed, so that a module built for a generic
   setup can be loaded.

 - add a new file arch/sh/kernel/mach_fred.c. This contains the definition
   of the machine vector. When building the machine specific version, this
   will be the real machine vector (via an alias), while in the generic
   version is used to initialise the machine vector, and then freed, by
   making it initdata. This should be defined as:

     struct sh_machine_vector mv_fred __initmv = {
       mv_name: "Fred"
     }
     ALIAS_MV(se)

 - finally add a file arch/sh/kernel/io_fred.c, which contains
   definitions of the machine specific io functions.

A note about initialisation functions. Three initialisation functions are
provided in the machine vector:
 - mv_arch_init - called very early on from setup_arch
 - mv_init_irq - called from init_IRQ, after the generic SH interrupt
   initialisation
 - mv_init_pci - currently not used

Any other remaining functions which need to be called at start up can be
added to the list using the __initcalls macro (or module_init if the code
can be built as a module). Many generic drivers probe to see if the device
they are targeting is present, however this may not always be appropriate,
so a flag can be added to the machine vector which will be set on those
machines which have the hardware in question, reducing the probe to a
single conditional.