Userland, including the C library and the dynamic linker, is keen to use
the SYNC instruction, even for "generic" MIPS I binaries these days.
Which makes it less than useful on MIPS I processors.

This change adds the emulation, but as our do_ri() infrastructure was not
really prepared to take yet another instruction, I have rewritten it and
its callees slightly as follows.

Now there is only a single place a possible signal is thrown from.  The
place is at the end of do_ri().  The instruction word is fetched in
do_ri() and passed down to handlers.  The handlers are called in sequence
and return a result that lets the caller decide upon further processing.
If the result is positive, then the handler has picked the instruction,
but a signal should be thrown and the result is the signal number.  If the
result is zero, then the handler has successfully simulated the
instruction.  If the result is negative, then the handler did not handle
the instruction; to make it more obvious the calls do not follow the usual
0/-Exxx result convention they now return -1 instead of -EFAULT.

The calculation of the return EPC is now at the beginning.  The reason is
it is easier to handle it there as emulation callees may modify a register
and an instruction may be located in delay slot of a branch whose result
depends on the register.  It has to be undone if a signal is to be raised,
but it is not a problem as this is the slow-path case, and both actions
are done in single places now rather than the former being scattered
through emulation handlers.

The part of do_cpu() being covered follows the changes to do_ri().
Signed-off-by: Maciej W. Rozycki <macro@linux-mips.org>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>

---