This was found when running linux crypto algorithm selftests used by
wireguard.  We found that randomly the tests would fail.  We found
through investigation that a combination of a tick timer interrupt,
raised when executing a delay slot instruction at a page boundary caused
the issue.

This was caused when handling the TB_EXIT_REQUESTED case in cpu_tb_exec.
On OpenRISC, which doesn't implement synchronize_from_tb, set_pc was
being used as a fallback.  The OpenRISC set_pc implementation clears
dflag, which caused the exception handling logic to not account for the
delay slot.  This was the bug, because it meant when execution resumed
after the interrupt was handling it resumed in the wrong place.

Fix this by implementing synchronize_from_tb which simply updates pc,
and not clear the delay slot flag.

Reported-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Stafford Horne <shorne@gmail.com>