The messing around with partial pointers for rllpack/rllunpack turned
out to be a source of bugs.  Instead, have all the values be 32 bits
wide, and require the callers to pass them accordingly.

Signed-off-by: H. Peter Anvin <hpa@zytor.com>

Clear the flags register while we still have a stack to do so.
This makes the state in the target OS more predictable.  It's worth
noting that none of the instructions we use in either the PM or RM
trampolines change the flags, either (except the IF setting in the RM
trampoline, but that is intentional.)

Signed-off-by: H. Peter Anvin <hpa@zytor.com>

Align the relocated bcopyxx code to a 16-byte boundary.  There are
CPUs which have errata relating to GDTs which are not 16-byte aligned.

Signed-off-by: H. Peter Anvin <hpa@zytor.com>

Since searchdir() uses getfssec() to read the directory it needs to set
file_bytesleft. Else it loops forever if you have a large directory (and/or
rock ridge on).

Fix minor bugs reported by pfranz73@tiscali.it.

Signed-off-by: H. Peter Anvin <hpa@zytor.com>

Rather than simply incrementing A20Test, advance it by a large prime
number every time.  The reason is to minimize false aliasing as
quickly as possible.

Signed-off-by: H. Peter Anvin <hpa@zytor.com>

We haven't enabled DO_WBINVD for a very long time.  A20 is implemented
on the inside of the L1 cache for 486+ (and 386 didn't have WBINVD),
so an I/O delay is the only thing we can rely on anyway.

Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>

The io_delay operation is only two instructions, might as well inline
it.

Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>

A20Test is now a dword; this makes it less likely that we'll have a
case of false aliasing resulting in the io_delay path being taken.

Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>