free_area() ended up looking for any area of memory, not necessarily a
*free* area of memory.  This had predictably disastrous consequences.

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

Make it a bit easier to write user-space test cases; use fgets+sscanf
so comment lines end up getting ignored.

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

One-to-many relationships, in which one chunk of a file is used in
more than one place, tends to naturally show up in decoding certain
fileformats, including (but not limited to) Microsoft SDI.  Make the
shuffler library handle those cases correctly, and remove a
special-purpose hack in sdi.c.

This is based on the observation that all one-to-many relationships
can be treated as a one-to-one shuffle followed by
destination-to-destination copies; i.e. one copy is (arbitrarily)
assigned the "master copy" status, and all aliases are then copied
from the master copy when the master copy is already in its final
place.  All other copies can then be simply ignored for the duration
of the shuffle, just as zero-memory is.

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

Rewrite the algorithm to prefer entries which can be directly moved
into their target slots; this should reduce the number of descriptors
in most cases (although not necessarily *all* cases.)

Try to clean up the code some while we're at it... the code is
confusing enough as it is.

Use the syslinux_memmap data structure for the free memory list.  This
means we get range coalescing; this sometimes generates lists that are
vastly shorter than without range coalescing.

Handle the case of an upward move when there is something in the way.
This happens when loading an SDI image.

Allocate high memory out of the way to hold the shuffle descriptors,
and generate continuation descriptors as needed.

The shuffle and boot interface, including the library support, should now
work as advertised.  Add an ELF-loading module as a demo, and it's probably
useful for someone, too.