At present each invocation of run_steps() updates OUTPUT_FILES_COMMON,
since it does not make a copy of the dict. This is fine for a single
invocation, but for tests, run_steps() is invoked many times.

As a result it may include unwanted items from the previous run, if it
happens that a test runs twice on the same CPU. The problem has not been
noticied previously, as there are few enough tests and enough CPUs that
is is rare for the 'wrong' combination of tests to run together.

Fix this by making a copy of the dict, before updating it. Update the
tests to suit, taking account of the files that are no-longer generated.

With this fix, we no-longer generate files which are not needed for a
particular state of OF_PLATDATA_INST, so the check_instantiate() function
is not needed anymore. It has become dead code and so fails the
code-coverage test (dtoc -T). Remove it.

Signed-off-by: Simon Glass <sjg@chromium.org>

This existing code assumes that a reg property is larger than one cell,
but this is not always the case. Fix this assumption.

Also if a node's parent is missing the #address-cells and #size-cells
properties we use 2 as a default for each. But this should not happen in
practice. More likely the properties were removed for SPL due to there
being no 'u-boot,dm-pre-reloc' property, or similar. Add a warning for
this as the failure can be very confusing.

Signed-off-by: Simon Glass <sjg@chromium.org>

At present an empty size is considered to be a 64-bit value. This does not
seem useful and wastes space. Limit the 64-bit detection to where one or
both of the addr/size is two cells or more.

Signed-off-by: Simon Glass <sjg@chromium.org>

At present warnings are shown as soon as they are discovered in the
source scannner. But the function that detects them may be called multiple
times.

Collect all the warnings and show them at the end.

Signed-off-by: Simon Glass <sjg@chromium.org>

Add a few more internal checks to make sure offsets are correct, before
updating the dtb.

To make this easier, update the functions which add a property to return
that property,.

Signed-off-by: Simon Glass <sjg@chromium.org>

So far we have only needed to add subnodes to empty notds, so have not
had to deal with ordering. However this feature is needed for binman's
expanded nodes, since there may be another node in the same section.

While libfdt adds new properties after existing properties, it adds new
subnodes before existing subnodes. This means that we must reorder the
nodes in the cached version, so that the ordering remains consistent.

Update the sync implementation to sync existing subnodes first, then
add new ones, then tidy up the ordering in the cached version. Update the
test to cover this behaviour.

Also improve the comment about property syncing while we are here.

Signed-off-by: Simon Glass <sjg@chromium.org>

Add a new test that adds a subnode alongside an existing one, as well as
adding properties to a subnode. This will expand to adding multiple
subnodes in future patches. Put a node after the one we are adding to so
we can check that things sync correctly.

The testAddNode() test should be in the TestNode class since it is a node
test, so move it.

Signed-off-by: Simon Glass <sjg@chromium.org>

Once the tree has been synced, thus potentially moving things around in the
fdt, we set _cached_offsets to False so that a refresh will happen next
time a property is accessed.

This 'lazy' refresh doesn't really save much time, since refresh is a very
fast operation, just a single walk of the tree. Also, having the refresh
happen in the bowels of property access it makes it harder to figure out
what is going on.

Simplify the code by always doing a refresh before and after a sync. Set
_cached_offsets to True immediately after this, in the Refresh() function,
since this makes more sense than doing it in the caller.

Signed-off-by: Simon Glass <sjg@chromium.org>

If a property does not yet have an offset, then that means it exists in
the cache'd fdt but has not yet been synced back to the flat tree. Use
the dirty flag for this so we don't need to check the offset too. Improve
the comments for Prop and Node to make it clear what an offset of None
means.

Also clear the dirty flag after the property is synced.

Signed-off-by: Simon Glass <sjg@chromium.org>

Add the node name too so it is easy to see which node failed.

Signed-off-by: Simon Glass <sjg@chromium.org>