Currently we scan the /include/ directive as two tokens, the
"/include/" keyword itself, then the string giving the file name to
include. We use a special scanner state to keep the two linked
together, and use the scanner state stack to keep track of the
original state while we're parsing the two /include/ tokens.
This does mean that we need to enable the 'stack' option in flex,
which results in a not-easily-suppressed warning from the flex
boilerplate code. This is mildly irritating.
However, this two-token scanning of the /include/ directive also has
some extremely strange edge cases, because there are a variety of
tokens recognized in all scanner states, including INCLUDE. For
example the following strange dts file:
/include/ /dts-v1/;
/ {
/* ... */
};
Will be processed successfully with the /include/ being effectively
ignored: the '/dts-v1/' and ';' are recognized even in INCLUDE state,
then the ';' transitions us to PROPNODENAME state, throwing away
INCLUDE, and the previous state is never popped off the stack. Or
for another example this construct:
foo /include/ = "somefile.dts"
will be parsed as though it were:
foo = /include/ "somefile.dts"
Again, the '=' is scanned without leaving INCLUDE state, then the next
string triggers the include logic.
And finally, we use a different regexp for the string with the
included filename than the normal string regexpt, which is also
potentially weird.
This patch, therefore, cleans up the lexical handling of the /include/
directive. Instead of the INCLUDE state, we instead scan the whole
include directive, both keyword and filename as a single token. This
does mean a bit more complexity in extracting the filename out of
yytext, but I think it's worth it to avoid the strageness described
above. It also means it's no longer possible to put a comment between
the /include/ and the filename, but I'm really not very worried about
breaking files using such a strange construct.
I've recently worked with a FreeBSD developer, getting dtc and libfdt
working on FreeBSD. This showed up a number of portability problems
in the dtc package which this patch addresses. Changes are as
follows:
- the parent_offset and supernode_atdepth_offset testcases
used the glibc extension functions strchrnul() and strndupa(). Those
are removed, using slightly longer coding with standard C functions
instead.
- some other testcases had a #define _GNU_SOURCE for no
particular reason. This is removed.
- run_tests.sh has bash specific constructs removed, and the
interpreter changed to /bin/sh. This apparently now runs fine on
FreeBSD's /bin/sh, and I've also tested it with both ash and dash.
- convert-dtsv0-lexer.l has some extra #includes added. These
must have been included indirectly with Linux and glibc, but aren't on
FreeBSD.
- the endian handling functions in libfdt_env.h, based on
endian.h and byteswap.h are replaced with some portable open-coded
versions. Unfortunately, these result in fairly crappy code when
compiled, but as far as I can determine there doesn't seem to be any
POSIX, SUS or de facto standard way of determining endianness at
compile time, nor standard names for byteswapping functions.
- some more endian handling, from testdata.h using the
problematic endian.h is simply removed, since it wasn't actually being
used anyway.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This patch adds a new utility program, convert-dtsv0, to the dtc
sources. This program will convert dts files from v0 to v1,
preserving comments and spacing. It also includes some heuristics to
guess an appropriate base to use in the v1 output (so it will use hex
for the contents of reg properties and decimal for clock-frequency
properties, for example). They're limited and imperfect, but not
terrible.
The guts of the converter program is a modified version of the lexer
from dtc itself.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
David Gibson