deheader [-h ] [-m command] [-b builddir] [-i pattern] [-q ] [-r ] [-v ] [-x pattern] [-V ] [file-or-dir]
This tool takes a list of C or C++ sourcefiles and generates a report on which #includes can be omitted from them; also, what standard inclusions may be required for portability. The test, for each foo.c or foo.cc or foo.cpp, is simply whether "rm foo.o; make foo.o" returns a zero status (but the build command may be overridden).
Exception: Under cmake, foo.o is a phony target. Therefore, when a "CMakeList.txt" is detected, "make clean" is done rather than "rm foo.o".
Optionally, with the -r switch, the
unneeded headers are removed from the sourcefiles. Don't use
this option unless you have your sourcefiles safely under version
control and can revert!
If a sourcefile argument is a directory, the report is generated on all source files beneath it. Subdirectories beginning with a dot are assumed to be repository directories for version-control systems and ignored. If no arguments are given, the program runs as if the name of the current directory had been passed to it.
Inclusions within the scope of #if/#ifdef/#else/#endif directives are left alone, because trying to reason about potential combinations of -D and U options would be too complicated and prone to weird errors. One exception: headers protected only by S_SPLINT_S, the conditional for blocking scanning by the static analysis tool splint(1), are scanned normally.
The tool will also emit warnings about duplicate inclusions, and inclusions required for portability but not present.
It is recommended that you arrange to compile with options that will stop the compiler on warnings when using this tool; otherwise it will report headers that only declare prototypes and return types (and thus throw only warnings) as being not required. Under gcc the compiler options to accomplish this are -Werror -Wfatal-errors. If your makefile follows normal conventions, running with -m "make CFLAGS='-Werror -Wfatal-errors' %s" may do the right thing; you can check this by running with -v -v -v to see what compilation commands are actually emitted.
On each test compile, the original sourcefile is moved to a name
with an .orig suffix and restored on interrupt or after processing
with its original timestamp, unless the -r option was
given and headers removed.
If the -b option is given, it tells the program that generated .o files live in a file tree parallel to the source tree but rooted at the specified argument. If the argument is a relative path, it is interpreted relative to the directory in which deheader is run.
If the first test compilation from the top-level directory fails, deheader descends into the subdirectory of the source file and retries compiling inside there.
At verbosity level 0, only messages indicating removable headers are issued. At verbosity 1, test compilations are timed and progress indicated with a twirling-baton prompt. At verbosity level 2, you get verbose progress messages on the analysis. At verbosity level 3, you see the output from the make and compilation commands.
If the -q (--quiet) option flag was not set, the last line of the output will be a statistical summary.
Running deheader will leave a lot of binaries in your directory that were compiled in ways possibly not invoked by your normal build process. Running "make clean" afterwards (or the equivalent under whatever build system you are using) is strongly recommended.
Display some help and exit.
Set the build command used for test compiles. Defaults to 'make %s'. A %s in the build command is replaced with the make target. Targetless builders such as meson can be supported with, e.g. "-m 'ninja -C build'"
Set the build directory for object files.
Set a pattern for includes to be ignored. Takes a Python regular expression.
Suppress statistical summary.
Remove header inclusions from sourcefiles where they are not required.
Set verbosity.
Exclude files with names matching the specified Python regexp.
Show version of program and exit.
Returns 1 if unneeded includes were found, 0 otherwise. Thus, you can use it for pre-release sanity checking in Makefile.
Very rarely, test-compiling after running with
-r may show that this tool removed some headers that
are actually required for your build. This can happen because
deheader doesn't know about all the strange things
your build system gets up to, and the problem of analyzing your build
to understand them would be Turing-complete. Simply revert the
altered files and continue.
Due to minor variations in system headers, it is possible your program may not port correctly in older Unix variants after being deheadered. This is normally not a problem with the portion of the API specified by POSIX and ANSI C, but may be for headers that are not standardized or only weakly standardized. The sockets API (sys/select.h, sys/sockets.h, and friends such as sys/types.h and sys.stat.h) is perhaps the most serious trouble spot. deheader has an internal table of rules that heads off the most common problems by suppressing deletion of headers that are required for portability, but your mileage may vary.
The dependency scanner does not ignore the text of comments. This, e.g, a reference to "log10" in a comment will produce a spurious warning that <math.h> is required for portability.
Sufficiently perverse C++ can silently invalidate the brute-force algorithm this tool uses. Example: if an overloaded function has different overloads from two different files, removing one may expose the other, changing runtime semantics without a compile-time warning. Similarly, removing a later file containing a template specialization may lead to undefined behavior from a template defined in an earlier file. Use this with caution near such features, and test carefully.