[2/3] gen_compile_commands: wire up build rule to Makefile
diff mbox series

Message ID 20200812173958.2307251-3-masahiroy@kernel.org
State New
Headers show
Series
  • kbuild: clang-tidy
Related show

Commit Message

Masahiro Yamada Aug. 12, 2020, 5:39 p.m. UTC
Currently, you need to explicitly run scripts/gen_compile_commands.py
to create compile_commands.json. It traverses the object tree
(you need to pass the -d option to deal with a separate output tree),
and parses all the .*.cmd file found.

If you rebuild the kernel over again without 'make clean', stale
.*.cmd files from older builds will create invalid entries in
compile_commands.json.

This commit wires up the compile_commands.json rule to the top
Makefile, and makes it parse .*.cmd files only from the current build
to avoid stale entries.

It is possible to extract only relevant .*.cmd files by checking
$(KBUILD_VMLINUX_OBJS), $(KBUILD_VMLINUX_LIBS), and modules.order.
The objects or archives linked to vmlinux are listed in
$(KBUILD_VMLINUX_OBJS) or $(KBUILD_VMLINUX_LIBS). All the modules are
listed in modules.order.

You can create compile_commands.json from Make:

  $ make -j$(nproc) CC=clang compile_commands.json

Of course, you can build vmlinux, modules, and compile_commands.json
all together in a single command:

  $ make -j$(nproc) CC=clang all compile_commands.json

It works also for M= builds. In this case, compile_commands.json
is created in the top directory of the external module.

I hope this will be overall improvements, but it has a drawback;
the coverage of the compile_commands.json is reduced because only
the objects linked to vmlinux or modules are handled. For example,
the following C files are not included in compile_commands.json:

 - Decompressor source files (arch/*/boot/compressed/)
 - VDSO source files
 - C files used to generate intermediates (e.g. kernel/bounds.c)
 - standalone host programs

Here is a note for out-of-tree builds. 'make compile_commands.json'
works with O= option, but please notice compile_commands.json is
created in the object tree instead of the source tree.

Some people may want to have compile_commands.json in the source tree
because Clang Tools searches for it through all parent paths of the
first input source file.

However, you cannot do it for O= builds. Kbuild should never generate
any build artifact in the source tree when O= is given because the
source tree might be read-only. Any write attempt to the source tree
is monitored and the violation may be reported. See the commit log of
8ef14c2c41d9.

So, the only possible way it to create compile_commands.json in the
object tree, then specify '-p <build-path>' when you use clang-check,
clang-tidy, etc.

Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
---

 Makefile                        |  29 ++++++-
 scripts/gen_compile_commands.py | 146 +++++++++++++-------------------
 2 files changed, 82 insertions(+), 93 deletions(-)

Comments

Nick Desaulniers Aug. 12, 2020, 10:30 p.m. UTC | #1
On Wed, Aug 12, 2020 at 10:40 AM Masahiro Yamada <masahiroy@kernel.org> wrote:
>
> Currently, you need to explicitly run scripts/gen_compile_commands.py
> to create compile_commands.json. It traverses the object tree
> (you need to pass the -d option to deal with a separate output tree),
> and parses all the .*.cmd file found.
>
> If you rebuild the kernel over again without 'make clean', stale
> .*.cmd files from older builds will create invalid entries in
> compile_commands.json.

Definitely a problem; happy to see compile_commands.json added to
`make clean` target, too.

>
> This commit wires up the compile_commands.json rule to the top
> Makefile, and makes it parse .*.cmd files only from the current build
> to avoid stale entries.
>
> It is possible to extract only relevant .*.cmd files by checking
> $(KBUILD_VMLINUX_OBJS), $(KBUILD_VMLINUX_LIBS), and modules.order.
> The objects or archives linked to vmlinux are listed in
> $(KBUILD_VMLINUX_OBJS) or $(KBUILD_VMLINUX_LIBS). All the modules are
> listed in modules.order.
>
> You can create compile_commands.json from Make:
>
>   $ make -j$(nproc) CC=clang compile_commands.json
>
> Of course, you can build vmlinux, modules, and compile_commands.json
> all together in a single command:
>
>   $ make -j$(nproc) CC=clang all compile_commands.json
>
> It works also for M= builds. In this case, compile_commands.json
> is created in the top directory of the external module.
>
> I hope this will be overall improvements, but it has a drawback;
> the coverage of the compile_commands.json is reduced because only
> the objects linked to vmlinux or modules are handled. For example,
> the following C files are not included in compile_commands.json:
>
>  - Decompressor source files (arch/*/boot/compressed/)
>  - VDSO source files
>  - C files used to generate intermediates (e.g. kernel/bounds.c)
>  - standalone host programs

Oof, for an x86_64 defconfig, the difference in line count of
compile_commands.json
before: 12826
after: 12351

That's a loss of 475 (3.7% of 12826) coverage. Is there something more
we can do to preserve this functionality, while avoiding stale .cmd
files?

Is it that those aren't specified by `$(KBUILD_VMLINUX_OBJS)
$(KBUILD_VMLINUX_LIBS)` ?

>
> Here is a note for out-of-tree builds. 'make compile_commands.json'
> works with O= option, but please notice compile_commands.json is
> created in the object tree instead of the source tree.
>
> Some people may want to have compile_commands.json in the source tree
> because Clang Tools searches for it through all parent paths of the
> first input source file.
>
> However, you cannot do it for O= builds. Kbuild should never generate
> any build artifact in the source tree when O= is given because the
> source tree might be read-only. Any write attempt to the source tree
> is monitored and the violation may be reported. See the commit log of
> 8ef14c2c41d9.
>
> So, the only possible way it to create compile_commands.json in the
> object tree, then specify '-p <build-path>' when you use clang-check,
> clang-tidy, etc.
>
> Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
> ---
>
>  Makefile                        |  29 ++++++-
>  scripts/gen_compile_commands.py | 146 +++++++++++++-------------------
>  2 files changed, 82 insertions(+), 93 deletions(-)
>
> diff --git a/Makefile b/Makefile
> index 6844b848bfec..4d65affb6917 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -634,7 +634,7 @@ endif
>  # in addition to whatever we do anyway.
>  # Just "make" or "make all" shall build modules as well
>
> -ifneq ($(filter all modules nsdeps,$(MAKECMDGOALS)),)
> +ifneq ($(filter all modules nsdeps %compile_commands.json,$(MAKECMDGOALS)),)
>    KBUILD_MODULES := 1
>  endif
>
> @@ -1459,7 +1459,8 @@ endif # CONFIG_MODULES
>
>  # Directories & files removed with 'make clean'
>  CLEAN_FILES += include/ksym vmlinux.symvers \
> -              modules.builtin modules.builtin.modinfo modules.nsdeps
> +              modules.builtin modules.builtin.modinfo modules.nsdeps \
> +              compile_commands.json
>
>  # Directories & files removed with 'make mrproper'
>  MRPROPER_FILES += include/config include/generated          \
> @@ -1693,9 +1694,12 @@ KBUILD_MODULES := 1
>
>  build-dirs := $(KBUILD_EXTMOD)
>  PHONY += modules
> -modules: descend
> +modules: $(MODORDER)
>         $(Q)$(MAKE) -f $(srctree)/scripts/Makefile.modpost
>
> +$(MODORDER): descend
> +       @:
> +
>  PHONY += modules_install
>  modules_install: _emodinst_ _emodinst_post
>
> @@ -1709,8 +1713,12 @@ PHONY += _emodinst_post
>  _emodinst_post: _emodinst_
>         $(call cmd,depmod)
>
> +compile_commands.json: $(extmod-prefix)compile_commands.json
> +PHONY += compile_commands.json
> +
>  clean-dirs := $(KBUILD_EXTMOD)
> -clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps
> +clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps \
> +       $(KBUILD_EXTMOD)/compile_commands.json

So the `clean` target doesn't make use of `CLEAN_FILES`? It looks like
there's some duplication there?  Oh, this is dependent on
!KBUILD_EXTMOD, and is a new `clean` target. Do I understand that
correctly?

>
>  PHONY += help
>  help:
> @@ -1823,6 +1831,19 @@ nsdeps: export KBUILD_NSDEPS=1
>  nsdeps: modules
>         $(Q)$(CONFIG_SHELL) $(srctree)/scripts/nsdeps
>
> +# Clang Tooling
> +# ---------------------------------------------------------------------------
> +
> +quiet_cmd_gen_compile_commands = GEN     $@
> +      cmd_gen_compile_commands = $(PYTHON3) $< -a $(AR) -o $@ $(filter-out $<, $(real-prereqs))
> +
> +$(extmod-prefix)compile_commands.json: scripts/gen_compile_commands.py \
> +       $(if $(KBUILD_EXTMOD),,$(KBUILD_VMLINUX_OBJS) $(KBUILD_VMLINUX_LIBS)) \
> +       $(if $(CONFIG_MODULES), $(MODORDER)) FORCE
> +       $(call if_changed,gen_compile_commands)
> +
> +targets += $(extmod-prefix)compile_commands.json
> +
>  # Scripts to check various things for consistency
>  # ---------------------------------------------------------------------------
>
> diff --git a/scripts/gen_compile_commands.py b/scripts/gen_compile_commands.py
> index 19c7338740e7..d2ff0d982521 100755
> --- a/scripts/gen_compile_commands.py
> +++ b/scripts/gen_compile_commands.py
> @@ -9,80 +9,49 @@
>
>  import argparse
>  import json
> -import logging
>  import os
>  import re
> -
> -_DEFAULT_OUTPUT = 'compile_commands.json'
> -_DEFAULT_LOG_LEVEL = 'WARNING'
> -
> -_FILENAME_PATTERN = r'^\..*\.cmd$'
> -_LINE_PATTERN = r'^cmd_[^ ]*\.o := (.* )([^ ]*\.c)$'
> -_VALID_LOG_LEVELS = ['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL']
> -
> -# A kernel build generally has over 2000 entries in its compile_commands.json
> -# database. If this code finds 300 or fewer, then warn the user that they might
> -# not have all the .cmd files, and they might need to compile the kernel.
> -_LOW_COUNT_THRESHOLD = 300
> +import subprocess
>
>
>  def parse_arguments():
>      """Sets up and parses command-line arguments.
>
>      Returns:
> -        log_level: A logging level to filter log output.
> -        directory: The directory to search for .cmd files.
> +        ar: Command used for parsing .a archives
>          output: Where to write the compile-commands JSON file.
> +        files: Files to parse
>      """
>      usage = 'Creates a compile_commands.json database from kernel .cmd files'
>      parser = argparse.ArgumentParser(description=usage)
>
> -    directory_help = ('Path to the kernel source directory to search '
> -                      '(defaults to the working directory)')
> -    parser.add_argument('-d', '--directory', type=str, help=directory_help)
> +    ar_help = 'command used for parsing .a archives'
> +    parser.add_argument('-a', '--ar', type=str, default='ar', help=ar_help)

Might be nice to warn if run with no arguments? In case someone does:
$ ./scripts/clang-tools/gen_compile_commands.py

>
> -    output_help = ('The location to write compile_commands.json (defaults to '
> -                   'compile_commands.json in the search directory)')
> -    parser.add_argument('-o', '--output', type=str, help=output_help)
> +    output_help = 'output file for the compilation database'
> +    parser.add_argument('-o', '--output', type=str,
> +                        default='compile_commands.json', help=output_help)
>
> -    log_level_help = ('The level of log messages to produce (one of ' +
> -                      ', '.join(_VALID_LOG_LEVELS) + '; defaults to ' +
> -                      _DEFAULT_LOG_LEVEL + ')')
> -    parser.add_argument(
> -        '--log_level', type=str, default=_DEFAULT_LOG_LEVEL,
> -        help=log_level_help)
> +    files_help='files to parse (should be *.o, *.a, or modules.order)'
> +    parser.add_argument('files', type=str, nargs='*', help=files_help)
>
>      args = parser.parse_args()
>
> -    log_level = args.log_level
> -    if log_level not in _VALID_LOG_LEVELS:
> -        raise ValueError('%s is not a valid log level' % log_level)
> -
> -    directory = args.directory or os.getcwd()
> -    output = args.output or os.path.join(directory, _DEFAULT_OUTPUT)
> -    directory = os.path.abspath(directory)
> -
> -    return log_level, directory, output
> +    return args.ar, args.output, args.files
>
>
> -def process_line(root_directory, file_directory, command_prefix, relative_path):
> +def process_line(root_directory, command_prefix, file_path):
>      """Extracts information from a .cmd line and creates an entry from it.
>
>      Args:
>          root_directory: The directory that was searched for .cmd files. Usually
>              used directly in the "directory" entry in compile_commands.json.
> -        file_directory: The path to the directory the .cmd file was found in.
>          command_prefix: The extracted command line, up to the last element.
> -        relative_path: The .c file from the end of the extracted command.
> -            Usually relative to root_directory, but sometimes relative to
> -            file_directory and sometimes neither.
> +        file_path: The .c file from the end of the extracted command.
> +            It can be either relative or absolute.
>
>      Returns:
>          An entry to append to compile_commands.
> -
> -    Raises:
> -        ValueError: Could not find the extracted file based on relative_path and
> -            root_directory or file_directory.
>      """
>      # The .cmd files are intended to be included directly by Make, so they
>      # escape the pound sign '#', either as '\#' or '$(pound)' (depending on the
> @@ -90,60 +59,59 @@ def process_line(root_directory, file_directory, command_prefix, relative_path):
>      # by Make, so this code replaces the escaped version with '#'.
>      prefix = command_prefix.replace('\#', '#').replace('$(pound)', '#')
>
> -    cur_dir = root_directory
> -    expected_path = os.path.join(cur_dir, relative_path)
> -    if not os.path.exists(expected_path):
> -        # Try using file_directory instead. Some of the tools have a different
> -        # style of .cmd file than the kernel.
> -        cur_dir = file_directory
> -        expected_path = os.path.join(cur_dir, relative_path)
> -        if not os.path.exists(expected_path):
> -            raise ValueError('File %s not in %s or %s' %
> -                             (relative_path, root_directory, file_directory))
>      return {
> -        'directory': cur_dir,
> -        'file': relative_path,
> -        'command': prefix + relative_path,
> +        'directory': root_directory,
> +        'file': file_path,
> +        'command': prefix + file_path,
>      }
>
>
>  def main():
> -    """Walks through the directory and finds and parses .cmd files."""
> -    log_level, directory, output = parse_arguments()
> -
> -    level = getattr(logging, log_level)
> -    logging.basicConfig(format='%(levelname)s: %(message)s', level=level)
> -
> -    filename_matcher = re.compile(_FILENAME_PATTERN)
> -    line_matcher = re.compile(_LINE_PATTERN)
> +    """Find and parse .cmd files for vmlinux and modules"""
> +    ar, output, files = parse_arguments()
> +
> +    line_matcher = re.compile(r'^cmd_[^ ]*\.o := (.* )([^ ]*\.c)$')
> +
> +    # Collect objects compiled for vmlinux or modules
> +    objects = []
> +    for file in files:
> +        if file.endswith('.o'):
> +            # Some objects (head-y) are linked to vmlinux directly
> +            objects.append(file)
> +        elif file.endswith('.a'):
> +            # Most of built-in objects are linked via built-in.a or lib.a.
> +            # Use 'ar -t' to get the list of the contained objects.
> +            objects += subprocess.check_output([ar, '-t', file]).decode().split()
> +        elif file.endswith('modules.order'):
> +           # modules.order lists all the modules.
> +            with open(file) as f:

`file` is another builtin (or at least was in Python2), perhaps `filename`?

> +                for line in f:
> +                    ko = line.rstrip()
> +                    base, ext = os.path.splitext(ko)
> +                    if ext != '.ko':
> +                        sys.exit('{}: mobule path must end with .ko'.format(ko))
> +                    mod = base + '.mod'
> +                   # The first line of *.mod lists the objects that
> +                   # compose the module.

This comment and the one above it uses tabs for indentation vs spaces
for the rest of the file.  I use
https://github.com/nickdesaulniers/dotfiles/blob/a90865a9ea48bbefa0082f7508607fdeb361e801/.vimrc#L37-L43
to help me catch these.

> +                    with open(mod) as mod_f:
> +                        objects += mod_f.readline().split()
> +        else:
> +            sys.exit('{}: unknown file type'.format(file))

Consider breaking up this one long function into multiple, perhaps the
above could just return `objects`?

>
>      compile_commands = []
> -    for dirpath, _, filenames in os.walk(directory):
> -        for filename in filenames:
> -            if not filename_matcher.match(filename):
> -                continue
> -            filepath = os.path.join(dirpath, filename)
> -
> -            with open(filepath, 'rt') as f:
> -                line = f.readline()
> -                result = line_matcher.match(line)
> -                if result:
> -                    try:
> -                        entry = process_line(directory, dirpath,
> -                                             result.group(1), result.group(2))
> -                        compile_commands.append(entry)
> -                    except ValueError as err:
> -                        logging.info('Could not add line from %s: %s',
> -                                     filepath, err)
> +    cwd = os.getcwd()
> +    for object in objects:
> +        dir, notdir = os.path.split(object)

`object` is a builtin Class in python.  I'm not sure if it's quite
considered a keyword, but maybe a different identifier would be nicer,
like `object_file` or something?

> +        cmd_file = os.path.join(dir, '.' + notdir + '.cmd')
> +        with open(cmd_file, 'rt') as f:
> +            line = f.readline()
> +            result = line_matcher.match(line)

^ combine statements.

> +            if result:
> +                entry = process_line(cwd, result.group(1), result.group(2))
> +                compile_commands.append(entry)
>
>      with open(output, 'wt') as f:
>          json.dump(compile_commands, f, indent=2, sort_keys=True)
>
> -    count = len(compile_commands)
> -    if count < _LOW_COUNT_THRESHOLD:
> -        logging.warning(
> -            'Found %s entries. Have you compiled the kernel?', count)
> -
> -
>  if __name__ == '__main__':
>      main()
> --
> 2.25.1
>

Thank you for your assistance and help enabling these tools.
Masahiro Yamada Aug. 13, 2020, 5:10 p.m. UTC | #2
On Thu, Aug 13, 2020 at 7:30 AM 'Nick Desaulniers' via Clang Built
Linux <clang-built-linux@googlegroups.com> wrote:
>
> On Wed, Aug 12, 2020 at 10:40 AM Masahiro Yamada <masahiroy@kernel.org> wrote:
> >
> > Currently, you need to explicitly run scripts/gen_compile_commands.py
> > to create compile_commands.json. It traverses the object tree
> > (you need to pass the -d option to deal with a separate output tree),
> > and parses all the .*.cmd file found.
> >
> > If you rebuild the kernel over again without 'make clean', stale
> > .*.cmd files from older builds will create invalid entries in
> > compile_commands.json.
>
> Definitely a problem; happy to see compile_commands.json added to
> `make clean` target, too.
>
> >
> > This commit wires up the compile_commands.json rule to the top
> > Makefile, and makes it parse .*.cmd files only from the current build
> > to avoid stale entries.
> >
> > It is possible to extract only relevant .*.cmd files by checking
> > $(KBUILD_VMLINUX_OBJS), $(KBUILD_VMLINUX_LIBS), and modules.order.
> > The objects or archives linked to vmlinux are listed in
> > $(KBUILD_VMLINUX_OBJS) or $(KBUILD_VMLINUX_LIBS). All the modules are
> > listed in modules.order.
> >
> > You can create compile_commands.json from Make:
> >
> >   $ make -j$(nproc) CC=clang compile_commands.json
> >
> > Of course, you can build vmlinux, modules, and compile_commands.json
> > all together in a single command:
> >
> >   $ make -j$(nproc) CC=clang all compile_commands.json
> >
> > It works also for M= builds. In this case, compile_commands.json
> > is created in the top directory of the external module.
> >
> > I hope this will be overall improvements, but it has a drawback;
> > the coverage of the compile_commands.json is reduced because only
> > the objects linked to vmlinux or modules are handled. For example,
> > the following C files are not included in compile_commands.json:
> >
> >  - Decompressor source files (arch/*/boot/compressed/)
> >  - VDSO source files
> >  - C files used to generate intermediates (e.g. kernel/bounds.c)
> >  - standalone host programs
>
> Oof, for an x86_64 defconfig, the difference in line count of
> compile_commands.json
> before: 12826


I think some lines of 'before'
are not so important.

Files suffixed with *.mod.c
are generated sources for modules.
There is no point to check them by Clang tools.


Some entries appear twice:

For example, 'before' contains two entries of
"file": "lib/cmdline.c"
Which entry is used by 'clang-tidy lib/cmdline.c',
the first one, the second one, or both?



Having said that, there is still a loss of more than 3%, yes.


> after: 12351
>
> That's a loss of 475 (3.7% of 12826) coverage. Is there something more
> we can do to preserve this functionality, while avoiding stale .cmd
> files?


I have no idea how to do this correctly.

> Is it that those aren't specified by `$(KBUILD_VMLINUX_OBJS)
> $(KBUILD_VMLINUX_LIBS)` ?

These variables contain only objects and archives
linked to vmlinux.




For example, VDSO is built as a prerequisite of
another object that wraps it.

See line 61 of arch/arm64/kernel/vdso/Makefile:
$(obj)/vdso.o : $(obj)/vdso.so


I do not know how to get the full list of active objects,
some of which are built on demand
in the dependency chain.


Idea 1)
Merge this series, and accept the loss.


Idea 2)
Add Makefile targets,
and also keep the previous work-flow.

When you run it from Make,
only objects for vmlinux and modules are handled.

When you need the full coverage, including non-kernel-space
sources, run scripts manually:

$ scripts/clang-tools/gen_compile_commands.py
$ scripts/clang-tools/run-clang-tools.py clang-tidy


Idea 3)
Give up supporting it from Makefile.
Instead, improve gen_scripts_commands.py
as a standalone program.


Maybe we can check whether the compiler is Clang or not.
We can run '<compiler> --version' and drop the
entry if it is GCC.

Usually, the compiler is the first word of
the "command" field in compile_commands.json,
but there are exceptions because
people may do CC="ccache clang".


If there are still stale entries causing troubles,
you need to run 'make clean', and rebuild the tree.


We were trying to have separate scripts,
gen_compile_commands.py and run-clang-tools.py,
and to add Makefile targets to run them in a row.

I think unifying the two scripts
might be handier.


Add two options, -t, -a,
to scripts/gen_compile_commands.py

If they are given,
scripts/gen_compilile_commands.py
generates compile_commands.json,
and immediately runs clang-tidy against it.


-t, --tidy
   Run 'clang-tidy -checks=-*,linuxkernel-*' after generating
compilation database
-a, --analyzer
   Run 'clang-tidy -checks=-*,clang-analyzer-*' after generating
compilation database


Both -a and -t are given,
it runs
'clang-tidy -checks=-*,linuxkernel-*,clang-analyzer-*'

This works more efficiently
if you want to check everything.


'make clang-tidy clang-analyzer'
will invoke clang-tidy twice for each file,
which is not very efficient.




> >  clean-dirs := $(KBUILD_EXTMOD)
> > -clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps
> > +clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps \
> > +       $(KBUILD_EXTMOD)/compile_commands.json
>
> So the `clean` target doesn't make use of `CLEAN_FILES`? It looks like
> there's some duplication there?  Oh, this is dependent on
> !KBUILD_EXTMOD, and is a new `clean` target. Do I understand that
> correctly?

Correct.

We can move CLEAN_FILES to a common part
so external module builds can use it.

> >      """
> >      usage = 'Creates a compile_commands.json database from kernel .cmd files'
> >      parser = argparse.ArgumentParser(description=usage)
> >
> > -    directory_help = ('Path to the kernel source directory to search '
> > -                      '(defaults to the working directory)')
> > -    parser.add_argument('-d', '--directory', type=str, help=directory_help)
> > +    ar_help = 'command used for parsing .a archives'
> > +    parser.add_argument('-a', '--ar', type=str, default='ar', help=ar_help)
>
> Might be nice to warn if run with no arguments? In case someone does:
> $ ./scripts/clang-tools/gen_compile_commands.py

Right.

nargs='+' seems to work.



> > +    # Collect objects compiled for vmlinux or modules
> > +    objects = []
> > +    for file in files:
> > +        if file.endswith('.o'):
> > +            # Some objects (head-y) are linked to vmlinux directly
> > +            objects.append(file)
> > +        elif file.endswith('.a'):
> > +            # Most of built-in objects are linked via built-in.a or lib.a.
> > +            # Use 'ar -t' to get the list of the contained objects.
> > +            objects += subprocess.check_output([ar, '-t', file]).decode().split()
> > +        elif file.endswith('modules.order'):
> > +           # modules.order lists all the modules.
> > +            with open(file) as f:
>
> `file` is another builtin (or at least was in Python2), perhaps `filename`?
>
> > +                for line in f:
> > +                    ko = line.rstrip()
> > +                    base, ext = os.path.splitext(ko)
> > +                    if ext != '.ko':
> > +                        sys.exit('{}: mobule path must end with .ko'.format(ko))
> > +                    mod = base + '.mod'
> > +                   # The first line of *.mod lists the objects that
> > +                   # compose the module.
>
> This comment and the one above it uses tabs for indentation vs spaces
> for the rest of the file.  I use
> https://github.com/nickdesaulniers/dotfiles/blob/a90865a9ea48bbefa0082f7508607fdeb361e801/.vimrc#L37-L43
> to help me catch these.

Awesome. Copied to mine.



> > +                    with open(mod) as mod_f:
> > +                        objects += mod_f.readline().split()
> > +        else:
> > +            sys.exit('{}: unknown file type'.format(file))
>
> Consider breaking up this one long function into multiple, perhaps the
> above could just return `objects`?



I thought that returning a big list causes needless memory-copy.
If we do not need to be worried too much,
I can make it a helper function.


>
> >
> >      compile_commands = []
> > -    for dirpath, _, filenames in os.walk(directory):
> > -        for filename in filenames:
> > -            if not filename_matcher.match(filename):
> > -                continue
> > -            filepath = os.path.join(dirpath, filename)
> > -
> > -            with open(filepath, 'rt') as f:
> > -                line = f.readline()
> > -                result = line_matcher.match(line)
> > -                if result:
> > -                    try:
> > -                        entry = process_line(directory, dirpath,
> > -                                             result.group(1), result.group(2))
> > -                        compile_commands.append(entry)
> > -                    except ValueError as err:
> > -                        logging.info('Could not add line from %s: %s',
> > -                                     filepath, err)
> > +    cwd = os.getcwd()
> > +    for object in objects:
> > +        dir, notdir = os.path.split(object)
>
> `object` is a builtin Class in python.  I'm not sure if it's quite
> considered a keyword, but maybe a different identifier would be nicer,
> like `object_file` or something?


Not a keyword, but 'object' is a class, yes.
Not sure about 'file'.


$ python
Python 3.8.2 (default, Jul 16 2020, 14:00:26)
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import keyword
>>> keyword.iskeyword("import")
True
>>> keyword.iskeyword("if")
True
>>> keyword.iskeyword("file")
False
>>> keyword.iskeyword("object")
False
>>> object
<class 'object'>
>>> file
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
NameError: name 'file' is not defined
>>>


If this is a problem, I can rename it.




> > +        cmd_file = os.path.join(dir, '.' + notdir + '.cmd')
> > +        with open(cmd_file, 'rt') as f:
> > +            line = f.readline()
> > +            result = line_matcher.match(line)
>
> ^ combine statements.

OK.


> > +            if result:
> > +                entry = process_line(cwd, result.group(1), result.group(2))
> > +                compile_commands.append(entry)
> >
> >      with open(output, 'wt') as f:
> >          json.dump(compile_commands, f, indent=2, sort_keys=True)
> >
> > -    count = len(compile_commands)
> > -    if count < _LOW_COUNT_THRESHOLD:
> > -        logging.warning(
> > -            'Found %s entries. Have you compiled the kernel?', count)
> > -
> > -
> >  if __name__ == '__main__':
> >      main()
> > --
> > 2.25.1
> >
>
> Thank you for your assistance and help enabling these tools.
>
> --
> Thanks,
> ~Nick Desaulniers
>
> --
> You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to clang-built-linux+unsubscribe@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/CAKwvOdkL%3D667%2Bcw_Rxq_5zaOKeTTptsMaxkkSXBic9QxozOWVg%40mail.gmail.com.



--
Best Regards
Masahiro Yamada
Masahiro Yamada Aug. 19, 2020, 4:29 a.m. UTC | #3
Hi Nick,

On Fri, Aug 14, 2020 at 2:10 AM Masahiro Yamada <masahiroy@kernel.org> wrote:
>
> On Thu, Aug 13, 2020 at 7:30 AM 'Nick Desaulniers' via Clang Built
> Linux <clang-built-linux@googlegroups.com> wrote:
> >
> > On Wed, Aug 12, 2020 at 10:40 AM Masahiro Yamada <masahiroy@kernel.org> wrote:
> > >
> > > Currently, you need to explicitly run scripts/gen_compile_commands.py
> > > to create compile_commands.json. It traverses the object tree
> > > (you need to pass the -d option to deal with a separate output tree),
> > > and parses all the .*.cmd file found.
> > >
> > > If you rebuild the kernel over again without 'make clean', stale
> > > .*.cmd files from older builds will create invalid entries in
> > > compile_commands.json.
> >
> > Definitely a problem; happy to see compile_commands.json added to
> > `make clean` target, too.
> >
> > >
> > > This commit wires up the compile_commands.json rule to the top
> > > Makefile, and makes it parse .*.cmd files only from the current build
> > > to avoid stale entries.
> > >
> > > It is possible to extract only relevant .*.cmd files by checking
> > > $(KBUILD_VMLINUX_OBJS), $(KBUILD_VMLINUX_LIBS), and modules.order.
> > > The objects or archives linked to vmlinux are listed in
> > > $(KBUILD_VMLINUX_OBJS) or $(KBUILD_VMLINUX_LIBS). All the modules are
> > > listed in modules.order.
> > >
> > > You can create compile_commands.json from Make:
> > >
> > >   $ make -j$(nproc) CC=clang compile_commands.json
> > >
> > > Of course, you can build vmlinux, modules, and compile_commands.json
> > > all together in a single command:
> > >
> > >   $ make -j$(nproc) CC=clang all compile_commands.json
> > >
> > > It works also for M= builds. In this case, compile_commands.json
> > > is created in the top directory of the external module.
> > >
> > > I hope this will be overall improvements, but it has a drawback;
> > > the coverage of the compile_commands.json is reduced because only
> > > the objects linked to vmlinux or modules are handled. For example,
> > > the following C files are not included in compile_commands.json:
> > >
> > >  - Decompressor source files (arch/*/boot/compressed/)
> > >  - VDSO source files
> > >  - C files used to generate intermediates (e.g. kernel/bounds.c)
> > >  - standalone host programs
> >
> > Oof, for an x86_64 defconfig, the difference in line count of
> > compile_commands.json
> > before: 12826
>
>
> I think some lines of 'before'
> are not so important.
>
> Files suffixed with *.mod.c
> are generated sources for modules.
> There is no point to check them by Clang tools.
>
>
> Some entries appear twice:
>
> For example, 'before' contains two entries of
> "file": "lib/cmdline.c"
> Which entry is used by 'clang-tidy lib/cmdline.c',
> the first one, the second one, or both?
>
>
>
> Having said that, there is still a loss of more than 3%, yes.
>
>
> > after: 12351
> >
> > That's a loss of 475 (3.7% of 12826) coverage. Is there something more
> > we can do to preserve this functionality, while avoiding stale .cmd
> > files?
>
>
> I have no idea how to do this correctly.
>
> > Is it that those aren't specified by `$(KBUILD_VMLINUX_OBJS)
> > $(KBUILD_VMLINUX_LIBS)` ?
>
> These variables contain only objects and archives
> linked to vmlinux.
>
>
>
>
> For example, VDSO is built as a prerequisite of
> another object that wraps it.
>
> See line 61 of arch/arm64/kernel/vdso/Makefile:
> $(obj)/vdso.o : $(obj)/vdso.so
>
>
> I do not know how to get the full list of active objects,
> some of which are built on demand
> in the dependency chain.
>
>
> Idea 1)
> Merge this series, and accept the loss.
>
>
> Idea 2)
> Add Makefile targets,
> and also keep the previous work-flow.
>
> When you run it from Make,
> only objects for vmlinux and modules are handled.
>
> When you need the full coverage, including non-kernel-space
> sources, run scripts manually:
>
> $ scripts/clang-tools/gen_compile_commands.py
> $ scripts/clang-tools/run-clang-tools.py clang-tidy



Do you have any idea to cope with
the 3% loss problem?

If it is a problem, maybe I should try Idea 2).


Thanks.








>
> Idea 3)
> Give up supporting it from Makefile.
> Instead, improve gen_scripts_commands.py
> as a standalone program.
>
>
> Maybe we can check whether the compiler is Clang or not.
> We can run '<compiler> --version' and drop the
> entry if it is GCC.
>
> Usually, the compiler is the first word of
> the "command" field in compile_commands.json,
> but there are exceptions because
> people may do CC="ccache clang".
>
>
> If there are still stale entries causing troubles,
> you need to run 'make clean', and rebuild the tree.
>
>
> We were trying to have separate scripts,
> gen_compile_commands.py and run-clang-tools.py,
> and to add Makefile targets to run them in a row.
>
> I think unifying the two scripts
> might be handier.
>
>
> Add two options, -t, -a,
> to scripts/gen_compile_commands.py
>
> If they are given,
> scripts/gen_compilile_commands.py
> generates compile_commands.json,
> and immediately runs clang-tidy against it.
>
>
> -t, --tidy
>    Run 'clang-tidy -checks=-*,linuxkernel-*' after generating
> compilation database
> -a, --analyzer
>    Run 'clang-tidy -checks=-*,clang-analyzer-*' after generating
> compilation database
>
>
> Both -a and -t are given,
> it runs
> 'clang-tidy -checks=-*,linuxkernel-*,clang-analyzer-*'
>
> This works more efficiently
> if you want to check everything.
>
>
> 'make clang-tidy clang-analyzer'
> will invoke clang-tidy twice for each file,
> which is not very efficient.
>
>
>
>
> > >  clean-dirs := $(KBUILD_EXTMOD)
> > > -clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps
> > > +clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps \
> > > +       $(KBUILD_EXTMOD)/compile_commands.json
> >
> > So the `clean` target doesn't make use of `CLEAN_FILES`? It looks like
> > there's some duplication there?  Oh, this is dependent on
> > !KBUILD_EXTMOD, and is a new `clean` target. Do I understand that
> > correctly?
>
> Correct.
>
> We can move CLEAN_FILES to a common part
> so external module builds can use it.
>
> > >      """
> > >      usage = 'Creates a compile_commands.json database from kernel .cmd files'
> > >      parser = argparse.ArgumentParser(description=usage)
> > >
> > > -    directory_help = ('Path to the kernel source directory to search '
> > > -                      '(defaults to the working directory)')
> > > -    parser.add_argument('-d', '--directory', type=str, help=directory_help)
> > > +    ar_help = 'command used for parsing .a archives'
> > > +    parser.add_argument('-a', '--ar', type=str, default='ar', help=ar_help)
> >
> > Might be nice to warn if run with no arguments? In case someone does:
> > $ ./scripts/clang-tools/gen_compile_commands.py
>
> Right.
>
> nargs='+' seems to work.
>
>
>
> > > +    # Collect objects compiled for vmlinux or modules
> > > +    objects = []
> > > +    for file in files:
> > > +        if file.endswith('.o'):
> > > +            # Some objects (head-y) are linked to vmlinux directly
> > > +            objects.append(file)
> > > +        elif file.endswith('.a'):
> > > +            # Most of built-in objects are linked via built-in.a or lib.a.
> > > +            # Use 'ar -t' to get the list of the contained objects.
> > > +            objects += subprocess.check_output([ar, '-t', file]).decode().split()
> > > +        elif file.endswith('modules.order'):
> > > +           # modules.order lists all the modules.
> > > +            with open(file) as f:
> >
> > `file` is another builtin (or at least was in Python2), perhaps `filename`?
> >
> > > +                for line in f:
> > > +                    ko = line.rstrip()
> > > +                    base, ext = os.path.splitext(ko)
> > > +                    if ext != '.ko':
> > > +                        sys.exit('{}: mobule path must end with .ko'.format(ko))
> > > +                    mod = base + '.mod'
> > > +                   # The first line of *.mod lists the objects that
> > > +                   # compose the module.
> >
> > This comment and the one above it uses tabs for indentation vs spaces
> > for the rest of the file.  I use
> > https://github.com/nickdesaulniers/dotfiles/blob/a90865a9ea48bbefa0082f7508607fdeb361e801/.vimrc#L37-L43
> > to help me catch these.
>
> Awesome. Copied to mine.
>
>
>
> > > +                    with open(mod) as mod_f:
> > > +                        objects += mod_f.readline().split()
> > > +        else:
> > > +            sys.exit('{}: unknown file type'.format(file))
> >
> > Consider breaking up this one long function into multiple, perhaps the
> > above could just return `objects`?
>
>
>
> I thought that returning a big list causes needless memory-copy.
> If we do not need to be worried too much,
> I can make it a helper function.
>
>
> >
> > >
> > >      compile_commands = []
> > > -    for dirpath, _, filenames in os.walk(directory):
> > > -        for filename in filenames:
> > > -            if not filename_matcher.match(filename):
> > > -                continue
> > > -            filepath = os.path.join(dirpath, filename)
> > > -
> > > -            with open(filepath, 'rt') as f:
> > > -                line = f.readline()
> > > -                result = line_matcher.match(line)
> > > -                if result:
> > > -                    try:
> > > -                        entry = process_line(directory, dirpath,
> > > -                                             result.group(1), result.group(2))
> > > -                        compile_commands.append(entry)
> > > -                    except ValueError as err:
> > > -                        logging.info('Could not add line from %s: %s',
> > > -                                     filepath, err)
> > > +    cwd = os.getcwd()
> > > +    for object in objects:
> > > +        dir, notdir = os.path.split(object)
> >
> > `object` is a builtin Class in python.  I'm not sure if it's quite
> > considered a keyword, but maybe a different identifier would be nicer,
> > like `object_file` or something?
>
>
> Not a keyword, but 'object' is a class, yes.
> Not sure about 'file'.
>
>
> $ python
> Python 3.8.2 (default, Jul 16 2020, 14:00:26)
> [GCC 9.3.0] on linux
> Type "help", "copyright", "credits" or "license" for more information.
> >>> import keyword
> >>> keyword.iskeyword("import")
> True
> >>> keyword.iskeyword("if")
> True
> >>> keyword.iskeyword("file")
> False
> >>> keyword.iskeyword("object")
> False
> >>> object
> <class 'object'>
> >>> file
> Traceback (most recent call last):
>   File "<stdin>", line 1, in <module>
> NameError: name 'file' is not defined
> >>>
>
>
> If this is a problem, I can rename it.
>
>
>
>
> > > +        cmd_file = os.path.join(dir, '.' + notdir + '.cmd')
> > > +        with open(cmd_file, 'rt') as f:
> > > +            line = f.readline()
> > > +            result = line_matcher.match(line)
> >
> > ^ combine statements.
>
> OK.
>
>
> > > +            if result:
> > > +                entry = process_line(cwd, result.group(1), result.group(2))
> > > +                compile_commands.append(entry)
> > >
> > >      with open(output, 'wt') as f:
> > >          json.dump(compile_commands, f, indent=2, sort_keys=True)
> > >
> > > -    count = len(compile_commands)
> > > -    if count < _LOW_COUNT_THRESHOLD:
> > > -        logging.warning(
> > > -            'Found %s entries. Have you compiled the kernel?', count)
> > > -
> > > -
> > >  if __name__ == '__main__':
> > >      main()
> > > --
> > > 2.25.1
> > >
> >
> > Thank you for your assistance and help enabling these tools.
> >
> > --
> > Thanks,
> > ~Nick Desaulniers
> >
> > --
> > You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
> > To unsubscribe from this group and stop receiving emails from it, send an email to clang-built-linux+unsubscribe@googlegroups.com.
> > To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/CAKwvOdkL%3D667%2Bcw_Rxq_5zaOKeTTptsMaxkkSXBic9QxozOWVg%40mail.gmail.com.
>
>
>
> --
> Best Regards
> Masahiro Yamada



--
Best Regards
Masahiro Yamada
Nick Desaulniers Aug. 20, 2020, 2:29 a.m. UTC | #4
On Tue, Aug 18, 2020 at 9:30 PM Masahiro Yamada <masahiroy@kernel.org> wrote:
>
> Do you have any idea to cope with
> the 3% loss problem?
>
> If it is a problem, maybe I should try Idea 2).

I think it will be ok.
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>

Patch
diff mbox series

diff --git a/Makefile b/Makefile
index 6844b848bfec..4d65affb6917 100644
--- a/Makefile
+++ b/Makefile
@@ -634,7 +634,7 @@  endif
 # in addition to whatever we do anyway.
 # Just "make" or "make all" shall build modules as well
 
-ifneq ($(filter all modules nsdeps,$(MAKECMDGOALS)),)
+ifneq ($(filter all modules nsdeps %compile_commands.json,$(MAKECMDGOALS)),)
   KBUILD_MODULES := 1
 endif
 
@@ -1459,7 +1459,8 @@  endif # CONFIG_MODULES
 
 # Directories & files removed with 'make clean'
 CLEAN_FILES += include/ksym vmlinux.symvers \
-	       modules.builtin modules.builtin.modinfo modules.nsdeps
+	       modules.builtin modules.builtin.modinfo modules.nsdeps \
+	       compile_commands.json
 
 # Directories & files removed with 'make mrproper'
 MRPROPER_FILES += include/config include/generated          \
@@ -1693,9 +1694,12 @@  KBUILD_MODULES := 1
 
 build-dirs := $(KBUILD_EXTMOD)
 PHONY += modules
-modules: descend
+modules: $(MODORDER)
 	$(Q)$(MAKE) -f $(srctree)/scripts/Makefile.modpost
 
+$(MODORDER): descend
+	@:
+
 PHONY += modules_install
 modules_install: _emodinst_ _emodinst_post
 
@@ -1709,8 +1713,12 @@  PHONY += _emodinst_post
 _emodinst_post: _emodinst_
 	$(call cmd,depmod)
 
+compile_commands.json: $(extmod-prefix)compile_commands.json
+PHONY += compile_commands.json
+
 clean-dirs := $(KBUILD_EXTMOD)
-clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps
+clean: rm-files := $(KBUILD_EXTMOD)/Module.symvers $(KBUILD_EXTMOD)/modules.nsdeps \
+	$(KBUILD_EXTMOD)/compile_commands.json
 
 PHONY += help
 help:
@@ -1823,6 +1831,19 @@  nsdeps: export KBUILD_NSDEPS=1
 nsdeps: modules
 	$(Q)$(CONFIG_SHELL) $(srctree)/scripts/nsdeps
 
+# Clang Tooling
+# ---------------------------------------------------------------------------
+
+quiet_cmd_gen_compile_commands = GEN     $@
+      cmd_gen_compile_commands = $(PYTHON3) $< -a $(AR) -o $@ $(filter-out $<, $(real-prereqs))
+
+$(extmod-prefix)compile_commands.json: scripts/gen_compile_commands.py \
+	$(if $(KBUILD_EXTMOD),,$(KBUILD_VMLINUX_OBJS) $(KBUILD_VMLINUX_LIBS)) \
+	$(if $(CONFIG_MODULES), $(MODORDER)) FORCE
+	$(call if_changed,gen_compile_commands)
+
+targets += $(extmod-prefix)compile_commands.json
+
 # Scripts to check various things for consistency
 # ---------------------------------------------------------------------------
 
diff --git a/scripts/gen_compile_commands.py b/scripts/gen_compile_commands.py
index 19c7338740e7..d2ff0d982521 100755
--- a/scripts/gen_compile_commands.py
+++ b/scripts/gen_compile_commands.py
@@ -9,80 +9,49 @@ 
 
 import argparse
 import json
-import logging
 import os
 import re
-
-_DEFAULT_OUTPUT = 'compile_commands.json'
-_DEFAULT_LOG_LEVEL = 'WARNING'
-
-_FILENAME_PATTERN = r'^\..*\.cmd$'
-_LINE_PATTERN = r'^cmd_[^ ]*\.o := (.* )([^ ]*\.c)$'
-_VALID_LOG_LEVELS = ['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL']
-
-# A kernel build generally has over 2000 entries in its compile_commands.json
-# database. If this code finds 300 or fewer, then warn the user that they might
-# not have all the .cmd files, and they might need to compile the kernel.
-_LOW_COUNT_THRESHOLD = 300
+import subprocess
 
 
 def parse_arguments():
     """Sets up and parses command-line arguments.
 
     Returns:
-        log_level: A logging level to filter log output.
-        directory: The directory to search for .cmd files.
+        ar: Command used for parsing .a archives
         output: Where to write the compile-commands JSON file.
+        files: Files to parse
     """
     usage = 'Creates a compile_commands.json database from kernel .cmd files'
     parser = argparse.ArgumentParser(description=usage)
 
-    directory_help = ('Path to the kernel source directory to search '
-                      '(defaults to the working directory)')
-    parser.add_argument('-d', '--directory', type=str, help=directory_help)
+    ar_help = 'command used for parsing .a archives'
+    parser.add_argument('-a', '--ar', type=str, default='ar', help=ar_help)
 
-    output_help = ('The location to write compile_commands.json (defaults to '
-                   'compile_commands.json in the search directory)')
-    parser.add_argument('-o', '--output', type=str, help=output_help)
+    output_help = 'output file for the compilation database'
+    parser.add_argument('-o', '--output', type=str,
+                        default='compile_commands.json', help=output_help)
 
-    log_level_help = ('The level of log messages to produce (one of ' +
-                      ', '.join(_VALID_LOG_LEVELS) + '; defaults to ' +
-                      _DEFAULT_LOG_LEVEL + ')')
-    parser.add_argument(
-        '--log_level', type=str, default=_DEFAULT_LOG_LEVEL,
-        help=log_level_help)
+    files_help='files to parse (should be *.o, *.a, or modules.order)'
+    parser.add_argument('files', type=str, nargs='*', help=files_help)
 
     args = parser.parse_args()
 
-    log_level = args.log_level
-    if log_level not in _VALID_LOG_LEVELS:
-        raise ValueError('%s is not a valid log level' % log_level)
-
-    directory = args.directory or os.getcwd()
-    output = args.output or os.path.join(directory, _DEFAULT_OUTPUT)
-    directory = os.path.abspath(directory)
-
-    return log_level, directory, output
+    return args.ar, args.output, args.files
 
 
-def process_line(root_directory, file_directory, command_prefix, relative_path):
+def process_line(root_directory, command_prefix, file_path):
     """Extracts information from a .cmd line and creates an entry from it.
 
     Args:
         root_directory: The directory that was searched for .cmd files. Usually
             used directly in the "directory" entry in compile_commands.json.
-        file_directory: The path to the directory the .cmd file was found in.
         command_prefix: The extracted command line, up to the last element.
-        relative_path: The .c file from the end of the extracted command.
-            Usually relative to root_directory, but sometimes relative to
-            file_directory and sometimes neither.
+        file_path: The .c file from the end of the extracted command.
+            It can be either relative or absolute.
 
     Returns:
         An entry to append to compile_commands.
-
-    Raises:
-        ValueError: Could not find the extracted file based on relative_path and
-            root_directory or file_directory.
     """
     # The .cmd files are intended to be included directly by Make, so they
     # escape the pound sign '#', either as '\#' or '$(pound)' (depending on the
@@ -90,60 +59,59 @@  def process_line(root_directory, file_directory, command_prefix, relative_path):
     # by Make, so this code replaces the escaped version with '#'.
     prefix = command_prefix.replace('\#', '#').replace('$(pound)', '#')
 
-    cur_dir = root_directory
-    expected_path = os.path.join(cur_dir, relative_path)
-    if not os.path.exists(expected_path):
-        # Try using file_directory instead. Some of the tools have a different
-        # style of .cmd file than the kernel.
-        cur_dir = file_directory
-        expected_path = os.path.join(cur_dir, relative_path)
-        if not os.path.exists(expected_path):
-            raise ValueError('File %s not in %s or %s' %
-                             (relative_path, root_directory, file_directory))
     return {
-        'directory': cur_dir,
-        'file': relative_path,
-        'command': prefix + relative_path,
+        'directory': root_directory,
+        'file': file_path,
+        'command': prefix + file_path,
     }
 
 
 def main():
-    """Walks through the directory and finds and parses .cmd files."""
-    log_level, directory, output = parse_arguments()
-
-    level = getattr(logging, log_level)
-    logging.basicConfig(format='%(levelname)s: %(message)s', level=level)
-
-    filename_matcher = re.compile(_FILENAME_PATTERN)
-    line_matcher = re.compile(_LINE_PATTERN)
+    """Find and parse .cmd files for vmlinux and modules"""
+    ar, output, files = parse_arguments()
+
+    line_matcher = re.compile(r'^cmd_[^ ]*\.o := (.* )([^ ]*\.c)$')
+
+    # Collect objects compiled for vmlinux or modules
+    objects = []
+    for file in files:
+        if file.endswith('.o'):
+            # Some objects (head-y) are linked to vmlinux directly
+            objects.append(file)
+        elif file.endswith('.a'):
+            # Most of built-in objects are linked via built-in.a or lib.a.
+            # Use 'ar -t' to get the list of the contained objects.
+            objects += subprocess.check_output([ar, '-t', file]).decode().split()
+        elif file.endswith('modules.order'):
+	    # modules.order lists all the modules.
+            with open(file) as f:
+                for line in f:
+                    ko = line.rstrip()
+                    base, ext = os.path.splitext(ko)
+                    if ext != '.ko':
+                        sys.exit('{}: mobule path must end with .ko'.format(ko))
+                    mod = base + '.mod'
+		    # The first line of *.mod lists the objects that
+		    # compose the module.
+                    with open(mod) as mod_f:
+                        objects += mod_f.readline().split()
+        else:
+            sys.exit('{}: unknown file type'.format(file))
 
     compile_commands = []
-    for dirpath, _, filenames in os.walk(directory):
-        for filename in filenames:
-            if not filename_matcher.match(filename):
-                continue
-            filepath = os.path.join(dirpath, filename)
-
-            with open(filepath, 'rt') as f:
-                line = f.readline()
-                result = line_matcher.match(line)
-                if result:
-                    try:
-                        entry = process_line(directory, dirpath,
-                                             result.group(1), result.group(2))
-                        compile_commands.append(entry)
-                    except ValueError as err:
-                        logging.info('Could not add line from %s: %s',
-                                     filepath, err)
+    cwd = os.getcwd()
+    for object in objects:
+        dir, notdir = os.path.split(object)
+        cmd_file = os.path.join(dir, '.' + notdir + '.cmd')
+        with open(cmd_file, 'rt') as f:
+            line = f.readline()
+            result = line_matcher.match(line)
+            if result:
+                entry = process_line(cwd, result.group(1), result.group(2))
+                compile_commands.append(entry)
 
     with open(output, 'wt') as f:
         json.dump(compile_commands, f, indent=2, sort_keys=True)
 
-    count = len(compile_commands)
-    if count < _LOW_COUNT_THRESHOLD:
-        logging.warning(
-            'Found %s entries. Have you compiled the kernel?', count)
-
-
 if __name__ == '__main__':
     main()