Added perf.py a wrapper around Linux's perf tool for perf sampling

This provides 2 things:

1. perf integration with the bench/test runners - This is a bit tricky
   with perf as it doesn't have its own way to combine perf measurements
   across multiple processes. perf.py works around this by writing
   everything to a zip file, using flock to synchronize. As a plus, free
   compression!

2. Parsing and presentation of perf results in a format consistent with
   the other CSV-based tools. This actually ran into a surprising number of
   issues:

   - We need to process raw events to get the information we want, this
     ends up being a lot of data (~16MiB at 100Hz uncompressed), so we
     paralellize the parsing of each decompressed perf file.

   - perf reports raw addresses post-ASLR. It does provide sym+off which
     is very useful, but to find the source of static functions we need to
     reverse the ASLR by finding the delta the produces the best
     symbol<->addr matches.

   - This isn't related to perf, but decoding dwarf line-numbers is
     really complicated. You basically need to write a tiny VM.

This also turns on perf measurement by default for the bench-runner, but at a
low frequency (100 Hz). This can be decreased or removed in the future
if it causes any slowdown.

13 files changed, 1972 insertions, 225 deletions

diff --git a/scripts/bench.py b/scripts/bench.py
index e401d7c..61db83e 100755
--- a/scripts/bench.py
+++ b/scripts/bench.py
@@ -27,9 +27,13 @@ import time
 import toml
 
 
-RUNNER_PATH = 'runners/bench_runner'
+RUNNER_PATH = './runners/bench_runner'
 HEADER_PATH = 'runners/bench_runner.h'
 
+GDB_TOOL = ['gdb']
+VALGRIND_TOOL = ['valgrind']
+PERF_SCRIPT = ['./scripts/perf.py']
+
 
 def openio(path, mode='r', buffering=-1, nb=False):
     if path == '-':
@@ -502,12 +506,25 @@ def find_runner(runner, **args):
 
     # run under valgrind?
     if args.get('valgrind'):
-        cmd[:0] = filter(None, [
-            'valgrind',
+        cmd[:0] = args['valgrind_tool'] + [
             '--leak-check=full',
             '--track-origins=yes',
             '--error-exitcode=4',
-            '-q'])
+            '-q']
+
+    # run under perf?
+    if args.get('perf'):
+        cmd[:0] = args['perf_script'] + list(filter(None, [
+            '-R',
+            '--perf-freq=%s' % args['perf_freq']
+                if args.get('perf_freq') else None,
+            '--perf-period=%s' % args['perf_period']
+                if args.get('perf_period') else None,
+            '--perf-events=%s' % args['perf_events']
+                if args.get('perf_events') else None,
+            '--perf-tool=%s' % args['perf_tool']
+                if args.get('perf_tool') else None,
+            '-o%s' % args['perf']]))
 
     # other context
     if args.get('geometry'):
@@ -789,9 +806,9 @@ def run_stage(name, runner_, ids, output_, **args):
                 try:
                     line = mpty.readline()
                 except OSError as e:
-                    if e.errno == errno.EIO:
-                        break
-                    raise
+                    if e.errno != errno.EIO:
+                        raise
+                    break
                 if not line:
                     break
                 last_stdout.append(line)
@@ -1126,24 +1143,24 @@ def run(runner, bench_ids=[], **args):
         cmd = runner_ + [failure.id]
 
         if args.get('gdb_main'):
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'break main',
                 '-ex', 'run',
                 '--args']
         elif args.get('gdb_case'):
             path, lineno = find_path(runner_, failure.id, **args)
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'break %s:%d' % (path, lineno),
                 '-ex', 'run',
                 '--args']
         elif failure.assert_ is not None:
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'run',
                 '-ex', 'frame function raise',
                 '-ex', 'up 2',
                 '--args']
         else:
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'run',
                 '--args']
 
@@ -1187,6 +1204,7 @@ if __name__ == "__main__":
     argparse._ArgumentGroup._handle_conflict_ignore = lambda *_: None
     parser = argparse.ArgumentParser(
         description="Build and run benches.",
+        allow_abbrev=False,
         conflict_handler='ignore')
     parser.add_argument(
         '-v', '--verbose',
@@ -1316,6 +1334,11 @@ if __name__ == "__main__":
         help="Drop into gdb on bench failure but stop at the beginning "
             "of main.")
     bench_parser.add_argument(
+        '--gdb-tool',
+        type=lambda x: x.split(),
+        default=GDB_TOOL,
+        help="Path to gdb tool to use. Defaults to %r." % GDB_TOOL)
+    bench_parser.add_argument(
         '--exec',
         type=lambda e: e.split(),
         help="Run under another executable.")
@@ -1324,6 +1347,37 @@ if __name__ == "__main__":
         action='store_true',
         help="Run under Valgrind to find memory errors. Implicitly sets "
             "--isolate.")
+    bench_parser.add_argument(
+        '--valgrind-tool',
+        type=lambda x: x.split(),
+        default=VALGRIND_TOOL,
+        help="Path to Valgrind tool to use. Defaults to %r." % VALGRIND_TOOL)
+    bench_parser.add_argument(
+        '--perf',
+        help="Run under Linux's perf to sample performance counters, writing "
+            "samples to this file.")
+    bench_parser.add_argument(
+        '--perf-freq',
+        help="perf sampling frequency. This is passed directly to the perf "
+            "script.")
+    bench_parser.add_argument(
+        '--perf-period',
+        help="perf sampling period. This is passed directly to the perf "
+            "script.")
+    bench_parser.add_argument(
+        '--perf-events',
+        help="perf events to record. This is passed directly to the perf "
+            "script.")
+    bench_parser.add_argument(
+        '--perf-script',
+        type=lambda x: x.split(),
+        default=PERF_SCRIPT,
+        help="Path to the perf script to use. Defaults to %r." % PERF_SCRIPT)
+    bench_parser.add_argument(
+        '--perf-tool',
+        type=lambda x: x.split(),
+        help="Path to the perf tool to use. This is passed directly to the "
+            "perf script")
 
     # compilation flags
     comp_parser = parser.add_argument_group('compilation options')
@@ -1348,7 +1402,7 @@ if __name__ == "__main__":
         '-o', '--output',
         help="Output file.")
 
-    # runner + bench_ids overlaps bench_paths, so we need to do some munging here
+    # runner/bench_paths overlap, so need to do some munging here
     args = parser.parse_intermixed_args()
     args.bench_paths = [' '.join(args.runner or [])] + args.bench_ids
     args.runner = args.runner or [RUNNER_PATH]
diff --git a/scripts/code.py b/scripts/code.py
index 6b373fc..7e7e960 100755
--- a/scripts/code.py
+++ b/scripts/code.py
@@ -5,7 +5,7 @@
 # by Linux's Bloat-O-Meter.
 #
 # Example:
-# ./scripts/code.py lfs.o lfs_util.o -S
+# ./scripts/code.py lfs.o lfs_util.o -Ssize
 #
 # Copyright (c) 2022, The littlefs authors.
 # Copyright (c) 2020, Arm Limited. All rights reserved.
@@ -14,6 +14,7 @@
 
 import collections as co
 import csv
+import difflib
 import glob
 import itertools as it
 import math as m
@@ -25,7 +26,8 @@ import subprocess as sp
 
 OBJ_PATHS = ['*.o']
 NM_TOOL = ['nm']
-TYPE = 'tTrRdD'
+NM_TYPES = 'tTrRdD'
+OBJDUMP_TOOL = ['objdump']
 
 
 # integer fields
@@ -135,21 +137,32 @@ def openio(path, mode='r'):
 
 def collect(paths, *,
         nm_tool=NM_TOOL,
-        type=TYPE,
-        build_dir=None,
+        nm_types=NM_TYPES,
+        objdump_tool=OBJDUMP_TOOL,
+        sources=None,
         everything=False,
         **args):
-    results = []
-    pattern = re.compile(
+    size_pattern = re.compile(
         '^(?P<size>[0-9a-fA-F]+)' +
-        ' (?P<type>[%s])' % re.escape(type) +
+        ' (?P<type>[%s])' % re.escape(nm_types) +
         ' (?P<func>.+?)$')
+    line_pattern = re.compile(
+        '^\s+(?P<no>[0-9]+)\s+'
+            '(?:(?P<dir>[0-9]+)\s+)?'
+            '.*\s+'
+            '(?P<path>[^\s]+)$')
+    info_pattern = re.compile(
+        '^(?:.*(?P<tag>DW_TAG_[a-z_]+).*'
+            '|^.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
+            '|^.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*)$')
+
+    results = []
     for path in paths:
-        # map to source file
-        src_path = re.sub('\.o$', '.c', path)
-        if build_dir:
-            src_path = re.sub('%s/*' % re.escape(build_dir), '',
-                src_path)
+        # guess the source, if we have debug-info we'll replace this later
+        file = re.sub('(\.o)?$', '.c', path, 1)
+
+        # find symbol sizes
+        results_ = []
         # note nm-tool may contain extra args
         cmd = nm_tool + ['--size-sort', path]
         if args.get('verbose'):
@@ -158,21 +171,18 @@ def collect(paths, *,
             stdout=sp.PIPE,
             stderr=sp.PIPE if not args.get('verbose') else None,
             universal_newlines=True,
-            errors='replace')
+            errors='replace',
+            close_fds=False)
         for line in proc.stdout:
-            m = pattern.match(line)
+            m = size_pattern.match(line)
             if m:
                 func = m.group('func')
                 # discard internal functions
                 if not everything and func.startswith('__'):
                     continue
-                # discard .8449 suffixes created by optimizer
-                func = re.sub('\.[0-9]+', '', func)
-
-                results.append(CodeResult(
-                    src_path, func,
+                results_.append(CodeResult(
+                    file, func,
                     int(m.group('size'), 16)))
-
         proc.wait()
         if proc.returncode != 0:
             if not args.get('verbose'):
@@ -180,6 +190,121 @@ def collect(paths, *,
                     sys.stdout.write(line)
             sys.exit(-1)
 
+
+        # try to figure out the source file if we have debug-info
+        dirs = {}
+        files = {}
+        # note objdump-tool may contain extra args
+        cmd = objdump_tool + ['--dwarf=rawline', path]
+        if args.get('verbose'):
+            print(' '.join(shlex.quote(c) for c in cmd))
+        proc = sp.Popen(cmd,
+            stdout=sp.PIPE,
+            stderr=sp.PIPE if not args.get('verbose') else None,
+            universal_newlines=True,
+            errors='replace',
+            close_fds=False)
+        for line in proc.stdout:
+            # note that files contain references to dirs, which we
+            # dereference as soon as we see them as each file table follows a
+            # dir table
+            m = line_pattern.match(line)
+            if m:
+                if not m.group('dir'):
+                    # found a directory entry
+                    dirs[int(m.group('no'))] = m.group('path')
+                else:
+                    # found a file entry
+                    dir = int(m.group('dir'))
+                    if dir in dirs:
+                        files[int(m.group('no'))] = os.path.join(
+                            dirs[dir],
+                            m.group('path'))
+                    else:
+                        files[int(m.group('no'))] = m.group('path')
+        proc.wait()
+        if proc.returncode != 0:
+            if not args.get('verbose'):
+                for line in proc.stderr:
+                    sys.stdout.write(line)
+            # do nothing on error, we don't need objdump to work, source files
+            # may just be inaccurate
+            pass
+
+        defs = {}
+        is_func = False
+        f_name = None
+        f_file = None
+        # note objdump-tool may contain extra args
+        cmd = objdump_tool + ['--dwarf=info', path]
+        if args.get('verbose'):
+            print(' '.join(shlex.quote(c) for c in cmd))
+        proc = sp.Popen(cmd,
+            stdout=sp.PIPE,
+            stderr=sp.PIPE if not args.get('verbose') else None,
+            universal_newlines=True,
+            errors='replace',
+            close_fds=False)
+        for line in proc.stdout:
+            # state machine here to find definitions
+            m = info_pattern.match(line)
+            if m:
+                if m.group('tag'):
+                    if is_func:
+                        defs[f_name] = files.get(f_file, '?')
+                    is_func = (m.group('tag') == 'DW_TAG_subprogram')
+                elif m.group('name'):
+                    f_name = m.group('name')
+                elif m.group('file'):
+                    f_file = int(m.group('file'))
+        proc.wait()
+        if proc.returncode != 0:
+            if not args.get('verbose'):
+                for line in proc.stderr:
+                    sys.stdout.write(line)
+            # do nothing on error, we don't need objdump to work, source files
+            # may just be inaccurate
+            pass
+
+        for r in results_:
+            # find best matching debug symbol, this may be slightly different
+            # due to optimizations
+            if defs:
+                # exact match? avoid difflib if we can for speed
+                if r.function in defs:
+                    file = defs[r.function]
+                else:
+                    _, file = max(
+                        defs.items(),
+                        key=lambda d: difflib.SequenceMatcher(None,
+                            d[0],
+                            r.function, False).ratio())
+            else:
+                file = r.file
+
+            # ignore filtered sources
+            if sources is not None:
+                if not any(
+                        os.path.abspath(file) == os.path.abspath(s)
+                        for s in sources):
+                    continue
+            else:
+                # default to only cwd
+                if not everything and not os.path.commonpath([
+                        os.getcwd(),
+                        os.path.abspath(file)]) == os.getcwd():
+                    continue
+
+            # simplify path
+            if os.path.commonpath([
+                    os.getcwd(),
+                    os.path.abspath(file)]) == os.getcwd():
+                file = os.path.relpath(file)
+            else:
+                file = os.path.abspath(file)
+
+            results.append(CodeResult(file, r.function, r.size))
+
     return results
 
 
@@ -437,7 +562,7 @@ def main(obj_paths, *,
                 paths.append(path)
 
         if not paths:
-            print("error: no .obj files found in %r?" % obj_paths)
+            print("error: no .o files found in %r?" % obj_paths)
             sys.exit(-1)
 
         results = collect(paths, **args)
@@ -469,13 +594,16 @@ def main(obj_paths, *,
     # write results to CSV
     if args.get('output'):
         with openio(args['output'], 'w') as f:
-            writer = csv.DictWriter(f, CodeResult._by
+            writer = csv.DictWriter(f,
+                (by if by is not None else CodeResult._by)
                 + ['code_'+k for k in CodeResult._fields])
             writer.writeheader()
             for r in results:
                 writer.writerow(
-                    {k: getattr(r, k) for k in CodeResult._by}
-                    | {'code_'+k: getattr(r, k) for k in CodeResult._fields})
+                    {k: getattr(r, k)
+                        for k in (by if by is not None else CodeResult._by)}
+                    | {'code_'+k: getattr(r, k)
+                        for k in CodeResult._fields})
 
     # find previous results?
     if args.get('diff'):
@@ -512,7 +640,8 @@ if __name__ == "__main__":
     import argparse
     import sys
     parser = argparse.ArgumentParser(
-        description="Find code size at the function level.")
+        description="Find code size at the function level.",
+        allow_abbrev=False)
     parser.add_argument(
         'obj_paths',
         nargs='*',
@@ -579,23 +708,30 @@ if __name__ == "__main__":
         action='store_true',
         help="Only show the total.")
     parser.add_argument(
-        '-A', '--everything',
+        '-F', '--source',
+        dest='sources',
+        action='append',
+        help="Only consider definitions in this file. Defaults to anything "
+            "in the current directory.")
+    parser.add_argument(
+        '--everything',
         action='store_true',
         help="Include builtin and libc specific symbols.")
     parser.add_argument(
-        '--type',
-        default=TYPE,
+        '--nm-types',
+        default=NM_TYPES,
         help="Type of symbols to report, this uses the same single-character "
-            "type-names emitted by nm. Defaults to %r." % TYPE)
+            "type-names emitted by nm. Defaults to %r." % NM_TYPES)
     parser.add_argument(
         '--nm-tool',
         type=lambda x: x.split(),
         default=NM_TOOL,
         help="Path to the nm tool to use. Defaults to %r." % NM_TOOL)
     parser.add_argument(
-        '--build-dir',
-        help="Specify the relative build directory. Used to map object files "
-            "to the correct source files.")
+        '--objdump-tool',
+        type=lambda x: x.split(),
+        default=OBJDUMP_TOOL,
+        help="Path to the objdump tool to use. Defaults to %r." % OBJDUMP_TOOL)
     sys.exit(main(**{k: v
         for k, v in vars(parser.parse_intermixed_args()).items()
         if v is not None}))
diff --git a/scripts/coverage.py b/scripts/coverage.py
index 7d36a47..f74dda8 100755
--- a/scripts/coverage.py
+++ b/scripts/coverage.py
@@ -3,7 +3,9 @@
 # Script to find coverage info after running tests.
 #
 # Example:
-# ./scripts/coverage.py lfs.t.a.gcda lfs_util.t.a.gcda -s
+# ./scripts/coverage.py \
+#     lfs.t.a.gcda lfs_util.t.a.gcda \
+#     -Flfs.c -Flfs_util.c -slines
 #
 # Copyright (c) 2022, The littlefs authors.
 # Copyright (c) 2020, Arm Limited. All rights reserved.
@@ -209,19 +211,13 @@ def openio(path, mode='r'):
     else:
         return open(path, mode)
 
-def collect(paths, *,
+def collect(gcda_paths, *,
         gcov_tool=GCOV_TOOL,
-        build_dir=None,
+        sources=None,
         everything=False,
         **args):
     results = []
-    for path in paths:
-        # map to source file
-        src_path = re.sub('\.t\.a\.gcda$', '.c', path)
-        if build_dir:
-            src_path = re.sub('%s/*' % re.escape(build_dir), '',
-                src_path)
-
+    for path in gcda_paths:
         # get coverage info through gcov's json output
         # note, gcov-tool may contain extra args
         cmd = GCOV_TOOL + ['-b', '-t', '--json-format', path]
@@ -231,7 +227,8 @@ def collect(paths, *,
             stdout=sp.PIPE,
             stderr=sp.PIPE if not args.get('verbose') else None,
             universal_newlines=True,
-            errors='replace')
+            errors='replace',
+            close_fds=False)
         data = json.load(proc.stdout)
         proc.wait()
         if proc.returncode != 0:
@@ -242,12 +239,30 @@ def collect(paths, *,
 
         # collect line/branch coverage
         for file in data['files']:
-            if file['file'] != src_path:
-                continue
+            # ignore filtered sources
+            if sources is not None:
+                if not any(
+                        os.path.abspath(file['file']) == os.path.abspath(s)
+                        for s in sources):
+                    continue
+            else:
+                # default to only cwd
+                if not everything and not os.path.commonpath([
+                        os.getcwd(),
+                        os.path.abspath(file['file'])]) == os.getcwd():
+                    continue
+
+            # simplify path
+            if os.path.commonpath([
+                    os.getcwd(),
+                    os.path.abspath(file['file'])]) == os.getcwd():
+                file_name = os.path.relpath(file['file'])
+            else:
+                file_name = os.path.abspath(file['file'])
 
             for func in file['functions']:
                 func_name = func.get('name', '(inlined)')
-                # discard internal function (this includes injected test cases)
+                # discard internal functions (this includes injected test cases)
                 if not everything:
                     if func_name.startswith('__'):
                         continue
@@ -255,7 +270,7 @@ def collect(paths, *,
                 # go ahead and add functions, later folding will merge this if
                 # there are other hits on this line
                 results.append(CoverageResult(
-                    src_path, func_name, func['start_line'],
+                    file_name, func_name, func['start_line'],
                     func['execution_count'], 0,
                     Frac(1 if func['execution_count'] > 0 else 0, 1),
                     0,
@@ -271,7 +286,7 @@ def collect(paths, *,
                 # go ahead and add lines, later folding will merge this if
                 # there are other hits on this line
                 results.append(CoverageResult(
-                    src_path, func_name, line['line_number'],
+                    file_name, func_name, line['line_number'],
                     0, line['count'],
                     0,
                     Frac(1 if line['count'] > 0 else 0, 1),
@@ -519,31 +534,25 @@ def table(Result, results, diff_results=None, *,
             line[-1]))
 
 
-def annotate(Result, results, paths, *,
+def annotate(Result, results, *,
         annotate=False,
         lines=False,
         branches=False,
-        build_dir=None,
         **args):
     # if neither branches/lines specified, color both
     if annotate and not lines and not branches:
         lines, branches = True, True
 
-    for path in paths:
-        # map to source file
-        src_path = re.sub('\.t\.a\.gcda$', '.c', path)
-        if build_dir:
-            src_path = re.sub('%s/*' % re.escape(build_dir), '',
-                src_path)
-
+    for path in co.OrderedDict.fromkeys(r.file for r in results).keys():
         # flatten to line info
         results = fold(Result, results, by=['file', 'line'])
-        table = {r.line: r for r in results if r.file == src_path}
+        table = {r.line: r for r in results if r.file == path}
 
         # calculate spans to show
         if not annotate:
             spans = []
             last = None
+            func = None
             for line, r in sorted(table.items()):
                 if ((lines and int(r.hits) == 0)
                         or (branches and r.branches.a < r.branches.b)):
@@ -553,27 +562,29 @@ def annotate(Result, results, paths, *,
                             line+1+args['context'])
                     else:
                         if last is not None:
-                            spans.append(last)
+                            spans.append((last, func))
                         last = range(
                             line-args['context'],
                             line+1+args['context'])
+                        func = r.function
             if last is not None:
-                spans.append(last)
+                spans.append((last, func))
 
-        with open(src_path) as f:
+        with open(path) as f:
             skipped = False
             for i, line in enumerate(f):
                 # skip lines not in spans?
-                if not annotate and not any(i+1 in s for s in spans):
+                if not annotate and not any(i+1 in s for s, _ in spans):
                     skipped = True
                     continue
 
                 if skipped:
                     skipped = False
-                    print('%s@@ %s:%d @@%s' % (
+                    print('%s@@ %s:%d: %s @@%s' % (
                         '\x1b[36m' if args['color'] else '',
-                        src_path,
+                        path,
                         i+1,
+                        next(iter(f for _, f in spans)),
                         '\x1b[m' if args['color'] else ''))
 
                 # build line
@@ -659,12 +670,14 @@ def main(gcda_paths, *,
     # write results to CSV
     if args.get('output'):
         with openio(args['output'], 'w') as f:
-            writer = csv.DictWriter(f, CoverageResult._by
+            writer = csv.DictWriter(f,
+                (by if by is not None else CoverageResult._by)
                 + ['coverage_'+k for k in CoverageResult._fields])
             writer.writeheader()
             for r in results:
                 writer.writerow(
-                    {k: getattr(r, k) for k in CoverageResult._by}
+                    {k: getattr(r, k)
+                        for k in (by if by is not None else CoverageResult._by)}
                     | {'coverage_'+k: getattr(r, k)
                         for k in CoverageResult._fields})
 
@@ -698,8 +711,7 @@ def main(gcda_paths, *,
                 or args.get('lines')
                 or args.get('branches')):
             # annotate sources
-            annotate(CoverageResult, results, paths,
-                **args)
+            annotate(CoverageResult, results, **args)
         else:
             # print table
             table(CoverageResult, results,
@@ -724,7 +736,8 @@ if __name__ == "__main__":
     import argparse
     import sys
     parser = argparse.ArgumentParser(
-        description="Find coverage info after running tests.")
+        description="Find coverage info after running tests.",
+        allow_abbrev=False)
     parser.add_argument(
         'gcda_paths',
         nargs='*',
@@ -791,15 +804,21 @@ if __name__ == "__main__":
         action='store_true',
         help="Only show the total.")
     parser.add_argument(
-        '-A', '--everything',
+        '-F', '--source',
+        dest='sources',
+        action='append',
+        help="Only consider definitions in this file. Defaults to anything "
+            "in the current directory.")
+    parser.add_argument(
+        '--everything',
         action='store_true',
         help="Include builtin and libc specific symbols.")
     parser.add_argument(
-        '-H', '--hits',
+        '--hits',
         action='store_true',
         help="Show total hits instead of coverage.")
     parser.add_argument(
-        '-l', '--annotate',
+        '-A', '--annotate',
         action='store_true',
         help="Show source files annotated with coverage info.")
     parser.add_argument(
@@ -814,7 +833,7 @@ if __name__ == "__main__":
         '-c', '--context',
         type=lambda x: int(x, 0),
         default=3,
-        help="Show a additional lines of context. Defaults to 3.")
+        help="Show n additional lines of context. Defaults to 3.")
     parser.add_argument(
         '-W', '--width',
         type=lambda x: int(x, 0),
@@ -838,10 +857,6 @@ if __name__ == "__main__":
         default=GCOV_TOOL,
         type=lambda x: x.split(),
         help="Path to the gcov tool to use. Defaults to %r." % GCOV_TOOL)
-    parser.add_argument(
-        '--build-dir',
-        help="Specify the relative build directory. Used to map object files "
-            "to the correct source files.")
     sys.exit(main(**{k: v
         for k, v in vars(parser.parse_intermixed_args()).items()
         if v is not None}))
diff --git a/scripts/data.py b/scripts/data.py
index 05ef868..f95540f 100755
--- a/scripts/data.py
+++ b/scripts/data.py
@@ -5,7 +5,7 @@
 # by Linux's Bloat-O-Meter.
 #
 # Example:
-# ./scripts/data.py lfs.o lfs_util.o -S
+# ./scripts/data.py lfs.o lfs_util.o -Ssize
 #
 # Copyright (c) 2022, The littlefs authors.
 # Copyright (c) 2020, Arm Limited. All rights reserved.
@@ -14,6 +14,7 @@
 
 import collections as co
 import csv
+import difflib
 import glob
 import itertools as it
 import math as m
@@ -25,7 +26,8 @@ import subprocess as sp
 
 OBJ_PATHS = ['*.o']
 NM_TOOL = ['nm']
-TYPE = 'dDbB'
+NM_TYPES = 'dDbB'
+OBJDUMP_TOOL = ['objdump']
 
 
 # integer fields
@@ -135,21 +137,32 @@ def openio(path, mode='r'):
 
 def collect(paths, *,
         nm_tool=NM_TOOL,
-        type=TYPE,
-        build_dir=None,
+        nm_types=NM_TYPES,
+        objdump_tool=OBJDUMP_TOOL,
+        sources=None,
         everything=False,
         **args):
-    results = []
-    pattern = re.compile(
+    size_pattern = re.compile(
         '^(?P<size>[0-9a-fA-F]+)' +
-        ' (?P<type>[%s])' % re.escape(type) +
+        ' (?P<type>[%s])' % re.escape(nm_types) +
         ' (?P<func>.+?)$')
+    line_pattern = re.compile(
+        '^\s+(?P<no>[0-9]+)\s+'
+            '(?:(?P<dir>[0-9]+)\s+)?'
+            '.*\s+'
+            '(?P<path>[^\s]+)$')
+    info_pattern = re.compile(
+        '^(?:.*(?P<tag>DW_TAG_[a-z_]+).*'
+            '|^.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
+            '|^.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*)$')
+
+    results = []
     for path in paths:
-        # map to source file
-        src_path = re.sub('\.o$', '.c', path)
-        if build_dir:
-            src_path = re.sub('%s/*' % re.escape(build_dir), '',
-                src_path)
+        # guess the source, if we have debug-info we'll replace this later
+        file = re.sub('(\.o)?$', '.c', path, 1)
+
+        # find symbol sizes
+        results_ = []
         # note nm-tool may contain extra args
         cmd = nm_tool + ['--size-sort', path]
         if args.get('verbose'):
@@ -158,21 +171,18 @@ def collect(paths, *,
             stdout=sp.PIPE,
             stderr=sp.PIPE if not args.get('verbose') else None,
             universal_newlines=True,
-            errors='replace')
+            errors='replace',
+            close_fds=False)
         for line in proc.stdout:
-            m = pattern.match(line)
+            m = size_pattern.match(line)
             if m:
                 func = m.group('func')
                 # discard internal functions
                 if not everything and func.startswith('__'):
                     continue
-                # discard .8449 suffixes created by optimizer
-                func = re.sub('\.[0-9]+', '', func)
-
-                results.append(DataResult(
-                    src_path, func,
+                results_.append(DataResult(
+                    file, func,
                     int(m.group('size'), 16)))
-
         proc.wait()
         if proc.returncode != 0:
             if not args.get('verbose'):
@@ -180,6 +190,121 @@ def collect(paths, *,
                     sys.stdout.write(line)
             sys.exit(-1)
 
+
+        # try to figure out the source file if we have debug-info
+        dirs = {}
+        files = {}
+        # note objdump-tool may contain extra args
+        cmd = objdump_tool + ['--dwarf=rawline', path]
+        if args.get('verbose'):
+            print(' '.join(shlex.quote(c) for c in cmd))
+        proc = sp.Popen(cmd,
+            stdout=sp.PIPE,
+            stderr=sp.PIPE if not args.get('verbose') else None,
+            universal_newlines=True,
+            errors='replace',
+            close_fds=False)
+        for line in proc.stdout:
+            # note that files contain references to dirs, which we
+            # dereference as soon as we see them as each file table follows a
+            # dir table
+            m = line_pattern.match(line)
+            if m:
+                if not m.group('dir'):
+                    # found a directory entry
+                    dirs[int(m.group('no'))] = m.group('path')
+                else:
+                    # found a file entry
+                    dir = int(m.group('dir'))
+                    if dir in dirs:
+                        files[int(m.group('no'))] = os.path.join(
+                            dirs[dir],
+                            m.group('path'))
+                    else:
+                        files[int(m.group('no'))] = m.group('path')
+        proc.wait()
+        if proc.returncode != 0:
+            if not args.get('verbose'):
+                for line in proc.stderr:
+                    sys.stdout.write(line)
+            # do nothing on error, we don't need objdump to work, source files
+            # may just be inaccurate
+            pass
+
+        defs = {}
+        is_func = False
+        f_name = None
+        f_file = None
+        # note objdump-tool may contain extra args
+        cmd = objdump_tool + ['--dwarf=info', path]
+        if args.get('verbose'):
+            print(' '.join(shlex.quote(c) for c in cmd))
+        proc = sp.Popen(cmd,
+            stdout=sp.PIPE,
+            stderr=sp.PIPE if not args.get('verbose') else None,
+            universal_newlines=True,
+            errors='replace',
+            close_fds=False)
+        for line in proc.stdout:
+            # state machine here to find definitions
+            m = info_pattern.match(line)
+            if m:
+                if m.group('tag'):
+                    if is_func:
+                        defs[f_name] = files.get(f_file, '?')
+                    is_func = (m.group('tag') == 'DW_TAG_subprogram')
+                elif m.group('name'):
+                    f_name = m.group('name')
+                elif m.group('file'):
+                    f_file = int(m.group('file'))
+        proc.wait()
+        if proc.returncode != 0:
+            if not args.get('verbose'):
+                for line in proc.stderr:
+                    sys.stdout.write(line)
+            # do nothing on error, we don't need objdump to work, source files
+            # may just be inaccurate
+            pass
+
+        for r in results_:
+            # find best matching debug symbol, this may be slightly different
+            # due to optimizations
+            if defs:
+                # exact match? avoid difflib if we can for speed
+                if r.function in defs:
+                    file = defs[r.function]
+                else:
+                    _, file = max(
+                        defs.items(),
+                        key=lambda d: difflib.SequenceMatcher(None,
+                            d[0],
+                            r.function, False).ratio())
+            else:
+                file = r.file
+
+            # ignore filtered sources
+            if sources is not None:
+                if not any(
+                        os.path.abspath(file) == os.path.abspath(s)
+                        for s in sources):
+                    continue
+            else:
+                # default to only cwd
+                if not everything and not os.path.commonpath([
+                        os.getcwd(),
+                        os.path.abspath(file)]) == os.getcwd():
+                    continue
+
+            # simplify path
+            if os.path.commonpath([
+                    os.getcwd(),
+                    os.path.abspath(file)]) == os.getcwd():
+                file = os.path.relpath(file)
+            else:
+                file = os.path.abspath(file)
+
+            results.append(DataResult(file, r.function, r.size))
+
     return results
 
 
@@ -437,7 +562,7 @@ def main(obj_paths, *,
                 paths.append(path)
 
         if not paths:
-            print("error: no .obj files found in %r?" % obj_paths)
+            print("error: no .o files found in %r?" % obj_paths)
             sys.exit(-1)
 
         results = collect(paths, **args)
@@ -469,13 +594,16 @@ def main(obj_paths, *,
     # write results to CSV
     if args.get('output'):
         with openio(args['output'], 'w') as f:
-            writer = csv.DictWriter(f, DataResult._by
+            writer = csv.DictWriter(f,
+                (by if by is not None else DataResult._by)
                 + ['data_'+k for k in DataResult._fields])
             writer.writeheader()
             for r in results:
                 writer.writerow(
-                    {k: getattr(r, k) for k in DataResult._by}
-                    | {'data_'+k: getattr(r, k) for k in DataResult._fields})
+                    {k: getattr(r, k)
+                        for k in (by if by is not None else DataResult._by)}
+                    | {'data_'+k: getattr(r, k)
+                        for k in DataResult._fields})
 
     # find previous results?
     if args.get('diff'):
@@ -512,7 +640,8 @@ if __name__ == "__main__":
     import argparse
     import sys
     parser = argparse.ArgumentParser(
-        description="Find data size at the function level.")
+        description="Find data size at the function level.",
+        allow_abbrev=False)
     parser.add_argument(
         'obj_paths',
         nargs='*',
@@ -579,23 +708,30 @@ if __name__ == "__main__":
         action='store_true',
         help="Only show the total.")
     parser.add_argument(
-        '-A', '--everything',
+        '-F', '--source',
+        dest='sources',
+        action='append',
+        help="Only consider definitions in this file. Defaults to anything "
+            "in the current directory.")
+    parser.add_argument(
+        '--everything',
         action='store_true',
         help="Include builtin and libc specific symbols.")
     parser.add_argument(
-        '--type',
-        default=TYPE,
+        '--nm-types',
+        default=NM_TYPES,
         help="Type of symbols to report, this uses the same single-character "
-            "type-names emitted by nm. Defaults to %r." % TYPE)
+            "type-names emitted by nm. Defaults to %r." % NM_TYPES)
     parser.add_argument(
         '--nm-tool',
         type=lambda x: x.split(),
         default=NM_TOOL,
         help="Path to the nm tool to use. Defaults to %r." % NM_TOOL)
     parser.add_argument(
-        '--build-dir',
-        help="Specify the relative build directory. Used to map object files "
-            "to the correct source files.")
+        '--objdump-tool',
+        type=lambda x: x.split(),
+        default=OBJDUMP_TOOL,
+        help="Path to the objdump tool to use. Defaults to %r." % OBJDUMP_TOOL)
     sys.exit(main(**{k: v
         for k, v in vars(parser.parse_intermixed_args()).items()
         if v is not None}))
diff --git a/scripts/perf.py b/scripts/perf.py
new file mode 100755
index 0000000..3eb3dbc
--- /dev/null
+++ b/scripts/perf.py
@@ -0,0 +1,1263 @@
+#!/usr/bin/env python3
+#
+# Script to aggregate and report Linux perf results.
+#
+# Example:
+# ./scripts/perf.py -R -obench.perf ./runners/bench_runner
+# ./scripts/perf.py bench.perf -Flfs.c -Flfs_util.c -Scycles
+#
+# Copyright (c) 2022, The littlefs authors.
+# SPDX-License-Identifier: BSD-3-Clause
+#
+
+import bisect
+import collections as co
+import csv
+import errno
+import fcntl
+import functools as ft
+import glob
+import itertools as it
+import math as m
+import multiprocessing as mp
+import os
+import re
+import shlex
+import shutil
+import subprocess as sp
+import tempfile
+import zipfile
+
+
+PERF_PATHS = ['*.perf']
+PERF_TOOL = ['perf']
+PERF_EVENTS = 'cycles,branch-misses,branches,cache-misses,cache-references'
+PERF_FREQ = 100
+OBJDUMP_TOOL = ['objdump']
+THRESHOLD = (0.5, 0.85)
+
+
+# integer fields
+class Int(co.namedtuple('Int', 'x')):
+    __slots__ = ()
+    def __new__(cls, x=0):
+        if isinstance(x, Int):
+            return x
+        if isinstance(x, str):
+            try:
+                x = int(x, 0)
+            except ValueError:
+                # also accept +-∞ and +-inf
+                if re.match('^\s*\+?\s*(?:∞|inf)\s*$', x):
+                    x = m.inf
+                elif re.match('^\s*-\s*(?:∞|inf)\s*$', x):
+                    x = -m.inf
+                else:
+                    raise
+        assert isinstance(x, int) or m.isinf(x), x
+        return super().__new__(cls, x)
+
+    def __str__(self):
+        if self.x == m.inf:
+            return '∞'
+        elif self.x == -m.inf:
+            return '-∞'
+        else:
+            return str(self.x)
+
+    def __int__(self):
+        assert not m.isinf(self.x)
+        return self.x
+
+    def __float__(self):
+        return float(self.x)
+
+    none = '%7s' % '-'
+    def table(self):
+        return '%7s' % (self,)
+
+    diff_none = '%7s' % '-'
+    diff_table = table
+
+    def diff_diff(self, other):
+        new = self.x if self else 0
+        old = other.x if other else 0
+        diff = new - old
+        if diff == +m.inf:
+            return '%7s' % '+∞'
+        elif diff == -m.inf:
+            return '%7s' % '-∞'
+        else:
+            return '%+7d' % diff
+
+    def ratio(self, other):
+        new = self.x if self else 0
+        old = other.x if other else 0
+        if m.isinf(new) and m.isinf(old):
+            return 0.0
+        elif m.isinf(new):
+            return +m.inf
+        elif m.isinf(old):
+            return -m.inf
+        elif not old and not new:
+            return 0.0
+        elif not old:
+            return 1.0
+        else:
+            return (new-old) / old
+
+    def __add__(self, other):
+        return self.__class__(self.x + other.x)
+
+    def __sub__(self, other):
+        return self.__class__(self.x - other.x)
+
+    def __mul__(self, other):
+        return self.__class__(self.x * other.x)
+
+# perf results
+class PerfResult(co.namedtuple('PerfResult', [
+        'file', 'function', 'line',
+        'self_cycles',
+        'self_bmisses', 'self_branches',
+        'self_cmisses', 'self_caches',
+        'cycles',
+        'bmisses', 'branches',
+        'cmisses', 'caches',
+        'children', 'parents'])):
+    _by = ['file', 'function', 'line']
+    _fields = [
+        'self_cycles',
+        'self_bmisses', 'self_branches',
+        'self_cmisses', 'self_caches',
+        'cycles',
+        'bmisses', 'branches',
+        'cmisses', 'caches']
+    _types = {
+        'self_cycles': Int,
+        'self_bmisses': Int, 'self_branches': Int,
+        'self_cmisses': Int, 'self_caches': Int,
+        'cycles': Int,
+        'bmisses': Int, 'branches': Int,
+        'cmisses': Int, 'caches': Int}
+
+    __slots__ = ()
+    def __new__(cls, file='', function='', line=0,
+            self_cycles=0,
+            self_bmisses=0, self_branches=0,
+            self_cmisses=0, self_caches=0,
+            cycles=0,
+            bmisses=0, branches=0,
+            cmisses=0, caches=0,
+            children=set(), parents=set()):
+        return super().__new__(cls, file, function, int(Int(line)),
+            Int(self_cycles),
+            Int(self_bmisses), Int(self_branches),
+            Int(self_cmisses), Int(self_caches),
+            Int(cycles),
+            Int(bmisses), Int(branches),
+            Int(cmisses), Int(caches),
+            children, parents)
+
+    def __add__(self, other):
+        return PerfResult(self.file, self.function, self.line,
+            self.self_cycles + other.self_cycles,
+            self.self_bmisses + other.self_bmisses,
+            self.self_branches + other.self_branches,
+            self.self_cmisses + other.self_cmisses,
+            self.self_caches + other.self_caches,
+            self.cycles + other.cycles,
+            self.bmisses + other.bmisses,
+            self.branches + other.branches,
+            self.cmisses + other.cmisses,
+            self.caches + other.caches,
+            self.children | other.children,
+            self.parents | other.parents)
+
+
+def openio(path, mode='r'):
+    if path == '-':
+        if mode == 'r':
+            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+        else:
+            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+    else:
+        return open(path, mode)
+
+# run perf as a subprocess, storing measurements into a zip file
+def record(command, *,
+        output=None,
+        perf_freq=PERF_FREQ,
+        perf_period=None,
+        perf_events=PERF_EVENTS,
+        perf_tool=PERF_TOOL,
+        **args):
+    if not command:
+        print('error: no command specified?')
+        sys.exit(-1)
+
+    if not output:
+        print('error: no output file specified?')
+        sys.exit(-1)
+
+    # create a temporary file for perf to write to, as far as I can tell
+    # this is strictly needed because perf's pipe-mode only works with stdout
+    with tempfile.NamedTemporaryFile('rb') as f:
+        # figure out our perf invocation
+        perf = perf_tool + list(filter(None, [
+            'record',
+            '-F%s' % perf_freq
+                if perf_freq is not None
+                and perf_period is None else None,
+            '-c%s' % perf_period
+                if perf_period is not None else None,
+            '-B',
+            '-g',
+            '--all-user',
+            '-e%s' % perf_events,
+            '-o%s' % f.name]))
+
+        # run our command
+        try:
+            if args.get('verbose'):
+                print(' '.join(shlex.quote(c) for c in perf + command))
+            err = sp.call(perf + command, close_fds=False)
+
+        except KeyboardInterrupt:
+            err = errno.EOWNERDEAD
+
+        # synchronize access
+        z = os.open(output, os.O_RDWR | os.O_CREAT)
+        fcntl.flock(z, fcntl.LOCK_EX)
+
+        # copy measurements into our zip file
+        with os.fdopen(z, 'r+b') as z:
+            with zipfile.ZipFile(z, 'a',
+                    compression=zipfile.ZIP_DEFLATED,
+                    compresslevel=1) as z:
+                with z.open('perf.%d' % os.getpid(), 'w') as g:
+                    shutil.copyfileobj(f, g)
+
+    # forward the return code
+    return err
+
+
+def collect_decompressed(path, *,
+        perf_tool=PERF_TOOL,
+        everything=False,
+        depth=0,
+        **args):
+    sample_pattern = re.compile(
+        '(?P<comm>\w+)'
+        '\s+(?P<pid>\w+)'
+        '\s+(?P<time>[\w.]+):'
+        '\s*(?P<period>\w+)'
+        '\s+(?P<event>[^:]+):')
+    frame_pattern = re.compile(
+        '\s+(?P<addr>\w+)'
+        '\s+(?P<sym>[^\s]+)'
+        '\s+\((?P<dso>[^\)]+)\)')
+    events = {
+        'cycles':           'cycles',
+        'branch-misses':    'bmisses',
+        'branches':         'branches',
+        'cache-misses':     'cmisses',
+        'cache-references': 'caches'}
+
+    # note perf_tool may contain extra args
+    cmd = perf_tool + [
+        'script',
+        '-i%s' % path]
+    if args.get('verbose'):
+        print(' '.join(shlex.quote(c) for c in cmd))
+    proc = sp.Popen(cmd,
+        stdout=sp.PIPE,
+        stderr=sp.PIPE if not args.get('verbose') else None,
+        universal_newlines=True,
+        errors='replace',
+        close_fds=False)
+
+    last_filtered = False
+    last_has_frame = False
+    last_event = ''
+    last_period = 0
+    results = co.defaultdict(lambda: co.defaultdict(lambda: (0, 0)))
+
+    for line in proc.stdout:
+        # we need to process a lot of data, so wait to use regex as late
+        # as possible
+        if not line:
+            continue
+        if not line.startswith('\t'):
+            m = sample_pattern.match(line)
+            if m:
+                last_event = m.group('event')
+                last_filtered = last_event in events
+                last_period = int(m.group('period'), 0)
+                last_has_frame = False
+        elif last_filtered:
+            m = frame_pattern.match(line)
+            if m:
+                # filter out internal/kernel functions
+                if not everything and (
+                        m.group('sym').startswith('__')
+                        or m.group('dso').startswith('/usr/lib')
+                        or not m.group('sym')[:1].isalpha()):
+                    continue
+
+                name = (
+                    m.group('dso'),
+                    m.group('sym'),
+                    int(m.group('addr'), 16))
+                self, total = results[name][last_event]
+                if not last_has_frame:
+                    results[name][last_event] = (
+                        self + last_period,
+                        total + last_period)
+                    last_has_frame = True
+                else:
+                    results[name][last_event] = (
+                        self,
+                        total + last_period)
+
+    proc.wait()
+    if proc.returncode != 0:
+        if not args.get('verbose'):
+            for line in proc.stderr:
+                sys.stdout.write(line)
+        sys.exit(-1)
+
+    # rearrange results into result type
+    results_ = []
+    for name, r in results.items():
+        results_.append(PerfResult(*name,
+            **{'self_'+events[e]: s for e, (s, _) in r.items()},
+            **{        events[e]: t for e, (_, t) in r.items()}))
+    results = results_
+
+    return results
+
+def collect_job(path, i, **args):
+    # decompress into a temporary file, this is to work around
+    # some limitations of perf
+    with zipfile.ZipFile(path) as z:
+        with z.open(i) as f:
+            with tempfile.NamedTemporaryFile('wb') as g:
+                shutil.copyfileobj(f, g)
+                g.flush()
+
+                return collect_decompressed(g.name, **args)
+
+def starapply(args):
+    f, args, kwargs = args
+    return f(*args, **kwargs)
+
+def collect(paths, *,
+        jobs=None,
+        objdump_tool=None,
+        sources=None,
+        everything=False,
+        **args):
+    symbol_pattern = re.compile(
+        '^(?P<addr>[0-9a-fA-F]+)\s.*\s(?P<name>[^\s]+)\s*$')
+    line_pattern = re.compile(
+        '^\s+(?:'
+            # matches dir/file table
+            '(?P<no>[0-9]+)\s+'
+                '(?:(?P<dir>[0-9]+)\s+)?'
+                '.*\s+'
+                '(?P<path>[^\s]+)'
+            # matches line opcodes
+            '|' '\[[^\]]*\]\s+'
+                '(?:'
+                    '(?P<op_special>Special)'
+                    '|' '(?P<op_copy>Copy)'
+                    '|' '(?P<op_end>End of Sequence)'
+                    '|' 'File .*?to (?:entry )?(?P<op_file>\d+)'
+                    '|' 'Line .*?to (?P<op_line>[0-9]+)'
+                    '|' '(?:Address|PC) .*?to (?P<op_addr>[0x0-9a-fA-F]+)'
+                    '|' '.' ')*'
+            ')$', re.IGNORECASE)
+
+    records = []
+    for path in paths:
+        # each .perf file is actually a zip file containing perf files from
+        # multiple runs
+        with zipfile.ZipFile(path) as z:
+            records.extend((path, i) for i in z.infolist())
+
+    # we're dealing with a lot of data but also surprisingly
+    # parallelizable
+    dsos = {}
+    results = []
+    with mp.Pool(jobs or len(os.sched_getaffinity(0))) as p:
+        for results_ in p.imap_unordered(
+                starapply,
+                ((collect_job, (path, i), dict(
+                    everything=everything,
+                    **args))
+                    for path, i in records)):
+
+            # organize by dso
+            results__ = {}
+            for r in results_:
+                if r.file not in results__:
+                    results__[r.file] = []
+                results__[r.file].append(r)
+            results_ = results__
+
+            for dso, results_ in results_.items():
+                if dso not in dsos:
+                    # find file+line ranges for dsos
+                    #
+                    # do this here so we only process each dso once
+                    syms = {}
+                    sym_at = []
+                    cmd = objdump_tool + ['-t', dso]
+                    if args.get('verbose'):
+                        print(' '.join(shlex.quote(c) for c in cmd))
+                    proc = sp.Popen(cmd,
+                        stdout=sp.PIPE,
+                        stderr=sp.PIPE if not args.get('verbose') else None,
+                        universal_newlines=True,
+                        errors='replace',
+                        close_fds=False)
+                    for line in proc.stdout:
+                        m = symbol_pattern.match(line)
+                        if m:
+                            name = m.group('name')
+                            addr = int(m.group('addr'), 16)
+                            # note multiple symbols can share a name
+                            if name not in syms:
+                                syms[name] = set()
+                            syms[name].add(addr)
+                            sym_at.append((addr, name))
+                    proc.wait()
+                    if proc.returncode != 0:
+                        if not args.get('verbose'):
+                            for line in proc.stderr:
+                                sys.stdout.write(line)
+                        # assume no debug-info on failure
+                        pass
+
+                    # sort and keep first when duplicates
+                    sym_at.sort()
+                    sym_at_ = []
+                    for addr, name in sym_at:
+                        if len(sym_at_) == 0 or sym_at_[-1][0] != addr:
+                            sym_at_.append((addr, name))
+                    sym_at = sym_at_
+
+                    # state machine for dwarf line numbers, note that objdump's
+                    # decodedline seems to have issues with multiple dir/file
+                    # tables, which is why we need this
+                    line_at = []
+                    dirs = {}
+                    files = {}
+                    op_file = 1
+                    op_line = 1
+                    op_addr = 0
+                    cmd = objdump_tool + ['--dwarf=rawline', dso]
+                    if args.get('verbose'):
+                        print(' '.join(shlex.quote(c) for c in cmd))
+                    proc = sp.Popen(cmd,
+                        stdout=sp.PIPE,
+                        stderr=sp.PIPE if not args.get('verbose') else None,
+                        universal_newlines=True,
+                        errors='replace',
+                        close_fds=False)
+                    for line in proc.stdout:
+                        m = line_pattern.match(line)
+                        if m:
+                            if m.group('no') and not m.group('dir'):
+                                # found a directory entry
+                                dirs[int(m.group('no'))] = m.group('path')
+                            elif m.group('no'):
+                                # found a file entry
+                                dir = int(m.group('dir'))
+                                if dir in dirs:
+                                    files[int(m.group('no'))] = os.path.join(
+                                        dirs[dir],
+                                        m.group('path'))
+                                else:
+                                    files[int(m.group('no'))] = m.group('path')
+                            else:
+                                # found a state machine update
+                                if m.group('op_file'):
+                                    op_file = int(m.group('op_file'), 0)
+                                if m.group('op_line'):
+                                    op_line = int(m.group('op_line'), 0)
+                                if m.group('op_addr'):
+                                    op_addr = int(m.group('op_addr'), 0)
+
+                                if (m.group('op_special')
+                                        or m.group('op_copy')
+                                        or m.group('op_end')):
+                                    line_at.append((
+                                        op_addr,
+                                        files.get(op_file, '?'),
+                                        op_line))
+
+                                if m.group('op_end'):
+                                    op_file = 1
+                                    op_line = 1
+                                    op_addr = 0
+                    proc.wait()
+                    if proc.returncode != 0:
+                        if not args.get('verbose'):
+                            for line in proc.stderr:
+                                sys.stdout.write(line)
+                        # assume no debug-info on failure
+                        pass
+
+                    # sort and keep first when duplicates
+                    #
+                    # I think dwarf requires this to be sorted but just in case
+                    line_at.sort()
+                    line_at_ = []
+                    for addr, file, line in line_at:
+                        if len(line_at_) == 0 or line_at_[-1][0] != addr:
+                            line_at_.append((addr, file, line))
+                    line_at = line_at_
+
+                    # discard lines outside of the range of the containing
+                    # function, these are introduced by dwarf for inlined
+                    # functions but don't map to elf-level symbols
+                    sym_at_ = []
+                    for addr, sym in sym_at:
+                        i = bisect.bisect(line_at, addr, key=lambda x: x[0])
+                        if i > 0:
+                            _, file, line = line_at[i-1]
+                            sym_at_.append((file, line, sym))
+                    sym_at_.sort()
+
+                    line_at_ = []
+                    for addr, file, line in line_at:
+                        # only keep if sym-at-addr and sym-at-line match
+                        i = bisect.bisect(
+                            sym_at, addr, key=lambda x: x[0])
+                        j = bisect.bisect(
+                            sym_at_, (file, line), key=lambda x: (x[0], x[1]))
+                        if i > 0 and j > 0 and (
+                                sym_at[i-1][1] == sym_at_[j-1][2]):
+                            line_at_.append((addr, file, line))
+                    line_at = line_at_
+
+                    dsos[dso] = (syms, sym_at, line_at)
+
+                syms, _, line_at = dsos[dso]
+
+                # first try to reverse ASLR
+                def deltas(r, d):
+                    if '+' in r.function:
+                        sym, off = r.function.split('+', 1)
+                        off = int(off, 0)
+                    else:
+                        sym, off = r.function, 0
+                    addr = r.line - off + d
+
+                    for addr_ in syms.get(sym, []):
+                        yield addr_ - addr
+
+                delta = min(
+                    it.chain.from_iterable(
+                        deltas(r, 0) for r in results_),
+                    key=lambda d: sum(it.chain.from_iterable(
+                        deltas(r, d) for r in results_)),
+                    default=0)
+
+                # then try to map addrs -> file+line
+                for r in results_:
+                    addr = r.line + delta
+                    i = bisect.bisect(line_at, addr, key=lambda x: x[0])
+                    if i > 0:
+                        _, file, line = line_at[i-1]
+                    else:
+                        file, line = re.sub('(\.o)?$', '.c', r.file, 1), 0
+
+                    # ignore filtered sources
+                    if sources is not None:
+                        if not any(
+                                os.path.abspath(file) == os.path.abspath(s)
+                                for s in sources):
+                            continue
+                    else:
+                        # default to only cwd
+                        if not everything and not os.path.commonpath([
+                                os.getcwd(),
+                                os.path.abspath(file)]) == os.getcwd():
+                            continue
+
+                    # simplify path
+                    if os.path.commonpath([
+                            os.getcwd(),
+                            os.path.abspath(file)]) == os.getcwd():
+                        file = os.path.relpath(file)
+                    else:
+                        file = os.path.abspath(file)
+
+                    function, *_ = r.function.split('+', 1)
+                    results.append(PerfResult(file, function, line,
+                        **{k: getattr(r, k) for k in PerfResult._fields}))
+
+    return results
+
+
+def fold(Result, results, *,
+        by=None,
+        defines=None,
+        **_):
+    if by is None:
+        by = Result._by
+
+    for k in it.chain(by or [], (k for k, _ in defines or [])):
+        if k not in Result._by and k not in Result._fields:
+            print("error: could not find field %r?" % k)
+            sys.exit(-1)
+
+    # filter by matching defines
+    if defines is not None:
+        results_ = []
+        for r in results:
+            if all(getattr(r, k) in vs for k, vs in defines):
+                results_.append(r)
+        results = results_
+
+    # organize results into conflicts
+    folding = co.OrderedDict()
+    for r in results:
+        name = tuple(getattr(r, k) for k in by)
+        if name not in folding:
+            folding[name] = []
+        folding[name].append(r)
+
+    # merge conflicts
+    folded = []
+    for name, rs in folding.items():
+        folded.append(sum(rs[1:], start=rs[0]))
+
+    return folded
+
+def table(Result, results, diff_results=None, *,
+        by=None,
+        fields=None,
+        sort=None,
+        summary=False,
+        all=False,
+        percent=False,
+        **_):
+    all_, all = all, __builtins__.all
+
+    if by is None:
+        by = Result._by
+    if fields is None:
+        fields = Result._fields
+    types = Result._types
+
+    # fold again
+    results = fold(Result, results, by=by)
+    if diff_results is not None:
+        diff_results = fold(Result, diff_results, by=by)
+
+    # organize by name
+    table = {
+        ','.join(str(getattr(r, k) or '') for k in by): r
+        for r in results}
+    diff_table = {
+        ','.join(str(getattr(r, k) or '') for k in by): r
+        for r in diff_results or []}
+    names = list(table.keys() | diff_table.keys())
+
+    # sort again, now with diff info, note that python's sort is stable
+    names.sort()
+    if diff_results is not None:
+        names.sort(key=lambda n: tuple(
+            types[k].ratio(
+                getattr(table.get(n), k, None),
+                getattr(diff_table.get(n), k, None))
+            for k in fields),
+            reverse=True)
+    if sort:
+        for k, reverse in reversed(sort):
+            names.sort(key=lambda n: (getattr(table[n], k),)
+                if getattr(table.get(n), k, None) is not None else (),
+                reverse=reverse ^ (not k or k in Result._fields))
+
+
+    # build up our lines
+    lines = []
+
+    # header
+    line = []
+    line.append('%s%s' % (
+        ','.join(by),
+        ' (%d added, %d removed)' % (
+            sum(1 for n in table if n not in diff_table),
+            sum(1 for n in diff_table if n not in table))
+            if diff_results is not None and not percent else '')
+        if not summary else '')
+    if diff_results is None:
+        for k in fields:
+            line.append(k)
+    elif percent:
+        for k in fields:
+            line.append(k)
+    else:
+        for k in fields:
+            line.append('o'+k)
+        for k in fields:
+            line.append('n'+k)
+        for k in fields:
+            line.append('d'+k)
+    line.append('')
+    lines.append(line)
+
+    # entries
+    if not summary:
+        for name in names:
+            r = table.get(name)
+            if diff_results is not None:
+                diff_r = diff_table.get(name)
+                ratios = [
+                    types[k].ratio(
+                        getattr(r, k, None),
+                        getattr(diff_r, k, None))
+                    for k in fields]
+                if not any(ratios) and not all_:
+                    continue
+
+            line = []
+            line.append(name)
+            if diff_results is None:
+                for k in fields:
+                    line.append(getattr(r, k).table()
+                        if getattr(r, k, None) is not None
+                        else types[k].none)
+            elif percent:
+                for k in fields:
+                    line.append(getattr(r, k).diff_table()
+                        if getattr(r, k, None) is not None
+                        else types[k].diff_none)
+            else:
+                for k in fields:
+                    line.append(getattr(diff_r, k).diff_table()
+                        if getattr(diff_r, k, None) is not None
+                        else types[k].diff_none)
+                for k in fields:
+                    line.append(getattr(r, k).diff_table()
+                        if getattr(r, k, None) is not None
+                        else types[k].diff_none)
+                for k in fields:
+                    line.append(types[k].diff_diff(
+                            getattr(r, k, None),
+                            getattr(diff_r, k, None)))
+            if diff_results is None:
+                line.append('')
+            elif percent:
+                line.append(' (%s)' % ', '.join(
+                    '+∞%' if t == +m.inf
+                    else '-∞%' if t == -m.inf
+                    else '%+.1f%%' % (100*t)
+                    for t in ratios))
+            else:
+                line.append(' (%s)' % ', '.join(
+                        '+∞%' if t == +m.inf
+                        else '-∞%' if t == -m.inf
+                        else '%+.1f%%' % (100*t)
+                        for t in ratios
+                        if t)
+                    if any(ratios) else '')
+            lines.append(line)
+
+    # total
+    r = next(iter(fold(Result, results, by=[])), None)
+    if diff_results is not None:
+        diff_r = next(iter(fold(Result, diff_results, by=[])), None)
+        ratios = [
+            types[k].ratio(
+                getattr(r, k, None),
+                getattr(diff_r, k, None))
+            for k in fields]
+
+    line = []
+    line.append('TOTAL')
+    if diff_results is None:
+        for k in fields:
+            line.append(getattr(r, k).table()
+                if getattr(r, k, None) is not None
+                else types[k].none)
+    elif percent:
+        for k in fields:
+            line.append(getattr(r, k).diff_table()
+                if getattr(r, k, None) is not None
+                else types[k].diff_none)
+    else:
+        for k in fields:
+            line.append(getattr(diff_r, k).diff_table()
+                if getattr(diff_r, k, None) is not None
+                else types[k].diff_none)
+        for k in fields:
+            line.append(getattr(r, k).diff_table()
+                if getattr(r, k, None) is not None
+                else types[k].diff_none)
+        for k in fields:
+            line.append(types[k].diff_diff(
+                    getattr(r, k, None),
+                    getattr(diff_r, k, None)))
+    if diff_results is None:
+        line.append('')
+    elif percent:
+        line.append(' (%s)' % ', '.join(
+            '+∞%' if t == +m.inf
+            else '-∞%' if t == -m.inf
+            else '%+.1f%%' % (100*t)
+            for t in ratios))
+    else:
+        line.append(' (%s)' % ', '.join(
+                '+∞%' if t == +m.inf
+                else '-∞%' if t == -m.inf
+                else '%+.1f%%' % (100*t)
+                for t in ratios
+                if t)
+            if any(ratios) else '')
+    lines.append(line)
+
+    # find the best widths, note that column 0 contains the names and column -1
+    # the ratios, so those are handled a bit differently
+    widths = [
+        ((max(it.chain([w], (len(l[i]) for l in lines)))+1+4-1)//4)*4-1
+        for w, i in zip(
+            it.chain([23], it.repeat(7)),
+            range(len(lines[0])-1))]
+
+    # print our table
+    for line in lines:
+        print('%-*s  %s%s' % (
+            widths[0], line[0],
+            ' '.join('%*s' % (w, x)
+                for w, x in zip(widths[1:], line[1:-1])),
+            line[-1]))
+
+
+def annotate(Result, results, *,
+        annotate=None,
+        threshold=None,
+        branches=False,
+        caches=False,
+        **args):
+    # figure out the threshold
+    if threshold is None:
+        t0, t1 = THRESHOLD
+    elif len(threshold) == 1:
+        t0, t1 = threshold[0], threshold[0]
+    else:
+        t0, t1 = threshold
+    t0, t1 = min(t0, t1), max(t0, t1)
+
+    if not branches and not caches:
+        tk = 'self_cycles'
+    elif branches:
+        tk = 'self_bmisses'
+    else:
+        tk = 'self_cmisses'
+
+    # find max cycles
+    max_ = max(it.chain((float(getattr(r, tk)) for r in results), [1]))
+
+    for path in co.OrderedDict.fromkeys(r.file for r in results).keys():
+        # flatten to line info
+        results = fold(Result, results, by=['file', 'line'])
+        table = {r.line: r for r in results if r.file == path}
+
+        # calculate spans to show
+        if not annotate:
+            spans = []
+            last = None
+            func = None
+            for line, r in sorted(table.items()):
+                if float(getattr(r, tk)) / max_ >= t0:
+                    if last is not None and line - last.stop <= args['context']:
+                        last = range(
+                            last.start,
+                            line+1+args['context'])
+                    else:
+                        if last is not None:
+                            spans.append((last, func))
+                        last = range(
+                            line-args['context'],
+                            line+1+args['context'])
+                        func = r.function
+            if last is not None:
+                spans.append((last, func))
+
+        with open(path) as f:
+            skipped = False
+            for i, line in enumerate(f):
+                # skip lines not in spans?
+                if not annotate and not any(i+1 in s for s, _ in spans):
+                    skipped = True
+                    continue
+
+                if skipped:
+                    skipped = False
+                    print('%s@@ %s:%d: %s @@%s' % (
+                        '\x1b[36m' if args['color'] else '',
+                        path,
+                        i+1,
+                        next(iter(f for _, f in spans)),
+                        '\x1b[m' if args['color'] else ''))
+
+                # build line
+                if line.endswith('\n'):
+                    line = line[:-1]
+
+                r = table.get(i+1)
+                if r is not None and (
+                        float(r.self_cycles) > 0
+                        if not branches and not caches
+                        else float(r.self_bmisses) > 0
+                            or float(r.self_branches) > 0
+                        if branches
+                        else float(r.self_cmisses) > 0
+                            or float(r.self_caches) > 0):
+                    line = '%-*s // %s' % (
+                        args['width'],
+                        line,
+                        '%s cycles' % r.self_cycles
+                        if not branches and not caches
+                        else '%s bmisses, %s branches' % (
+                            r.self_bmisses, r.self_branches)
+                        if branches
+                        else '%s cmisses, %s caches' % (
+                            r.self_cmisses, r.self_caches))
+
+                    if args['color']:
+                        if float(getattr(r, tk)) / max_ >= t1:
+                            line = '\x1b[1;31m%s\x1b[m' % line
+                        elif float(getattr(r, tk)) / max_ >= t0:
+                            line = '\x1b[35m%s\x1b[m' % line
+
+                print(line)
+
+
+def report(perf_paths, *,
+        by=None,
+        fields=None,
+        defines=None,
+        sort=None,
+        self=False,
+        branches=False,
+        caches=False,
+        tree=False,
+        depth=None,
+        **args):
+    # figure out what color should be
+    if args.get('color') == 'auto':
+        args['color'] = sys.stdout.isatty()
+    elif args.get('color') == 'always':
+        args['color'] = True
+    else:
+        args['color'] = False
+
+    # it doesn't really make sense to not have a depth with tree,
+    # so assume depth=inf if tree by default
+    if args.get('depth') is None:
+        args['depth'] = m.inf if tree else 1
+    elif args.get('depth') == 0:
+        args['depth'] = m.inf
+
+    # find sizes
+    if not args.get('use', None):
+        # find .o files
+        paths = []
+        for path in perf_paths:
+            if os.path.isdir(path):
+                path = path + '/*.perf'
+
+            for path in glob.glob(path):
+                paths.append(path)
+
+        if not paths:
+            print("error: no .perf files found in %r?" % perf_paths)
+            sys.exit(-1)
+
+        results = collect(paths, **args)
+    else:
+        results = []
+        with openio(args['use']) as f:
+            reader = csv.DictReader(f, restval='')
+            for r in reader:
+                try:
+                    results.append(PerfResult(
+                        **{k: r[k] for k in PerfResult._by
+                            if k in r and r[k].strip()},
+                        **{k: r['perf_'+k] for k in PerfResult._fields
+                            if 'perf_'+k in r and r['perf_'+k].strip()}))
+                except TypeError:
+                    pass
+
+    # fold
+    results = fold(PerfResult, results, by=by, defines=defines)
+
+    # sort, note that python's sort is stable
+    results.sort()
+    if sort:
+        for k, reverse in reversed(sort):
+            results.sort(key=lambda r: (getattr(r, k),)
+                if getattr(r, k) is not None else (),
+                reverse=reverse ^ (not k or k in PerfResult._fields))
+
+    # write results to CSV
+    if args.get('output'):
+        with openio(args['output'], 'w') as f:
+            writer = csv.DictWriter(f,
+                (by if by is not None else PerfResult._by)
+                + ['perf_'+k for k in PerfResult._fields])
+            writer.writeheader()
+            for r in results:
+                writer.writerow(
+                    {k: getattr(r, k)
+                        for k in (by if by is not None else PerfResult._by)}
+                    | {'perf_'+k: getattr(r, k)
+                        for k in PerfResult._fields})
+
+    # find previous results?
+    if args.get('diff'):
+        diff_results = []
+        try:
+            with openio(args['diff']) as f:
+                reader = csv.DictReader(f, restval='')
+                for r in reader:
+                    try:
+                        diff_results.append(PerfResult(
+                            **{k: r[k] for k in PerfResult._by
+                                if k in r and r[k].strip()},
+                            **{k: r['perf_'+k] for k in PerfResult._fields
+                                if 'perf_'+k in r and r['perf_'+k].strip()}))
+                    except TypeError:
+                        pass
+        except FileNotFoundError:
+            pass
+
+        # fold
+        diff_results = fold(PerfResult, diff_results, by=by, defines=defines)
+
+    # print table
+    if not args.get('quiet'):
+        if args.get('annotate') or args.get('threshold'):
+            # annotate sources
+            annotate(PerfResult, results,
+                branches=branches,
+                caches=caches,
+                **args)
+        else:
+            # print table
+            table(PerfResult, results,
+                diff_results if args.get('diff') else None,
+                by=by if by is not None else ['function'],
+                fields=fields if fields is not None else [
+                    'self_'+k if self else k
+                    for k in (
+                        ['cycles'] if not branches and not caches
+                        else ['bmisses', 'branches'] if branches
+                        else ['cmisses', 'caches'])],
+                sort=sort,
+                **args)
+
+
+def main(**args):
+    if args.get('record'):
+        return record(**args)
+    else:
+        return report(**args)
+
+
+if __name__ == "__main__":
+    import argparse
+    import sys
+
+    # bit of a hack, but parse_intermixed_args and REMAINDER are
+    # incompatible, so we need to figure out what we want before running
+    # argparse
+    if '-R' in sys.argv or '--record' in sys.argv:
+        nargs = argparse.REMAINDER
+    else:
+        nargs = '*'
+
+    argparse.ArgumentParser._handle_conflict_ignore = lambda *_: None
+    argparse._ArgumentGroup._handle_conflict_ignore = lambda *_: None
+    parser = argparse.ArgumentParser(
+        description="Aggregate and report Linux perf results.",
+        allow_abbrev=False,
+        conflict_handler='ignore')
+    parser.add_argument(
+        'perf_paths',
+        nargs=nargs,
+        help="Description of where to find *.perf files. May be a directory "
+            "or a list of paths. Defaults to %r." % PERF_PATHS)
+    parser.add_argument(
+        '-v', '--verbose',
+        action='store_true',
+        help="Output commands that run behind the scenes.")
+    parser.add_argument(
+        '-q', '--quiet',
+        action='store_true',
+        help="Don't show anything, useful with -o.")
+    parser.add_argument(
+        '-o', '--output',
+        help="Specify CSV file to store results.")
+    parser.add_argument(
+        '-u', '--use',
+        help="Don't parse anything, use this CSV file.")
+    parser.add_argument(
+        '-d', '--diff',
+        help="Specify CSV file to diff against.")
+    parser.add_argument(
+        '-a', '--all',
+        action='store_true',
+        help="Show all, not just the ones that changed.")
+    parser.add_argument(
+        '-p', '--percent',
+        action='store_true',
+        help="Only show percentage change, not a full diff.")
+    parser.add_argument(
+        '-b', '--by',
+        action='append',
+        choices=PerfResult._by,
+        help="Group by this field.")
+    parser.add_argument(
+        '-f', '--field',
+        dest='fields',
+        action='append',
+        choices=PerfResult._fields,
+        help="Show this field.")
+    parser.add_argument(
+        '-D', '--define',
+        dest='defines',
+        action='append',
+        type=lambda x: (lambda k,v: (k, set(v.split(','))))(*x.split('=', 1)),
+        help="Only include results where this field is this value.")
+    class AppendSort(argparse.Action):
+        def __call__(self, parser, namespace, value, option):
+            if namespace.sort is None:
+                namespace.sort = []
+            namespace.sort.append((value, True if option == '-S' else False))
+    parser.add_argument(
+        '-s', '--sort',
+        action=AppendSort,
+        help="Sort by this fields.")
+    parser.add_argument(
+        '-S', '--reverse-sort',
+        action=AppendSort,
+        help="Sort by this fields, but backwards.")
+    parser.add_argument(
+        '-Y', '--summary',
+        action='store_true',
+        help="Only show the total.")
+    parser.add_argument(
+        '-F', '--source',
+        dest='sources',
+        action='append',
+        help="Only consider definitions in this file. Defaults to anything "
+            "in the current directory.")
+    parser.add_argument(
+        '--everything',
+        action='store_true',
+        help="Include builtin and libc specific symbols.")
+    parser.add_argument(
+        '--self',
+        action='store_true',
+        help="Show samples before propagation up the call-chain.")
+    parser.add_argument(
+        '--branches',
+        action='store_true',
+        help="Show branches and branch misses.")
+    parser.add_argument(
+        '--caches',
+        action='store_true',
+        help="Show cache accesses and cache misses.")
+    parser.add_argument(
+        '-A', '--annotate',
+        action='store_true',
+        help="Show source files annotated with coverage info.")
+    parser.add_argument(
+        '-T', '--threshold',
+        nargs='?',
+        type=lambda x: tuple(float(x) for x in x.split(',')),
+        const=THRESHOLD,
+        help="Show lines wth samples above this threshold as a percent of "
+            "all lines. Defaults to %s." % ','.join(str(t) for t in THRESHOLD))
+    parser.add_argument(
+        '-c', '--context',
+        type=lambda x: int(x, 0),
+        default=3,
+        help="Show n additional lines of context. Defaults to 3.")
+    parser.add_argument(
+        '-W', '--width',
+        type=lambda x: int(x, 0),
+        default=80,
+        help="Assume source is styled with this many columns. Defaults to 80.")
+    parser.add_argument(
+        '--color',
+        choices=['never', 'always', 'auto'],
+        default='auto',
+        help="When to use terminal colors. Defaults to 'auto'.")
+    parser.add_argument(
+        '-j', '--jobs',
+        nargs='?',
+        type=lambda x: int(x, 0),
+        const=0,
+        help="Number of processes to use. 0 spawns one process per core.")
+    parser.add_argument(
+        '--perf-tool',
+        type=lambda x: x.split(),
+        help="Path to the perf tool to use. Defaults to %r." % PERF_TOOL)
+    parser.add_argument(
+        '--objdump-tool',
+        type=lambda x: x.split(),
+        default=OBJDUMP_TOOL,
+        help="Path to the objdump tool to use. Defaults to %r." % OBJDUMP_TOOL)
+
+    # record flags
+    record_parser = parser.add_argument_group('record options')
+    record_parser.add_argument(
+        'command',
+        nargs=nargs,
+        help="Command to run.")
+    record_parser.add_argument(
+        '-R', '--record',
+        action='store_true',
+        help="Run a command and aggregate perf measurements.")
+    record_parser.add_argument(
+        '-o', '--output',
+        help="Output file. Uses flock to synchronize. This is stored as a "
+            "zip-file of multiple perf results.")
+    record_parser.add_argument(
+        '--perf-freq',
+        help="perf sampling frequency. This is passed directly to perf. "
+            "Defaults to %r." % PERF_FREQ)
+    record_parser.add_argument(
+        '--perf-period',
+        help="perf sampling period. This is passed directly to perf.")
+    record_parser.add_argument(
+        '--perf-events',
+        help="perf events to record. This is passed directly to perf. "
+            "Defaults to %r." % PERF_EVENTS)
+    record_parser.add_argument(
+        '--perf-tool',
+        type=lambda x: x.split(),
+        help="Path to the perf tool to use. Defaults to %r." % PERF_TOOL)
+
+    # avoid intermixed/REMAINDER conflict, see above
+    if nargs == argparse.REMAINDER:
+        args = parser.parse_args()
+    else:
+        args = parser.parse_intermixed_args()
+
+    # perf_paths/command overlap, so need to do some munging here
+    args.command = args.perf_paths
+    args.perf_paths = args.perf_paths or PERF_PATHS
+
+    sys.exit(main(**{k: v
+        for k, v in vars(args).items()
+        if v is not None}))
diff --git a/scripts/plot.py b/scripts/plot.py
index 15ec846..52faf81 100755
--- a/scripts/plot.py
+++ b/scripts/plot.py
@@ -727,7 +727,8 @@ if __name__ == "__main__":
     import sys
     import argparse
     parser = argparse.ArgumentParser(
-        description="Plot CSV files in terminal.")
+        description="Plot CSV files in terminal.",
+        allow_abbrev=False)
     parser.add_argument(
         'csv_paths',
         nargs='*',
diff --git a/scripts/prettyasserts.py b/scripts/prettyasserts.py
index 73a43a1..f69c5ca 100755
--- a/scripts/prettyasserts.py
+++ b/scripts/prettyasserts.py
@@ -424,7 +424,8 @@ if __name__ == "__main__":
     import argparse
     import sys
     parser = argparse.ArgumentParser(
-        description="Preprocessor that makes asserts easier to debug.")
+        description="Preprocessor that makes asserts easier to debug.",
+        allow_abbrev=False)
     parser.add_argument(
         'input',
         help="Input C file.")
diff --git a/scripts/stack.py b/scripts/stack.py
index 6cb20ff..4da0f13 100755
--- a/scripts/stack.py
+++ b/scripts/stack.py
@@ -4,7 +4,7 @@
 # report as infinite stack usage.
 #
 # Example:
-# ./scripts/stack.py lfs.ci lfs_util.ci -S
+# ./scripts/stack.py lfs.ci lfs_util.ci -Slimit
 #
 # Copyright (c) 2022, The littlefs authors.
 # SPDX-License-Identifier: BSD-3-Clause
@@ -131,6 +131,7 @@ def openio(path, mode='r'):
         return open(path, mode)
 
 def collect(paths, *,
+        sources=None,
         everything=False,
         **args):
     # parse the vcg format
@@ -181,8 +182,9 @@ def collect(paths, *,
                         if (not args.get('quiet')
                                 and 'static' not in type
                                 and 'bounded' not in type):
-                            print("warning: found non-static stack for %s (%s)"
-                                % (function, type, size))
+                            print("warning: "
+                                "found non-static stack for %s (%s, %s)" % (
+                                function, type, size))
                         _, _, _, targets = callgraph[info['title']]
                         callgraph[info['title']] = (
                             file, function, int(size), targets)
@@ -193,11 +195,48 @@ def collect(paths, *,
                 else:
                     continue
 
+    callgraph_ = co.defaultdict(lambda: (None, None, 0, set()))
+    for source, (s_file, s_function, frame, targets) in callgraph.items():
+        # discard internal functions
+        if not everything and s_function.startswith('__'):
+            continue
+        # ignore filtered sources
+        if sources is not None:
+            if not any(
+                    os.path.abspath(s_file) == os.path.abspath(s)
+                    for s in sources):
+                continue
+        else:
+            # default to only cwd
+            if not everything and not os.path.commonpath([
+                    os.getcwd(),
+                    os.path.abspath(s_file)]) == os.getcwd():
+                continue
+
+        # smiplify path
+        if os.path.commonpath([
+                os.getcwd(),
+                os.path.abspath(s_file)]) == os.getcwd():
+            s_file = os.path.relpath(s_file)
+        else:
+            s_file = os.path.abspath(s_file)
+
+        callgraph_[source] = (s_file, s_function, frame, targets)
+    callgraph = callgraph_
+
     if not everything:
-        for source, (s_file, s_function, _, _) in list(callgraph.items()):
+        callgraph_ = co.defaultdict(lambda: (None, None, 0, set()))
+        for source, (s_file, s_function, frame, targets) in callgraph.items():
+            # discard filtered sources
+            if sources is not None and not any(
+                    os.path.abspath(s_file) == os.path.abspath(s)
+                    for s in sources):
+                continue
             # discard internal functions
-            if s_file.startswith('<') or s_file.startswith('/usr/include'):
-                del callgraph[source]
+            if s_function.startswith('__'):
+                continue
+            callgraph_[source] = (s_file, s_function, frame, targets)
+        callgraph = callgraph_
 
     # find maximum stack size recursively, this requires also detecting cycles
     # (in case of recursion)
@@ -278,7 +317,7 @@ def table(Result, results, diff_results=None, *,
         all=False,
         percent=False,
         tree=False,
-        depth=None,
+        depth=1,
         **_):
     all_, all = all, __builtins__.all
 
@@ -467,15 +506,8 @@ def table(Result, results, diff_results=None, *,
     # adjust the name width based on the expected call depth, though
     # note this doesn't really work with unbounded recursion
     if not summary:
-        # it doesn't really make sense to not have a depth with tree,
-        # so assume depth=inf if tree by default
-        if depth is None:
-            depth = m.inf if tree else 0
-        elif depth == 0:
-            depth = m.inf
-
         if not m.isinf(depth):
-            widths[0] += 4*depth
+            widths[0] += 4*(depth-1)
 
     # print our table with optional call info
     #
@@ -528,7 +560,7 @@ def table(Result, results, diff_results=None, *,
                          prefixes[2+is_last] + "'-> ",
                          prefixes[2+is_last] + "|   ",
                          prefixes[2+is_last] + "    "))
-        recurse(names, depth)
+        recurse(names, depth-1)
 
     if not tree:
         print('%-*s  %s%s' % (
@@ -544,6 +576,13 @@ def main(ci_paths,
         defines=None,
         sort=None,
         **args):
+    # it doesn't really make sense to not have a depth with tree,
+    # so assume depth=inf if tree by default
+    if args.get('depth') is None:
+        args['depth'] = m.inf if args['tree'] else 1
+    elif args.get('depth') == 0:
+        args['depth'] = m.inf
+
     # find sizes
     if not args.get('use', None):
         # find .ci files
@@ -588,13 +627,16 @@ def main(ci_paths,
     # write results to CSV
     if args.get('output'):
         with openio(args['output'], 'w') as f:
-            writer = csv.DictWriter(f, StackResult._by
+            writer = csv.DictWriter(f,
+                (by if by is not None else StackResult._by)
                 + ['stack_'+k for k in StackResult._fields])
             writer.writeheader()
             for r in results:
                 writer.writerow(
-                    {k: getattr(r, k) for k in StackResult._by}
-                    | {'stack_'+k: getattr(r, k) for k in StackResult._fields})
+                    {k: getattr(r, k)
+                        for k in (by if by is not None else StackResult._by)}
+                    | {'stack_'+k: getattr(r, k)
+                        for k in StackResult._fields})
 
     # find previous results?
     if args.get('diff'):
@@ -636,7 +678,8 @@ if __name__ == "__main__":
     import argparse
     import sys
     parser = argparse.ArgumentParser(
-        description="Find stack usage at the function level.")
+        description="Find stack usage at the function level.",
+        allow_abbrev=False)
     parser.add_argument(
         'ci_paths',
         nargs='*',
@@ -703,7 +746,13 @@ if __name__ == "__main__":
         action='store_true',
         help="Only show the total.")
     parser.add_argument(
-        '-A', '--everything',
+        '-F', '--source',
+        dest='sources',
+        action='append',
+        help="Only consider definitions in this file. Defaults to anything "
+            "in the current directory.")
+    parser.add_argument(
+        '--everything',
         action='store_true',
         help="Include builtin and libc specific symbols.")
     parser.add_argument(
@@ -711,20 +760,16 @@ if __name__ == "__main__":
         action='store_true',
         help="Only show the function call tree.")
     parser.add_argument(
-        '-L', '--depth',
+        '-Z', '--depth',
         nargs='?',
         type=lambda x: int(x, 0),
         const=0,
-        help="Depth of function calls to show. 0 show all calls but may not "
+        help="Depth of function calls to show. 0 shows all calls but may not "
             "terminate!")
     parser.add_argument(
         '-e', '--error-on-recursion',
         action='store_true',
         help="Error if any functions are recursive.")
-    parser.add_argument(
-        '--build-dir',
-        help="Specify the relative build directory. Used to map object files "
-            "to the correct source files.")
     sys.exit(main(**{k: v
         for k, v in vars(parser.parse_intermixed_args()).items()
         if v is not None}))
diff --git a/scripts/struct_.py b/scripts/struct_.py
index ed6584c..76a77ba 100755
--- a/scripts/struct_.py
+++ b/scripts/struct_.py
@@ -3,7 +3,7 @@
 # Script to find struct sizes.
 #
 # Example:
-# ./scripts/struct_.py lfs.o lfs_util.o -S
+# ./scripts/struct_.py lfs.o lfs_util.o -Ssize
 #
 # Copyright (c) 2022, The littlefs authors.
 # SPDX-License-Identifier: BSD-3-Clause
@@ -11,6 +11,7 @@
 
 import collections as co
 import csv
+import difflib
 import glob
 import itertools as it
 import math as m
@@ -128,26 +129,28 @@ def openio(path, mode='r'):
     else:
         return open(path, mode)
 
-def collect(paths, *,
+def collect(obj_paths, *,
         objdump_tool=OBJDUMP_TOOL,
-        build_dir=None,
+        sources=None,
         everything=False,
+        internal=False,
         **args):
-    decl_pattern = re.compile(
-        '^\s+(?P<no>[0-9]+)'
-            '\s+(?P<dir>[0-9]+)'
-            '\s+.*'
-            '\s+(?P<file>[^\s]+)$')
-    struct_pattern = re.compile(
-        '^(?:.*DW_TAG_(?P<tag>[a-z_]+).*'
+    line_pattern = re.compile(
+        '^\s+(?P<no>[0-9]+)\s+'
+            '(?:(?P<dir>[0-9]+)\s+)?'
+            '.*\s+'
+            '(?P<path>[^\s]+)$')
+    info_pattern = re.compile(
+        '^(?:.*(?P<tag>DW_TAG_[a-z_]+).*'
             '|^.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
-            '|^.*DW_AT_decl_file.*:\s*(?P<decl>[0-9]+)\s*'
+            '|^.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*'
             '|^.*DW_AT_byte_size.*:\s*(?P<size>[0-9]+)\s*)$')
 
     results = []
-    for path in paths:
-        # find decl, we want to filter by structs in .h files
-        decls = {}
+    for path in obj_paths:
+        # find files, we want to filter by structs in .h files
+        dirs = {}
+        files = {}
         # note objdump-tool may contain extra args
         cmd = objdump_tool + ['--dwarf=rawline', path]
         if args.get('verbose'):
@@ -156,12 +159,26 @@ def collect(paths, *,
             stdout=sp.PIPE,
             stderr=sp.PIPE if not args.get('verbose') else None,
             universal_newlines=True,
-            errors='replace')
+            errors='replace',
+            close_fds=False)
         for line in proc.stdout:
-            # find file numbers
-            m = decl_pattern.match(line)
+            # note that files contain references to dirs, which we
+            # dereference as soon as we see them as each file table follows a
+            # dir table
+            m = line_pattern.match(line)
             if m:
-                decls[int(m.group('no'))] = m.group('file')
+                if not m.group('dir'):
+                    # found a directory entry
+                    dirs[int(m.group('no'))] = m.group('path')
+                else:
+                    # found a file entry
+                    dir = int(m.group('dir'))
+                    if dir in dirs:
+                        files[int(m.group('no'))] = os.path.join(
+                            dirs[dir],
+                            m.group('path'))
+                    else:
+                        files[int(m.group('no'))] = m.group('path')
         proc.wait()
         if proc.returncode != 0:
             if not args.get('verbose'):
@@ -170,11 +187,11 @@ def collect(paths, *,
             sys.exit(-1)
 
         # collect structs as we parse dwarf info
-        found = False
-        name = None
-        decl = None
-        size = None
-
+        results_ = []
+        is_struct = False
+        s_name = None
+        s_file = None
+        s_size = None
         # note objdump-tool may contain extra args
         cmd = objdump_tool + ['--dwarf=info', path]
         if args.get('verbose'):
@@ -183,38 +200,23 @@ def collect(paths, *,
             stdout=sp.PIPE,
             stderr=sp.PIPE if not args.get('verbose') else None,
             universal_newlines=True,
-            errors='replace')
+            errors='replace',
+            close_fds=False)
         for line in proc.stdout:
             # state machine here to find structs
-            m = struct_pattern.match(line)
+            m = info_pattern.match(line)
             if m:
                 if m.group('tag'):
-                    if (name is not None
-                            and decl is not None
-                            and size is not None):
-                        file = decls.get(decl, '?')
-                        # map to source file
-                        file = re.sub('\.o$', '.c', file)
-                        if build_dir:
-                            file = re.sub(
-                                '%s/*' % re.escape(build_dir), '',
-                                file)
-                        # only include structs declared in header files in the
-                        # current directory, ignore internal-only structs (
-                        # these are represented in other measurements)
-                        if everything or file.endswith('.h'):
-                            results.append(StructResult(file, name, size))
-
-                    found = (m.group('tag') == 'structure_type')
-                    name = None
-                    decl = None
-                    size = None
-                elif found and m.group('name'):
-                    name = m.group('name')
-                elif found and name and m.group('decl'):
-                    decl = int(m.group('decl'))
-                elif found and name and m.group('size'):
-                    size = int(m.group('size'))
+                    if is_struct:
+                        file = files.get(s_file, '?')
+                        results_.append(StructResult(file, s_name, s_size))
+                    is_struct = (m.group('tag') == 'DW_TAG_structure_type')
+                elif m.group('name'):
+                    s_name = m.group('name')
+                elif m.group('file'):
+                    s_file = int(m.group('file'))
+                elif m.group('size'):
+                    s_size = int(m.group('size'))
         proc.wait()
         if proc.returncode != 0:
             if not args.get('verbose'):
@@ -222,6 +224,34 @@ def collect(paths, *,
                     sys.stdout.write(line)
             sys.exit(-1)
 
+        for r in results_:
+            # ignore filtered sources
+            if sources is not None:
+                if not any(
+                        os.path.abspath(r.file) == os.path.abspath(s)
+                        for s in sources):
+                    continue
+            else:
+                # default to only cwd
+                if not everything and not os.path.commonpath([
+                        os.getcwd(),
+                        os.path.abspath(r.file)]) == os.getcwd():
+                    continue
+
+                # limit to .h files unless --internal
+                if not internal and not r.file.endswith('.h'):
+                    continue
+
+            # simplify path
+            if os.path.commonpath([
+                    os.getcwd(),
+                    os.path.abspath(r.file)]) == os.getcwd():
+                file = os.path.relpath(r.file)
+            else:
+                file = os.path.abspath(r.file)
+
+            results.append(StructResult(r.file, r.struct, r.size))
+
     return results
 
 
@@ -479,7 +509,7 @@ def main(obj_paths, *,
                 paths.append(path)
 
         if not paths:
-            print("error: no .obj files found in %r?" % obj_paths)
+            print("error: no .o files found in %r?" % obj_paths)
             sys.exit(-1)
 
         results = collect(paths, **args)
@@ -513,12 +543,14 @@ def main(obj_paths, *,
     # write results to CSV
     if args.get('output'):
         with openio(args['output'], 'w') as f:
-            writer = csv.DictWriter(f, StructResult._by
+            writer = csv.DictWriter(f,
+                (by if by is not None else StructResult._by)
                 + ['struct_'+k for k in StructResult._fields])
             writer.writeheader()
             for r in results:
                 writer.writerow(
-                    {k: getattr(r, k) for k in StructResult._by}
+                    {k: getattr(r, k)
+                        for k in (by if by is not None else StructResult._by)}
                     | {'struct_'+k: getattr(r, k)
                         for k in StructResult._fields})
 
@@ -559,7 +591,8 @@ if __name__ == "__main__":
     import argparse
     import sys
     parser = argparse.ArgumentParser(
-        description="Find struct sizes.")
+        description="Find struct sizes.",
+        allow_abbrev=False)
     parser.add_argument(
         'obj_paths',
         nargs='*',
@@ -626,18 +659,24 @@ if __name__ == "__main__":
         action='store_true',
         help="Only show the total.")
     parser.add_argument(
-        '-A', '--everything',
+        '-F', '--source',
+        dest='sources',
+        action='append',
+        help="Only consider definitions in this file. Defaults to anything "
+            "in the current directory.")
+    parser.add_argument(
+        '--everything',
         action='store_true',
         help="Include builtin and libc specific symbols.")
     parser.add_argument(
+        '--internal',
+        action='store_true',
+        help="Also show structs in .c files.")
+    parser.add_argument(
         '--objdump-tool',
         type=lambda x: x.split(),
         default=OBJDUMP_TOOL,
         help="Path to the objdump tool to use. Defaults to %r." % OBJDUMP_TOOL)
-    parser.add_argument(
-        '--build-dir',
-        help="Specify the relative build directory. Used to map object files "
-            "to the correct source files.")
     sys.exit(main(**{k: v
         for k, v in vars(parser.parse_intermixed_args()).items()
         if v is not None}))
diff --git a/scripts/summary.py b/scripts/summary.py
index 9a13703..ba7fd40 100755
--- a/scripts/summary.py
+++ b/scripts/summary.py
@@ -750,7 +750,8 @@ if __name__ == "__main__":
     import argparse
     import sys
     parser = argparse.ArgumentParser(
-        description="Summarize measurements in CSV files.")
+        description="Summarize measurements in CSV files.",
+        allow_abbrev=False)
     parser.add_argument(
         'csv_paths',
         nargs='*',
diff --git a/scripts/tailpipe.py b/scripts/tailpipe.py
index 7e8c454..69e6455 100755
--- a/scripts/tailpipe.py
+++ b/scripts/tailpipe.py
@@ -115,7 +115,8 @@ if __name__ == "__main__":
     import sys
     import argparse
     parser = argparse.ArgumentParser(
-        description="Efficiently displays the last n lines of a file/pipe.")
+        description="Efficiently displays the last n lines of a file/pipe.",
+        allow_abbrev=False)
     parser.add_argument(
         'path',
         nargs='?',
diff --git a/scripts/test.py b/scripts/test.py
index eaab3e3..358f0e8 100755
--- a/scripts/test.py
+++ b/scripts/test.py
@@ -27,9 +27,13 @@ import time
 import toml
 
 
-RUNNER_PATH = 'runners/test_runner'
+RUNNER_PATH = './runners/test_runner'
 HEADER_PATH = 'runners/test_runner.h'
 
+GDB_TOOL = ['gdb']
+VALGRIND_TOOL = ['valgrind']
+PERF_SCRIPT = ['./scripts/perf.py']
+
 
 def openio(path, mode='r', buffering=-1, nb=False):
     if path == '-':
@@ -516,12 +520,25 @@ def find_runner(runner, **args):
 
     # run under valgrind?
     if args.get('valgrind'):
-        cmd[:0] = filter(None, [
-            'valgrind',
+        cmd[:0] = args['valgrind_tool'] + [
             '--leak-check=full',
             '--track-origins=yes',
             '--error-exitcode=4',
-            '-q'])
+            '-q']
+
+    # run under perf?
+    if args.get('perf'):
+        cmd[:0] = args['perf_script'] + list(filter(None, [
+            '-R',
+            '--perf-freq=%s' % args['perf_freq']
+                if args.get('perf_freq') else None,
+            '--perf-period=%s' % args['perf_period']
+                if args.get('perf_period') else None,
+            '--perf-events=%s' % args['perf_events']
+                if args.get('perf_events') else None,
+            '--perf-tool=%s' % args['perf_tool']
+                if args.get('perf_tool') else None,
+            '-o%s' % args['perf']]))
 
     # other context
     if args.get('geometry'):
@@ -799,9 +816,9 @@ def run_stage(name, runner_, ids, output_, **args):
                 try:
                     line = mpty.readline()
                 except OSError as e:
-                    if e.errno == errno.EIO:
-                        break
-                    raise
+                    if e.errno != errno.EIO:
+                        raise
+                    break
                 if not line:
                     break
                 last_stdout.append(line)
@@ -1126,24 +1143,24 @@ def run(runner, test_ids=[], **args):
         cmd = runner_ + [failure.id]
 
         if args.get('gdb_main'):
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'break main',
                 '-ex', 'run',
                 '--args']
         elif args.get('gdb_case'):
             path, lineno = find_path(runner_, failure.id, **args)
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'break %s:%d' % (path, lineno),
                 '-ex', 'run',
                 '--args']
         elif failure.assert_ is not None:
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'run',
                 '-ex', 'frame function raise',
                 '-ex', 'up 2',
                 '--args']
         else:
-            cmd[:0] = ['gdb',
+            cmd[:0] = args['gdb_tool'] + [
                 '-ex', 'run',
                 '--args']
 
@@ -1188,6 +1205,7 @@ if __name__ == "__main__":
     argparse._ArgumentGroup._handle_conflict_ignore = lambda *_: None
     parser = argparse.ArgumentParser(
         description="Build and run tests.",
+        allow_abbrev=False,
         conflict_handler='ignore')
     parser.add_argument(
         '-v', '--verbose',
@@ -1324,6 +1342,11 @@ if __name__ == "__main__":
         help="Drop into gdb on test failure but stop at the beginning "
             "of main.")
     test_parser.add_argument(
+        '--gdb-tool',
+        type=lambda x: x.split(),
+        default=GDB_TOOL,
+        help="Path to gdb tool to use. Defaults to %r." % GDB_TOOL)
+    test_parser.add_argument(
         '--exec',
         type=lambda e: e.split(),
         help="Run under another executable.")
@@ -1332,6 +1355,37 @@ if __name__ == "__main__":
         action='store_true',
         help="Run under Valgrind to find memory errors. Implicitly sets "
             "--isolate.")
+    test_parser.add_argument(
+        '--valgrind-tool',
+        type=lambda x: x.split(),
+        default=VALGRIND_TOOL,
+        help="Path to Valgrind tool to use. Defaults to %r." % VALGRIND_TOOL)
+    test_parser.add_argument(
+        '--perf',
+        help="Run under Linux's perf to sample performance counters, writing "
+            "samples to this file.")
+    test_parser.add_argument(
+        '--perf-freq',
+        help="perf sampling frequency. This is passed directly to the perf "
+            "script.")
+    test_parser.add_argument(
+        '--perf-period',
+        help="perf sampling period. This is passed directly to the perf "
+            "script.")
+    test_parser.add_argument(
+        '--perf-events',
+        help="perf events to record. This is passed directly to the perf "
+            "script.")
+    test_parser.add_argument(
+        '--perf-script',
+        type=lambda x: x.split(),
+        default=PERF_SCRIPT,
+        help="Path to the perf script to use. Defaults to %r." % PERF_SCRIPT)
+    test_parser.add_argument(
+        '--perf-tool',
+        type=lambda x: x.split(),
+        help="Path to the perf tool to use. This is passed directly to the "
+            "perf script")
 
     # compilation flags
     comp_parser = parser.add_argument_group('compilation options')
@@ -1356,7 +1410,7 @@ if __name__ == "__main__":
         '-o', '--output',
         help="Output file.")
 
-    # runner + test_ids overlaps test_paths, so we need to do some munging here
+    # runner/test_paths overlap, so need to do some munging here
     args = parser.parse_intermixed_args()
     args.test_paths = [' '.join(args.runner or [])] + args.test_ids
     args.runner = args.runner or [RUNNER_PATH]
diff --git a/scripts/tracebd.py b/scripts/tracebd.py
index d69c132..556cd0d 100755
--- a/scripts/tracebd.py
+++ b/scripts/tracebd.py
@@ -830,7 +830,8 @@ if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser(
         description="Display operations on block devices based on "
-            "trace output.")
+            "trace output.",
+        allow_abbrev=False)
     parser.add_argument(
         'path',
         nargs='?',