blob: f06884f8dafe1f7ba54c8586d18882c0e0ec5227 [file] [log] [blame]
#!/usr/bin/env python
# Copyright 2014 The Chromium Authors. All rights reserved.
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
"""Generate a spatial analysis against an arbitrary library.
To use, build the 'binary_size_tool' target. Then run this tool, passing
in the location of the library to be analyzed along with any other options
you desire.
import collections
import json
import logging
import multiprocessing
import optparse
import os
import re
import shutil
import subprocess
import sys
import tempfile
import time
import binary_size_utils
# This path changee is not beautiful. Temporary (I hope) measure until
# the chromium project has figured out a proper way to organize the
# library of python tools.
elf_symbolizer_path = os.path.abspath(os.path.join(
import symbols.elf_symbolizer as elf_symbolizer # pylint: disable=F0401
# Node dictionary keys. These are output in json read by the webapp so
# keep them short to save file size.
# Note: If these change, the webapp must also change.
NODE_CHILDREN_KEY = 'children'
# The display name of the bucket where we put symbols without path.
# Try to keep data buckets smaller than this to avoid killing the
# graphing lib.
# TODO(andrewhayden): Only used for legacy reports. Delete.
def FormatBytes(byte_count):
"""Pretty-print a number of bytes."""
if byte_count > 1e6:
byte_count = byte_count / 1.0e6
return '%.1fm' % byte_count
if byte_count > 1e3:
byte_count = byte_count / 1.0e3
return '%.1fk' % byte_count
return str(byte_count)
# TODO(andrewhayden): Only used for legacy reports. Delete.
def SymbolTypeToHuman(symbol_type):
"""Convert a symbol type as printed by nm into a human-readable name."""
return {'b': 'bss',
'd': 'data',
'r': 'read-only data',
't': 'code',
'w': 'weak symbol',
'v': 'weak symbol'}[symbol_type]
def _MkChild(node, name):
child = node[NODE_CHILDREN_KEY].get(name)
if child is None:
child = {NODE_NAME_KEY: name,
node[NODE_CHILDREN_KEY][name] = child
return child
def SplitNoPathBucket(node):
"""NAME_NO_PATH_BUCKET can be too large for the graphing lib to
handle. Split it into sub-buckets in that case."""
root_children = node[NODE_CHILDREN_KEY]
if NAME_NO_PATH_BUCKET in root_children:
no_path_bucket = root_children[NAME_NO_PATH_BUCKET]
old_children = no_path_bucket[NODE_CHILDREN_KEY]
count = 0
for symbol_type, symbol_bucket in old_children.iteritems():
count += len(symbol_bucket[NODE_CHILDREN_KEY])
if count > BIG_BUCKET_LIMIT:
new_children = {}
no_path_bucket[NODE_CHILDREN_KEY] = new_children
current_bucket = None
index = 0
for symbol_type, symbol_bucket in old_children.iteritems():
for symbol_name, value in symbol_bucket[NODE_CHILDREN_KEY].iteritems():
if index % BIG_BUCKET_LIMIT == 0:
group_no = (index / BIG_BUCKET_LIMIT) + 1
current_bucket = _MkChild(no_path_bucket,
'%s subgroup %d' % (NAME_NO_PATH_BUCKET,
assert not NODE_TYPE_KEY in node or node[NODE_TYPE_KEY] == 'p'
node[NODE_TYPE_KEY] = 'p' # p for path
index += 1
symbol_size = value[NODE_SYMBOL_SIZE_KEY]
AddSymbolIntoFileNode(current_bucket, symbol_type,
symbol_name, symbol_size)
def MakeChildrenDictsIntoLists(node):
largest_list_len = 0
largest_list_len = len(node[NODE_CHILDREN_KEY])
child_list = []
for child in node[NODE_CHILDREN_KEY].itervalues():
child_largest_list_len = MakeChildrenDictsIntoLists(child)
if child_largest_list_len > largest_list_len:
largest_list_len = child_largest_list_len
node[NODE_CHILDREN_KEY] = child_list
return largest_list_len
def AddSymbolIntoFileNode(node, symbol_type, symbol_name, symbol_size):
"""Puts symbol into the file path node |node|.
Returns the number of added levels in tree. I.e. returns 2."""
# 'node' is the file node and first step is to find its symbol-type bucket.
node = _MkChild(node, symbol_type)
assert not NODE_TYPE_KEY in node or node[NODE_TYPE_KEY] == 'b'
node[NODE_SYMBOL_TYPE_KEY] = symbol_type
node[NODE_TYPE_KEY] = 'b' # b for bucket
# 'node' is now the symbol-type bucket. Make the child entry.
node = _MkChild(node, symbol_name)
logging.warning('A container node used as symbol for %s.' % symbol_name)
# This is going to be used as a leaf so no use for child list.
node[NODE_SYMBOL_SIZE_KEY] = symbol_size
node[NODE_SYMBOL_TYPE_KEY] = symbol_type
node[NODE_TYPE_KEY] = 's' # s for symbol
return 2 # Depth of the added subtree.
def MakeCompactTree(symbols):
result = {NODE_NAME_KEY: '/',
seen_symbol_with_path = False
for symbol_name, symbol_type, symbol_size, file_path in symbols:
if 'vtable for ' in symbol_name:
symbol_type = '@' # hack to categorize these separately
# Take path like '/foo/bar/baz', convert to ['foo', 'bar', 'baz']
if file_path:
file_path = os.path.normpath(file_path)
seen_symbol_with_path = True
if file_path.startswith('/'):
file_path = file_path[1:]
path_parts = file_path.split('/')
# Find pre-existing node in tree, or update if it already exists
node = result
depth = 0
while len(path_parts) > 0:
path_part = path_parts.pop(0)
if len(path_part) == 0:
depth += 1
node = _MkChild(node, path_part)
assert not NODE_TYPE_KEY in node or node[NODE_TYPE_KEY] == 'p'
node[NODE_TYPE_KEY] = 'p' # p for path
depth += AddSymbolIntoFileNode(node, symbol_type, symbol_name, symbol_size)
result[NODE_MAX_DEPTH_KEY] = max(result[NODE_MAX_DEPTH_KEY], depth)
if not seen_symbol_with_path:
logging.warning('Symbols lack paths. Data will not be structured.')
# The (no path) bucket can be extremely large if we failed to get
# path information. Split it into subgroups if needed.
largest_list_len = MakeChildrenDictsIntoLists(result)
if largest_list_len > BIG_BUCKET_LIMIT:
logging.warning('There are sections with %d nodes. '
'Results might be unusable.' % largest_list_len)
return result
# TODO(andrewhayden): Only used for legacy reports. Delete.
def TreeifySymbols(symbols):
"""Convert symbols into a path-based tree, calculating size information
along the way.
The result is a dictionary that contains two kinds of nodes:
1. Leaf nodes, representing source code locations (e.g., c++ files)
These nodes have the following dictionary entries:
sizes: a dictionary whose keys are categories (such as code, data,
vtable, etceteras) and whose values are the size, in bytes, of
those categories;
size: the total size, in bytes, of all the entries in the sizes dict
2. Non-leaf nodes, representing directories
These nodes have the following dictionary entries:
children: a dictionary whose keys are names (path entries; either
directory or file names) and whose values are other nodes;
size: the total size, in bytes, of all the leaf nodes that are
contained within the children dict (recursively expanded)
The result object is itself a dictionary that represents the common ancestor
of all child nodes, e.g. a path to which all other nodes beneath it are
relative. The 'size' attribute of this dict yields the sum of the size of all
leaf nodes within the data structure.
dirs = {'children': {}, 'size': 0}
for sym, symbol_type, size, path in symbols:
dirs['size'] += size
if path:
path = os.path.normpath(path)
if path.startswith('/'):
path = path[1:]
parts = None
if path:
parts = path.split('/')
if parts:
assert path
file_key = parts.pop()
tree = dirs
# Traverse the tree to the parent of the file node, creating as needed
for part in parts:
assert part != ''
if part not in tree['children']:
tree['children'][part] = {'children': {}, 'size': 0}
tree = tree['children'][part]
tree['size'] += size
# Get (creating if necessary) the node for the file
# This node doesn't have a 'children' attribute
if file_key not in tree['children']:
tree['children'][file_key] = {'sizes': collections.defaultdict(int),
'size': 0}
tree = tree['children'][file_key]
tree['size'] += size
# Accumulate size into a bucket within the file
symbol_type = symbol_type.lower()
if 'vtable for ' in sym:
tree['sizes']['[vtable]'] += size
elif 'r' == symbol_type:
tree['sizes']['[rodata]'] += size
elif 'd' == symbol_type:
tree['sizes']['[data]'] += size
elif 'b' == symbol_type:
tree['sizes']['[bss]'] += size
elif 't' == symbol_type:
# 'text' in binary parlance means 'code'.
tree['sizes']['[code]'] += size
elif 'w' == symbol_type:
tree['sizes']['[weak]'] += size
tree['sizes']['[other]'] += size
print >> sys.stderr, sym, parts, file_key
key = 'symbols without paths'
if key not in dirs['children']:
dirs['children'][key] = {'sizes': collections.defaultdict(int),
'size': 0}
tree = dirs['children'][key]
subkey = 'misc'
if (sym.endswith('::__FUNCTION__') or
subkey = '__FUNCTION__'
elif sym.startswith('CSWTCH.'):
subkey = 'CSWTCH'
elif '::' in sym:
subkey = sym[0:sym.find('::') + 2]
tree['sizes'][subkey] = tree['sizes'].get(subkey, 0) + size
tree['size'] += size
return dirs
# TODO(andrewhayden): Only used for legacy reports. Delete.
def JsonifyTree(tree, name):
"""Convert TreeifySymbols output to a JSON treemap.
The format is very similar, with the notable exceptions being
lists of children instead of maps and some different attribute names."""
children = []
css_class_map = {
'[vtable]': 'vtable',
'[rodata]': 'read-only_data',
'[data]': 'data',
'[bss]': 'bss',
'[code]': 'code',
'[weak]': 'weak_symbol'
if 'children' in tree:
# Non-leaf node. Recurse.
for child_name, child in tree['children'].iteritems():
children.append(JsonifyTree(child, child_name))
# Leaf node; dump per-file stats as entries in the treemap
for kind, size in tree['sizes'].iteritems():
child_json = {'name': kind + ' (' + FormatBytes(size) + ')',
'data': { '$area': size }}
css_class = css_class_map.get(kind)
if css_class is not None:
child_json['data']['$symbol'] = css_class
# Sort children by size, largest to smallest.
children.sort(key=lambda child: -child['data']['$area'])
# For leaf nodes, the 'size' attribute is the size of the leaf;
# Non-leaf nodes don't really have a size, but their 'size' attribute is
# the sum of the sizes of all their children.
return {'name': name + ' (' + FormatBytes(tree['size']) + ')',
'data': { '$area': tree['size'] },
'children': children }
def DumpCompactTree(symbols, outfile):
tree_root = MakeCompactTree(symbols)
with open(outfile, 'w') as out:
out.write('var tree_data = ')
json.dump(tree_root, out)
print('Writing %d bytes json' % os.path.getsize(outfile))
# TODO(andrewhayden): Only used for legacy reports. Delete.
def DumpTreemap(symbols, outfile):
dirs = TreeifySymbols(symbols)
out = open(outfile, 'w')
out.write('var kTree = ' + json.dumps(JsonifyTree(dirs, '/')))
# TODO(andrewhayden): Only used for legacy reports. Delete.
def DumpLargestSymbols(symbols, outfile, n):
# a list of (sym, symbol_type, size, path); sort by size.
symbols = sorted(symbols, key=lambda x: -x[2])
dumped = 0
out = open(outfile, 'w')
out.write('var largestSymbols = [\n')
for sym, symbol_type, size, path in symbols:
if symbol_type in ('b', 'w'):
continue # skip bss and weak symbols
if path is None:
path = ''
entry = {'size': FormatBytes(size),
'symbol': sym,
'type': SymbolTypeToHuman(symbol_type),
'location': path }
dumped += 1
if dumped >= n:
def MakeSourceMap(symbols):
sources = {}
for _sym, _symbol_type, size, path in symbols:
key = None
if path:
key = os.path.normpath(path)
key = '[no path]'
if key not in sources:
sources[key] = {'path': path, 'symbol_count': 0, 'size': 0}
record = sources[key]
record['size'] += size
record['symbol_count'] += 1
return sources
# TODO(andrewhayden): Only used for legacy reports. Delete.
def DumpLargestSources(symbols, outfile, n):
source_map = MakeSourceMap(symbols)
sources = sorted(source_map.values(), key=lambda x: -x['size'])
dumped = 0
out = open(outfile, 'w')
out.write('var largestSources = [\n')
for record in sources:
entry = {'size': FormatBytes(record['size']),
'symbol_count': str(record['symbol_count']),
'location': record['path']}
dumped += 1
if dumped >= n:
# TODO(andrewhayden): Only used for legacy reports. Delete.
def DumpLargestVTables(symbols, outfile, n):
vtables = []
for symbol, _type, size, path in symbols:
if 'vtable for ' in symbol:
vtables.append({'symbol': symbol, 'path': path, 'size': size})
vtables = sorted(vtables, key=lambda x: -x['size'])
dumped = 0
out = open(outfile, 'w')
out.write('var largestVTables = [\n')
for record in vtables:
entry = {'size': FormatBytes(record['size']),
'symbol': record['symbol'],
'location': record['path']}
dumped += 1
if dumped >= n:
# Regex for parsing "nm" output. A sample line looks like this:
# 0167b39c 00000018 t ACCESS_DESCRIPTION_free /path/file.c:95
# The fields are: address, size, type, name, source location
# Regular expression explained ( see also: ):
# ([0-9a-f]{8,}+) The address
# [\s]+ Whitespace separator
# ([0-9a-f]{8,}+) The size. From here on out it's all optional.
# [\s]+ Whitespace separator
# (\S?) The symbol type, which is any non-whitespace char
# [\s*] Whitespace separator
# ([^\t]*) Symbol name, any non-tab character (spaces ok!)
# [\t]? Tab separator
# (.*) The location (filename[:linennum|?][ (discriminator n)]
sNmPattern = re.compile(
class Progress():
def __init__(self):
self.count = 0
self.skip_count = 0
self.collisions = 0
self.time_last_output = time.time()
self.count_last_output = 0
def RunElfSymbolizer(outfile, library, addr2line_binary, nm_binary, jobs):
nm_output = RunNm(library, nm_binary)
nm_output_lines = nm_output.splitlines()
nm_output_lines_len = len(nm_output_lines)
address_symbol = {}
progress = Progress()
def map_address_symbol(symbol, addr):
progress.count += 1
if addr in address_symbol:
# 'Collision between %s and %s.' % (str(,
# str(address_symbol[addr].name))
progress.collisions += 1
address_symbol[addr] = symbol
progress_chunk = 100
if progress.count % progress_chunk == 0:
time_now = time.time()
time_spent = time_now - progress.time_last_output
if time_spent > 1.0:
# Only output at most once per second.
progress.time_last_output = time_now
chunk_size = progress.count - progress.count_last_output
progress.count_last_output = progress.count
if time_spent > 0:
speed = chunk_size / time_spent
speed = 0
progress_percent = (100.0 * (progress.count + progress.skip_count) /
print('%.1f%%: Looked up %d symbols (%d collisions) - %.1f lookups/s.' %
(progress_percent, progress.count, progress.collisions, speed))
symbolizer = elf_symbolizer.ELFSymbolizer(library, addr2line_binary,
user_interrupted = False
for line in nm_output_lines:
match = sNmPattern.match(line)
if match:
location =
if not location:
addr = int(, 16)
size = int(, 16)
if addr in address_symbol: # Already looked up, shortcut
# ELFSymbolizer.
map_address_symbol(address_symbol[addr], addr)
elif size == 0:
# Save time by not looking up empty symbols (do they even exist?)
print('Empty symbol: ' + line)
symbolizer.SymbolizeAsync(addr, addr)
progress.skip_count += 1
except KeyboardInterrupt:
user_interrupted = True
print('Interrupting - killing subprocesses. Please wait.')
except KeyboardInterrupt:
# Don't want to abort here since we will be finished in a few seconds.
user_interrupted = True
print('Patience you must have my young padawan.')
if user_interrupted:
print('Skipping the rest of the file mapping. '
'Output will not be fully classified.')
with open(outfile, 'w') as out:
for line in nm_output_lines:
match = sNmPattern.match(line)
if match:
location =
if not location:
addr = int(, 16)
symbol = address_symbol.get(addr)
if symbol is not None:
path = '??'
if symbol.source_path is not None:
path = symbol.source_path
line_number = 0
if symbol.source_line is not None:
line_number = symbol.source_line
out.write('%s\t%s:%d\n' % (line, path, line_number))
out.write('%s\n' % line)
print('%d symbols in the results.' % len(address_symbol))
def RunNm(binary, nm_binary):
print('Starting nm')
cmd = [nm_binary, '-C', '--print-size', '--size-sort', '--reverse-sort',
nm_process = subprocess.Popen(cmd,
(process_output, err_output) = nm_process.communicate()
if nm_process.returncode != 0:
if err_output:
raise Exception, err_output
raise Exception, process_output
print('Finished nm')
return process_output
def GetNmSymbols(nm_infile, outfile, library, jobs, verbose,
addr2line_binary, nm_binary):
if nm_infile is None:
if outfile is None:
outfile = tempfile.NamedTemporaryFile(delete=False).name
if verbose:
print 'Running parallel addr2line, dumping symbols to ' + outfile
RunElfSymbolizer(outfile, library, addr2line_binary, nm_binary, jobs)
nm_infile = outfile
elif verbose:
print 'Using nm input from ' + nm_infile
with file(nm_infile, 'r') as infile:
return list(binary_size_utils.ParseNm(infile))
def _find_in_system_path(binary):
"""Locate the full path to binary in the system path or return None
if not found."""
system_path = os.environ["PATH"].split(os.pathsep)
for path in system_path:
binary_path = os.path.join(path, binary)
if os.path.isfile(binary_path):
return binary_path
return None
def CheckDebugFormatSupport(library, addr2line_binary):
"""Kills the program if debug data is in an unsupported format.
There are two common versions of the DWARF debug formats and
since we are right now transitioning from DWARF2 to newer formats,
it's possible to have a mix of tools that are not compatible. Detect
that and abort rather than produce meaningless output."""
tool_output = subprocess.check_output([addr2line_binary, '--version'])
version_re = re.compile(r'^GNU [^ ]+ .* (\d+).(\d+).*?$', re.M)
parsed_output = version_re.match(tool_output)
major = int(
minor = int(
supports_dwarf4 = major > 2 or major == 2 and minor > 22
if supports_dwarf4:
print('Checking version of debug information in %s.' % library)
debug_info = subprocess.check_output(['readelf', '--debug-dump=info',
'--dwarf-depth=1', library])
dwarf_version_re = re.compile(r'^\s+Version:\s+(\d+)$', re.M)
parsed_dwarf_format_output =
version = int(
if version > 2:
print('The supplied tools only support DWARF2 debug data but the binary\n' +
'uses DWARF%d. Update the tools or compile the binary\n' % version +
'with -gdwarf-2.')
def main():
usage = """%prog [options]
Runs a spatial analysis on a given library, looking up the source locations
of its symbols and calculating how much space each directory, source file,
and so on is taking. The result is a report that can be used to pinpoint
sources of large portions of the binary, etceteras.
Under normal circumstances, you only need to pass two arguments, thusly:
%prog --library /path/to/library --destdir /path/to/output
In this mode, the program will dump the symbols from the specified library
and map those symbols back to source locations, producing a web-based
report in the specified output directory.
Other options are available via '--help'.
parser = optparse.OptionParser(usage=usage)
parser.add_option('--nm-in', metavar='PATH',
help='if specified, use nm input from <path> instead of '
'generating it. Note that source locations should be '
'present in the file; i.e., no addr2line symbol lookups '
'will be performed when this option is specified. '
'Mutually exclusive with --library.')
parser.add_option('--destdir', metavar='PATH',
help='write output to the specified directory. An HTML '
'report is generated here along with supporting files; '
'any existing report will be overwritten.')
parser.add_option('--library', metavar='PATH',
help='if specified, process symbols in the library at '
'the specified path. Mutually exclusive with --nm-in.')
help='use the specified nm binary to analyze library. '
'This is to be used when the nm in the path is not for '
'the right architecture or of the right version.')
help='use the specified addr2line binary to analyze '
'library. This is to be used when the addr2line in '
'the path is not for the right architecture or '
'of the right version.')
parser.add_option('--jobs', type='int',
help='number of jobs to use for the parallel '
'addr2line processing pool; defaults to 1. More '
'jobs greatly improve throughput but eat RAM like '
'popcorn, and take several gigabytes each. Start low '
'and ramp this number up until your machine begins to '
'struggle with RAM. '
'This argument is only valid when using --library.')
parser.add_option('-v', dest='verbose', action='store_true',
help='be verbose, printing lots of status information.')
parser.add_option('--nm-out', metavar='PATH',
help='keep the nm output file, and store it at the '
'specified path. This is useful if you want to see the '
'fully processed nm output after the symbols have been '
'mapped to source locations. By default, a tempfile is '
'used and is deleted when the program terminates.'
'This argument is only valid when using --library.')
parser.add_option('--legacy', action='store_true',
help='emit legacy binary size report instead of modern')
opts, _args = parser.parse_args()
if ((not opts.library) and (not opts.nm_in)) or (opts.library and opts.nm_in):
parser.error('exactly one of --library or --nm-in is required')
if (opts.nm_in):
print >> sys.stderr, ('WARNING: --jobs has no effect '
'when used with --nm-in')
if not opts.destdir:
parser.error('--destdir is required argument')
if not
# Use the number of processors but cap between 2 and 4 since raw
# CPU power isn't the limiting factor. It's I/O limited, memory
# bus limited and available-memory-limited. Too many processes and
# the computer will run out of memory and it will be slow. = max(2, min(4, str(multiprocessing.cpu_count())))
if opts.addr2line_binary:
assert os.path.isfile(opts.addr2line_binary)
addr2line_binary = opts.addr2line_binary
addr2line_binary = _find_in_system_path('addr2line')
assert addr2line_binary, 'Unable to find addr2line in the path. '\
'Use --addr2line-binary to specify location.'
if opts.nm_binary:
assert os.path.isfile(opts.nm_binary)
nm_binary = opts.nm_binary
nm_binary = _find_in_system_path('nm')
assert nm_binary, 'Unable to find nm in the path. Use --nm-binary '\
'to specify location.'
print('addr2line: %s' % addr2line_binary)
print('nm: %s' % nm_binary)
CheckDebugFormatSupport(opts.library, addr2line_binary)
symbols = GetNmSymbols(opts.nm_in, opts.nm_out, opts.library,, opts.verbose is True,
addr2line_binary, nm_binary)
if not os.path.exists(opts.destdir):
os.makedirs(opts.destdir, 0755)
if opts.legacy: # legacy report
DumpTreemap(symbols, os.path.join(opts.destdir, 'treemap-dump.js'))
os.path.join(opts.destdir, 'largest-symbols.js'), 100)
os.path.join(opts.destdir, 'largest-sources.js'), 100)
os.path.join(opts.destdir, 'largest-vtables.js'), 100)
treemap_out = os.path.join(opts.destdir, 'webtreemap')
if not os.path.exists(treemap_out):
os.makedirs(treemap_out, 0755)
treemap_src = os.path.join('third_party', 'webtreemap', 'src')
shutil.copy(os.path.join(treemap_src, 'COPYING'), treemap_out)
shutil.copy(os.path.join(treemap_src, 'webtreemap.js'), treemap_out)
shutil.copy(os.path.join(treemap_src, 'webtreemap.css'), treemap_out)
shutil.copy(os.path.join('tools', 'binary_size', 'legacy_template',
'index.html'), opts.destdir)
else: # modern report
DumpCompactTree(symbols, os.path.join(opts.destdir, 'data.js'))
d3_out = os.path.join(opts.destdir, 'd3')
if not os.path.exists(d3_out):
os.makedirs(d3_out, 0755)
d3_src = os.path.join(os.path.dirname(__file__),
'third_party', 'd3', 'src')
template_src = os.path.join(os.path.dirname(__file__),
shutil.copy(os.path.join(d3_src, 'LICENSE'), d3_out)
shutil.copy(os.path.join(d3_src, 'd3.js'), d3_out)
shutil.copy(os.path.join(template_src, 'index.html'), opts.destdir)
shutil.copy(os.path.join(template_src, 'D3SymbolTreeMap.js'), opts.destdir)
print 'Report saved to ' + opts.destdir + '/index.html'
if __name__ == '__main__':