pytorch

1
"""A tool to inspect the binary size of a built binary file.
2

3
This script prints out a tree of symbols and their corresponding sizes, using
4
Linux's nm functionality.
5

6
Usage:
7

8
    python binary_size.py -- \
9
            --target=/path/to/your/target/binary \
10
            [--nm_command=/path/to/your/custom/nm] \
11
            [--max_depth=10] [--min_size=1024] \
12
            [--color] \
13

14
To assist visualization, pass in '--color' to make the symbols color coded to
15
green, assuming that you have a xterm connection that supports color.
16
"""
17

18

19

20

21

22
import argparse
23
import subprocess
24
import sys
25

26

27
class Trie:
28
    """A simple class that represents a Trie."""
29

30
    def __init__(self, name):
31
        """Initializes a Trie object."""
32
        self.name = name
33
        self.size = 0
34
        self.dictionary = {}
35

36

37
def GetSymbolTrie(target, nm_command, max_depth):
38
    """Gets a symbol trie with the passed in target.
39

40
    Args:
41
            target: the target binary to inspect.
42
            nm_command: the command to run nm.
43
            max_depth: the maximum depth to create the trie.
44
    """
45
    # Run nm to get a dump on the strings.
46
    proc = subprocess.Popen(
47
        [nm_command, '--radix=d', '--size-sort', '--print-size', target],
48
        stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
49
    nm_out, _ = proc.communicate()
50
    if proc.returncode != 0:
51
        print('NM command failed. Output is as follows:')
52
        print(nm_out)
53
        sys.exit(1)
54
    # Run c++filt to get proper symbols.
55
    proc = subprocess.Popen(['c++filt'],
56
                            stdin=subprocess.PIPE, stdout=subprocess.PIPE,
57
                            stderr=subprocess.STDOUT)
58
    out, _ = proc.communicate(input=nm_out)
59
    if proc.returncode != 0:
60
        print('c++filt failed. Output is as follows:')
61
        print(out)
62
        sys.exit(1)
63
    # Splits the output to size and function name.
64
    data = []
65
    for line in out.split('\n'):
66
        if line:
67
            content = line.split(' ')
68
            if len(content) < 4:
69
                # This is a line not representing symbol sizes. skip.
70
                continue
71
            data.append([int(content[1]), ' '.join(content[3:])])
72
    symbol_trie = Trie('')
73
    for size, name in data:
74
        curr = symbol_trie
75
        for c in name:
76
            if c not in curr.dictionary:
77
                curr.dictionary[c] = Trie(curr.name + c)
78
            curr = curr.dictionary[c]
79
            curr.size += size
80
            if len(curr.name) > max_depth:
81
                break
82
    symbol_trie.size = sum(t.size for t in symbol_trie.dictionary.values())
83
    return symbol_trie
84

85

86
def MaybeAddColor(s, color):
87
    """Wrap the input string to the xterm green color, if color is set.
88
    """
89
    if color:
90
        return '\033[92m{0}\033[0m'.format(s)
91
    else:
92
        return s
93

94

95
def ReadableSize(num):
96
    """Get a human-readable size."""
97
    for unit in ['B', 'KB', 'MB', 'GB']:
98
        if abs(num) <= 1024.0:
99
            return '%3.2f%s' % (num, unit)
100
        num /= 1024.0
101
    return '%.1f TB' % (num,)
102

103

104
# Note(jiayq): I know, I know, this is a recursive function, but it is
105
# convenient to write.
106
def PrintTrie(trie, prefix, max_depth, min_size, color):
107
    """Prints the symbol trie in a readable manner.
108
    """
109
    if len(trie.name) == max_depth or not trie.dictionary.keys():
110
        # If we are reaching a leaf node or the maximum depth, we will print the
111
        # result.
112
        if trie.size > min_size:
113
            print('{0}{1} {2}'.format(
114
                  prefix,
115
                  MaybeAddColor(trie.name, color),
116
                  ReadableSize(trie.size)))
117
    elif len(trie.dictionary.keys()) == 1:
118
        # There is only one child in this dictionary, so we will just delegate
119
        # to the downstream trie to print stuff.
120
        PrintTrie(
121
            trie.dictionary.values()[0], prefix, max_depth, min_size, color)
122
    elif trie.size > min_size:
123
        print('{0}{1} {2}'.format(
124
              prefix,
125
              MaybeAddColor(trie.name, color),
126
              ReadableSize(trie.size)))
127
        keys_with_sizes = [
128
            (k, trie.dictionary[k].size) for k in trie.dictionary.keys()]
129
        keys_with_sizes.sort(key=lambda x: x[1])
130
        for k, _ in keys_with_sizes[::-1]:
131
            PrintTrie(
132
                trie.dictionary[k], prefix + ' |', max_depth, min_size, color)
133

134

135
def main(argv):
136
    if not sys.platform.startswith('linux'):
137
        raise RuntimeError('Currently this tool only supports Linux.')
138
    parser = argparse.ArgumentParser(
139
        description="Tool to inspect binary size.")
140
    parser.add_argument(
141
        '--max_depth', type=int, default=10,
142
        help='The maximum depth to print the symbol tree.')
143
    parser.add_argument(
144
        '--min_size', type=int, default=1024,
145
        help='The mininum symbol size to print.')
146
    parser.add_argument(
147
        '--nm_command', type=str, default='nm',
148
        help='The path to the nm command that the tool needs.')
149
    parser.add_argument(
150
        '--color', action='store_true',
151
        help='If set, use ascii color for output.')
152
    parser.add_argument(
153
        '--target', type=str,
154
        help='The binary target to inspect.')
155
    args = parser.parse_args(argv)
156
    if not args.target:
157
        raise RuntimeError('You must specify a target to inspect.')
158
    symbol_trie = GetSymbolTrie(
159
        args.target, args.nm_command, args.max_depth)
160
    PrintTrie(symbol_trie, '', args.max_depth, args.min_size, args.color)
161

162

163
if __name__ == '__main__':
164
    main(sys.argv[1:])
165