• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7#     http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15
16"""create_def_file.py - tool to create a windows def file.
17
18The def file can be used to export symbols from the tensorflow dll to enable
19tf.load_library().
20
21Because the linker allows only 64K symbols to be exported per dll
22we filter the symbols down to the essentials. The regular expressions
23we use for this are specific to tensorflow.
24
25TODO: this works fine but there is an issue with exporting
26'const char * const' and importing it from a user_ops. The problem is
27on the importing end and using __declspec(dllimport) works around it.
28"""
29from __future__ import absolute_import
30from __future__ import division
31from __future__ import print_function
32
33import argparse
34import codecs
35import os
36import re
37import subprocess
38import sys
39import tempfile
40
41# External tools we use that come with visual studio sdk and
42# we assume that the caller has the correct PATH to the sdk
43UNDNAME = "undname.exe"
44DUMPBIN = "dumpbin.exe"
45
46# Exclude if matched
47EXCLUDE_RE = re.compile(r"RTTI|deleting destructor|::internal::|Internal|"
48                        r"python_op_gen_internal|grappler")
49
50# Include if matched before exclude
51INCLUDEPRE_RE = re.compile(r"google::protobuf::internal::ExplicitlyConstructed|"
52                           r"tensorflow::internal::LogMessage|"
53                           r"tensorflow::internal::LogString|"
54                           r"tensorflow::internal::CheckOpMessageBuilder|"
55                           r"tensorflow::internal::PickUnusedPortOrDie|"
56                           r"tensorflow::internal::ValidateDevice|"
57                           r"tensorflow::ops::internal::Enter|"
58                           r"tensorflow::strings::internal::AppendPieces|"
59                           r"tensorflow::strings::internal::CatPieces|"
60                           r"tensorflow::errors::Internal|"
61                           r"tensorflow::Tensor::CopyFromInternal|"
62                           r"tensorflow::kernel_factory::"
63                           r"OpKernelRegistrar::InitInternal|"
64                           r"tensorflow::io::internal::JoinPathImpl")
65
66# Include if matched after exclude
67INCLUDE_RE = re.compile(r"^(TF_\w*)$|"
68                        r"^(TFE_\w*)$|"
69                        r"tensorflow::|"
70                        r"functor::|"
71                        r"\?nsync_|"
72                        r"stream_executor::")
73
74# We want to identify data members explicitly in the DEF file, so that no one
75# can implicitly link against the DLL if they use one of the variables exported
76# from the DLL and the header they use does not decorate the symbol with
77# __declspec(dllimport). It is easier to detect what a data symbol does
78# NOT look like, so doing it with the below regex.
79DATA_EXCLUDE_RE = re.compile(r"[)(]|"
80                             r"vftable|"
81                             r"vbtable|"
82                             r"vcall|"
83                             r"RTTI|"
84                             r"protobuf::internal::ExplicitlyConstructed")
85
86def get_args():
87  """Parse command line."""
88  filename_list = lambda x: x.split(";")
89  parser = argparse.ArgumentParser()
90  parser.add_argument("--input", type=filename_list,
91                      help="paths to input libraries separated by semicolons",
92                      required=True)
93  parser.add_argument("--output", help="output deffile", required=True)
94  parser.add_argument("--target", help="name of the target", required=True)
95  parser.add_argument("--bitness", help="build target bitness", required=True)
96  args = parser.parse_args()
97  return args
98
99
100def main():
101  """main."""
102  args = get_args()
103
104  # Pipe dumpbin to extract all linkable symbols from libs.
105  # Good symbols are collected in candidates and also written to
106  # a temp file.
107  candidates = []
108  tmpfile = tempfile.NamedTemporaryFile(mode="w", delete=False)
109  for lib_path in args.input:
110    proc = subprocess.Popen([DUMPBIN, "/nologo", "/linkermember:1", lib_path],
111                            stdout=subprocess.PIPE)
112    for line in codecs.getreader("utf-8")(proc.stdout):
113      cols = line.split()
114      if len(cols) < 2:
115        continue
116      sym = cols[1]
117      tmpfile.file.write(sym + "\n")
118      candidates.append(sym)
119    exit_code = proc.wait()
120    if exit_code != 0:
121      print("{} failed, exit={}".format(DUMPBIN, exit_code))
122      return exit_code
123  tmpfile.file.close()
124
125  # Run the symbols through undname to get their undecorated name
126  # so we can filter on something readable.
127  with open(args.output, "w") as def_fp:
128    # track dupes
129    taken = set()
130
131    # Header for the def file.
132    def_fp.write("LIBRARY " + args.target + "\n")
133    def_fp.write("EXPORTS\n")
134    if args.bitness == "64":
135      def_fp.write("\t??1OpDef@tensorflow@@UEAA@XZ\n")
136    else:
137      def_fp.write("\t??1OpDef@tensorflow@@UAE@XZ\n")
138
139    # Each symbols returned by undname matches the same position in candidates.
140    # We compare on undname but use the decorated name from candidates.
141    dupes = 0
142    proc = subprocess.Popen([UNDNAME, tmpfile.name], stdout=subprocess.PIPE)
143    for idx, line in enumerate(codecs.getreader("utf-8")(proc.stdout)):
144      decorated = candidates[idx]
145      if decorated in taken:
146        # Symbol is already in output, done.
147        dupes += 1
148        continue
149
150      if not INCLUDEPRE_RE.search(line):
151        if EXCLUDE_RE.search(line):
152          continue
153        if not INCLUDE_RE.search(line):
154          continue
155
156      if "deleting destructor" in line:
157        # Some of the symbols convered by INCLUDEPRE_RE export deleting
158        # destructor symbols, which is a bad idea.
159        # So we filter out such symbols here.
160        continue
161
162      if DATA_EXCLUDE_RE.search(line):
163        def_fp.write("\t" + decorated + "\n")
164      else:
165        def_fp.write("\t" + decorated + " DATA\n")
166      taken.add(decorated)
167  exit_code = proc.wait()
168  if exit_code != 0:
169    print("{} failed, exit={}".format(UNDNAME, exit_code))
170    return exit_code
171
172  os.unlink(tmpfile.name)
173
174  print("symbols={}, taken={}, dupes={}"
175        .format(len(candidates), len(taken), dupes))
176  return 0
177
178
179if __name__ == "__main__":
180  sys.exit(main())
181