binary_size_tool: fix for ambiguous addr2line output

build/android/pylib/symbols/elf_symbolizer.py

 # 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.
 
 import collections
 import datetime
 import logging
 import multiprocessing
 import os
 import posixpath
 import Queue
 import re
 import subprocess
 import sys
 import threading
-
+from sets import Set

[Primiano Tucci (use gerrit), 2014/06/18 09:49:11]

just use the builtin set. We don't target python < 2.6

 
 # addr2line builds a possibly infinite memory cache that can exhaust
 # the computer's memory if allowed to grow for too long. This constant
 # controls how many lookups we do before restarting the process. 4000
 # gives near peak performance without extreme memory usage.
 ADDR2LINE_RECYCLE_LIMIT = 4000
 
 
 class ELFSymbolizer(object):
   """An uber-fast (multiprocessing, pipelined and asynchronous) ELF symbolizer.
@@ skipping 41 matching lines @@
   |max_queue_size| bound, a new addr2line instance is kicked in.
   In the case of a very eager producer (i.e. all |max_concurrent_jobs| instances
   have a backlog of |max_queue_size|), back-pressure is applied on the caller by
   blocking the SymbolizeAsync method.
 
   This module has been deliberately designed to be dependency free (w.r.t. of
   other modules in this project), to allow easy reuse in external projects.
   """
 
   def __init__(self, elf_file_path, addr2line_path, callback, inlines=False,
-      max_concurrent_jobs=None, addr2line_timeout=30, max_queue_size=50):
+      max_concurrent_jobs=None, addr2line_timeout=30, max_queue_size=50,
+      disambiguate=False, disambiguation_source_path=''):

[Primiano Tucci (use gerrit), 2014/06/18 10:48:53]

I had a quick chat with Andrew and we feel the best thing to do inside here is
pass two arguments:
1. source_root_path (optional): the path to the source tree on the local disk
(for disambiguation)
2. strip_base_path: the base path to strip off the symbol information (to
address the problem mentioned in my earlier comment).

In essence the behavior should be the following: in presence of both these
variables, elf_symbolizer will strive to produce paths which are absolute and
refer to existing stuff in the local filesystem. Here is en example:

src_root_path = /home/primiano/chrome
strip_base_path = /foo/chrome

elf_symbolizer will translate the paths as follows:
bar.c -> /home/primiano/chrome/somewhere/bar.c if there was one and only one hit
for bar.c in source_root_path
baz.c -> baz.c if we couldn't disambiguate any (or found than one hits).
/foo/chrome/bar.c -> /home/primiano/chrome/bar.c (strip off strip_base_path and
prepend source_root_path)

     """Args:
       elf_file_path: path of the elf file to be symbolized.
       addr2line_path: path of the toolchain's addr2line binary.
       callback: a callback which will be invoked for each resolved symbol with
           the two args (sym_info, callback_arg). The former is an instance of
           |ELFSymbolInfo| and contains the symbol information. The latter is an
           embedder-provided argument which is passed to SymbolizeAsync().
       inlines: when True, the ELFSymbolInfo will contain also the details about
           the outer inlining functions. When False, only the innermost function
           will be provided.
       max_concurrent_jobs: Max number of addr2line instances spawned.
           Parallelize responsibly, addr2line is a memory and I/O monster.
       max_queue_size: Max number of outstanding requests per addr2line instance.
       addr2line_timeout: Max time (in seconds) to wait for a addr2line response.
           After the timeout, the instance will be considered hung and respawned.
+      disambiguate: Whether to run a disambiguation process or not.

[Primiano Tucci (use gerrit), 2014/06/18 09:49:11]

Can we use just one variable (call it source_root_path), which is None by
default, and when set kicks in the disambiguation logic?
We definitely don't need both (is there any sense by setting disambiguate=True
and disambiguation_source_path = '' ?)

+          Disambiguation means to resolve ambiguous source_paths, for
+          example turn addr2line output "unicode.cc" into a full and absolute
+          path. In some toolchains only the name of the source file is output,

[Primiano Tucci (use gerrit), 2014/06/18 09:49:11]

I'd love to know more about this btw. In which cases do you get symbols which
have only a file name and not the path? This sounds like broken compiler /
compiler flags.

+          without any path information; disambiguation searches through the
+          source directory specified by 'disambiguate_source_path' argument
+          for files whose name matches. If there are multiple files with the
+          same name, disambiguation will fail.
+      disambiguate_source_path: The path to the directory where the source
+          files are located, used for disambiguating paths.
     """
     assert(os.path.isfile(addr2line_path)), 'Cannot find ' + addr2line_path
     self.elf_file_path = elf_file_path
     self.addr2line_path = addr2line_path
     self.callback = callback
     self.inlines = inlines
     self.max_concurrent_jobs = (max_concurrent_jobs or
                                 min(multiprocessing.cpu_count(), 4))
     self.max_queue_size = max_queue_size
     self.addr2line_timeout = addr2line_timeout
     self.requests_counter = 0  # For generating monotonic request IDs.
     self._a2l_instances = []  # Up to |max_concurrent_jobs| _Addr2Line inst.
 
+    # If necessary, create disambiguation lookup table
+    self.disambiguate = disambiguate
+    self.commonprefix = ''
+    self.lookup_table = {}
+    if(self.disambiguate):
+      self._CreateDisambiguationTable(disambiguation_source_path)
+
     # Create one addr2line instance. More instances will be created on demand
     # (up to |max_concurrent_jobs|) depending on the rate of the requests.
     self._CreateNewA2LInstance()
 
   def SymbolizeAsync(self, addr, callback_arg=None):
     """Requests symbolization of a given address.
 
     This method is not guaranteed to return immediately. It generally does, but
     in some scenarios (e.g. all addr2line instances have full queues) it can
     block to create back-pressure.
@@ skipping 37 matching lines @@
     for a2l in self._a2l_instances:
       a2l.WaitForIdle()
       a2l.Terminate()
 
   def _CreateNewA2LInstance(self):
     assert(len(self._a2l_instances) < self.max_concurrent_jobs)
     a2l = ELFSymbolizer.Addr2Line(self)
     self._a2l_instances.append(a2l)
     return a2l
 
+  def _CreateDisambiguationTable(self, src_root_path):

[Primiano Tucci (use gerrit), 2014/06/18 09:49:11]

I don't really like this approach. This is trying to index all source files in
the root path. In the case of large code bases like chrome/android this feels
overkilling. (Also this pretends to know the  extensions which are relevant for
the caller, which looks terrible).
At very last, can we do the other way round?
That is:  keep only a single hashtable 'file.c' -> '/path/file.c', and whenever
you encounter a bare filename with no path which is not present in the
hashtable, you walk the oath looking for the file and populate the table (OS FS
cache will be your friend).
I'd blow away this entire table.

[Primiano Tucci (use gerrit), 2014/06/18 10:48:53]

After talking with Andrew, given your use case, I'm ok with indexing the paths
here. Just we have to simplify this code. I don't like caring at all about the
extensions. Indexing everything will not be that worse that indexing the c/C++
files, and will save a lot of complexity.
In other words, this thing should be as easy and legible as the following four
lines of python:

src_paths_index = {}
for dpath, _, fnames in os.walk(src_root_path):
  for fname in fnames:
    src_paths_index[fname] = os.path.join(dpath, fname) if fname not in
src_paths_index else None

This will produce a table which keys are file names and entries are either a
full path (if unique) or None (if ambiguous)

+    """ Creates a table of files used for disambiguation later
+        Disambiguation:

[Primiano Tucci (use gerrit), 2014/06/18 09:49:11]

This comment is redundant. You already explained the concept in __init__

+          addr2line sometimes return an ambigous file-name rather than the
+          full path of the file where the symbol is located.
+
+        adopted from andrewhaydens implementation in earlier commits """
+    interesting_file_endings = { ".c", ".cc", ".h", ".cp", ".cpp", ".cxx",
+                                 ".c++", ".asm", ".inc", ".s", ".hxx" }
+    duplicates = Set()
+    self.lookup_table = {}
+    src_root_path = os.path.abspath(src_root_path)
+
+    for root, _, filenames in os.walk(src_root_path):
+      for f in filenames:
+        _, ext = os.path.splitext(f)
+        if not ext in interesting_file_endings:
+          continue
+
+        base = os.path.basename(f)  # Just in case
+        if self.lookup_table.get(base) is None:
+          self.lookup_table[base] = "%s/%s" % (root, f)
+        else:
+          duplicates.add(base)
+
+    # Duplicates can not be used for disambiguation, as we can not determine
+    # the true source if we have more than one to choose from
+    for d in duplicates:
+      del self.lookup_table[d]
+
+    # Get the common prefix for the source paths
+    self.commonprefix = os.path.commonprefix(self.lookup_table.values())
 
   class Addr2Line(object):
     """A python wrapper around an addr2line instance.
 
     The communication with the addr2line process looks as follows:
       [STDIN]         [STDOUT]  (from addr2line's viewpoint)
     > f001111
     > f002222
                     < Symbol::Name(foo, bar) for f001111
                     < /path/to/source/file.c:line_number
@@ skipping 131 matching lines @@
         source_line = None
         m = ELFSymbolizer.Addr2Line.SYM_ADDR_RE.match(line2)
         if m:
           if not m.group(1).startswith('?'):
             source_path = m.group(1)
             if not m.group(2).startswith('?'):
               source_line = int(m.group(2))
         else:
           logging.warning('Got invalid symbol path from addr2line: %s' % line2)
 
-        sym_info = ELFSymbolInfo(name, source_path, source_line)
+        # In case disambiguation is on, and needed
+        disambiguated = False
+        failed_disambiguation = False
+        if self._symbolizer.disambiguate:
+          if not source_path is None and not source_path.startswith('/'):

[Andrew Hayden (chromium.org), 2014/06/18 09:10:59]

Again, let's avoid double-negation:
if source_path and not source_path.startswith('/')

This is clearer than
if not source_path is None...

+            source_path = self._symbolizer.lookup_table.get(source_path)
+            failed_disambiguation = source_path is None
+            disambiguated = not failed_disambiguation

[Andrew Hayden (chromium.org), 2014/06/18 09:10:59]

As written, you don't need two booleans. They're just being assigned inverse
values, so diambiguated implies !failed_disambiguated and failed_disambiguated
implies !disambiguated.

I think what you want here is more like:
was_ambiguous: True iff the path was a bare file, no path information.
disambiguated: True iff was_ambiguous == True AND we were able to successfully
disambiguate the path.

Then you have this truth table
(was_ambiguous,dismabiguated):
(false, false): full path info produced by nm
(false, true): full path info produced by nm
(true, false): raw filename produced by nm, failed to disambiguate
(true, true): raw filename produced by nm, succeeded in disambiguation attempt


Edit: OK, you could also use failed_disambiguation (like in later code), and
flip the meaning in the truth table above.

+
+        # Use the absolute path
+        if not source_path is None:

[Andrew Hayden (chromium.org), 2014/06/18 09:10:59]

Again let's avoid double negation:
if source_path

or if you really prefer I think python allows:
if source_path is not None

+          source_path = os.path.abspath(source_path)

[Primiano Tucci (use gerrit), 2014/06/18 09:49:11]

What is this for?
Looks like that even if you're not disambiguating, you force any relative path
to be translated using cwd as base? This looks a very arguable assumption.

[Primiano Tucci (use gerrit), 2014/06/18 10:48:53]

At this point you also check if source_path.startswith(strip_base_path), and if
so replace it with the src_root_path

+
+        sym_info = ELFSymbolInfo(name, source_path, source_line, disambiguated,
+                                 failed_disambiguation)
         if prev_sym_info:
           prev_sym_info.inlined_by = sym_info
         if not innermost_sym_info:
           innermost_sym_info = sym_info
 
       self._processed_symbols_count += 1
       self._symbolizer.callback(innermost_sym_info, callback_arg)
 
     def _RestartAddr2LineProcess(self):
       if self._proc:
@@ skipping 60 matching lines @@
 
     @property
     def first_request_id(self):
       """Returns the request_id of the oldest pending request in the queue."""
       return self._request_queue[0][2] if self._request_queue else 0
 
 
 class ELFSymbolInfo(object):
   """The result of the symbolization passed as first arg. of each callback."""
 
-  def __init__(self, name, source_path, source_line):
+  def __init__(self, name, source_path, source_line, disambiguated=False,

[Primiano Tucci (use gerrit), 2014/06/18 09:49:11]

Do you need to pass these booleans at all? Can't you infer that in the caller by
looking at the presence of source_path?

+               failed_disambiguation=False):
     """All the fields here can be None (if addr2line replies with '??')."""
     self.name = name
     self.source_path = source_path
     self.source_line = source_line
     # In the case of |inlines|=True, the |inlined_by| points to the outer
     # function inlining the current one (and so on, to form a chain).
     self.inlined_by = None
+    self.disambiguated = disambiguated
+    self.failed_disambiguation = failed_disambiguation
 
   def __str__(self):
     return '%s [%s:%d]' % (
         self.name or '??', self.source_path or '??', self.source_line or 0)