supersize: Tweak diff symbol matching for generated symbols

tools/binary_size/libsupersize/diff.py

 # Copyright 2017 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.
 """Logic for diffing two SizeInfo objects."""
 
 import collections
+import re
 
 import models
 
 
+def _SymbolKey(symbol):
+  """Returns a tuple that can be used to see if two Symbol are the same.
+
+  Keys are not guaranteed to be unique within a SymbolGroup. When multiple
+  symbols have the same key, they will be matched up in order of appearance.
+
+  Examples of symbols with shared keys:
+    "** merge strings"
+    "** symbol gap 3", "** symbol gap 5"
+    "foo() [clone ##]"
+    "CSWTCH.61", "CSWTCH.62"
+    "._468", "._467"
+    ".L__unnamed_1193", ".L__unnamed_712"
+  """
+  name = symbol.full_name or symbol.name
+  if name:

[estevenson, 2017/05/03 17:29:06]

nit: this check seems superfluous since the rest of the function depends on
|name| anyways.

[agrieve, 2017/05/03 18:27:08]

Done.

+    clone_idx = name.find(' [clone ')
+    if clone_idx != -1:
+      name = name[:clone_idx]
+  if name.startswith('*'):
+    # "symbol gap 3 (bar)" -> "symbol gaps"
+    name = re.sub(r'\s+\d+( \(.*\))?$', 's', name)
+
+  if '.' not in name:
+    return (symbol.section_name, name)
+  # Compiler or Linker generated symbol.
+  name = re.sub(r'[.0-9]', '', name)  # Strip out all numbers and dots.

[estevenson, 2017/05/03 17:29:06]

Just to confirm I understand:

On Linux there are a bunch of .L__unnamed_XXXX, and since the names presumably
aren't stable there's no way for us to capture changes in these symbols. So
we're just assuming based on order that two symbols with the same keys are
deltas of each other right?

[agrieve, 2017/05/03 18:27:07]

Updated the pydoc to make confirm the "why".

But yeah, I don't know that it will work all that well for the linux symbols
since the example you pasted showed them also varying on size. It should work
well for symbols found on android like:

+       44 r@0x2a13ba8  24     
obj/third_party/WebKit/Source/platform/platform/RecordingImageBufferSurface.o
               CSWTCH.61
-       20 r@0x2a0f9b8  -24    
obj/third_party/WebKit/Source/platform/platform/RecordingImageBufferSurface.o
               CSWTCH.62
-        4 r@0x2b58888  -16    
obj/third_party/WebKit/Source/modules/vr/libvr.a(VRDisplay.o)
               ._468
+       20 r@0x2b5ca88  16     
obj/third_party/WebKit/Source/modules/vr/libvr.a(VRDisplay.o)
               ._470
-        4 r@0x2b7e148  -16    
obj/components/history/content/browser/libbrowser.a(download_conversions.o)
               CSWTCH.29
-      -12 r@0x2a101c0  -16    
obj/third_party/WebKit/Source/platform/platform/SkiaUtils.o
               CSWTCH.31
+        4 r@0x2a143a0  16     
obj/third_party/WebKit/Source/platform/platform/SkiaUtils.o
               CSWTCH.32
+       20 r@0x2b7e120  16      obj/chrome/browser/ui/libui.a(page_info_ui.o)
               CSWTCH.34

+  return (symbol.section_name, name, symbol.object_path)
+
+
 def _CloneSymbol(sym, size):
   """Returns a copy of |sym| with an updated |size|.
 
   Padding and aliases are not copied.
   """
   return models.Symbol(
       sym.section_name, size, address=sym.address, name=sym.name,
       source_path=sym.source_path, object_path=sym.object_path,
       full_name=sym.full_name, flags=sym.flags)
 
@@ skipping 73 matching lines @@
     else:
       cloned = _CloneSymbol(sym, -sym.size_without_padding)
       cloned.padding = -sym.padding
     ret[i] = cloned
   return ret
 
 
 def _DiffSymbolGroups(before, after):
   before_symbols_by_key = collections.defaultdict(list)
   for s in before:
-    before_symbols_by_key[s._Key()].append(s)
+    before_symbols_by_key[_SymbolKey(s)].append(s)
 
   similar = []
   diffed_symbol_by_after_aliases = {}
   matched_before_aliases = set()
   unmatched_after_syms = []
   # For similar symbols, padding is zeroed out. In order to not lose the
   # information entirely, store it in aggregate.
   padding_by_section_name = collections.defaultdict(int)
 
   # Step 1: Create all delta symbols and record unmatched symbols.
   for after_sym in after:
-    matching_syms = before_symbols_by_key.get(after_sym._Key())
+    matching_syms = before_symbols_by_key.get(_SymbolKey(after_sym))
     if matching_syms:
       before_sym = matching_syms.pop(0)
       if before_sym.IsGroup() and after_sym.IsGroup():
         similar.append(_DiffSymbolGroups(before_sym, after_sym))
       else:
         if before_sym.aliases:
           matched_before_aliases.add(id(before_sym.aliases))
         similar.append(
             _DiffSymbol(before_sym, after_sym, diffed_symbol_by_after_aliases,
                         padding_by_section_name))
@@ skipping 31 matching lines @@
   """
   assert isinstance(before, models.SizeInfo)
   assert isinstance(after, models.SizeInfo)
   section_sizes = {k: after.section_sizes[k] - v
                    for k, v in before.section_sizes.iteritems()}
   symbol_diff = _DiffSymbolGroups(before.symbols, after.symbols)
   if cluster:
     symbol_diff = symbol_diff.Cluster()
   return models.SizeInfoDiff(section_sizes, symbol_diff, before.metadata,
                              after.metadata)