[SuperSize] Ignore LLVM promoted global names suffix (.llvm.1234).

Under LLD ThinLTO, if thin_lto_enable_optimizations is enabled then
"promoted global names" can appear. For these, mangled symbol names get
the suffix '.llvm.1234' or '.1.llvm.1234', where '1' is a small number,
and '1234' is a hash. When demangled, the suffix becomes ' (.llvm.1324)'
or ' (.1.llvm.1324)' if successful; and is left alone otherwise.

These suffix interferes with name comparison, leading to secondary
problems like symbols losing path info.

This CL makes SuperSize detect and strip these suffixes, to reduce
spurious diffs from enabling ThinLTO optimization. The transformation
is applied when parsing .map and nm output. Details:
* Add demangle.StripLlvmPromotedGlobalNames(), which performs a
  quick-reject substring test for '.llvm.', followed by RegExp
  string substitution.
* Print the number of promoted global names found in .map file.
* Minor updates to comments and tests.

Bug: 902723
Change-Id: I4ca4142c891c948894c10180b31ff65b759cccfe
Reviewed-on: https://chromium-review.googlesource.com/c/1324192
Commit-Queue: Samuel Huang <huangs@chromium.org>
Reviewed-by: Eric Stevenson <estevenson@chromium.org>
Reviewed-by: Samuel Huang <huangs@chromium.org>
Cr-Commit-Position: refs/heads/master@{#606487}

tools/binary_size/libsupersize/demangle.py

@@ -6,17 +6,39 @@
 
 import collections
 import logging
+import re
 import subprocess
 
 import path_util
 
+_PROMOTED_GLOBAL_NAME_DEMANGLED_PATTERN = re.compile(
+    r' \((\.\d+)?\.llvm\.\d+\)$')
+_PROMOTED_GLOBAL_NAME_RAW_PATTERN = re.compile(r'(\.\d+)?\.llvm\.\d+$')
+
+def StripLlvmPromotedGlobalNames(name):
+  """Strips LLVM promoted global names suffix, and returns the result.
+
+  LLVM can promote global names by adding the suffix '.llvm.1234', or
+  '.1.llvm.1234', where the last numeric suffix is a hash. If demangle is
+  sucessful, the suffix transforms into, e.g., ' (.llvm.1234)' or
+  ' (.1.llvm.1234)'. Otherwise the suffix is left as is. This function strips
+  the suffix to prevent it from intefering with name comparison.
+  """
+  llvm_pos = name.find('.llvm.')
+  if llvm_pos < 0:
+    return name  # Handles most cases.
+  if name.endswith(')'):
+    return _PROMOTED_GLOBAL_NAME_DEMANGLED_PATTERN.sub('', name)
+  return _PROMOTED_GLOBAL_NAME_RAW_PATTERN.sub('', name)
+
+
 def _DemangleNames(names, tool_prefix):
   """Uses c++filt to demangle a list of names."""
   proc = subprocess.Popen([path_util.GetCppFiltPath(tool_prefix)],
                           stdin=subprocess.PIPE, stdout=subprocess.PIPE)
   stdout = proc.communicate('\n'.join(names))[0]
   assert proc.returncode == 0
-  ret = stdout.splitlines()
+  ret = [StripLlvmPromotedGlobalNames(line) for line in stdout.splitlines()]
   if logging.getLogger().isEnabledFor(logging.INFO):
     fail_count = sum(1 for s in ret if s.startswith('_Z'))
     if fail_count:

tools/binary_size/libsupersize/linker_map_parser.py

@@ -2,6 +2,7 @@
 # Use of this source code is governed by a BSD-style license that can be
 # found in the LICENSE file.
 
+import demangle
 import logging
 import os
 import re
@@ -320,6 +321,11 @@ class MapFileParserLld(object):
 #     600      600       14     4         obj/...:(.text.OUTLINED_FUNCTION_0)
 #     600      600        0     1                 $x.3
 #     600      600       14     1                 OUTLINED_FUNCTION_0
+#  123800   123800    20000   256 .rodata
+#  123800   123800       4      4         ...:o:(.rodata._ZN3fooE.llvm.1234)
+#  123800   123800       4      1                 foo (.llvm.1234)
+#  123804   123804       4      4         ...:o:(.rodata.bar.llvm.1234)
+#  123804   123804       4      1                 bar.llvm.1234
 # Older format:
 # Address          Size             Align Out     In      Symbol
 # 00000000002002a8 000000000000001c     1 .interp
@@ -343,6 +349,7 @@ class MapFileParserLld(object):
     sym_maker = _SymbolMaker()
     cur_section = None
     cur_section_is_useful = None
+    promoted_name_count = 0
 
     for line in lines:
       m = pattern.match(line)
@@ -399,12 +406,17 @@ class MapFileParserLld(object):
             # renaming them to '** outlined function'.
             if tok.startswith('OUTLINED_FUNCTION_'):
               tok = '** outlined function'
+            stripped_tok = demangle.StripLlvmPromotedGlobalNames(tok)
+            if len(tok) != len(stripped_tok):
+              promoted_name_count += 1
+              tok = stripped_tok
             sym_maker.cur_sym.full_name = tok
-
         else:
           logging.error('Problem line: %r', line)
 
     sym_maker.Flush()
+    if promoted_name_count:
+      logging.info('Found %d promoted global names', promoted_name_count)
     return self._section_sizes, sym_maker.syms
 
 

tools/binary_size/libsupersize/testdata/mock_toolchain/mock_nm.py

@@ -80,14 +80,14 @@ _OBJECT_OUTPUTS = {
         '01010101 t extFromUUseMapping(signed char, unsigned int, int)',
         '01010101 t Name',
         '01010101 v vtable for mojo::MessageReceiver',
-        '01010101 r kMethodsAnimationFrameTimeHistogram',
+        '01010101 r kMethodsAnimationFrameTimeHistogram (.llvm.12341234)',
         '01010101 r google::protobuf::internal::pLinuxKernelCmpxchg',
     ],
     'obj/third_party/ffmpeg/libffmpeg_internal.a': [
         '',
         'fft_float.o:',
         '01010101 b ff_cos_65536',
-        '01010101 b ff_cos_131072',
+        '01010101 b ff_cos_131072.1.llvm.43214321',
         '002a0010 t FooAlias()',
         '002b6bb8 t $t',
         '002a0010 t BarAlias()',