Don't depend on llvm-nm to detect archive indexes (#10300)

There is an llvm but that mean that certain binary files get mistaken
for COFF object which means llvm-nm will error out:
  https://bugs.llvm.org/show_bug.cgi?id=44683

Instead write out own trivial ar parser to detect the presence of that
index.

Fixes: #10195

Author: sbc100

emcc.py

@@ -480,21 +480,36 @@ def do_emscripten(infile, memfile):
   return outfile
 
 
+def is_ar_file_with_missing_index(archive_file):
+  # We parse the archive header outselves because llvm-nm --print-armap is slower and less
+  # reliable.
+  # See: https://github.com/emscripten-core/emscripten/issues/10195
+  archive_header = b'!<arch>\n'
+  file_header_size = 60
+
+  with open(archive_file, 'rb') as f:
+    header = f.read(len(archive_header))
+    if header != archive_header:
+      # This is not even an ar file
+      return False
+    file_header = f.read(file_header_size)
+    if len(file_header) != file_header_size:
+      # We don't have any file entires at all so we don't consider the index missing
+      return False
+
+  name = file_header[:16].strip()
+  # If '/' is the name of the first file we have an index
+  return name != '/'
+
+
 def ensure_archive_index(archive_file):
   # Fastcomp linking works without archive indexes.
   if not shared.Settings.WASM_BACKEND or not shared.Settings.AUTO_ARCHIVE_INDEXES:
     return
-  # Ignore stderr since llvm-nm prints "no symbols" to stderr for each object that has no symbols
-  stdout = run_process([shared.LLVM_NM, '--print-armap', archive_file], stdout=PIPE, stderr=PIPE).stdout
-  stdout = stdout.strip()
-  # Ignore empty archives
-  if not stdout:
-    return
-  if stdout.startswith('Archive map\n') or stdout.startswith('Archive index\n'):
-    return
-  shared.warning('%s: archive is missing an index; Use emar when creating libraries to ensure an index is created', archive_file)
-  shared.warning('%s: adding index', archive_file)
-  run_process([shared.LLVM_RANLIB, archive_file])
+  if is_ar_file_with_missing_index(archive_file):
+    shared.warning('%s: archive is missing an index; Use emar when creating libraries to ensure an index is created', archive_file)
+    shared.warning('%s: adding index', archive_file)
+    run_process([shared.LLVM_RANLIB, archive_file])
 
 
 #

tests/test_other.py

@@ -8728,6 +8728,19 @@ def test_archive_no_index(self):
     # The default behavior is to add archive indexes automatically.
     run_process([PYTHON, EMCC, 'libfoo.a', 'hello_world.o'])
 
+  @no_fastcomp('AUTO_ARCHIVE_INDEXES only applies to wasm backend')
+  def test_archive_non_objects(self):
+    create_test_file('file.txt', 'test file')
+    # llvm-nm has issues with files that start with two or more null bytes since it thinks they
+    # are COFF files.  Ensure that we correctly ignore such files when we process them.
+    create_test_file('zeros.bin', '\0\0\0\0')
+    run_process([PYTHON, EMCC, '-c', path_from_root('tests', 'hello_world.c')])
+    # No index added.
+    # --format=darwin (the default on OSX has a strange issue where it add extra
+    # newlines to files: https://bugs.llvm.org/show_bug.cgi?id=42562
+    run_process([PYTHON, EMAR, 'crS', '--format=gnu', 'libfoo.a', 'file.txt', 'zeros.bin', 'hello_world.o'])
+    run_process([PYTHON, EMCC, path_from_root('tests', 'hello_world.c'), 'libfoo.a'])
+
   def test_flag_aliases(self):
     def assert_aliases_match(flag1, flag2, flagarg, extra_args):
       results = {}