class LinkHelper:
"""
Helps with creating and running a custom linking command for YugabyteDB outside of the build
system.
"""
dep_graph: DependencyGraph
initial_node: Node
# Arguments to the linker in the original command that produces the initial_node target.
# Does not include the compiler driver executable.
original_link_args: List[str]
build_root: str
build_paths: BuildPaths
llvm_path: str
thirdparty_path: str
clang_cpp_path: str
# Dependency graph nodes corresponding to the object files present in the original linker
# command. Used for deduplication.
obj_file_graph_nodes: Set[Node]
new_args: LinkCommand
# Build directory of the Postgres backend.
pg_backend_build_dir: str
# The command for linking the yb_pgbackend library.
yb_pgbackend_link_cmd: List[str]
lto_output_suffix: Optional[str]
# Populated by consume_original_link_cmd.
final_output_name: str
def __init__(self, dep_graph: DependencyGraph, initial_node: Node,
lto_output_suffix: Optional[str]) -> None:
self.dep_graph = dep_graph
self.initial_node = initial_node
self.build_root = self.dep_graph.conf.build_root
self.build_paths = BuildPaths(self.build_root)
self.llvm_path = self.build_paths.get_llvm_path()
self.thirdparty_path = self.build_paths.get_thirdparty_path()
self.clang_cpp_path = self.build_paths.get_llvm_tool_path('clang++')
assert initial_node.link_cmd
self.original_link_args = process_original_link_cmd(
initial_node.link_cmd)
self.static_lib_paths = get_static_lib_paths(self.thirdparty_path)
self.new_args = LinkCommand([self.clang_cpp_path])
self.obj_file_graph_nodes = set()
self.pg_backend_build_dir, self.yb_pgbackend_link_cmd = get_yb_pgbackend_link_cmd(
self.build_root)
self.lto_output_suffix = lto_output_suffix
def convert_to_static_lib(self, arg: str) -> Optional[str]:
"""
Given an argument to the original linker command, try to interpret it as a library, either
specified as a shared library path, or using -l... syntax, and return the corresponding
static library path if available.
"""
if arg.startswith('/') and arg.endswith('.so'):
arg_static_prefix = arg[:-3]
static_found = False
for suffix in ['.a', '-s.a']:
arg_static = arg_static_prefix + suffix
if os.path.exists(arg_static):
logging.info(
"Using static library %s instead of shared library %s",
arg_static, arg)
return arg_static
logging.info("Did not find static library corresponding to %s",
arg)
if arg.startswith('-l'):
static_found = False
logging.info("Looking for static lib for: %s", arg)
lib_name = arg[2:]
for static_lib_path in self.static_lib_paths:
static_lib_basename = os.path.basename(static_lib_path)
if (static_lib_basename == 'lib' + lib_name + '.a'
or static_lib_basename == 'lib' + lib_name + '-s.a'):
logging.info("Found static lib for %s: %s", lib_name,
static_lib_path)
return static_lib_path
logging.info("Did not find a static lib for %s", lib_name)
if arg.endswith('.so') or '.so.' in arg:
logging.info("Still using a shared library: %s", arg)
return None
def process_arg(self, arg: str) -> None:
if arg in SKIPPED_ARGS:
logging.info("Skipping argument: %s", arg)
return
new_arg = self.convert_to_static_lib(arg)
if new_arg:
if not self.new_args.contains(new_arg):
self.new_args.add_new_arg(new_arg)
else:
self.new_args.add_new_arg(arg)
def consume_original_link_cmd(self) -> None:
"""
Goes over the original linker command and reuses some of its arguments for the new command.
"""
with WorkDirContext(self.build_root):
expect_output_name = False
output_name: Optional[str] = None
for arg in self.original_link_args:
if arg == '-o':
expect_output_name = True
continue
if expect_output_name:
if output_name:
raise ValueError(
"Found multiple output names in the original link command: "
"%s and %s" % (output_name, arg))
output_name = arg
expect_output_name = False
continue
expect_output_name = False
if is_yb_library(arg):
logging.info("Skipping YB library: %s", arg)
continue
if arg.endswith('.cc.o'):
# E.g. tablet_server_main.cc.o.
# Remember this node for later deduplication.
self.obj_file_graph_nodes.add(
self.dep_graph.find_node(os.path.realpath(arg)))
self.process_arg(arg)
if not output_name:
raise ValueError(
"Did not find an output name in the original link command")
self.final_output_name = os.path.abspath(output_name)
logging.info("Final output file name: %s", self.final_output_name)
if self.lto_output_suffix is not None:
self.final_output_name += self.lto_output_suffix
self.new_args.extend(['-o', self.final_output_name])
def add_leaf_object_files(self) -> None:
"""
Goes over all the object files that the original node transitively depends on, and adds
them to the link command if they have not already been added.
"""
transitive_deps = self.initial_node.get_recursive_deps(
skip_node_types=set([NodeType.EXECUTABLE]))
with WorkDirContext(self.build_root):
# Sort nodes by path for determinism.
for node in sorted(list(transitive_deps),
key=lambda dep: dep.path):
if node in self.obj_file_graph_nodes:
# Dedup .cc.o files already existing on the command line.
continue
if node.node_type == NodeType.OBJECT:
self.new_args.add_new_arg(node.path)
for arg in self.yb_pgbackend_link_cmd:
if arg.endswith('.o'):
if os.path.basename(arg) == 'main_cpp_wrapper.cc.o':
# TOOD: why is this file even linked into libyb_pgbackend?
continue
self.new_args.append(
os.path.join(self.pg_backend_build_dir, arg))
continue
if (arg.startswith('-l') and not self.new_args.contains(arg)
and not arg.startswith('-lyb_')):
self.process_arg(arg)
def add_final_args(self, lto_type: str) -> None:
assert lto_type in ['full', 'thin']
self.new_args.extend([
'-L%s' % os.path.join(self.build_root, 'postgres', 'lib'),
'-l:libpgcommon.a',
'-l:libpgport.a',
'-l:libpq.a',
'-fwhole-program',
'-Wl,-v',
'-nostdlib++',
'-flto=' + lto_type,
])
for lib_name in ['libc++.a', 'libc++abi.a']:
self.new_args.append(
os.path.join(self.thirdparty_path, 'installed',
'uninstrumented', 'libcxx', 'lib', lib_name))
with WorkDirContext(self.build_root):
self.write_link_cmd_file(self.final_output_name +
'_lto_link_cmd_args.txt')
def run_linker(self) -> None:
with WorkDirContext(self.build_root):
start_time_sec = time.time()
logging.info("Running linker")
try:
subprocess.check_call(self.new_args.args)
except subprocess.CalledProcessError as ex:
# Avoid printing the extremely long command line.
logging.error("Linker returned exit code %d", ex.returncode)
elapsed_time_sec = time.time() - start_time_sec
logging.info("Linking finished in %.1f sec", elapsed_time_sec)
def write_link_cmd_file(self, out_path: str) -> None:
logging.info(
"Writing the linker command line (one argument per line) to %s",
os.path.abspath(out_path))
write_file('\n'.join(self.new_args.args), out_path)
class LinkHelper:
"""
Helps with creating and running a custom linking command for YugabyteDB outside of the build
system.
"""
dep_graph: DependencyGraph
initial_node: Node
# Arguments to the linker in the original command that produces the initial_node target.
# Does not include the compiler driver executable.
original_link_args: List[str]
build_root: str
build_paths: BuildPaths
llvm_path: str
thirdparty_path: str
clang_cpp_path: str
# Dependency graph nodes corresponding to the object files present in the original linker
# command. Used for deduplication.
obj_file_graph_nodes: Set[Node]
new_args: LinkCommand
# Build directory of the Postgres backend.
pg_backend_build_dir: str
# The command for linking the yb_pgbackend library.
yb_pgbackend_link_cmd: List[str]
lto_output_suffix: Optional[str]
# Populated by consume_original_link_cmd.
final_output_name: str
# We look at shared library dependencies (detected using ldd) of the libraries we add, and for
# those dependencies that fall within the third-party directory, we determine the corresponding
# static libraries and add them to the list below so we can link with them. This is necessary
# because in some cases, e.g. for libgssapi_krb5, the static libraries we need to add cannot be
# determined in any other way. The dictionary below maps the static library to the list of
# shared libraries that necessitated its addition.
static_libs_from_ldd: Dict[str, Set[str]]
# Set of shared library file paths for which we have already examined dependencies using ldd
# as described above.
processed_shared_lib_deps_for: Set[str]
def __init__(self, dep_graph: DependencyGraph, initial_node: Node,
lto_output_suffix: Optional[str]) -> None:
self.dep_graph = dep_graph
self.initial_node = initial_node
self.build_root = self.dep_graph.conf.build_root
self.build_paths = BuildPaths(self.build_root)
self.llvm_path = self.build_paths.get_llvm_path()
self.thirdparty_path = self.build_paths.get_thirdparty_path()
self.clang_cpp_path = self.build_paths.get_llvm_tool_path('clang++')
assert initial_node.link_cmd
self.original_link_args = process_original_link_cmd(
initial_node.link_cmd)
self.static_lib_paths = get_static_lib_paths(self.thirdparty_path)
self.new_args = LinkCommand([self.clang_cpp_path])
self.obj_file_graph_nodes = set()
self.pg_backend_build_dir, self.yb_pgbackend_link_cmd = get_yb_pgbackend_link_cmd(
self.build_root)
self.lto_output_suffix = lto_output_suffix
self.static_libs_from_ldd = {}
self.processed_shared_lib_deps_for = set()
def convert_to_static_lib(self, arg: str) -> Optional[str]:
"""
Given an argument to the original linker command, try to interpret it as a library, either
specified as a shared library path, or using -l... syntax, and return the corresponding
static library path if available.
"""
assert not is_system_lib(arg)
if os.path.isabs(arg):
lib_path_prefix, shared_lib_suffix = split_shared_lib_ext(arg)
if lib_path_prefix is not None:
static_found = False
lib_path_prefixes: List[str] = [
item for item in [
lib_path_prefix,
remove_dash_numeric_suffix(lib_path_prefix)
] if item is not None
]
static_lib_candidates = [
lib_path_prefix + suffix
for lib_path_prefix in lib_path_prefixes
for suffix in STATIC_LIBRARY_SUFFIXES
]
for static_lib_path in static_lib_candidates:
if os.path.exists(static_lib_path):
logging.info(
"Using static library %s instead of shared library %s",
static_lib_path, arg)
return static_lib_path
raise ValueError(
"Did not find static library corresponding to %s" % arg)
if arg.startswith('-l'):
static_found = False
logging.info("Looking for static lib for: %s", arg)
lib_name = arg[2:]
for static_lib_path in self.static_lib_paths:
static_lib_basename = os.path.basename(static_lib_path)
if any(static_lib_basename == 'lib' + lib_name + suffix
for suffix in STATIC_LIBRARY_SUFFIXES):
logging.info("Found static lib for %s: %s", lib_name,
static_lib_path)
self.add_shared_library_dependencies(
find_shared_lib_from_static(static_lib_path))
return static_lib_path
logging.info("Did not find a static lib for %s", lib_name)
if is_shared_lib(arg):
raise ValueError("Still using a shared library: %s" % arg)
return None
def add_shared_library_dependencies(self,
shared_library_path: str) -> None:
if shared_library_path in self.processed_shared_lib_deps_for:
return
self.processed_shared_lib_deps_for.add(shared_library_path)
if not os.path.exists(shared_library_path):
logging.info("File does ont exist, not running ldd: %s",
shared_library_path)
return
ldd_output = subprocess.check_output(['ldd', shared_library_path
]).decode('utf-8')
for line in ldd_output.split('\n'):
line = line.strip()
ldd_output_line_match = LDD_OUTPUT_LINE_RE.match(line)
if ldd_output_line_match:
so_name = ldd_output_line_match.group(1)
so_path = ldd_output_line_match.group(2)
if so_path.startswith(self.thirdparty_path + '/'):
static_lib_path = self.convert_to_static_lib(so_path)
if static_lib_path:
if os.path.basename(
static_lib_path) in LIBCXX_STATIC_LIB_NAMES:
# Skip libc++ and libc++abi, we will add them explicitly later.
# All third-party libraries written in C++ will depend on these and it
# is not very useful to include that in the output.
continue
if static_lib_path not in self.static_libs_from_ldd:
self.static_libs_from_ldd[static_lib_path] = set()
self.static_libs_from_ldd[static_lib_path].add(
os.path.realpath(shared_library_path))
def process_arg(self, arg: str) -> None:
assert arg is not None
if arg in SKIPPED_ARGS:
logging.info("Skipping argument: %s", arg)
return
if is_system_lib(arg):
if is_static_lib(arg):
raise ValueError(
"Linking with a system static library is not allowed: %s" %
arg)
if is_shared_lib(arg):
name = os.path.basename(arg)
if name == 'librt.so':
arg = '-lrt'
else:
raise ValueError("System shared library: %s" % arg)
else:
if is_shared_lib(arg):
self.add_shared_library_dependencies(arg)
arg = self.convert_to_static_lib(arg) or arg
self.new_args.add_new_arg(arg)
def consume_original_link_cmd(self) -> None:
"""
Goes over the original linker command and reuses some of its arguments for the new command.
"""
with WorkDirContext(self.build_root):
expect_output_name = False
output_name: Optional[str] = None
for arg in self.original_link_args:
if arg == '-o':
expect_output_name = True
continue
if expect_output_name:
if output_name:
raise ValueError(
"Found multiple output names in the original link command: "
"%s and %s" % (output_name, arg))
output_name = arg
expect_output_name = False
continue
expect_output_name = False
if is_yb_library(arg):
logging.info("Skipping YB library: %s", arg)
continue
if arg.endswith('.cc.o'):
# E.g. tablet_server_main.cc.o.
# Remember this node for later deduplication.
self.obj_file_graph_nodes.add(
self.dep_graph.find_node(os.path.realpath(arg)))
self.process_arg(arg)
if not output_name:
raise ValueError(
"Did not find an output name in the original link command")
self.final_output_name = os.path.abspath(output_name)
logging.info("Final output file name: %s", self.final_output_name)
if self.lto_output_suffix is not None:
self.final_output_name += self.lto_output_suffix
self.new_args.extend(['-o', self.final_output_name])
def add_leaf_object_files(self) -> None:
"""
Goes over all the object files that the original node transitively depends on, and adds
them to the link command if they have not already been added.
"""
transitive_deps = self.initial_node.get_recursive_deps(
skip_node_types=set([NodeType.EXECUTABLE]))
with WorkDirContext(self.build_root):
# Sort nodes by path for determinism.
for node in sorted(list(transitive_deps),
key=lambda dep: dep.path):
if node in self.obj_file_graph_nodes:
# Dedup .cc.o files already existing on the command line.
continue
if node.node_type == NodeType.OBJECT:
self.new_args.add_new_arg(node.path)
for arg in self.yb_pgbackend_link_cmd:
if arg.endswith('.o'):
if os.path.basename(arg) == 'main_cpp_wrapper.cc.o':
# TOOD: why is this file even linked into libyb_pgbackend?
continue
self.new_args.append(
os.path.join(self.pg_backend_build_dir, arg))
continue
if (arg.startswith('-l') and not self.new_args.contains(arg)
and not arg.startswith('-lyb_')):
self.process_arg(arg)
def add_final_args(self, lto_type: str) -> None:
assert lto_type in ['full', 'thin']
for static_lib_path in sorted(self.static_libs_from_ldd):
if not self.new_args.contains(static_lib_path):
logging.info(
"Adding a static library determined using shared library dependencies: %s "
"(needed by: %s)",
static_lib_path,
# The static_libs_from_ldd dictionary stores the set of shared library paths
# that caused us to add a particular static library dependency as the value
# corresponding to that static library's path in the key.
', '.join(
sorted(self.static_libs_from_ldd[static_lib_path])))
self.new_args.append(static_lib_path)
self.new_args.extend([
'-L%s' % os.path.join(self.build_root, 'postgres', 'lib'),
'-l:libpgcommon.a',
'-l:libpgport.a',
'-l:libpq.a',
'-fwhole-program',
'-Wl,-v',
'-nostdlib++',
# For __res_nsearch, ns_initparse, ns_parserr, ns_name_uncompress.
# See https://github.com/yugabyte/yugabyte-db/issues/12738 for details.
'-lresolv',
'-flto=' + lto_type,
])
for lib_name in LIBCXX_STATIC_LIB_NAMES:
self.new_args.append(
os.path.join(self.thirdparty_path, 'installed',
'uninstrumented', 'libcxx', 'lib', lib_name))
with WorkDirContext(self.build_root):
self.write_link_cmd_file(self.final_output_name +
'_lto_link_cmd_args.txt')
def run_linker(self) -> None:
with WorkDirContext(self.build_root):
start_time_sec = time.time()
logging.info("Running linker")
try:
subprocess.check_call(self.new_args.as_list())
except subprocess.CalledProcessError as ex:
# Avoid printing the extremely long command line.
logging.error("Linker returned exit code %d", ex.returncode)
elapsed_time_sec = time.time() - start_time_sec
logging.info("Linking finished in %.1f sec", elapsed_time_sec)
def write_link_cmd_file(self, out_path: str) -> None:
logging.info(
"Writing the linker command line (one argument per line) to %s",
os.path.abspath(out_path))
write_file('\n'.join(self.new_args.as_list()), out_path)