Corrupts the DWARF section when relocating multiple objects into a single one

[Danil42Russia]

Environment:

• System: Debian GNU/Linux 10 (buster)
• mold version: mold 2.31.0 (20fa8d56f5e0c47d1f4bbf7b829c12d3f43298e1; compatible with GNU ld)
• g++ version: g++ (Debian 8.3.0-6) 8.3.0
• lld version: Debian LLD 16.0.6 (compatible with GNU linkers)

Problem:

After we moved to mold, we found that the functions have strange names in the addr2line output. Their beginnings may
be cut off.

How we build the final binary:

1. Codegen from PHP to c++. This turns into about 200k c++ files that lie in 100 folders
   (about 2k c++ files in one folder);
2. All c++ are compiled into o files;
3. From the o files from each folder, we make a "middle" o file via relocating
   (2k o files from a folder into one "average" o file. In the end, there will be 100 of them.);
4. 100 "average" o files are collected into a final binary.

Reproducer:

Ready code, it will run the g++ compiler and the two linkers mold and lld by itself

import shlex
import shutil
import subprocess
from pathlib import Path

CPP_BLOCK_TEMPLATE = """
int uninit_{file_id}_{iter_id};
int zeroed_{file_id}_{iter_id} = 0;
int init_{file_id}_{iter_id} = {count_id};
static int static_{file_id}_{iter_id} = {count_id};
int foo_{file_id}_{iter_id}() {{ std::cout << {count_id}; return {count_id}; }}"""

CPP_BLOCK_ITERS = 2
CPP_FILES_COUNT = 10

FOLDER_REPRODUCER = Path("/tmp/problem_reproducer")
OBJ_FOLDER = FOLDER_REPRODUCER / "obj_1"

COMPILER_PATH = "/usr/bin/g++"
BASE_LINKER_PATH = "/usr/bin"


def try_command_execute(cmd: str) -> None:
    try:
        pid = subprocess.run(shlex.split(cmd))
        return_code = pid.returncode
    except Exception:
        return_code = 1  # :D

    if return_code != 0:
        print(f"There was a startup error:", shlex.quote(cmd))


def prepare_target() -> None:
    if OBJ_FOLDER.exists():
        shutil.rmtree(OBJ_FOLDER)

    OBJ_FOLDER.mkdir(parents=True)


def gen2cpp_target(file_id: int) -> None:
    cpp_file_name = f"file_{file_id}.cpp"
    print(f"STAGE 1.1: Start codegen for file '{cpp_file_name}'")

    cpp_code = "#include <iostream>"
    for iter_id in range(1, CPP_BLOCK_ITERS + 1):
        cpp_code += f"\n// BLOCK {iter_id}"
        cpp_code += CPP_BLOCK_TEMPLATE.format(file_id=file_id, iter_id=iter_id, count_id=file_id * iter_id)

    file_path = OBJ_FOLDER / cpp_file_name
    file_path.write_text(cpp_code)


def cpp2obj_target() -> None:
    compiler_flags = "-g1 -ffunction-sections -fdata-sections -c"

    for cpp_file in OBJ_FOLDER.glob("*.cpp"):
        output_obj_file = cpp_file.parent / f"{cpp_file.stem}.o"
        cmd = f"{COMPILER_PATH} {compiler_flags} -o {output_obj_file} {cpp_file}"

        print(f"STAGE 2.2: Start compiler for file '{cpp_file.name}'")
        try_command_execute(cmd)


def objs2obj_target(linker_name: str) -> None:
    base_linker_path = f"{BASE_LINKER_PATH}/{linker_name}"
    linker_flags = "-r -nostdlib"

    target_objs_list = ""
    for obj_file in OBJ_FOLDER.glob("*.o"):
        target_objs_list += f"{obj_file} "
    target_objs_list = target_objs_list.strip()

    output_object_file = FOLDER_REPRODUCER / f"{OBJ_FOLDER.name}_{linker_name}.o"
    cmd = f"{base_linker_path} {linker_flags} -o {output_object_file} {target_objs_list}"

    print(f"STAGE 3.1: Start linker '{linker_name}'")
    try_command_execute(cmd)


def main() -> None:
    print("STAGE 0: (prepare)")
    prepare_target()
    print("STAGE 1: (generate c++)")
    for file_id in range(1, CPP_FILES_COUNT + 1):
        gen2cpp_target(file_id)

    print("STAGE 2: (cpp to obj)")
    cpp2obj_target()

    print("STAGE 3: (objs to obj)")
    linkers = ["mold", "lld"]
    for linker_name in linkers:
        objs2obj_target(linker_name)


if __name__ == "__main__":
    main()

Research:

If you run the command, you will see that the beginning of some attributes has been cut off. And part of the attributes
contains information that shouldn't be there. :(

objdump --dwarf=info obj_1_mold.o | grep "\.cpp"

objdump: Warning: DW_FORM_strp offset too big: 542
objdump: Warning:     <c>   DW_AT_producer    : (indirect string, offset: 0x1a2): /tmp/problem_reproducer/obj_1/file_7.cpp
DW_FORM_strp offset too big: 4d5
    <41>   DW_AT_name        : (indirect string, offset: 0x178): m_reproducer/obj_1/file_6.cpp
objdump: Warning: DW_FORM_strp offset too big: 4fe
objdump: Warning: DW_FORM_strp offset too big: 518
objdump: Warning: DW_FORM_strp offset too big: 59f
objdump: Warning: DW_FORM_strp offset too big: 5a7
objdump: Warning: DW_FORM_strp offset too big: 597
    <134>   DW_AT_producer    : (indirect string, offset: 0x37f): .cpp
    <139>   DW_AT_name        : (indirect string, offset: 0x32c): mp/problem_reproducer/obj_1/file_5.cpp
    <19e>   DW_AT_name        : (indirect string, offset: 0x3e0): em_reproducer/obj_1/file_1.cpp
    <2f5>   DW_AT_name        : (indirect string, offset: 0x2be): /tmp/problem_reproducer/obj_1/file_9.cpp
    <2f9>   DW_AT_comp_dir    : (indirect string, offset: 0x362): oblem_reproducer/obj_1/file_3.cpp
    <325>   DW_AT_name        : (indirect string, offset: 0x379): file_3.cpp
    <41d>   DW_AT_name        : (indirect string, offset: 0x3f1): _1/file_1.cpp
    <436>   DW_AT_name        : (indirect string, offset: 0x338): eproducer/obj_1/file_5.cpp
    <44d>   DW_AT_name        : (indirect string, offset: 0x36a): producer/obj_1/file_3.cpp
    <46b>   DW_AT_linkage_name: (indirect string, offset: 0x35e): p/problem_reproducer/obj_1/file_3.cpp
    <482>   DW_AT_name        : (indirect string, offset: 0x3e9): ucer/obj_1/file_1.cpp
    <4ac>   DW_AT_producer    : (indirect string, offset: 0x178): m_reproducer/obj_1/file_6.cpp
    <540>   DW_AT_producer    : (indirect string, offset: 0x429): roducer/obj_1/file_2.cpp
    <545>   DW_AT_name        : (indirect string, offset: 0x3ac): /tmp/problem_reproducer/obj_1/file_10.cpp
    <549>   DW_AT_comp_dir    : (indirect string, offset: 0x3d6): /tmp/problem_reproducer/obj_1/file_1.cpp
    <58c>   DW_AT_name        : (indirect string, offset: 0x3f6): le_1.cpp
    <5aa>   DW_AT_name        : (indirect string, offset: 0x3ed): /obj_1/file_1.cpp
    <5b1>   DW_AT_linkage_name: (indirect string, offset: 0x47e): e_8.cpp

I checked what's going on at lld, so it's fine.

objdump --dwarf=info obj_1_lld.o | grep "\.cpp"

    <11>   DW_AT_name        : (indirect string, offset: 0x499): /tmp/problem_reproducer/obj_1/file_1.cpp
    <a5>   DW_AT_name        : (indirect string, offset: 0x0): /tmp/problem_reproducer/obj_1/file_2.cpp
    <139>   DW_AT_name        : (indirect string, offset: 0xc9): /tmp/problem_reproducer/obj_1/file_3.cpp
    <1cd>   DW_AT_name        : (indirect string, offset: 0x1d1): /tmp/problem_reproducer/obj_1/file_4.cpp
    <261>   DW_AT_name        : (indirect string, offset: 0xfe): /tmp/problem_reproducer/obj_1/file_5.cpp
    <2f5>   DW_AT_name        : (indirect string, offset: 0x127): /tmp/problem_reproducer/obj_1/file_6.cpp
    <389>   DW_AT_name        : (indirect string, offset: 0x380): /tmp/problem_reproducer/obj_1/file_7.cpp
    <41d>   DW_AT_name        : (indirect string, offset: 0x2c0): /tmp/problem_reproducer/obj_1/file_8.cpp
    <4b1>   DW_AT_name        : (indirect string, offset: 0x41b): /tmp/problem_reproducer/obj_1/file_9.cpp
    <545>   DW_AT_name        : (indirect string, offset: 0x62): /tmp/problem_reproducer/obj_1/file_10.cpp

[rui314]

I can reproduce the issue. Thank you for the reproducer. I'll fix the issue.

[rui314]

The above commit should fix the issue. Please try again with git head.

[Danil42Russia]

Thanks for the fix! Tested it on ours, no problems noticed so far