[inductor] add _get_inductor_debug_symbol_cflags for debug symbol control. (#159938)

We need to add inductor debug symbol support for crash case debug. When we turn on generate debug symbol.
On Windows, it should create a [module_name].pdb file. It helps debug by WinDBG.
On Linux, it should create some debug sections in binary file.

I added UT for it also.

It works well on Windows inductor debug.
<img width="1648" height="833" alt="image" src="https://github.com/user-attachments/assets/5282a7de-cef3-4a38-9cd4-a0e63482c8b6" />

Pull Request resolved: https://github.com/pytorch/pytorch/pull/159938
Approved by: https://github.com/jansel, https://github.com/angelayi

test/inductor/test_compile.py

@@ -1,6 +1,14 @@
 # Owner(s): ["module: inductor"]
+import os
+import shlex
+import subprocess
+import sys
+from unittest import mock
+
 import torch
 from torch import _dynamo as dynamo, _inductor as inductor
+from torch._inductor.codecache import write
+from torch._inductor.cpp_builder import CppBuilder, CppOptions
 from torch._inductor.test_case import run_tests, TestCase
 from torch._inductor.utils import gen_gm_and_inputs
 from torch.fx import symbolic_trace
@@ -8,6 +16,25 @@ from torch.fx.experimental.proxy_tensor import make_fx
 from torch.testing._internal.inductor_utils import HAS_CPU
 
 
+_IS_MACOS = sys.platform.startswith("darwin")
+_IS_WINDOWS = sys.platform == "win32"
+
+
+def safe_command_output(cmd, timeout=30):
+    try:
+        return subprocess.check_output(
+            cmd,
+            stderr=subprocess.STDOUT,
+            text=True,
+            timeout=timeout,
+            shell=isinstance(cmd, str),
+        ).strip()
+    except subprocess.CalledProcessError as e:
+        return f"run failed（error code {e.returncode}）: {e.output.strip()}"
+    except subprocess.TimeoutExpired:
+        return "runt timeout"
+
+
 class MyModule(torch.nn.Module):
     def __init__(self) -> None:
         super().__init__()
@@ -109,6 +136,53 @@ class TestStandaloneInductor(TestCase):
         mod_opt = inductor.compile(mod, inp)
         self.assertEqual(mod(*inp), mod_opt(*inp))
 
+    @mock.patch.dict(os.environ, {"TORCHINDUCTOR_DEBUG_SYMBOL": "1"})
+    def test_inductor_generate_debug_symbol(self):
+        cpp_code = """
+int main(){
+    return 0;
+}
+        """
+
+        _, source_path = write(
+            cpp_code,
+            "cpp",
+        )
+        build_option = CppOptions()
+        cpp_builder = CppBuilder(
+            name="test_symbol",
+            sources=source_path,
+            output_dir=os.path.dirname(source_path),
+            BuildOption=build_option,
+        )
+        cpp_builder.build()
+        binary_path = cpp_builder.get_target_file_path()
+
+        """
+        When we turn on generate debug symbol.
+        On Windows, it should create a [module_name].pdb file. It helps debug by WinDBG.
+        On Linux, it should create some debug sections in binary file.
+        """
+
+        def check_linux_debug_section(module_path: str):
+            check_cmd = shlex.split(f"readelf -S {module_path}")
+            output = safe_command_output(check_cmd)
+            has_debug_sym = ".debug_info" in output
+            self.assertEqual(has_debug_sym, True)
+
+        def check_windows_pdb_exist(module_path: str):
+            file_name_no_ext = os.path.splitext(module_path)[0]
+            file_name_pdb = f"{file_name_no_ext}.pdb"
+            has_pdb_file = os.path.exists(file_name_pdb)
+            self.assertEqual(has_pdb_file, True)
+
+        if _IS_WINDOWS:
+            check_windows_pdb_exist(binary_path)
+        elif _IS_MACOS:
+            pass  # MacOS not sure that if it should be works.
+        else:
+            check_linux_debug_section(binary_path)
+
 
 if __name__ == "__main__":
     if HAS_CPU:

torch/_inductor/cpp_builder.py

@@ -637,7 +637,7 @@ def _get_optimization_cflags(
         return cflags
 
 
-def _get_shared_cflag(do_link: bool) -> list[str]:
+def _get_shared_cflags(do_link: bool) -> list[str]:
     if _IS_WINDOWS:
         """
         MSVC `/MD` using python `ucrtbase.dll` lib as runtime.
@@ -652,6 +652,25 @@ def _get_shared_cflag(do_link: bool) -> list[str]:
     return ["shared", "fPIC"]
 
 
+def _get_inductor_debug_symbol_cflags() -> tuple[list[str], list[str]]:
+    """
+    When we turn on generate debug symbol.
+    On Windows, it should create a [module_name].pdb file. It helps debug by WinDBG.
+    On Linux, it should create some debug sections in binary file.
+    """
+    cflags: list[str] = []
+    ldflags: list[str] = []
+    b_enable_debug_symbol = os.environ.get("TORCHINDUCTOR_DEBUG_SYMBOL", "0") == "1"
+    if b_enable_debug_symbol:
+        if _IS_WINDOWS:
+            cflags = ["Z7", "_DEBUG", "OD"]
+            ldflags = ["DEBUG", "OPT:REF", "OPT:ICF"]
+        else:
+            cflags.append("g")
+
+    return cflags, ldflags
+
+
 def get_cpp_options(
     cpp_compiler: str,
     do_link: bool,
@@ -667,12 +686,15 @@ def get_cpp_options(
     libraries: list[str] = []
     passthrough_args: list[str] = []
 
+    dbg_cflags, dbg_ldflags = _get_inductor_debug_symbol_cflags()
+
     cflags = (
-        _get_shared_cflag(do_link)
+        _get_shared_cflags(do_link)
         + _get_optimization_cflags(cpp_compiler, min_optimize)
         + _get_warning_all_cflag(warning_all)
         + _get_cpp_std_cflag()
         + _get_os_related_cpp_cflags(cpp_compiler)
+        + dbg_cflags
     )
 
     if not _IS_WINDOWS and config.aot_inductor.enable_lto and _is_clang(cpp_compiler):
@@ -685,7 +707,7 @@ def get_cpp_options(
         definitions,
         include_dirs,
         cflags,
-        ldflags,
+        ldflags + dbg_ldflags,
         libraries_dirs,
         libraries,
         passthrough_args,