[MLIR, Python] Make it easy to run tests with ASan on mac (#115524)

There are two things that make running MLIR tests with ASan on a mac
tedious:

1. The `DYLD_INSERT_LIBRARIES` environment variable needs to be set to
point to `libclang_rt.asan_osx_dynamic.dylib`
2. Mac is wrapping Python, which means that the `DYLD_INSERT_LIBRARIES`
environment variable is not being respected in the Python tests. The
solution is to find and use a non-wrapped Python binary.

With the above two changes, ASan works out of the box on mac's by
setting the `-DLLVM_USE_SANITIZER=Address` cmake flag.

I have stolen most of the code in this PR from other LLVM projects. It
may be a good idea to reconcile it somewhere.

Author: kasper0406

mlir/test/get_darwin_real_python.py

@@ -0,0 +1,16 @@
+# On macOS, system python binaries like /usr/bin/python and $(xcrun -f python3)
+# are shims. They do some light validation work and then spawn the "real" python
+# binary. Find the "real" python by asking dyld -- sys.executable reports the
+# wrong thing more often than not. This is also useful when we're running under
+# a Homebrew python3 binary, which also appears to be some kind of shim.
+def getDarwinRealPythonExecutable():
+    import ctypes
+
+    dyld = ctypes.cdll.LoadLibrary("/usr/lib/system/libdyld.dylib")
+    namelen = ctypes.c_ulong(1024)
+    name = ctypes.create_string_buffer(b"\000", namelen.value)
+    dyld._NSGetExecutablePath(ctypes.byref(name), ctypes.byref(namelen))
+    return name.value.decode("utf-8").strip()
+
+
+print(getDarwinRealPythonExecutable())

mlir/test/lit.cfg.py

@@ -3,6 +3,7 @@
 import os
 import platform
 import re
+import shutil
 import subprocess
 import tempfile
 
@@ -77,6 +78,75 @@ def add_runtime(name):
     return ToolSubst(f"%{name}", find_runtime(name))
 
 
+# Provide the path to asan runtime lib 'libclang_rt.asan_osx_dynamic.dylib' if
+# available. This is darwin specific since it's currently only needed on darwin.
+# Stolen from llvm/test/lit.cfg.py with a few modifications
+def get_asan_rtlib():
+    if not "asan" in config.available_features or not "Darwin" in config.host_os:
+        return ""
+    # Find the asan rt lib
+    resource_dir = (
+        subprocess.check_output([config.host_cc.strip(), "-print-resource-dir"])
+        .decode("utf-8")
+        .strip()
+    )
+    return os.path.join(
+        resource_dir, "lib", "darwin", "libclang_rt.asan_osx_dynamic.dylib"
+    )
+
+
+# On macOS, we can't do the DYLD_INSERT_LIBRARIES trick with a shim python
+# binary as the ASan interceptors get loaded too late. Also, when SIP is
+# enabled, we can't inject libraries into system binaries at all, so we need a
+# copy of the "real" python to work with.
+# Stolen from lldb/test/API/lit.cfg.py with a few modifications
+def find_real_python_interpreter():
+    # If we're running in a virtual environment, we have to copy Python into
+    # the virtual environment for it to work.
+    if sys.prefix != sys.base_prefix:
+        copied_python = os.path.join(sys.prefix, "bin", "copied-python")
+    else:
+        copied_python = os.path.join(config.lldb_build_directory, "copied-python")
+
+    # Avoid doing any work if we already copied the binary.
+    if os.path.isfile(copied_python):
+        return copied_python
+
+    # Find the "real" python binary.
+    real_python = (
+        subprocess.check_output(
+            [
+                config.python_executable,
+                os.path.join(
+                    os.path.dirname(os.path.realpath(__file__)),
+                    "get_darwin_real_python.py",
+                ),
+            ]
+        )
+        .decode("utf-8")
+        .strip()
+    )
+
+    shutil.copy(real_python, copied_python)
+
+    # Now make sure the copied Python works. The Python in Xcode has a relative
+    # RPATH and cannot be copied.
+    try:
+        # We don't care about the output, just make sure it runs.
+        subprocess.check_call([copied_python, "-V"])
+    except subprocess.CalledProcessError:
+        # The copied Python didn't work. Assume we're dealing with the Python
+        # interpreter in Xcode. Given that this is not a system binary SIP
+        # won't prevent us form injecting the interceptors, but when running in
+        # a virtual environment, we can't use it directly. Create a symlink
+        # instead.
+        os.remove(copied_python)
+        os.symlink(real_python, copied_python)
+
+    # The copied Python works.
+    return copied_python
+
+
 llvm_config.with_system_environment(["HOME", "INCLUDE", "LIB", "TMP", "TEMP"])
 
 llvm_config.use_default_substitutions()
@@ -91,6 +161,7 @@ def add_runtime(name):
     "LICENSE.txt",
     "lit.cfg.py",
     "lit.site.cfg.py",
+    "get_darwin_real_python.py",
 ]
 
 # Tweak the PATH to include the tools dir.
@@ -172,10 +243,30 @@ def add_runtime(name):
 )
 
 python_executable = config.python_executable
-# Python configuration with sanitizer requires some magic preloading. This will only work on clang/linux.
-# TODO: detect Darwin/Windows situation (or mark these tests as unsupported on these platforms).
-if "asan" in config.available_features and "Linux" in config.host_os:
-    python_executable = f"LD_PRELOAD=$({config.host_cxx} -print-file-name=libclang_rt.asan-{config.host_arch}.so) {config.python_executable}"
+# Python configuration with sanitizer requires some magic preloading. This will only work on clang/linux/darwin.
+# TODO: detect Windows situation (or mark these tests as unsupported on these platforms).
+if "asan" in config.available_features:
+    if "Linux" in config.host_os:
+        python_executable = f"LD_PRELOAD=$({config.host_cxx} -print-file-name=libclang_rt.asan-{config.host_arch}.so) {config.python_executable}"
+    if "Darwin" in config.host_os:
+        # Ensure we use a non-shim Python executable, for the `DYLD_INSERT_LIBRARIES`
+        # env variable to take effect
+        real_python_executable = find_real_python_interpreter()
+        if real_python_executable:
+            python_executable = real_python_executable
+            lit_config.note(
+                "Using {} instead of {}".format(
+                    python_executable, config.python_executable
+                )
+            )
+
+        asan_rtlib = get_asan_rtlib()
+        lit_config.note("Using ASan rtlib {}".format(asan_rtlib))
+        config.environment["MallocNanoZone"] = "0"
+        config.environment["ASAN_OPTIONS"] = "detect_stack_use_after_return=1"
+        config.environment["DYLD_INSERT_LIBRARIES"] = asan_rtlib
+
+
 # On Windows the path to python could contains spaces in which case it needs to be provided in quotes.
 # This is the equivalent of how %python is setup in llvm/utils/lit/lit/llvm/config.py.
 elif "Windows" in config.host_os: