Link-time-optimization with clang is broken

[matthiasgoergens]

I tried to enable link time optimizations for clang.

export CC="clang"
export LD="clang"

nice "${here}/configure" \
    --enable-optimizations \
    --with-lto=thin

nice make -j8

Eventually building reaches this step:

clang -fprofile-instr-generate -o _bootstrap_python Modules/getbuildinfo.o Parser/token.o Parser/pegen.o Parser/pegen_errors.o Parser/action_helpers.o Parser/parser.o Parser/string_parser.o Parser/peg_api.o Parser/myreadline.o Parser/tokenizer.o Objects/abstract.o Objects/boolobject.o Objects/bytes_methods.o Objects/bytearrayobject.o Objects/bytesobject.o Objects/call.o Objects/capsule.o Objects/cellobject.o Objects/classobject.o Objects/codeobject.o Objects/complexobject.o Objects/descrobject.o Objects/enumobject.o Objects/exceptions.o Objects/genericaliasobject.o Objects/genobject.o Objects/fileobject.o Objects/floatobject.o Objects/frameobject.o Objects/funcobject.o Objects/interpreteridobject.o Objects/iterobject.o Objects/listobject.o Objects/longobject.o Objects/dictobject.o Objects/odictobject.o Objects/memoryobject.o Objects/methodobject.o Objects/moduleobject.o Objects/namespaceobject.o Objects/object.o Objects/obmalloc.o Objects/picklebufobject.o Objects/rangeobject.o Objects/setobject.o Objects/sliceobject.o Objects/structseq.o Objects/tupleobject.o Objects/typeobject.o Objects/unicodeobject.o Objects/unicodectype.o Objects/unionobject.o Objects/weakrefobject.o Objects/perf_trampoline.o Objects/asm_trampoline.o Python/_warnings.o Python/Python-ast.o Python/Python-tokenize.o Python/asdl.o Python/ast.o Python/ast_opt.o Python/ast_unparse.o Python/bltinmodule.o Python/ceval.o Python/codecs.o Python/compile.o Python/context.o Python/dynamic_annotations.o Python/errors.o Python/frame.o Python/frozenmain.o Python/future.o Python/getargs.o Python/getcompiler.o Python/getcopyright.o Python/getplatform.o Python/getversion.o Python/ceval_gil.o Python/hamt.o Python/hashtable.o Python/import.o Python/importdl.o Python/initconfig.o Python/marshal.o Python/modsupport.o Python/mysnprintf.o Python/mystrtoul.o Python/pathconfig.o Python/preconfig.o Python/pyarena.o Python/pyctype.o Python/pyfpe.o Python/pyhash.o Python/pylifecycle.o Python/pymath.o Python/pystate.o Python/pythonrun.o Python/pytime.o Python/bootstrap_hash.o Python/specialize.o Python/structmember.o Python/symtable.o Python/sysmodule.o Python/thread.o Python/traceback.o Python/getopt.o Python/pystrcmp.o Python/pystrtod.o Python/pystrhex.o Python/dtoa.o Python/formatter_unicode.o Python/fileutils.o Python/suggestions.o Python/dynload_shlib.o Modules/config.o Modules/main.o Modules/gcmodule.o Modules/atexitmodule.o Modules/faulthandler.o Modules/posixmodule.o Modules/signalmodule.o Modules/_tracemalloc.o Modules/_codecsmodule.o Modules/_collectionsmodule.o Modules/errnomodule.o Modules/_io/_iomodule.o Modules/_io/iobase.o Modules/_io/fileio.o Modules/_io/bytesio.o Modules/_io/bufferedio.o Modules/_io/textio.o Modules/_io/stringio.o Modules/itertoolsmodule.o Modules/_sre/sre.o Modules/_threadmodule.o Modules/timemodule.o Modules/_weakref.o Modules/_abc.o Modules/_functoolsmodule.o Modules/_localemodule.o Modules/_operator.o Modules/_stat.o Modules/symtablemodule.o Modules/pwdmodule.o Programs/_bootstrap_python.o Modules/getpath.o -ldl -lm

And fails:

Modules/getbuildinfo.o: file not recognized: file format not recognized
clang-14: error: linker command failed with exit code 1 (use -v to see invocation)

Your environment

I'm on Archlinux x86-64.

$ clang --version
clang version 14.0.6
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

I tested against main.

[corona10]

-no-lto flag was first introduced at https://github.com/python/cpython/pull/29859 by @tiran Would you like to take a look at this issue please?

[matthiasgoergens]

@corona10 Thanks! That seems to be mostly about speeding up the GCC build. LLVM is already pretty fast with thin-lto.

[corona10]

I will try to reproduce the issue in my Linux environment with clang14 before reviewing the PR. cc @tiran

[corona10]

Okay, I am able to reproduce the issue even with the lld.

[matthiasgoergens]

Thanks for confirming that it's not just some idiosyncratic weirdness of my environment.

[corona10]

For @matthiasgoergens, @tiran @vstinner cc @pablogsal as release manager.

My conclusion

• At this moment Modules/*.o are actually LLVM bitcode not object files.
• so the LTO processor needs to transform the LLVM bitcode into the object file for the traditional linker but looks like without -flto it doesn't transform to object files for the traditional linker.
• This issue not only happens for ThinLTO but also in full LTO because both processes have the phase for transforming the LLVM bitcode into the object file.
• see http://blog.llvm.org/2016/06/thinlto-scalable-and-incremental-lto.html
• IMHO we can add the transformation phase for -fno-lto but it could be complicated, I prefer using https://github.com/python/cpython/pull/96762 as a workaround approach and we need to backport to Python 3.11 too.
• But I am still curious about the fact why apple clang does not raise this issue.

Reference

file format

corona10@python-dev:~/cpython$ file  Modules/getbuildinfo.o
Modules/getbuildinfo.o: LLVM IR bitcode

clang full LTO process

clang ThinLTO process

[pablogsal]

Why is this only affecting 3.11 and not previous version?

[corona10]

Why is this only affecting 3.11 and not previous version?

because https://github.com/python/cpython/pull/29859 was first introduced in Python 3.11 for _bootstrap_python.

[corona10]

When building the final binary for CPython, we actually pass the -flto flags that the process will transform the LLVM bitcode into object files, so this issue only affects while building the _bootstrap_python. I didn't check if the issue will affect to old versions (< 3.10)

[corona10]

Manual transformation example:

corona10@python-dev:~/cpython$ llc-14 -filetype=obj Modules/getbuildinfo.o -o Modules/getbuildinfo.o
corona10@python-dev:~/cpython$ file Modules/getbuildinfo.o
Modules/getbuildinfo.o: ELF 64-bit LSB relocatable, x86-64, version 1 (GNU/Linux), with debug_info, not stripped

[pablogsal]

Thanks for the quick answer and great analysis @corona10. I think the fix for this should go to 3.11.1 as this likely involve configure changes unless someone fundamentally disagrees.

[matthiasgoergens]

As far as I can tell, building _bootstrap_python with different flags than the full Python binary was only a hack to speed up the build, wasn't it?

Thin LTO is pretty fast, and compatible with the other LTO settings (apart perhaps from no-LTO), so we can build the bootstrap with that one, if any LTO is requested for the main build?

[corona10]

Thin LTO is pretty fast, and compatible with the other LTO settings (apart perhaps from no-LTO), so we can build the bootstrap with that one, if any LTO is requested for the main build?

Please follow my last review: https://github.com/python/cpython/pull/96762#discussion_r973207014 Since we should support all C11 compilers including old clangs which doesn't support ThinLTO, the configuration should use ThinLTO only if the compiler supports it. We should not break the build for old compilers too. see: https://clang.llvm.org/c_status.html and https://peps.python.org/pep-0007/#c-dialect

[matthiasgoergens]

@corona10 Yes, definitely. I need to add checks.

[corona10]

All patches are merged. Feel free to reopen if we need other solutions. cc @tiran @pablogsal