clang's memory sanitizer (-fsanitize=memory) turns up useful problems in code.  I'm working on getting a CPython buildbot running it setup but would like our build to be cleaner to start with before I run that.

These are the initial fixes required for most of CPython to pass in an msan build.  We've been using these with our interpreters at Google.  (PR coming)

New changeset 1584a0081500d35dc93ff88e5836df35faf3e3e2 by Gregory P. Smith in branch 'master':
bpo-35214: Initial clang MemorySanitizer support (GH-10479)
https://github.com/python/cpython/commit/1584a0081500d35dc93ff88e5836df35faf3e3e2

New changeset 5f4d05d83fde05fed0b6554d0beee4f1c1f1d8f1 by Gregory P. Smith in branch '3.7':
[3.7] bpo-35214: Initial clang MemorySanitizer support (GH-10479) (GH-10492)
https://github.com/python/cpython/commit/5f4d05d83fde05fed0b6554d0beee4f1c1f1d8f1

New changeset 3b5b1c0931cd32253bbf2f3bd74e90ff821e026e by Gregory P. Smith in branch '3.6':
[3.6] bpo-35214: Initial clang MemorySanitizer support (GH-10479) (GH-10493)
https://github.com/python/cpython/commit/3b5b1c0931cd32253bbf2f3bd74e90ff821e026e

Can we prefix MEMORY_SANITIZER with _Py_?

I cannot initialize the interpreter after compiling with --with-memory-sanitizer:

❯ CC=clang ./configure  --with-memory-sanitizer && make -j
❯ ./python
Python 3.8.0a0 (heads/master:1584a00815, Nov 13 2018, 03:29:18)
[Clang 7.0.0 (tags/RELEASE_700/final)] on linux
Type "help", "copyright", "credits" or "license" for more information.
==10989==WARNING: MemorySanitizer: use-of-uninitialized-value
    #0 0x5592f18005c3  (/home/pablogsal/cpython/python+0x9a55c3)
    #1 0x5592f175c176  (/home/pablogsal/cpython/python+0x901176)
    #2 0x5592f17592da  (/home/pablogsal/cpython/python+0x8fe2da)
    #3 0x5592f1750f82  (/home/pablogsal/cpython/python+0x8f5f82)
    #4 0x5592f174a336  (/home/pablogsal/cpython/python+0x8ef336)
    #5 0x5592f174c906  (/home/pablogsal/cpython/python+0x8f1906)
    #6 0x5592f14ae214  (/home/pablogsal/cpython/python+0x653214)
    #7 0x5592f14a6915  (/home/pablogsal/cpython/python+0x64b915)
    #8 0x5592f14a293c  (/home/pablogsal/cpython/python+0x64793c)
    #9 0x5592f0f5ad88  (/home/pablogsal/cpython/python+0xffd88)
    #10 0x5592f0f5ce73  (/home/pablogsal/cpython/python+0x101e73)
    #11 0x5592f0f4d908  (/home/pablogsal/cpython/python+0xf2908)
    #12 0x7fd1a7381222  (/usr/lib/libc.so.6+0x24222)
    #13 0x5592f0ed3cdd  (/home/pablogsal/cpython/python+0x78cdd)

  Uninitialized value was created by a heap allocation
    #0 0x5592f0f02a0d  (/home/pablogsal/cpython/python+0xa7a0d)
    #1 0x7fd1a73cd790  (/usr/lib/libc.so.6+0x70790)

SUMMARY: MemorySanitizer: use-of-uninitialized-value (/home/pablogsal/github/cpython/python+0x9a55c3)
Exiting

❯ clang --version
clang version 7.0.0 (tags/RELEASE_700/final)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

❯ /lib/libc.so.6
GNU C Library (GNU libc) stable release version 2.28.
Copyright (C) 2018 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.
Compiled by GNU CC version 8.2.1 20180831.
libc ABIs: UNIQUE IFUNC ABSOLUTE
For bug reporting instructions, please see:
<https://bugs.archlinux.org/>.

I am missing something or is this expected (as there are more PRs to come)?

All libraries that are linked against, including libc, need to be compiled with MSAN. MSAN is not for the faint of heart.

Thanks @alex! I will try again after linking against ASAN-compiled version of my libraries.

yeah, i've been surprised how far i was able to get from an oss CPython tree and pre-built clang 7 binary installation.  If you have headers installed for common libraries like libssl-dev and libreadline-dev you can't produce an interpreter that'll work as those both lead to quick crashes.  also, in my experience optimized builds tend to be more problematic (too early to say if what i'm seeing are actual problems).

also, make sure llvm-symbolize (no suffixes) is in your path for it to auto-symbolize the traces showing you where in the code it pointed out issues.  when it points within shared libraries instead of the Python source tree - that's a hint that the library probably may need msan compilation.

I'm plodding through things to see if I can get a _useful_ buildbot setup (i don't care of some extension modules can't be covered for now) out of all of this.

> Can we prefix MEMORY_SANITIZER with _Py_?

Yes, I wondered if I should do that.  not that I expect anyone would ever define it to mean anything else, but that seems like the right thing to do.  https://github.com/python/cpython/pull/10503

New changeset 3015fb8ce4d25603434b9b44bb7effb98a481532 by Gregory P. Smith in branch 'master':
bpo-35214: Add _Py_ prefix to MEMORY_SANITIZER def. (GH-10503)
https://github.com/python/cpython/commit/3015fb8ce4d25603434b9b44bb7effb98a481532

New changeset f6602f92664b782125a1ec5009706f9300fe7e80 by Miss Islington (bot) in branch '3.7':
bpo-35214: Add _Py_ prefix to MEMORY_SANITIZER def. (GH-10503)
https://github.com/python/cpython/commit/f6602f92664b782125a1ec5009706f9300fe7e80

New changeset 60cf26528b931215160a53fd61836dbe4f036009 by Miss Islington (bot) in branch '3.6':
bpo-35214: Add _Py_ prefix to MEMORY_SANITIZER def. (GH-10503)
https://github.com/python/cpython/commit/60cf26528b931215160a53fd61836dbe4f036009

New changeset 746b2d35ea47005054ed774fecaed64fab803d7d by Gregory P. Smith in branch 'master':
bpo-35214: Fix OOB memory access in unicode escape parser (GH-10506)
https://github.com/python/cpython/commit/746b2d35ea47005054ed774fecaed64fab803d7d

New changeset fdc485a5dff8508328b4f26265d0af1ba5004597 by Miss Islington (bot) in branch '3.6':
bpo-35214: Fix OOB memory access in unicode escape parser (GH-10506)
https://github.com/python/cpython/commit/fdc485a5dff8508328b4f26265d0af1ba5004597

New changeset 9fbcb1402efab4e287f25145a69ba14c9c6dbce9 by Miss Islington (bot) in branch '3.7':
[3.7] bpo-35214: Fix OOB memory access in unicode escape parser (GH-10506) (GH-10522)
https://github.com/python/cpython/commit/9fbcb1402efab4e287f25145a69ba14c9c6dbce9

New changeset b6f4472dc4190e2fd668490d86aeefd2ab0df935 by Gregory P. Smith in branch '2.7':
[2.7] bpo-35214: Fix OOB memory access in unicode escape parser (GH-10506) (GH-10538)
https://github.com/python/cpython/commit/b6f4472dc4190e2fd668490d86aeefd2ab0df935

Is this issue completely fixed?

I believe there are still some issues to deal with.  I don't want to close the issue until I've got my buildbot running.

New changeset b474e6774d60fa67d5373e361a0ed53c18b24f53 by Gregory P. Smith in branch 'master':
bpo-35214: MSan workarounds for socket, time, and test_faulthandler. (GH-11375)
https://github.com/python/cpython/commit/b474e6774d60fa67d5373e361a0ed53c18b24f53

New changeset 01b9664740307b39c2907bd84cbb0b2c35be9df4 by Gregory P. Smith (Miss Islington (bot)) in branch '3.7':
bpo-35214: MSan workarounds for socket, time, and test_faulthandler. (GH-11375) (GH-11378)
https://github.com/python/cpython/commit/01b9664740307b39c2907bd84cbb0b2c35be9df4

New changeset e5796c42c687e1454e84dcc50e6f67db48ff69a0 by Gregory P. Smith in branch 'master':
bpo-35214: Skip test_io tests that'd cause a huge malloc under msan (#11385)
https://github.com/python/cpython/commit/e5796c42c687e1454e84dcc50e6f67db48ff69a0

Status on my upcoming buildbot host after today's changes:

== Tests result: FAILURE ==

375 tests OK.

11 tests failed:
    test_asyncio test_builtin test_code test_ctypes test_ioctl
    test_openpty test_os test_posix test_pty test_shutil test_uuid

32 tests skipped:
    test_bz2 test_curses test_dbm_gnu test_dbm_ndbm test_devpoll
    test_gzip test_idle test_kqueue test_lzma test_msilib
    test_ossaudiodev test_readline test_smtpnet test_socketserver
    test_sqlite test_ssl test_startfile test_tcl test_timeout test_tix
    test_tk test_ttk_guionly test_ttk_textonly test_turtle
    test_urllib2net test_urllibnet test_winconsoleio test_winreg
    test_winsound test_xmlrpc_net test_zipfile64 test_zlib


Most of those are dying due to pty use (openpty, etc) which is not properly memory sanitizer traced.  test_posix appears to have something
I can fix by annotating in the code.

after that, I'll decide how to tell my buildbot not to run those tests so we can have a green buildbot memory sanitizing everything else.

New changeset 5d2e4b1ff2f01b6aeac2f2f302f363d3eed225fa by Miss Islington (bot) in branch '3.7':
bpo-35214: Skip test_io tests that'd cause a huge malloc under msan (GH-11385)
https://github.com/python/cpython/commit/5d2e4b1ff2f01b6aeac2f2f302f363d3eed225fa

New changeset 1d300ce1d8238136595c8fea76266a4755cd73a2 by Gregory P. Smith in branch 'master':
bpo-35214: Annotate posix calls for clang MSan. (#11389)
https://github.com/python/cpython/commit/1d300ce1d8238136595c8fea76266a4755cd73a2

New changeset efcf08d8ca9084d8248715f0634c21b705f02ca2 by Gregory P. Smith in branch '3.7':
[3.7] bpo-35214: Annotate posix calls for clang MSan. (GH-11389) (GH-11391)
https://github.com/python/cpython/commit/efcf08d8ca9084d8248715f0634c21b705f02ca2

Can this be closed now?

Nope, work remains to be done.  I've got an msan buildbot system waiting but haven't had time to follow up on figuring out what remains in a while.  (getting a functioning memory sanitizer build is... finnicky to say the least)