Build is broken with clang 3.8

[karkhaz]

I'm assigning @smowton because git blame told me to, but I doubt it's actually your fault. I'll bisect this later today and update the ticket.

> CXX=/usr/bin/clang++ cmake -GNinja -H. -Bbuild -DCMAKE_BUILD_TYPE=Debug
> cmake --build build
cbmc/src/pointer-analysis/value_set.h:227:19: error: 
      invalid operands to binary expression ('const object_mapt' (aka 'const
      reference_counting<map<unsigned long,
      nonstd::optional_lite::optional<BigInt> >, empty_object_map>') and
      'const object_mapt')
        object_map==other.object_map;
        ~~~~~~~~~~^ ~~~~~~~~~~~~~~~~

ws> /usr/bin/clang++ --version
clang version 3.8.0-2ubuntu4 (tags/RELEASE_380/final)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

ws> cat /etc/os-release
NAME="Ubuntu"
VERSION="16.04.6 LTS (Xenial Xerus)"
ID=ubuntu
ID_LIKE=debian
PRETTY_NAME="Ubuntu 16.04.6 LTS"
VERSION_ID="16.04"
HOME_URL="http://www.ubuntu.com/"
SUPPORT_URL="http://help.ubuntu.com/"
BUG_REPORT_URL="http://bugs.launchpad.net/ubuntu/"
VERSION_CODENAME=xenial
UBUNTU_CODENAME=xenial

ws> git show
commit e7944411a13177dfaccfc8667908f8fbe5c94a78
Merge: 0888187 e739fbb
Author: Michael Tautschnig <[email protected]>
Date:   Thu May 23 12:10:40 2019 +0100

    Merge pull request #4697 from tautschnig/fix-value-set-filtering
    
    Value-set filtering: gracefully handle pointer arithmetic

> git blame -L227,227 src/pointer-analysis/value_set.h
d4f1b294 (Chris Smowton 2017-10-11 11:35:04 +0100 227)         object_map==other.object_map;

[tautschnig]

@karkhaz My suspicion is that this is caused by #4694.

[smowton]

This (https://github.com/diffblue/cbmc/blob/develop/src/util/reference_counting.h#L184) should be providing the relevant operator==, any hints why Clang 3.8 uniquely thinks it doesn't?

[karkhaz]

@tautschnig: git-bisect was fast, but tautschnig-bisect was faster.

@smowton I've been scratching my head over this all afternoon. I finally tried building with clang 9, and it does not complain. I can only conclude that clang 3.8 is miscompiling. (I know saying that "the compiler is wrong" is a cop-out, and this is the second time I've found a clang bug while building CBMC)...

Not sure what course of action to take. Close as WONTFIX? But Ubuntu 16.04 is still very common in the wild, it's a long-term support distro.

[karkhaz]

I suppose we could at least modify the CMake and Makefile builds to require a non-ancient Clang, and maybe make the build tools print the URL for the official LLVM Ubuntu packages if the build is being run on Ubuntu...

[smowton]

I'm happy enough if we can find a workaround -- perhaps an explicit template instantiation to nudge clang in the right direction?

[karkhaz]

I spent some time on trying to add a template instantiation this afternoon without success. I agree that this solution would be good, but I'm unable to spend more time on this, sorry. I'm going to leave the issue open, up to Diffblue to decide how to close it; I'm just going to install a recent Clang on my machine or use GCC.

[allredj]

We stumbled upon this problem on TG as well. Do we want to keep support for clang 3.8? If so we should add a build on Travis.

[allredj]

^ @peterschrammel ?