pydrake doc: gcc and clang disagree on need for lambda capture for constexpr vars

[EricCousineau-TRI]

From here: #9571 (comment)

Workarounds:

• Use constexpr auto& doc, rather than auto& doc
• Use [&] in lambda capture
• Add copts = ["-Wno-unused-lambda-capture"] for pybind_py_library

Action items:

• Use a workaround in relevant sites (searching for doc_lambda)
• Find or file upstream bug report:
  • clang: https://bugs.llvm.org/show_bug.cgi?id=39202
  • gcc: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87559

\cc @m-chaturvedi @jamiesnape

[EricCousineau-TRI]

Possibly related; however, I am getting this with clang-6.0, and it'd seem that this should have been incorporated already:
https://bugs.llvm.org/show_bug.cgi?id=31815

EDIT: Nevermind; it seems exactly related. We have a generic lambda here. Will check versioning more closely.

EDIT 2: That was circumstantial; had to deal with storage duration inference via upstream auto& (determined before lambda definition).

[m-chaturvedi]

Maybe related: https://reviews.llvm.org/D33526

[EricCousineau-TRI]

Seems that this patch was incorporated before clang-6.0:

$ clang-6.0 --version
clang version 6.0.0-1ubuntu2~16.04.1 (tags/RELEASE_600/final)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

The branch in the Git repo release_60:
https://github.com/llvm-mirror/clang/tree/release_60
Does incorporate the corresponding git commit for the patch:
llvm-mirror/clang@6c70606

So perhaps it's still an issue?

[EricCousineau-TRI]

Huh, seems like you can also change auto& doc = ... to constexpr auto& doc = ..., and then neither clang nor gcc seem to care, due to how constexpr affects storage duration?

[jamiesnape]

Very Java-like (except that would be final). Makes a certain amount of sense.

[EricCousineau-TRI]

Reproduction code + output:
https://github.com/EricCousineau-TRI/repro/blob/c80dc2360068a256665275c52518ed2876f5d112/bug/clang_unused_lambda_capture/output.txt

[EricCousineau-TRI]

Filed bug with clang, will see if they know where it should go:
https://bugs.llvm.org/show_bug.cgi?id=39202

[EricCousineau-TRI]

From chat with @m-chaturvedi, seems that GCC 7.3.0 on Bionic decided to further deviate from clang; with the merge of #9606, we now get a Bionic failure:
https://drake-jenkins.csail.mit.edu/job/linux-bionic-gcc-bazel-experimental/11/console

bindings/pydrake/systems/framework_py_semantics.cc:72:37: note: the lambda has no capture-default
   auto bind_common_scalar_types = [m](auto dummy) {
                                     ^
bindings/pydrake/systems/framework_py_semantics.cc:31:19: note: 'constexpr const<unnamed struct>::<unnamed struct>::<unnamed struct>& doc' declared here
   constexpr auto& doc = pydrake_doc.drake.systems;
                   ^~~
bindings/pydrake/systems/framework_py_semantics.cc:310:12: error: 'doc' is not captured
            doc.Parameters.SetFrom.doc);

Will submit a chaser PR to use the second workaround :(

[EricCousineau-TRI]

I may have spoken too soon? I am unable to reproduce this with simple reproduction:
EricCousineau-TRI/repro@4109dc4?diff=split#diff-bc917a88a9cbeb890e61e5fb340a8ec5

Even trying to re-simulate the nested structs does not yield an issue:
EricCousineau-TRI/repro@bbc8f78?diff=split#diff-bc917a88a9cbeb890e61e5fb340a8ec5R18

Not sure what the discrepancy is... @m-chaturvedi @jamiesnape Might I ask if y'all have an inkling?

[EricCousineau-TRI]

Mmanu found that this error can be reproduced if we use a generic / templated lambda, such as using auto as a parameter:
EricCousineau-TRI/repro@fd025bc?diff=split#diff-bc917a88a9cbeb890e61e5fb340a8ec5

Seems related to old aforementioned bug with LLVM, but now specific to GCC: https://bugs.llvm.org/show_bug.cgi?id=31815

[EricCousineau-TRI]

Posted a GCC bug report: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87559