libdrake: Use alwayslink instead of legacy_whole_archive

[jwnimmer-tri]

As a side-effect, duplicate definitions (e.g., of main) are now an error.

Closes #12137.

Relates bazelbuild/bazel#7362.

---

This change is 

[jwnimmer-tri]

@drake-jenkins-bot mac-mojave-clang-bazel-experimental-release please
@drake-jenkins-bot mac-catalina-clang-bazel-experimental-debug please
@drake-jenkins-bot mac-catalina-clang-bazel-experimental-everything-release please

[jwnimmer-tri]

+@jamiesnape for feature review, please.
+@sammy-tri for platform review per schedule, please.
\CC @EricCousineau-TRI FYI

[jamiesnape]

Reviewed 5 of 5 files at r1.
Reviewable status: LGTM missing from assignee sammy-tri(platform), labeled "do not merge" (waiting on @sammy-tri)

[sammy-tri]

Reviewed 5 of 5 files at r1.
Reviewable status: labeled "do not merge"

[jwnimmer-tri]

The macOS debug error is "failed due to unexpected I/O exception: writing file failed". I wonder if that's a flake, or if this change substantially affects the build footprint.

[jamiesnape]

Not one of our usual flakes. Flakes during building rather than testing are rare. I would try a Mojave debug build.

[jamiesnape]

Disk usage on the Linux debug builds looks much higher (potentially up to 50 GB).

[jamiesnape]

Looks like the disk was full on the Mac, so not a flake.

[jwnimmer-tri]

Yup. Sounds like I can profile and debug on Ubunutu for now, though.

[jwnimmer-tri]

Testing with bazel build -c dbg //examples/pendulum/... from scratch. The problem seems to be simply that statically-linked programs now have all of the extra unnecessary symbols polluting them, which used to be culled when not referenced:

jwnimmer@call-cps:~/tools/bloaty$ ./bloaty \
  -d compileunits \
  --domain=file \
  ~/jwnimmer-tri/drake.12262/bazel-bin/examples/pendulum/print_symbolic_dynamics \
  -- \
  ~/jwnimmer-tri/drake.master/bazel-bin/examples/pendulum/print_symbolic_dynamics

    FILE SIZE   
 -------------- 
  [NEW] +32.8Mi    math/discrete_algebraic_riccati_equation.cc
  [NEW] +20.3Mi    math/continuous_lyapunov_equation.cc
  +3.5% +14.1Mi    [176 Others]
  [NEW] +11.5Mi    math/continuous_algebraic_riccati_equation.cc
  [NEW] +7.73Mi    math/rigid_transform.cc
  [NEW] +7.25Mi    math/discrete_lyapunov_equation.cc
  [NEW] +7.17Mi    multibody/tree/space_xyz_mobilizer.cc
  [NEW] +6.89Mi    geometry/proximity/distance_to_point_with_gradient.cc
  [NEW] +6.29Mi    math/rotation_matrix.cc
  [NEW] +5.50Mi    geometry/query_results/contact_surface.cc
  [NEW] +5.01Mi    math/roll_pitch_yaw.cc
  [NEW] +4.07Mi    multibody/tree/linear_spring_damper.cc
  [NEW] +2.35Mi    math/quadratic_form.cc
   +22% +1.71Mi    [section .debug_aranges]
  [NEW] +1.62Mi    multibody/tree/revolute_spring.cc
  [NEW] +1.59Mi    systems/framework/single_output_vector_source.cc
  [NEW] +1.57Mi    systems/framework/vector_system.cc
  [NEW] +1.51Mi    multibody/tree/body_node_impl.cc
  -2.4%  -914Ki    multibody/plant/tamsi_solver.cc
 -34.7% -1.09Mi    multibody/tree/weld_joint.cc
 -70.8% -7.98Mi    common/polynomial.cc
   +23%  +129Mi    TOTAL

In my initial draft of the PR, I tried to apply alwayslink = True to the intermediate cc_library targets on the way to libdrake.so, but it seemed to only take effect when places on the targets with srcs = [] directly, so I moved it up to drake_cc_library. This (extra linking) is not a good outcome, though.

I'll have to investigate more.

[jwnimmer-tri]

This is pretty close, but is still failing some installed-image acceptance tests due to losing some DT_NEEDED on the shared library.

[jwnimmer-tri]

@drake-jenkins-bot mac-mojave-clang-bazel-experimental-release please
@drake-jenkins-bot mac-catalina-clang-bazel-experimental-debug please
@drake-jenkins-bot mac-catalina-clang-bazel-experimental-everything-release please

[jwnimmer-tri]

Ugh. The macOS debug build filled up its disk again:
https://drake-jenkins.csail.mit.edu/job/mac-catalina-clang-bazel-experimental-debug/2/

[jwnimmer-tri]

It looks like bazelbuild/rules_cc@c2b692b might be on the way to fixing this upstream.

[jwnimmer-tri]

I finally found a solution that I am not unhappy with. It should be exactly the same as before for libdrake build time and disk footprint, and retain the improvement for link size and times for standalone binary programs.

[jwnimmer-tri]

@drake-jenkins-bot mac-mojave-clang-bazel-experimental-release please
@drake-jenkins-bot mac-catalina-clang-bazel-experimental-debug please
@drake-jenkins-bot mac-catalina-clang-bazel-experimental-everything-release please

[jwnimmer-tri]

Guess what +@sammy-tri and +@jamiesnape? This is ready for review again!

[jamiesnape]

Uses less than 100 MB extra space on Mac CI in debug as far as I can tell.

Reviewed 4 of 7 files at r2.
Reviewable status: labeled "do not merge"

[jwnimmer-tri]

-@sammy-tri +@ggould-tri for today's on-call platform review.

[ggould-tri]

Well this is pretty painful. Seems correct, though.

Reviewed 4 of 7 files at r2.
Reviewable status: labeled "do not merge"

[jwnimmer-tri]

I did a retest on Anzu; everything worked fine.

(I also fixed some nits I noticed while reading over this again with fresh eyes.)

[ggould-tri]

Reviewed 2 of 2 files at r3.
Reviewable status: complete! all discussions resolved, LGTM from assignees ggould-tri(platform),jamiesnape