Build target for 3C tests that reuses previous results if inputs haven't changed

[mattmccutchen-cci]

There should be a simple, documented script or build target to run all 3C tests that we believe should be run on every change. Obviously, a necessary condition for a test to be in this set is that it be fully automated (which the benchmark program tests aren't yet); beyond that, it's a policy decision of what tests are worth the running time, though we could conceivably have different modes.

It looks like clang/test/CheckedCRewriter/llvm-lit-wrapper.py is trying to be what we want, but I'm getting an error even after I manually add build/bin to $PATH and llvm/utils/lit to $PYTHONPATH:

llvm-lit: /home/user/3c/llvm/utils/lit/lit/TestingConfig.py:101: fatal: unable to parse config file '/home/user/3c/clang/test/lit.cfg.py', traceback: Traceback (most recent call last):
  File "/home/user/3c/llvm/utils/lit/lit/TestingConfig.py", line 90, in load_from_path
    exec(compile(data, path, 'exec'), cfg_globals, None)
  File "/home/user/3c/clang/test/lit.cfg.py", line 25, in <module>
    config.test_format = lit.formats.ShTest(not llvm_config.use_lit_shell)
AttributeError: 'NoneType' object has no attribute 'use_lit_shell'

It would be even better to have a build target that won't rerun the tests if they previously passed and none of the dependencies have changed since then.

I want to have this script/target as a necessary (but not sufficient) element of testing the renames in #286.

[mattmccutchen-cci]

Looks like ninja check-clang-checkedcrewriter in the build directory might be what I'm looking for. Waiting for the build to run so I can see if it works...

[mattmccutchen-cci]

Here's what I've learned so far:

• ninja check-clang-checkedcrewriter does work. In my test from a clean tree, it ran the regression tests properly, so it must have built everything needed. It exits 0 or 1 depending on whether all tests passed, so it should work for automated tools such as the upcoming 5C/3C sync tool to check the status of the tests.
• The build consumed ~36 GB of disk space, compared to ~25 GB for my previous build. Some of the difference might be due to changes in build flags I made in an attempt to make the build go faster, but I suspect most of it is because the code in clang/test/CMakeLists.txt naively makes each test suite depend on all the tools used by any of the test suites. Hopefully this is something that won't bother most of us too much, like the long full build time in general, but it does mean I may not be able to afford the ~108 GB on my personal laptop for the minimum of three build trees needed under the planned design of the 5C/3C sync tool. Fortunately, it looks like I may have enough other things to keep me busy until I get access to another computer to use for CCI work.
• I imagine the target doesn't support all the options we want to be able to use to troubleshoot the tests, so we should still fix the separate instructions to troubleshoot the tests in the 3C document. That's outside the scope of this issue; perhaps I can address it as part of the readme in repository reorg #286.
• The target re-runs the test suite even if no inputs have changed since the last run. If the target were smart enough to save the results to a file, re-display them, and exit with the proper exit code, that would improve the performance of tools such as the 5C/3C sync tool that want to check the status of the tests without those tools needing their own logic to detect whether any inputs have changed.

I've documented the existing build target in the 3C document, so the main goal of this issue is adequately achieved for now. I'll reuse the issue for the "reuse previous results" feature.