tests: Avoid dialog boxes, prevent stdout from being lost

[StephanTLavavej]

• Fix tests/std: Failing asserts emit assertion dialog boxes #781 (dialogs) and fix tests: log files aren't capturing test stdout #789 (stdout).
  • Add tests/std/include/force_include.hpp, derived from libcxx's msvc_stdlib_force_include.h (which I contributed).
    • Add /FI to tests/std/tests/prefix.lst.
    • Update tests/tr1/lit.site.cfg.in so it can access tests/std/include. (This is lazy/expedient. In the future, we might want to have a common test directory, or we might want to upstream this enhanced TestEnvironmentPreparer to libcxx for direct consumption.)
    • Add /FI to tests/tr1/prefix.lst.
  • _CrtSetReportMode and _CrtSetReportFile send assert messages to stderr instead of displaying dialog boxes in debug mode.
  • However, in release mode (/MT and /MD), assert simply calls abort which displays a crash dialog box from Windows Error Reporting. _set_abort_behavior prevents that. (Note 1)
  • Anything printing to stdout (like printf, puts, and cout) might not be captured when we assert/abort because of buffering. setvbuf disables buffering, preventing this. (Note 2)
• Print randomized test failures to stderr.
  • These are extremely important (as otherwise they aren't reproducible), so I think it's best to explicitly print to stderr instead of relying on the setvbuf fix here.
• Change wcout to wcerr in P0218R1_filesystem.
  • Similar reasoning: This might also fail at runtime and we really need its output to investigate.
• Update include_each_header_alone_matrix.lst.
  • Test iso646.h, which is still part of the repo.
  • Don't test CRT headers, which aren't part of the repo.
  • (This was problematic because the force-included header was macroizing errno, which was stomping over our attempt to test errno.h by forming the header name through macro replacement.)
• NO_TEST_ENVIRONMENT_PREPARER escape hatch.
  • This makes Dev11_0000000_include_each_header_alone cleaner.

I've manually tested each of these changes. Temporarily reverting #800 allowed me to verify that P0218R1_filesystem failure is sent to stderr, and that it doesn't emit assertion dialog boxes in debug mode or crash dialog boxes in release mode. Temporarily partially reverting #724 allowed me to verify that the randomized tests are printing to stderr. (As these changes were systematic, I didn't exercise the other randomized tests.) Finally, I took an ordinary test and added each of cout, puts, and printf to the end, followed by a failed assert, and verified that the test logs were capturing stdout properly.

Note 1: This wasn't a problem for libcxx because we always build it with /MTd:

STL/tests/libcxx/usual_matrix.lst

Line 6 in 0cdf5fb

PM_CL="/EHsc /MTd /std:c++latest /permissive- /FImsvc_stdlib_force_include.h /wd4643"

Note 2: This wasn't a problem for libcxx because it generally doesn't print things to stdout. We primarily experience compiler errors and failed assertions.

[CaseyCarter]

We should note that the importer will need to verify that std/include is made available to tr1 when running in the internal test harness.

[StephanTLavavej]

Excellent catch, I forgot that Microsoft-internal src/qa/distrib.yaml would need to be changed. Thanks!

[cbezault]

Thanks so much for dealing with this!