CHPL_UNWIND=libunwind nightly testing #5017
Assignees
Labels
Comments
|
We should probably decide, as a matter of policy, if stack traces are expected in .good files, or if the prediff should handle the case that the .good file has no stack trace. If the stack traces are included in .good files, many more internal/standard module line numbers can be exposed, so we presumably need a way to handle that intelligently. It would not be an improvement to have to update every test that halts when certain internal module code is moved to different lines. |
bradcray
added a commit
that referenced
this issue
Jan 19, 2017
Disable CHPL_UNWIND during start_test runs [reviewed by @mppf] This commit disables CHPL_UNWIND during start_test runs, since the output can be dramatically different (e.g., if an expected error message is called, which calls halt, a stack trace is printed). It's an open question as to how we'd do testing in a "CHPL_UNWIND on by default" world, but in the meantime, I'm trying to work with CHPL_UNWIND on by default to see how well it works. In the meantime, this mod seems like the nice thing to do for developers who use it. In the conversation on the PR, Michael points out that there is a prediff script that can be manually invoked to filter out CHPL_UNWIND output and that there are tests that will be skipped if CHPL_UNWIND is unset (which now it always will be). But, we're currently never running with CHPL_UNWIND set in our nightly testing, so these tests aren't being run anyway and Michael reports that their .good files are stale anyway. Meanwhile, Michael has opened issue #5017 to further discuss how CHPL_UNWIND testing should ultimately work.
4 tasks
mppf
added a commit
that referenced
this issue
Aug 20, 2020
Add ability to print stack trace when printing verbose comm This PR takes several steps to improve stack trace support and then improves verbose comm diagnostics support with a config param that can enable stack traces to be printed with comm events. To improve the stack trace support, it: * fixes problems with the linux-only call to `addr2line` to get specific line numbers. These were not working on my system because the address passed to `addr2line` included a random offset (from ASLR probably). So, I adjusted the code to use `dladdr` to compute the base address of the object containing the address. (And adding at program link-time `-ldl` in addition to the libunwind library). This is subtracted from the base address to get a relative address that works with `addr2line`. I ran into this because this code was core-dumping in gasnet configurations for me. * improves bounds checking on some operations & give ourselves a bit more buffer size working room * adds `chpl_stack_unwind_to_string` that saves a stack trace to an allocated string Within the comm diagnostics code, it is mainly a matter of getting the setting for `chpl_verbose_comm_stacktrace` from the config param and the calling `chpl_stack_unwind_to_string`. To use the stack tracing in verbose comms, make with `CHPL_UNWIND=libunwind` and then compile a program printing verbose comms with `-scommDiagsStacktrace=true`. Note that we don't currently have libunwind testing in nightly testing - see issue #5017. Reviewed by @ronawho and @gbtitus - thanks! - [x] test/runtime/stacktrace passes with gasnet and libunwind on linux - [x] test/runtime/stacktrace passes on mac os x with system libunwind - [x] full local testing - [x] full local gasnet testing (earlier version)
mppf
changed the title
aconsroe-hpe
added a commit
that referenced
this issue
Oct 18, 2021
Bump libunwind and add test configuration with CHPL_UNWIND=bundled reviewed by @ronawho Bumping libunwind version to address a build failure reported in #17597 and add a test configuration utilizing CHPL_UNWIND=bundled. Currently we have no nightly testing that uses CHPL_UNWIND=bundled, but we did have existing tests that were being skipped. This PR adds a cron configuration test-unwind.gasnet.bash with comm and unwind to support testing stacktraces in halts and stacktraces in verbose comm. In #5017 it was raised that we might use a system prediff script to remove any stacktraces but initially this configuration will only be testing that the compiler successfully builds and testing the tests in test/runtime/stacktrace. Relates to: * #5017 * #17597
|
We now have a basic |
|
Thanks, I'll close it. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
CHPL_UNWIND=libunwind should be tested nightly as a prerequisite for turning it on-by-default in some configurations.
Note that
util/test/prediff-for-stacktraceis supposed to help with the various forms of mismatch between a .good file and a .chpl file, in case one has a stacktrace and the other does not.Also note that stack traces might include call sites on linux64, but generally won't include call sites on Mac OS X. And also note that stack traces do currently include line numbers from internal modules.
The text was updated successfully, but these errors were encountered: