| Title | awlut test failure |
| Status | closed |
| Priority | essential |
| Assigned user | Gareth Rees |
| Organization | Ravenbrook |
| Description | On OS X: $ code/xc/Release/awlut 938401525 code/xc/Release/awlut: randomize(): resetting initial state (v3) to: 938401525. Strongly unreachable weak table entry found, slot 48. |
| Analysis | This test case fails in the HOT variety on OS X for about 20% of random states, but I didn't find any failures in the COOL variety after running $ while code/xc/Debug/awlut; do sleep 1; done for 250 iterations, nor did I find any failures on Ubuntu in either variety. I note that when the test case fails, it's always slot 48 (the last slot in the table) that has wrongly been kept alive. Moreover, when I edit the test case so that the last slot is never preserved: - if (rnd() % 2 == 0) { + if (rnd() % 2 == 0 && i + 1 != TABLE_SLOTS) { the test case now always fails (in the HOT variety on OS X). This strongly suggests that this slot is being kept alive by a local variable in the function. The test case is designed so that the local variable "string" is overwritten later, but on looking at the disassembly in GDB, it appears that LLVM has optimized away the local variable assignment. So there's no bug in the MPS. It's just a test case that isn't robust against LLVM's aggressive optimization. I made the test case reliable again, but ensuring that the last entry in the table is always preserved: + if (rnd() % 2 == 0 || i + 1 == TABLE_SLOTS) { which makes sure that the local variable is overwritten. |
| How found | automated_test |
| Evidence | None. |
| Observed in | 1.110.0 |
| Created by | Gareth Rees |
| Created on | 2013-03-08 10:09:04 |
| Last modified by | Richard Brooksby |
| Last modified on | 2016-03-05 12:31:32 |
| History | 2013-03-08 GDR Created. |
| Change | Effect | Date | User | Description |
|---|---|---|---|---|
| 189640 | closed | 2016-03-05 12:13:53 | Richard Brooksby | Populating awlut's tables in a thread to ensure there are no references left in registers. Imported from Git Author: Richard Brooksby <rb@ravenbrook.com> 1457179742 +0000 Committer: Richard Brooksby <rb@ravenbrook.com> 1457179742 +0000 sha1: 606e90b60111449bc1a6bc28169142cbcb4187b9 push-state: complete parent-changes: 0e17e88d94b1f7a23b0a0e3af959fb8cf12a7196=[189639] |
| 181088 | closed | 2013-03-08 11:36:38 | Gareth Rees | Ensure that the last entry in the table is preserved, so that we don't get a false positive due to the local variable 'string' keeping this entry alive. |