I'm running into this with gem5 now (the problem with canceled or aborted runs) and I think I've tracked down a possible explanation... basically running the test and comparing the outputs are set up as two separate tasks in scons, and the only thing the output comparison task is dependent on is the m5 stats output. So if you don't clear out the old stats file before the run starts, and you abort m5 before it starts writing a new one, the "compare outputs" task doesn't think it needs to rerun.
I'm wondering if there was some change Nate made to stats that caused it to stop clobbering the stats output file when the program aborts. Anyway, the right answer should be to force the task comparison to rerun when any of the m5 outputs changes. I'll give that a try soon and push it if it works. Anecdotally I've had problems with things getting rebuilt unnecessarily too, though usually it doesn't cause the regressions to rerun. It seems to be something to do with python and/or swig. Steve On Tue, Feb 24, 2009 at 1:39 AM, Gabe Black <[email protected]> wrote: > I mentioned this earlier, but scons and regressions are misbehaving, > and it's making updating the regressions very annoying. If a run is > canceled halfway, scons now assumes it actually finished and was just > wrong. I have to go and manually delete the old, incomplete run before > it's willing to try again. Also, it seems really anxious to rebuild and > rerun tests when it -doesn't- need to. This makes it really annoying to, > for instance, see which tests fail, look at the differences, update the > results, and verify that they took. If I don't make any mistakes like > loosing track of which tests failed, that takes three runs through all > of the regressions I'm interested in which takes three times as long as > it needs to. Unfortunately I haven't yet managed this minimum, so I've > resorted to just updating the stats the first time around and looking at > the patch post mortem which is not ideal. If somebody could look at this > I'd really appreciate it. If it seems like a local problem, like from > when I recently upgraded scons for example, that would be useful > information. > > Gabe > _______________________________________________ > m5-dev mailing list > [email protected] > http://m5sim.org/mailman/listinfo/m5-dev > _______________________________________________ m5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/m5-dev
