Re-remove the --no-timestamp option

[ldionne]

See #170

[ldionne]

@MatzeB @kbeyls In #170 I tried removing the --no-timestamp option, which should have been made obsolete by the existence of --build-dir. However, when I did that, we discovered that it broke the LNT jobs defined in llvm-zorg. Indeed, those are currently using --no-timestamp, e.g. here and here.

That makes me wonder -- do you folks remember what was the purpose of timestamping the build directory in the first place? Instead of removing --no-timestamp and always timestamping the build directory, would it make sense to instead remove --timestamp and never timestamp the build directory, thereby always producing a stable build directory?

I'm trying to see if there would be some subtle reason for it, such as allowing different CI processes to run on the same machine in the same directory with reduced risks of collision (if they timestamp their build dirs, the likelihood of build dir collisions is really small). But since --no-timestamp is used in the llvm-zorg definition of the LNT jobs, I wonder if there are in fact any jobs running with timestamped builds -- so maybe that isn't a concern.

Anyway, just trying to get more context around this.

[ldionne]

As a side note, timestamping the build directory in those llvm-zorg jobs has the (probably surmountable) challenge that the jobs currently rely on a stable build directory they can get log files from: https://github.com/llvm/llvm-zorg/blob/5b7570a883c6f9caf1531b7c721be1ff77f4448a/zorg/buildbot/builders/ClangBuilder.py#L638

[kbeyls]

It looks like timestamping directories has been present since the very first public version of LNT, see 62ac2aa.
So maybe only @ddunbar knows why it was introduced.
If I had to make a guess of why (or at least how I find adding timestamps can sometimes be valuable in directory names): When invoking LNT manually for evaluating/testing/benchmarking work in progress, it is often somewhat helpful that a timestamp gets added to your build dir, so that it captures some info about "I ran this experiment on this date/time". Also when running lnt test runs many times in quick succession, it can be useful to not have to manually change the build dir name before being able to start of the next test run.
Obviously, the usefulness of the timestamp is somewhat limited as it does not record most info needed to understand what was used to produce the test/benchmark results in the build dir.
I'd be OK with removing timestamps from build dirs; but it probably is an interface ("ABI") breaking change for some people's workflows, so might need some form of deprecation management. (I'm thinking just announcing you'll do this on Discourse and give people some time to object to the removal?).
Maybe the best way forward could be to disable this option by default, but keep the feature as an opt-in?