Drop per-run infrastructure

[odersky]

The per-run infrastructure caused a 10-20% performance hit. We could go back to the old
caching scheme but this is verbose and ugly. This commit drops the whole caching infra
structure instead. The downside is that we need a new compiler if any symbol cached
by Definitions has been edited and should be reloaded.

[odersky]

test performance please

[dottybot]

performance test scheduled: 1 job(s) in queue, 1 running.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/6965/ to see the changes.

Benchmarks is based on merging with master (52fe263)

[odersky]

Comparing with #6962, we seem to get about 2-3% speed improvement over the previous scheme that uses two entries per definition. So it's not only simpler but faster as well. My tendency would be to use this PR instead of #6962.

[smarter]

It might be possible to get the seqtype-cycle pickling test to pass by running it using -sourcepath

[odersky]

It might be possible to get the seqtype-cycle pickling test to pass by running it using -sourcepath

OK, can you or Nicolas try that out?

[smarter]

I had a quick look but it's likely that this would break in other ways, I would just drop the test since the way it works is a bit too fragile.

[nicolasstucki]

The downside is that we need a new compiler if any symbol cached by Definitions has been edited and should be reloaded.

In practice, when would we need a new compiler?

[odersky]

In practice, when would we need a new compiler?

It would be if we run with a resident compiler and edit a file touched by Definitions, such as Seq.scala. That would give us a StaleSymbol error. The most common scenario is if the compiler runs in an IDE. However, we already handle stale symbols in the IDE, so I am not sure where it would matter in practice.

I believe ~compile is not a true resident compiler, or is it? I mean, do we keep the same instance of Compiler across several runs with ~compile?

[smarter]

I believe ~compile is not a true resident compiler, or is it? I mean, do we keep the same instance of Compiler across several runs with ~compile?

No, it always creates a new compiler instance.

[odersky]

Then it's just the IDE and that one handles the error already. I.e. ``` if (ctx.acceptStale(symd)) return symd.currentSymbol.denot.orElse(symd).updateValidity() ``` in Denotations#bringForward. So, should we merge the PR?

…

On Fri, Aug 2, 2019 at 4:57 PM Guillaume Martres ***@***.***> wrote: I believe ~compile is not a true resident compiler, or is it? I mean, do we keep the same instance of Compiler across several runs with ~compile? No, it always creates a new compiler instance. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6965>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAGCKVRAGWNBIRF4WK77ISDQCRDNJANCNFSM4IH2UWIA> .

-- Prof. Martin Odersky LAMP/IC, EPFL