Try some more optimizations #9409

odersky · 2020-07-22T11:29:35Z

Based on #9405 and includes #9414.

We further reduce context fields, so that I count now 14 fields overall: 06756eb
We reduce contexts created with new periods by about 3/4ths: 5aa5c1f
We eliminate several allocation hotspots (remaining commits)

Together with #9405 and #9343, this reduces total memory allocated to contexts by 2/3rds and total memory allocated to TypeComparers to practically nothing.

odersky · 2020-07-22T18:39:15Z

test performance please

dottybot · 2020-07-22T18:40:57Z

performance test scheduled: 1 job(s) in queue, 0 running.

dottybot · 2020-07-22T20:18:10Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (0f1a23e)

odersky · 2020-07-23T09:47:51Z

test performance please

dottybot · 2020-07-23T09:48:23Z

performance test scheduled: 1 job(s) in queue, 0 running.

dottybot · 2020-07-23T11:25:05Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (0f1a23e)

liufengyun · 2020-07-23T11:28:46Z

Note that stdlib test failed, there's no point for that curve.

odersky · 2020-07-23T15:45:15Z

Why did stdlib fail? Was that a trasient failure?

odersky · 2020-07-23T17:41:01Z

test performance please

dottybot · 2020-07-23T17:42:12Z

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

liufengyun · 2020-07-23T19:58:36Z

Why did stdlib fail? Was that a trasient failure?

It is not transient, it has been failing since #8652 . I'll propose a fix ASAP.

dottybot · 2020-07-23T21:02:20Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (eb3b0ac)

retronym · 2020-07-24T01:50:48Z

FreshContext is down to 88 bytes... nice improvement!

dotty.tools.dotc.core.Contexts$FreshContext object internals: OFFSET SIZE TYPE DESCRIPTION VALUE 0 4 (object header) 05 00 00 00 (00000101 00000000 00000000 00000000) (5) 4 4 (object header) 00 00 00 00 (00000000 00000000 00000000 00000000) (0) 8 4 (object header) 40 70 12 00 (01000000 01110000 00010010 00000000) (1208384) 12 4 int Context._period 0 16 8 long Context.bitmap$0 0 24 4 int Context._mode 0 28 4 dotty.tools.dotc.core.Contexts.ContextBase Context.base null 32 4 dotty.tools.dotc.core.Contexts.Context Context.given_Context$lzy1 null 36 4 dotty.tools.dotc.core.Contexts.Context Context._outer null 40 4 dotty.tools.dotc.core.Symbols.Symbol Context._owner null 44 4 dotty.tools.dotc.ast.Trees.Tree Context._tree null 48 4 dotty.tools.dotc.core.Scopes.Scope Context._scope null 52 4 dotty.tools.dotc.core.TyperState Context._typerState null 56 4 dotty.tools.dotc.typer.TypeAssigner Context._typeAssigner null 60 4 dotty.tools.dotc.core.GadtConstraint Context._gadt null 64 4 dotty.tools.dotc.typer.SearchHistory Context._searchHistory null 68 4 dotty.tools.dotc.util.SourceFile Context._source null 72 4 scala.collection.immutable.Map Context._moreProperties null 76 4 java.lang.Object[] Context._store null 80 4 dotty.tools.dotc.typer.Implicits.ContextualImplicits Context.implicitsCache null 84 4 dotty.tools.dotc.util.SimpleIdentityMap Context.related null Instance size: 88 bytes Space losses: 0 bytes internal + 0 bytes external = 0 bytes total

odersky · 2020-07-24T16:00:08Z

@retronym I notice there's an unexpected lazy val bitmap and given_Context$lzy1 in there, which can both be avoided. If we could shave off one more word, we'd be down to 72.

odersky · 2020-07-24T18:01:01Z

test performance please

dottybot · 2020-07-24T18:01:41Z

performance test scheduled: 1 job(s) in queue, 0 running.

dottybot · 2020-07-24T19:36:20Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (fd18546)

odersky · 2020-07-24T20:02:16Z

test performance please

dottybot · 2020-07-24T20:04:09Z

performance test scheduled: 1 job(s) in queue, 0 running.

dottybot · 2020-07-24T21:41:33Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (fd18546)

odersky · 2020-07-25T08:23:40Z

test performance please

dottybot · 2020-07-25T08:24:56Z

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

dottybot · 2020-07-25T08:55:14Z

performance test failed:

Please check http://lamppc37.epfl.ch:8000/pull-9409-07-25-10.55.out for more information

odersky · 2020-07-25T09:40:01Z

test performance please

dottybot · 2020-07-25T09:41:15Z

performance test scheduled: 1 job(s) in queue, 0 running.

dottybot · 2020-09-07T09:36:29Z

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

dottybot · 2020-09-07T09:45:22Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (8c94870)

dottybot · 2020-09-07T11:01:53Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (8c94870)

Looking at these benchmarks, our own HashSet is clearly the fastest so we should use it wherever possible. First tests without GC: ``` java.util.HashMap took 1977.537756 ms java.util.IdentityHashMap took 2123.028843 ms scala.collection.HashMap took 2116.062172 ms scala.collection.HashSet took 1998.378418 ms dotty.tools.dotc.HashSet took 1429.826866 ms ``` Second tests with System.gc() run before every test: ``` java.util.HashMap took 1980.977419 ms java.util.IdentityHashMap took 2082.178216 ms scala.collection.HashMap took 2101.882151 ms scala.collection.HashSet took 1967.379402 ms dotty.tools.dotc.HashSet took 1450.755273 ms ```

Allow to record total sizes of collections on which some selected operation is performed. # Conflicts: # compiler/src/dotty/tools/dotc/transform/Instrumentation.scala

Add nestedExists extension method for List[List[T]] exists and optimize it as well as nestedMap.

Avoid closures for copied and simple. Avoid substitutions for simple.

Make all orElse methods that take a call-by-name argument inline methods.

Run stats only at last run. This allows one to warm up and JOT compile code before measurements start.

If Stats is enabled, track total time spent in - implicit search overall - searching implicits in context / in type scope - typedImplicit

See scala#9748 for why this is necessary

When tranisitioning from dense to hashing, we grow the table more than usual since otherwise we'd have to grow it again at the very next addEntry. The condition for this was wrong for HashMaps.

Make criterion when to fill a hole in a HashMap or HashSet remove more robust. The old criterion relied on fill factor always being less than 0.5. The new criterion works for arbitrary fill factors.

odersky · 2020-09-08T14:38:56Z

test performance please

dottybot · 2020-09-08T14:40:50Z

performance test scheduled: 4 job(s) in queue, 1 running.

dottybot · 2020-09-08T16:32:20Z

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9409/ to see the changes.

Benchmarks is based on merging with master (ce48f5a)

odersky force-pushed the optimize-more branch from d6646de to 06756eb Compare July 22, 2020 12:41

odersky changed the title ~~Streamline treatment of withPhase and withSource~~ Try some more optimizations Jul 23, 2020

odersky force-pushed the optimize-more branch from 491eee3 to a35b7a4 Compare July 23, 2020 17:13

nicolasstucki assigned odersky Jul 24, 2020

odersky force-pushed the optimize-more branch from a949072 to d737f6a Compare July 25, 2020 09:39

odersky force-pushed the optimize-more branch from 99898df to 5a37540 Compare September 8, 2020 11:45

odersky added 20 commits September 8, 2020 15:49

Instrument collection sizes

4d17739

Allow to record total sizes of collections on which some selected operation is performed. # Conflicts: # compiler/src/dotty/tools/dotc/transform/Instrumentation.scala

Avoid some expensive collection operations in backend

5cc5375

Avoid redundant filter in memberNames

e90da06

Add nestedExists extension method

10bf6fb

Add nestedExists extension method for List[List[T]] exists and optimize it as well as nestedMap.

Avoid more collection ops in backend

f4e6844

Avoid more collection ops in frontend

7ee2691

Split method types into general, copied, and simple

a84e755

Avoid closures for copied and simple. Avoid substitutions for simple.

Reduce SourcePosition allocations

b9e948f

Generate DottyPrimitives only once

ae9018a

Make all orElse methods inline methods

25ecbb1

Make all orElse methods that take a call-by-name argument inline methods.

Move more mutable.HashMaps to util.HashMaps (2)

a4c2768

Reduce context creations (1)

bb8b340

Avoid creating a fresh TyperState in isFullyDefined

4486a1a

Fix test

07aefa9

Change Bench to give better timings under -YdetailedStats

9f42ffa

Run stats only at last run. This allows one to warm up and JOT compile code before measurements start.

Track some implicit timings

31a4874

If Stats is enabled, track total time spent in - implicit search overall - searching implicits in context / in type scope - typedImplicit

Track time spent in resolveOverloaded

9982d19

Fix hashcodes of inner case classes in NameKinds

5db64c3

See scala#9748 for why this is necessary

Fix growTable function for HashMap

682ed70

When tranisitioning from dense to hashing, we grow the table more than usual since otherwise we'd have to grow it again at the very next addEntry. The condition for this was wrong for HashMaps.

odersky force-pushed the optimize-more branch from 5a37540 to 682ed70 Compare September 8, 2020 13:56

Fix HashMap and HashSet remove criterion

dd00621

Make criterion when to fill a hole in a HashMap or HashSet remove more robust. The old criterion relied on fill factor always being less than 0.5. The new criterion works for arbitrary fill factors.

odersky closed this Nov 16, 2020

Try some more optimizations #9409

Try some more optimizations #9409

Uh oh!

Conversation

odersky commented Jul 22, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

odersky commented Jul 22, 2020

dottybot commented Jul 22, 2020

dottybot commented Jul 22, 2020

odersky commented Jul 23, 2020

dottybot commented Jul 23, 2020

dottybot commented Jul 23, 2020

liufengyun commented Jul 23, 2020

odersky commented Jul 23, 2020

odersky commented Jul 23, 2020

dottybot commented Jul 23, 2020

liufengyun commented Jul 23, 2020

dottybot commented Jul 23, 2020

retronym commented Jul 24, 2020

odersky commented Jul 24, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

odersky commented Jul 24, 2020

dottybot commented Jul 24, 2020

dottybot commented Jul 24, 2020

odersky commented Jul 24, 2020

dottybot commented Jul 24, 2020

dottybot commented Jul 24, 2020

odersky commented Jul 25, 2020

dottybot commented Jul 25, 2020

dottybot commented Jul 25, 2020

odersky commented Jul 25, 2020

dottybot commented Jul 25, 2020

dottybot commented Sep 7, 2020

dottybot commented Sep 7, 2020

dottybot commented Sep 7, 2020

odersky commented Sep 8, 2020

dottybot commented Sep 8, 2020

dottybot commented Sep 8, 2020

Labels

6 participants

odersky commented Jul 22, 2020 •

edited

Loading

odersky commented Jul 24, 2020 •

edited

Loading