Improve performance of tuple operations

[brunnerant]

In this pull request, I made some extensive benchmarks of the current tuple operations (cons, concat, tail, map, zip, ...). I analyzed their performance and tried to understand what were the bottlenecks of the current implementation, and improved them.

Not surprisingly, it turns out that the most efficient way to implement them is to avoid conversions between tuple, iterators and arrays as much as possible because this is the main source of overhead.

For Tuple1 to Tuple22, making a big match on the tuple arity is what gives the best performance (but very ugly code unfortunately).
For TupleXXL, performing the operation on the array seems to be the way to go.

See @nicolasstucki, my semester project supervisor, for more information.

[brunnerant]

test performance with #tuple-runtime : 0a33ad3

[brunnerant]

dotty.tools.dotc.CompilationTests.tastyBootstrap doesn't pass, but all the other tests do. Has anyone an idea of what could be causing this (because I'm not familiar with tasty at all) ?

[nicolasstucki]

You may have found a bug in the compiler. I would suggest using git bisect to find the commit where it started failing.

[brunnerant]

Ok, I will do this.

[brunnerant]

I identified the commit where it starts failing, it is e1a01d2. It probably has something to do with polymorphic lambdas so I will try to see if I can fix it.

[brunnerant]

Ok, the problem was coming from dynamicMap. I was pattern matching using existential types, and then applying the polymorphic lambda to each tuple element. Maybe the compiler cannot infer the types correctly in this case, so what I did for now is that I cast everything to Object.

[brunnerant]

Should we try to test the performance using the profile tuple-runtime, comparing against master ?

[nicolasstucki]

test performance with #tuple-runtime

[dottybot]

performance test scheduled: 2 job(s) in queue, 1 running.

[nicolasstucki]

@liufengyun did the benchmark run?

[liufengyun]

There is a bug report lampepfl/bench#47 . However, I couldn't find any logs. I just rescheduled the job.

[liufengyun]

Currently, it does not support profile names with a hyphen.

unknown profile: tuple

I renamed tuple-runtime to tupleRuntime, now the job is running.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/7689/ to see the changes.

Benchmarks is based on merging with master (0e761ea)

[brunnerant]

Ok, thanks for fixing it, I didn't know about the hyphen limitation !
Is there a way to launch the benchmarks on the master branch ?

[nicolasstucki]

Otherwise LGTM

[nicolasstucki]

@liufengyun could you execute these benchmarks on master?

[brunnerant]

@liufengyun could you execute these benchmarks on master?

@liufengyun Is it possible to run them on both branches, to have a comparison in the same graph ? Thanks !

[liufengyun]

Now they are added to the default profile: lampepfl/bench@2acb4ef

[nicolasstucki]

test performance with #tupleRuntime: master 08613cb

[dottybot]

performance test scheduled for master 08613cb: 2 job(s) in queue, 1 running.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/7689/ to see the changes.

Benchmarks is based on merging with master (e036f46)