Make dotc faster

[odersky]

This issue collects items that should be looked at in the work to improve performance of dotc.

The following items should definitely be worked on:

• Port the improvements to classpath handling from scalac 2.13
• Port the improvements to classfile parsing from scalac 2.13
• Write class files in parallel, like scalac 2.13
• Finish the implementation of function specialization. Profiles show that we get quite a bit of unexpected boxings in the compiler due to the fact that function types are not yet specialized.
• Go through the previous benchmark results and identify the (top 20-30) commits that caused a performance loss.

The following items are more for exploration:

• See whether we can replace some usages of lists with arrays, with the aim of getting better locaility
• See whether we can parallelize more code

[liufengyun]

Some commits that cause regression

• Revert to the Scala 2 encoding for trait initialization and super calls #8652 causes ~30% regression in micro-test issue #1535
  • (the point for Revert to the Scala 2 encoding for trait initialization and super calls #8652 is missing as it's not made from dotty-staging, it's between Split isSplice into isExprSplice and isTypeSlpice #9393 and Move macro tests to macro test directories #9406 in the chart)
• Implement generic number literals #6919 causes ~30% regression in micro-test issue #1687
• Remove quoted.Type synthesization #9448 causes ~8% regression in micro-test implicit scope loop
  • also causes ~8% regression in micro-test implicit cache I - bootstrapped
• Fix #7442: Only compute line numbers if the file exists #8883 causes ~4% regression in stdlib (already fixed in Avoid unnecessary file IO when computing positions #8935 and SourcePosition#line: Avoid unnecessary IO #9351)