Store FramePointer in the IR and cache it

[JaroslavTulach]

This inquiry is inspired by this IrToTruffle code. First of all it, takes Occurance from the IR:

          val useInfo = name
            .unsafeGetMetadata(
              AliasAnalysis,
              "No occurrence on variable usage."
            )
            .unsafeAs[AliasInfo.Occurrence]

          val framePointer = scope.getFramePointer(useInfo.id)

and it converts it into FramePointer to do that it needs a scope also coming from the IR (some parent). The conversion of Occurence via LocalScope is non-trivial as observed by

• Analyze and improve the performance of AliasAnalysis compiler pass #9235

However, we don't have to do it at all! All we need is the FramePointer (a pair of integers), we don't need all the scopes & co. As such:

Let's compute FramePointer onces, before generating caches and then just load it.

Tasks

Beta Give feedback

1. Implement an IRPass to compute the FramePointer and store it as metadata
2. Change IrToTruffle to just read FramePointer
3. Ideally make sure AliasInfo & co. isn't read from caches at all
4. Verify some benchmark times have improved

[hubertp]

Could potentially eliminate a lot of deserialization that is currently done for AliasAnalysis, e.g.

[enso-bot]

Pavel Marek reports a new STANDUP for today (2024-08-01):

Progress: - Finished with first batch of tests

• The new FramePointerAnalysis pass is closely tight to the AliasAnalysis pass.
  • It tries to attach FramePointer as metadata to all the relevant occurences gathered from AliasAnalysis. It should be finished by 2024-08-13.

[enso-bot]

Pavel Marek reports a new STANDUP for yesterday (2024-08-01):

Progress: - Fixing tests

• Add --dump-irs cmd line option - Add org.enso.compiler.dumpIr system prop #10740 It should be finished by 2024-08-13.

[enso-bot]

Pavel Marek reports a new STANDUP for today (2024-08-06):

Progress: - Fixing more tests: making sure the metadata is attached to all synthetic variables and arguments.

• Do not make unnecessary copies of the IR. It should be finished by 2024-08-13.

[enso-bot]

Pavel Marek reports a new STANDUP for today (2024-08-07):

Progress: - Merged the warnings speedup PR #10555

• Discussions, meetings
• Ensure FramePointerAnalysis traverses case expressions It should be finished by 2024-08-13.

[enso-bot]

Pavel Marek reports a new STANDUP for today (2024-08-12):

Progress: - Catching up with emails and reviews.

• Small QoL improvements: Add --vm.D cmdline option #10798, Flatten object hierarchy of AliasAnalysis metadata so it is accessible from Java #10780
• Attach framepointers to expressions in annotations and argument default values.
• Discussion about private access from tests. It should be finished by 2024-08-13.

[enso-bot]

Pavel Marek reports a new STANDUP for today (2024-08-13):

Progress: - Fix rest of the test

• Put various temporary assertions that the new frame pointer metadata is correct.
  • So far, only use the old algorithm, and just print a custom error message if thing are wrong.
  • Everything seems fine, tomorrow, I will migrate to the new metadata. It should be finished by 2024-08-13.

[enso-bot]

Pavel Marek reports a new STANDUP for today (2024-08-14):

Progress: - In IrToTruffle, replaced all the usages of AliasAnalysis metadata with FramePointer metadata.

• I still need to think about how to minimalize LocalScope.
  • It contains references to Alias graph, which is huge and causes performance overhead when deserialized.
  • Theoretically, IRToTruffle does not need these references, but they are used later in different places.
  • I will probably leave this for a follow-up PR. It should be finished by 2024-08-16.