JIT struct work planned for .NET 8

[jakobbotsch]

This issue captures the planned work items for .NET 8 with respect to struct improvements. Smaller workitems that the larger workitems depend on may also be tracked here. This list is expected to change throughout the release cycle according to ongoing planning and discussions, with possible additions and subtractions to the scope.

Planned for .NET 8

Q4 2022

• JIT: Improve last-use copy omission for implicit byrefs #76069

Jan-Mar 2023

• Handle regressions from early liveness
  • JIT: Add a stress mode that poisons implicit byrefs #80691
  • JIT: Account for GT_JMP implicit uses in local morph ref counting #80734
  • JIT: Add Statement::m_treeListEnd #81031
  • JIT: Avoid xchg for resolution #81216
  • JIT: Update GTF_GLOB_REF after last-use copy omission #81430
  • JIT: Avoid omitting copies for struct fields passed implicitly by reference #81787
• Extend scope of early liveness
  • JIT: More cases for last-use copy elision  #80494
  • JIT: Switch tailcall implicit byref optimization to use last use information #81033
  • JIT: Avoid creating stores to dead fields in block morphing #81095
• JIT: Add a (disabled) prototype for a generalized promotion pass #83388

Apr-Jun 2023

See #76928.

Future/Interesting

• Struct SSA
• Remove existing promotion (or perhaps just remove existing dependent promotion)
• Support field reconstruction with bitwise ops at ABI boundaries
• Support promotion of non-overlapping parts of structs with overlapping fields (non-overlapping up to access)
• Support partial lifetime promotion (Regressions due to QJFL and OSR  #67594 (comment))
• [RyuJIT] Introduce a general reverse copy prop for struct temps #8887
  • JIT: consider optimizing constructor pattern for struct creation #9237

[dotnet-issue-labeler]

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

This issue captures the planned work items for .NET 8 with respect to struct improvements. Smaller workitems that the larger workitems depend on may also be tracked here. This list is expected to change throughout the release cycle according to ongoing planning and discussions, with possible additions and subtractions to the scope.

Planned for .NET 8

Q4 2022

• JIT: Improve last-use copy omission for implicit byrefs #76069

2023

• JIT: Generalized struct promotion #76928
  • JIT: Move call ABI determination earlier #76926 (Heuristics precursor)
  • Promote (scalar replace) structs with more than 4 fields #6534 (Non-whole promotion)
  • Span<T> wrapped in a struct isn't performing as fast as it could be #68797 (Recursive promotion)
• [RyuJIT] Introduce a general reverse copy prop for struct temps #8887
  • JIT: consider optimizing constructor pattern for struct creation #9237
  • CQ: Big Value Type Suboptimal Inlining Code Generation #39732

Interesting

• Struct SSA
• Removal existing promotion (or perhaps just dependent promotion)
• Support field reconstruction with bitwise ops at ABI boundaries
• Support promotion of non-overlapping parts of structs with overlapping fields
• Support partial lifetime promotion (Regressions due to QJFL and OSR  #67594 (comment))

Author:	jakobbotsch
Assignees:	-
Labels:	area-CodeGen-coreclr, User Story
Milestone:	-

[jakobbotsch]

cc @dotnet/jit-contrib

[SingleAccretion]

remove existing dependent promotion

This would imply teaching liveness to work with parts of non-promoted aggregates (this case is the only benefit of PROMOTION_TYPE_DEPENDENT as it exists today, AFAIK).

Struct SSA

What kind of enhancements?

#68797

Note this is just an odd limitation of the current scheme. It can easily be lifted by "skipping" all trivial struct wrappers, not just those that ultimately simplify down to a primitive.

[jakobbotsch]

This would imply teaching liveness to work with parts of non-promoted aggregates (this case is the only benefit of PROMOTION_TYPE_DEPENDENT as it exists today, AFAIK).

The idea (hope) would be that the new promotion pass would create IR that never needed DNER and then dependent promotion could be replaced by the general promotion. I'm sure lots of tuning and work would be required here.

Struct SSA

What kind of enhancements?

@AndyAyersMS and I were discussing this a while ago for various copy prop/promotion reasons. Also, some SSA forms allow representing some forms of aliasing, which could be interesting in relation to partial lifetime promotion. This one is probably pretty out there though.
Another thing I talked with Andy about recently is going into SSA form during import to let the inliner have SSA form at its disposal. But it does not seem very realistic given all the code we have that would need to be updated, that's why it's not under any immediate plans.

#68797

Note this is just an odd limitation of the current scheme. It can easily be lifted by "skipping" all trivial struct wrappers, not just > those that ultimately simplify down to a primitive.

Interesting. I guess you quickly hit the 4 field limit though, but I guess we could look into if it would make sense to raise this limit in some cases (e.g. to be able to promote Span<T> inside a ref struct).

[JulieLeeMSFT]

@jakobbotsch, I am closing this as completed. Let's plan remaining items for future or .NET 9.