JIT: Propagate class handle when split tree

[hez2010]

Propagate the class handle when we spill a tree into a local so that later devirtualization can see the more exact class.

Codegen for the snippet in the linked issue:

Before:

G_M000_IG01:                ;; offset=0x0000
       sub      rsp, 40
 
G_M000_IG02:                ;; offset=0x0004
       mov      rcx, 0x7FFB9190F410
       call     CORINFO_HELP_NEWSFAST
       mov      rcx, rax
       mov      byte  ptr [rcx+0x08], 0
       mov      r11, 0x7FFB91640078
       call     [r11]Program+IBar:Bar():this
       nop      
 
G_M000_IG03:                ;; offset=0x0028
       add      rsp, 40
       ret      
 
; Total bytes of code 45

After:

G_M24006_IG01:  ;; offset=0x0000
       sub      rsp, 40
                                                ;; size=4 bbWeight=1 PerfScore 0.25
G_M24006_IG02:  ;; offset=0x0004
       mov      ecx, 42
       call     [System.Console:WriteLine(int)]
       nop
                                                ;; size=12 bbWeight=1 PerfScore 3.50
G_M24006_IG03:  ;; offset=0x0010
       add      rsp, 40
       ret
                                                ;; size=5 bbWeight=1 PerfScore 1.25

; Total bytes of code 21

Resolves #128483

[hez2010]

@MihuBot

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates gtSplitTree to better annotate compiler-generated temps created during tree splitting, including marking them as single-definition locals and propagating reference type class metadata when available.

Changes:

• Marks the temporary local produced by gtNewTempStore as lvSingleDef.
• When the value is a TYP_REF, attempts to propagate its class handle to the temp via lvaSetClass.
• Refactors the splitting logic to use a named value temp for clarity.

Comments suppressed due to low confidence (2)

src/coreclr/jit/gentree.cpp:1

• isNonNull is computed via gtGetClassHandle(...) but not used. Either remove isNonNull entirely if it’s not needed here, or (preferred if available in this codebase) propagate the non-null information to the temp local (e.g., via the appropriate lva* API) so the extra work has a functional effect.
  src/coreclr/jit/gentree.cpp:1
• This directly mutates lvSingleDef at the point of inserting a store. If later phases can introduce additional defs for this temp (or if lvSingleDef is expected to be derived by an analysis pass instead of being set manually), this risks stale/incorrect single-def metadata. Consider setting this only when the temp is created (where its lifetime/def-count is guaranteed), or using an existing helper/annotation mechanism for ‘compiler temp known single-def’ (so invariants stay centralized).

[dotnet-policy-service]

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

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (2)

src/coreclr/jit/gentree.cpp:1

• isNonNull is populated by gtGetClassHandle but never used afterward. If nullability isn’t needed here, pass nullptr for that out-param and remove isNonNull; if it is needed, propagate it onto the local using the appropriate local/property so the information isn’t computed and then dropped.
  src/coreclr/jit/gentree.cpp:1
• Prefer assigning boolean fields using true/false rather than 1/0 to better communicate intent and avoid relying on the underlying representation of the field.