[wasm] Regalloc: Allow Non-exposed GC Ref Locals to Live in Wasm Locals by adamperlin · Pull Request #135543 · dotnet/runtime

adamperlin · 2026-10-09T21:35:36Z

Background

This PR introduces a pass which does some liveness analysis to determine GC-ref tracked locals which are meaningfully live at GC safepoints. Locals which are live at safepoints are treated as before and we continue to assign them pinned stack homes. However, the pass uses this list to then determine candidate gc ref locals which are NOT live at any safepoint, which RA can assign local homes since the GC does not need to see them.

The one exception here is temporary reads of locals. A tree temporary from a local var which is live across a call still needs a pinned stack slot, and `fgWasmSpillRefs has been updated to handle this case. This is also why this analysis was structured as a separate phase; it is a self-contained analysis and provides information that two subsequent phases need to consume.

Notes

For throwing code paths, we must account for any LIR node which could possibly throw (this includes all null checks, div-by-zero checks, range checks, etc.) and also track which live locals are actually live on the throwing path. A reference stored in a local is not actually exposed to GC if the only collecting path is the throwing path, unless it is also live into a handler for that site.

Performance

SPMI

Diff: -549,758/+1,245 bytes
16,424 size improvements, 205 size regressions
Total bytes of delta: -548,513 (-6.41 % of base)

Local benchmarks

Metric	Mono baseline (interpreter)	CoreCLR baseline (composite R2R)	CoreCLR local (composite R2R)	Local vs. CoreCLR baseline	Local vs. Mono baseline
Cold startup — mobile	3,320 ms	7,228 ms	6,539 ms	9.5% faster	97.0% slower
Walkthrough — desktop	9.23 s	5.95 s	5.56 s	6.6% faster	39.8% faster
Cold download — desktop	4.567 MB	10.323 MB	10.138 MB	1.8% smaller	122.0% larger
App publish	37.85 s	151.89 s	147.84 s	2.7% faster	290.6% slower

Validation

The new WasmGCLiveness tests pass which is a good start, but it is hard to force the GC to move objects to detect if something is wrong. If we had GC Stress for Wasm that might help. For several of the test cases (including live across finally/EH, mixed stack/reg homes, and others), I inspected JIT dumps manually to ensure that the pass was working as expected.

azure-pipelines · 2026-10-09T21:36:31Z

Azure Pipelines: Successfully started running 6 pipeline(s). 10 pipeline(s) were filtered out due to trigger conditions. There may be pipelines that require an authorized user to comment /azp run to run.

dotnet-policy-service · 2026-10-09T21:36:53Z

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

Copilot

🟡 Changes recommended

The tests do not automatically verify that eligible references are actually assigned Wasm-local homes.

3 open findings

What changed in this PR

Adds Wasm GC-reference liveness analysis so safe reference locals can use Wasm-local homes.

Changes:

Identifies tracked GC references not live across safepoints.
Enables their register allocation and preserves call arguments via spills.
Adds GC-liveness correctness scenarios.

File	Description
`src/coreclr/jit/fgwasm.cpp`	Implements candidate analysis and spill handling.
`src/coreclr/jit/regallocwasm.cpp`	Allows identified references into Wasm locals.
`src/coreclr/jit/compiler.cpp`	Runs the new analysis phase.
`src/coreclr/jit/compiler.h`	Declares candidate state and phase method.
`src/coreclr/jit/compphases.h`	Registers the new phase.
`src/coreclr/jit/compmemkind.h`	Adds candidate-set allocation tracking.
`src/tests/JIT/Directed/WasmGCLiveness/WasmGCLiveness.cs`	Adds GC-liveness scenarios.
`src/tests/JIT/Directed/WasmGCLiveness/WasmGCLiveness.csproj`	Configures the isolated priority-one test.

🧠 Review effort: Balanced

The current `fgWasmSpillRefs` pass spills all outstanding GC-ref typed temporary values, even if those temps are consumed by the most immediate call. However, for calls it should be the callee's responsibility to maintain GC safety of any incoming GC-ref typed args (either by spilling to a pinned stack home, spilling and reloading, or potentially avoiding spilling at all, see #135543). In the R2R fixup case, where a thunk will fix up the call target on first invocation, we should also be GC-protecting all incoming args by default (See the spilling logic in crossgen2 `WasmImportThunk.cs`). This PR removes the need for spilling in cases like: ``` MyClass x = Allocate(); Use(x); ``` Note that the spilling STILL needs to happen for nested calls, where a ref has an outstanding use which is NOT its most immediate parent. ```csharp MyClass x = Allocate(); M(x, F()) // <-- temporary value of x is an operand of the call to M, and may be live across the other sub-tree call, F where a GC might occur. ``` ## Results ### SuperPMI Total bytes of base: 8771651 Total bytes of diff: 8555629 Total bytes of delta: **-216,022 (-2.46% of base)** ### Local Benchmarks | Metric | CoreCLR baseline | CoreCLR after | Change | |---|---:|---:|---:| | Mobile cold startup | 7,474 ms | 6,901 ms | 7.7% faster | | Desktop walkthrough | 5.60 s | 6.05 s | 8.1% slower | | Desktop cold download | 10.324 MB | 10.215 MB | 1.1% smaller | | Havit publish | 150.81 s | 149.21 s | 1.1% faster | Cold download and startup improve as expected. I didn't put too much weight on the walkthrough benchmark because it is only run once and is pretty variable.

adamperlin · 2026-10-16T00:43:07Z

The regressions have mostly been mitigated by #135577, there were many more before. The ones that remain are expected given the current state; In these cases we have a local which is not considered live over a safepoint using local liveness, but its value is read into a temporary which is live across a safepoint. In that case, it is a bit more expensive to spill the temporary value and reload it than it would have been to assign the local a stack home to begin with. We could potentially try and do some more analysis to avoid even these cases, but that doesn't seem worth it to me due to the comparatively very small size of the regressions.

adamperlin · 2026-10-16T17:01:43Z

cc @dotnet/wasm-contrib; @EgorBo PTAL.

EgorBo

Approving to unblock the work, but this might need a 2nd look by @AndyAyersMS I think

pavelsavara · 2026-10-20T11:51:18Z

would this make the values/references invisible to managed debugger ? Is that OK for R2R code ? @radekdoulik

dotnet-policy-service · 2026-10-20T11:51:37Z

Tagging subscribers to 'arch-wasm': @lewing, @pavelsavara
See info in area-owners.md if you want to be subscribed.

…r a method and checks the state of local liveness at each potential GC safepoint to determine which locals are live across a safepoint. It distinguishes between normal GC safepoints triggered by a call or and exception-only GC safepoints, where an operation can only collect on its throwing path. This allows us to effectively discount locals as gc-exposed if they are only exposed to GC on a throwing path but are NOT live into a handler. The set of gc-safe locals is the set of gc exposed locals subracted from the set of all locals. This reduced set of gc-safe locals can be used by Wasm regalloc to make gc refs into wasm-local candidates.

…asses to allow safely placing select GC references in locals

adamperlin · 2026-10-21T00:59:06Z

/ba-g unrelated failures

adamperlin requested a balanced review from Copilot October 9, 2026 21:35

github-actions Bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Oct 9, 2026

Copilot started reviewing on behalf of adamperlin October 9, 2026 21:36

View session

Copilot AI reviewed Oct 9, 2026

View reviewed changes

Comment thread src/coreclr/jit/regallocwasm.cpp

Comment thread src/coreclr/jit/fgwasm.cpp Outdated

Comment thread src/tests/JIT/Directed/WasmGCLiveness/WasmGCLiveness.cs Outdated

adamperlin mentioned this pull request Oct 10, 2026

[wasm] Avoid Spilling Many GC-ref Typed Tree Temps #135577

Merged

adamperlin force-pushed the adamperlin/wasm-enregister-gcsafe-refs branch from 6ee0537 to f702924 Compare October 15, 2026 20:43

adamperlin marked this pull request as ready for review October 16, 2026 00:33

adamperlin requested review from EgorBo and davidwrighton October 16, 2026 00:33

EgorBo approved these changes Oct 16, 2026

View reviewed changes

adamperlin mentioned this pull request Oct 16, 2026

[User Story] RyuJIT Wasm Codegen Quality Improvements #135985

Open

1 of 13 tasks

adamperlin force-pushed the adamperlin/wasm-enregister-gcsafe-refs branch from 74537f1 to c25a9fa Compare October 16, 2026 23:11

build-analysis Bot mentioned this pull request Oct 17, 2026

[ci-scan] Test failure: SocketsHttpHandler_HttpClientHandler_ServerCertificates_Test bad-certificate cases on Android CoreCLR #135988

Closed

adamperlin force-pushed the adamperlin/wasm-enregister-gcsafe-refs branch from c25a9fa to e3d52f9 Compare October 19, 2026 22:26

pavelsavara added the arch-wasm WebAssembly architecture label Oct 20, 2026

adamperlin added 8 commits October 20, 2026 13:34

Add Wasm GC Liveness tests for a variety of scenarios created by copilot

1d61bb8

Propagate Wasm local candidate info to fgWasmSpillRefs and regalloc p…

ef5aa56

…asses to allow safely placing select GC references in locals

jit-format

1b8705b

Update comment

52c1b5c

Clean up unnecessary JITDUMPs and add helper macro

3dcd577

Update some comments

8affa92

Reformat and change some comments

4e7a64c

adamperlin force-pushed the adamperlin/wasm-enregister-gcsafe-refs branch from e3d52f9 to 4e7a64c Compare October 20, 2026 20:34

davidwrighton approved these changes Oct 20, 2026

View reviewed changes

adamperlin merged commit dedc9b3 into dotnet:main Oct 21, 2026
131 of 135 checks passed

dotnet-maestro Bot mentioned this pull request Oct 21, 2026

[main] Source code updates from dotnet/runtime dotnet/dotnet#10348

Merged

dotnet-milestone-bot Bot added this to the 12.0-preview1 milestone Oct 21, 2026

LoopedBard3 mentioned this pull request Oct 22, 2026

[Perf] Linux/x64: 1 Regression on 10/21/2026 4:49:01 AM +00:00 dotnet/perf-autofiling-issues#81400

Open

Navigation Menu

Repository navigation

[wasm] Regalloc: Allow Non-exposed GC Ref Locals to Live in Wasm Locals - #135543

[wasm] Regalloc: Allow Non-exposed GC Ref Locals to Live in Wasm Locals#135543
adamperlin merged 8 commits into
dotnet:mainfrom
adamperlin:adamperlin/wasm-enregister-gcsafe-refs

adamperlin commented Oct 9, 2026 •

edited

Loading

Uh oh!

azure-pipelines Bot commented Oct 9, 2026

Uh oh!

dotnet-policy-service Bot commented Oct 9, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

adamperlin commented Oct 16, 2026 •

edited

Loading

Uh oh!

adamperlin commented Oct 16, 2026 •

edited

Loading

Uh oh!

EgorBo left a comment

Uh oh!

pavelsavara commented Oct 20, 2026

Uh oh!

dotnet-policy-service Bot commented Oct 20, 2026

Uh oh!

adamperlin commented Oct 21, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

adamperlin commented Oct 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Background

Notes

Performance

SPMI

Local benchmarks

Validation

Uh oh!

azure-pipelines Bot commented Oct 9, 2026

Uh oh!

dotnet-policy-service Bot commented Oct 9, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Uh oh!

adamperlin commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

adamperlin commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

EgorBo left a comment

Choose a reason for hiding this comment

Uh oh!

pavelsavara commented Oct 20, 2026

Uh oh!

dotnet-policy-service Bot commented Oct 20, 2026

Uh oh!

adamperlin commented Oct 21, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

adamperlin commented Oct 9, 2026 •

edited

Loading

adamperlin commented Oct 16, 2026 •

edited

Loading

adamperlin commented Oct 16, 2026 •

edited

Loading