[Wasm RyuJIT] Spill live ref/byref values to pinned stack slots at calls

[kg]

GC refs/byrefs can end up temporarily living on the Wasm shadow stack - where the GC can't see them - across calls, which creates a GC hole. This PR adds a pass that 'spills' these live refs/byrefs to dedicated pinned locals so that the GC won't move or collect those refs.

[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

This PR introduces a Wasm-specific mechanism intended to make ref/byref values GC-visible across call sites by injecting a new IR node (GT_WASM_SPILL_REF) and a new Wasm phase (WasmSpillRefs) that inserts these nodes and allocates pinned stack spill slots used during Wasm codegen.

Changes:

• Add GT_WASM_SPILL_REF node kind and operand iteration support.
• Add Compiler::WasmSpillRefs phase to insert spill nodes around calls and allocate spill locals.
• Extend Wasm codegen/regalloc to track a spill index and (temporarily) force-enregister a scratch “splash zone” local.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/coreclr/jit/regallocwasm.cpp	Forces the “splash zone” local to be treated as a reg candidate.
src/coreclr/jit/gtlist.h	Adds new Wasm node WASM_SPILL_REF.
src/coreclr/jit/gentree.cpp	Updates operand-edge iterator to treat GT_WASM_SPILL_REF as unary.
src/coreclr/jit/fgwasm.cpp	Implements Compiler::WasmSpillRefs to insert spill nodes and allocate spill locals.
src/coreclr/jit/compphases.h	Adds PHASE_WASM_SPILL_REFS.
src/coreclr/jit/compmemkind.h	Adds WasmSpillRefs memory kind.
src/coreclr/jit/compiler.h	Adds m_wasmSpillSlots field and WasmSpillRefs declaration.
src/coreclr/jit/compiler.cpp	Wires WasmSpillRefs into the Wasm compilation pipeline.
src/coreclr/jit/codegenwasm.cpp	Emits Wasm for GT_WASM_SPILL_REF and resets spill index at calls.
src/coreclr/jit/codegenlinear.cpp	Resets spill index at block boundaries on Wasm.
src/coreclr/jit/codegen.h	Adds wasmSpillRefIndex state.

[AndyAyersMS]

Generally looks good.

I'm curious how this intersects/overlaps with LSRA's spill temp mechanism. Not saying we should use that here, but it likely serves a similar purpose.

[kg]

Still working on this, going to try moving to STORE_LCL_VAR/LCL_VAR like Jakob suggested.

[Copilot]

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated 4 comments.

[kg]

Updated to use STORE_LCL_VAR and LCL_VAR instead, removing the splash zone var.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

[kg]

I had to come up with a different test program to get failures without this PR's changes - the previous test program was only failing because of some other bugs that AndyA has since fixed.

publicclassC{publicstringS="";publicbyte[]?Trash;[MethodImpl(MethodImplOptions.NoInlining)]publicvoidAppendToS(stringvalue){Trash=newbyte[32768];S+=value;}}[MethodImpl(MethodImplOptions.NoInlining)]publicstaticintMain(){Console.WriteLine("Hello world!");strings="";for(inti=0;i<1000;i++){Cc=newC();for(intj=0;j<2000;j++)c.AppendToS(j.ToString());s=c.S;}Console.WriteLine(s);return42;}

With the spilling pass disabled this asserts inside the GC during execution, i.e.

Hello world!
ASSERT FAILED
Expression: xl <= end
Location: Z:/runtime/src/coreclr/gc\plan_phase.cpp:3194
Function: find_next_marked
Process: 42
Frame (DynamicHelperFrame): 0x4fc938
Skipping 0x4fc938
Frame (InterpreterFrame): 0x4fdeb0
0) System.String::FastAllocateString, IR_001d
1) System.String::Concat, IR_007f
Frame (InterpreterFrame): 0x4ff360
2) System.Environment::CallEntryPoint, IR_0097

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

[AndyAyersMS]

Think this is close, though we probably still need slot zeroing.

Another option would be to try and report these slots as tracked, since we know their lifetimes, but that can happen later.

Interested to see the code size impact.

[AndyAyersMS]

LGTM

@jakobbotsch you want to take another look?

[jakobbotsch]

In general I would have preferred to keep the slot reuse and pinning optimization separate if a simpler "just create a new local for the LIR edge" would have sufficed for correctness.

That aside this looks good to me, I left mainly nits as comments. Feel free to ignore them.

[AndyAyersMS]

Had to fix a merge conflict.

[AndyAyersMS]

@jakobbotsch can you re-approve (new repo policy)

[AndyAyersMS]

/ba-g persistent timeouts in some arm64 legs