JIT: Clear out return value references from continuations

[jakobbotsch]

Async1 deterministically clears out awaiters on resumption, meaning that the callee's Task and its result do not stay alive. This change similarly makes it so that we do not keep results alive in async2.

Async1 has similar clearing for locals based on lexical scope. I am hoping we can get away with not implementing something similar for runtime async (it would be expensive and impossible to guarantee similar behavior as async1).

Fix #126735

[Copilot]

Pull request overview

This PR updates the CoreCLR JIT’s runtime-async transformation to proactively clear GC references from the continuation’s stored return value after resumption copies the result out, reducing unintended object rooting when continuations are reused.

Changes:

• Adds AsyncTransformation::ClearReturnValueOnResumption and invokes it after copying the awaited call’s return value when continuation reuse is enabled.
• Implements two clearing strategies for struct returns with GC pointers: per-GC-slot clearing for small/ref-sparse structs, otherwise a bulk zeroing store.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/coreclr/jit/async.h	Declares the new helper used to clear return-value GC references on resumption.
src/coreclr/jit/async.cpp	Calls the helper after copying results and implements logic to clear GC refs in ref/struct return slots when continuations are reused.

[jakobbotsch]

@EgorBo Any suggestion on the heuristic to use here? Wondering if something in the backend does similar reasoning.
Could also always just be a struct clear.

[EgorBo]

@jakobbotsch I assume we don't care about non-gc slots? it's actually a very interesting question that might improve perf everywhere. Example: https://godbolt.org/z/xMTcqn8EW today in SkipLocalsInit context (which is default everywhere in BCL) we don't try to optimize it like you do here 🤔 so I assume a fix for that might improve perf in many places.

I wonder if we should introduce a "GT_UNINIT" (or GT_POISON?) node as RHL of GT_STORE_BLK and then lower will come up with the best strategy

[jakobbotsch]

Right, we only care about clearing GC slots to ensure they don't stay rooted.

Example: https://godbolt.org/z/xMTcqn8EW today in SkipLocalsInit context (which is default everywhere in BCL) we don't try to optimize it like you do here 🤔

We do have some logic it seems:

runtime/src/coreclr/jit/codegencommon.cpp

Lines 4097 to 4112 in 4b6b5d0

	if (varDsc->TypeIs(TYP_STRUCT) && !m_compiler->info.compInitMem &&
	(varDsc->lvExactSize() >= TARGET_POINTER_SIZE))
	{
	// We only initialize the GC variables in the TYP_STRUCT
	const unsigned slots = (unsigned)m_compiler->lvaLclStackHomeSize(varNum) / REGSIZE_BYTES;
	ClassLayout* layout = varDsc->GetLayout();
	for (unsigned i = 0; i < slots; i++)
	{
	if (layout->IsGCPtr(i))
	{
	GetEmitter()->emitIns_S_R(ins_Store(TYP_I_IMPL), EA_PTRSIZE,
	genGetZeroReg(initReg, pInitRegZeroed), varNum, i * REGSIZE_BYTES);
	}
	}
	}