fix(Heap2Local): Track RMW results to handle nested flows (#8850)

[ArkadySkv]

This PR fixes a bug in Heap2Local where nested atomic RMW operations
were incorrectly optimized.

When an inner cmpxchg result flows into the value operand of an outer
xchg, the pass now correctly tracks this data flow and preserves the
RMW operations while localizing their ref operands.

The fix adds RMW result tracking infrastructure (RMWResultInfo) and
extends EscapeAnalyzer, visitStructRMW, and visitStructCmpxchg
to handle nested flows.

Tested with:

• The original reproducer from Issue Heap2Local bug on cmpxchng #8850
• Updated heap2local-rmw.wast lit test

All tests pass (4 d8-related failures are unrelated and due to local
environment).

Fixes #8850.

[tlively]

Was this bug not already fixed by #8857?

[ArkadySkv]

@tlively, Given the diff shows only CHECK line changes, here is the honest reply. It preserves the structure of what you prepared but corrects the claim.

Thanks for the question. I looked into this more carefully and the picture is different from what I assumed when I opened the PR.

I compared the test file on main and on this branch:

git show main:test/lit/passes/heap2local-rmw.wast > /tmp/main.wast
git show fix-heap2local-rmw-tracking-8850:test/lit/passes/heap2local-rmw.wast > /tmp/pr.wast
diff /tmp/main.wast /tmp/pr.wast

The only differences are in ;; CHECK: lines. No new test cases were added on the branch that are not already on main.

I also ran the nested RMW reproducer on both branches:

./bin/wasm-opt nested-rmw.wast --heap2local --fuzz-exec -all

The fuzz-exec self-check — which runs the module before and after optimization and compares the results — passes on both main and this branch. The output is identical. So main already handles the nested case correctly.

You were right in your original question.#8857 did cover this, and the Heap2Local.cpp changes here do not change the execution behaviour of the nested pattern. What they do change is the IR shape, which is why the CHECK lines needed updating.

Since main already handles the nested case, I will close this PR. The Heap2Local.cpp changes add tracking infrastructure that does not change behaviour on any input I have tested, so it is not worth carrying. Sorry for the noise — I should have verified against main first.