Fix side-effect reordering when folding constant-zero-mask BlendVariable

[tannergooding]

Fixes #125431

gtFoldExprHWIntrinsic folds a BlendVariable/BlendVariableMask with a constant mask down to one of its operands. Operands evaluate in order op1 (selectFalse), op2 (selectTrue), op3 (mask). When the mask is a constant all-zero, the result is op1 and op2 is dropped. The old code did:

return gtWrapWithSideEffects(op1, op2, GTF_ALL_EFFECT);

gtWrapWithSideEffects(op1, op2) builds COMMA(sideEffects(op2), op1), which hoists op2's side effects ahead of op1. But op2 is evaluated after op1 in source order, so this reorders them. In the repro, op1 is CreateScalar(M43(...)) (a call) and op2 is a null-array load (throws NullReferenceException); the fold moved the null-load exception ahead of the M43 call, dropping the observable side effect of M43 (matching Debug prints 1 line, Release prints 0).

Fix: mirror the conservative maskIsAllBitsSet branch directly above -- only fold when op2 has no side effects and return op1 directly; otherwise break and let value numbering handle the value fold (which preserves side-effect ordering), the same way the neighboring maskIsAllBitsSet and ConditionalSelect handlers already do.

---

Added a deterministic regression test under JitBlue/Runtime_125431 (runs with DOTNET_TieredCompilation=0). Confirmed it fails on the unfixed JIT and passes with the fix; the original Fuzzlyn repro now prints the expected line before the NullReferenceException. jit-format reports no changes.

---

Known related gap (deferred, not fixed here): the same dropped-operand hazard exists in the shared side-effect extraction machinery. gtNodeHasSideEffects/gtExtractSideEffList only consider GTF_ASG | GTF_CALL | GTF_EXCEPT | GTF_MAKE_CSE and ignore GTF_ORDER_SIDEEFF/GTF_GLOB_REF, even when passed GTF_ALL_EFFECT. A non-faulting volatile load (GLOB_REF | ORDER_SIDEEFF, no EXCEPT) can therefore still be dropped by callers that discard a value-dead operand through that path -- e.g. fgMorphModToZero (x % 1) and the TernaryLogic unused-operand path (gtUnusedValNode). A per-site mask widen does not fix those (the global-morph extraction still drops the load); the correct fix is centralized in gtNodeHasSideEffects/gtExtractSideEffList and has a wider blast radius, so it is tracked separately.

Note

This PR description was drafted by Copilot.

[azure-pipelines]

Azure Pipelines:
Successfully started running 5 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]

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 updates RyuJIT’s gtFoldExprHWIntrinsic constant-folding for BlendVariable/BlendVariableMask when the mask is a constant all-zero, to avoid unsafe operand dropping / side-effect reordering, and adds a regression test under JitBlue.

Changes:

• Adjust BlendVariable* folding so maskIsZero only returns op1 when it’s safe to drop op2.
• Add a deterministic JitBlue regression test exercising the evaluation-order / exception ordering scenario.
• Add a test project file that forces DOTNET_TieredCompilation=0 for determinism.

Show a summary per file

File	Description
src/coreclr/jit/gentree.cpp	Updates BlendVariable* folding behavior for constant masks.
src/tests/JIT/Regression/JitBlue/Runtime_125431/Runtime_125431.cs	Adds regression coverage for operand evaluation order and exception behavior.
src/tests/JIT/Regression/JitBlue/Runtime_125431/Runtime_125431.csproj	Configures isolated execution + env var for deterministic tiering behavior.

Copilot's findings

• Files reviewed: 3/3 changed files
• Comments generated: 2

[Copilot]

Copilot's findings

• Files reviewed: 3/3 changed files
• Comments generated: 2

[Copilot]

Copilot's findings

• Files reviewed: 3/3 changed files
• Comments generated: 0 new

[tannergooding]

CC. @dotnet/jit-contrib, @dhartglassMSFT a small side-effect preservation fix

[tannergooding]

Found a couple more side effect checks that were technically incorrect, so fixed them. Given they require a non-faulting volatile load, getting a regression test added is difficult.

[Copilot]

Copilot's findings

• Files reviewed: 5/5 changed files
• Comments generated: 0 new

[Copilot]

Copilot's findings

• Files reviewed: 5/5 changed files
• Comments generated: 0 new