JIT: Allow optNarrowTree CAST case to match ULONG and LONG equivalently

[AndyAyersMS]

Fixes #111888

[AndyAyersMS]

@dhartglassMSFT PTAL
fyi @dotnet/jit-contrib

Not many diffs, and again clustered in tests, but not a very big change either.

[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

Updates CoreCLR JIT narrowing logic so optNarrowTree’s GT_CAST handling treats LONG and ULONG as equivalent (same actual type/width), enabling downstream ARM64 instruction combining to produce shifted-register forms in cases that previously got blocked by signedness mismatches in cast chains.

Changes:

• CoreCLR JIT: relax optNarrowTree’s GT_CAST gate from exact-type equality to genActualType(...) equality, and broaden the “cast long” match to include ULONG via genActualType(tree) == TYP_LONG.
• ARM64 codegen validation: update existing Neg/CMN instruction-combining tests to expect single-instruction shifted-register forms (neg ..., LSR #imm, cmn ..., LSR #imm).
• Add/adjust tests to cover an MVN cast-chain LSR case and assert elimination of a redundant ARM64 mov for (uint)(ulong)x.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/coreclr/jit/optimizer.cpp	Allows optNarrowTree cast narrowing to proceed when srct and CastToType() differ only by signedness (e.g., LONG vs ULONG).
src/tests/JIT/opt/InstructionCombining/Neg.cs	Updates ARM64 disasm expectations for NegLSR to the shifted-register neg form.
src/tests/JIT/opt/InstructionCombining/Cmn.cs	Updates ARM64 disasm expectations for CmnLSR to the shifted-register cmn form.
src/tests/JIT/opt/InstructionCombining/Mvn.cs	Adds MvnCastChainLSR test and runtime check validating MVN+LSR combining through a cast chain.
src/tests/JIT/opt/InstructionCombining/Casts.cs	Tightens ARM64 expectation for (uint)(ulong)x to ensure no redundant mov remains.

[tannergooding]

Changes LGTM, just a question about whether one change is actually doing anything and if we can maybe improve the comment.

[AndyAyersMS]

going to bounce to trigger new CI

[AndyAyersMS]

Interestingly this has minopts diffs, evidently optNarrowTree is reachable from fgOptimizeEqualityComparisonWithConst without any optimization enabled check.

[tannergooding]

Interestingly this has minopts diffs, evidently optNarrowTree is reachable from fgOptimizeEqualityComparisonWithConst without any optimization enabled check.

I think we have a few cases where opts like this aren't fully prevented in MinOpts, particularly via morph. We were discussing it on Discord a few weeks back and it'd probably be good to cleanup longer term, but obviously not critical to do right now.

[AndyAyersMS]

/ba-g unrelated test leg timing out