Improve the handling of popcnt and tzcnt on arm64

[tannergooding]

Previously popcnt wasn't handled at all and tzcnt was imported as two IR nodes which could negatively impact certain heuristics and pattern matching.

[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

Improves ARM64 codegen for int.PopCount/long.PopCount and int.TrailingZeroCount/long.TrailingZeroCount by representing them as a single GT_INTRINSIC IR node instead of leaving popcnt unhandled and emitting tzcnt as two separate hwintrinsic nodes (ReverseElementBits + LeadingZeroCount). Lowering to actual instructions (cnt/addv for popcnt, rbit/clz for tzcnt) is now done directly in the codegen, which should preserve more information for heuristics and pattern matching.

Changes:

• In the importer, emit a GenTreeIntrinsic(TYP_INT, ...) node for ARM64 PopCount (previously not handled) and for ARM64 TrailingZeroCount (previously decomposed into two hwintrinsics).
• Add ARM64 LSRA build logic for NI_PRIMITIVE_PopCount (requests an internal SIMD temp) and NI_PRIMITIVE_TrailingZeroCount.
• Add ARM64 codegen for the two intrinsics: cnt+addv over an 8-byte SIMD register for PopCount, and rbit+clz (sized via emitActualTypeSize(srcNode)) for TrailingZeroCount.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/coreclr/jit/importercalls.cpp	Build a single NI_PRIMITIVE_PopCount/NI_PRIMITIVE_TrailingZeroCount intrinsic node on ARM64; set baseType = TYP_INT so the existing widen-back cast handles long return types.
src/coreclr/jit/lsraarm64.cpp	Add LSRA cases: PopCount needs an internal float (SIMD) reg; TrailingZeroCount is a straight 1-src, 1-dst integer op.
src/coreclr/jit/codegenarmarch.cpp	Emit cnt/addv via a SIMD temp for PopCount and rbit/clz for TrailingZeroCount under TARGET_ARM64.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

[tannergooding]

CC. @dotnet/jit-contrib, @EgorBo for review. This ensures that tzcnt and popcnt are properly imported as intrinsic on Arm64, allowing other phases to handle those nodes and avoiding the inliner.

Diffs are generally positive, showing up to -7.4k bytes of codegen on Linux Arm64 and nearly as much on Windows Arm64. The overall PerfScore is an improvement as well.

There are a few regressions intermixed, these mostly look to be due to additional CSEs, although a couple are from extra casts inserted for small types.

[tannergooding]

There are a few regressions intermixed, these mostly look to be due to additional CSEs, although a couple are from extra casts inserted for small types.

The extra casts look easy enough to fix, seems to be because we aren't handling it in IntegralRange::ForNode like we did with ToScalar(). I'd prefer to do that in a follow up PR, however, since there's a few intrinsics to add there; not just tzcnt/popcnt

[EgorBo]

This feels like yet another case where I don't understand why the managed code in BitOperations

isn't inlined into exactly what you emit by hands here. My understanding that it happens because we don't do opts for AdvSimd.PopCount that we do for PRIMITIVE_PopCount later in JIT (meaning users code that calls AdvSimd.PopCount directly won't be optimized this much)

[tannergooding]

meaning users code that calls AdvSimd.PopCount directly won't be optimized this much

Correct, because it is a SIMD instruction and must be presumed to operate across all SIMD8/SIMD16 bytes. We have no way to trivially map that back, especially given the more complex IR and our limited ability to track ranges already.

This allows us to track it directly as an integer operation and ignore the fact that codegen actually involves SIMD operations due to a limitation in the Arm64 instruction set.