Remove duplicated SVE GatherVector APIs

[ylpoonlg]

These APIs are duplicated with the normal GatherVector methods, as 32-bit values doesn't need to be extended to 32-bit.

See #103370, which removed the Int32 extensions, but did not remove the UInt32 extensions.

@dotnet/arm64-contrib @a74nh

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-runtime-intrinsics
See info in area-owners.md if you want to be subscribed.

[a74nh]

Given this is removing released APIs, I think this might need a API break issue raised against it?

[tannergooding]

Given this is removing released APIs, I think this might need a API break issue raised against it?

All the APIs are [Experimental] so they aren't "released" and we don't strictly need a breaking change doc.

I don't have a preference if we want to have a general callout for all the fixes we've made in .NET 11 to the experimental surface or not, will leave that up to everyone else.

[stephentoub]

Code Review: PR #124033 - Remove duplicated SVE GatherVector APIs

Holistic Assessment

Motivation: This PR removes duplicated SVE GatherVectorUInt32 API overloads. The removed APIs that operate on 32-bit values (Vector/Vector) are semantically equivalent to the regular GatherVector methods since 32-bit values don't need to be "zero-extended to 32-bit" - they're already 32-bit. PR #103370 previously removed the Int32 versions, but the UInt32 versions were missed.

Approach: The PR:

1. Removes redundant API overloads from Sve.cs and Sve.PlatformNotSupported.cs
2. Updates the reference assembly (System.Runtime.Intrinsics.cs)
3. Updates the JIT intrinsic table (hwintrinsiclistarm64sve.h) to only enable instructions for 64-bit types
4. Adds API compatibility suppressions for the breaking change (CP0002)
5. Removes corresponding test cases

Net positive: ✅ Yes - this is a necessary API cleanup to remove semantically incorrect/duplicate APIs, following the precedent set in #103370.

---

Detailed Findings

✅ API Removals (Sve.cs, Sve.PlatformNotSupported.cs, System.Runtime.Intrinsics.cs)

• Correctly removes 16 overloads (4 methods × 4 overloads each) with Vector<int> or Vector<uint> as the return type
• Retains the 64-bit variants (Vector<long>, Vector<ulong>) which are semantically correct
• Parallel removals in both the implementation and PlatformNotSupported stub files

✅ JIT Intrinsic Table Changes (hwintrinsiclistarm64sve.h)

• Changes INS_sve_ld1w and INS_sve_ldff1w entries from supporting int/uint positions to INS_invalid
• This correctly restricts the 4 affected intrinsics to only long/ulong types

✅ API Compatibility Baseline (ApiCompatBaseline.NetCoreAppLatestStable.xml)

• All 16 removed API suppressions are correctly added with CP0002 diagnostic IDs
• Left/Right versions (net10.0 → net11.0) are correct for this .NET 11 timeframe

✅ Test Removals (SveTests.cs)

• Removes test cases for the removed APIs, keeping tests for the retained 64-bit variants

---

Summary

Verdict: ✅ Approve

This PR is a clean follow-up to #103370 that removes the remaining incorrectly-designed UInt32 "zero-extend" gather APIs. The change is breaking but correctly suppressed via the API compatibility baseline. All files are consistently updated and the approach matches the prior art.

No issues found.

[stephentoub]

🤖 Copilot Code Review — PR #124033

Holistic Assessment

Motivation: This PR removes 16 semantically incorrect API overloads that were duplicates of regular GatherVector methods. The removed APIs claim to "zero-extend 32-bit to 64-bit" but return Vector<int> or Vector<uint> — a 32-bit value cannot be meaningfully "zero-extended to 32-bit." PR #103370 previously removed the Int32-returning variants; this PR completes the cleanup by removing the UInt32-returning variants.

Approach: The PR correctly:

1. Removes the 16 redundant overloads from Sve.cs and Sve.PlatformNotSupported.cs
2. Updates the ref assembly to match
3. Restricts JIT intrinsic table entries to only support 64-bit types (INS_invalid for int/uint slots)
4. Adds API compatibility suppressions (CP0002) for the breaking change
5. Removes corresponding test data

Summary: ✅ LGTM — This is a clean API cleanup that follows established precedent from #103370. All file changes are consistent, the breaking change is correctly documented in the API compat baseline, and the APIs being removed are experimental ([Experimental]).

---

Detailed Findings

✅ API Removals — Correct and Complete

The 16 removed overloads fall into 4 methods × 4 overloads each:

• GatherVectorUInt32WithByteOffsetsZeroExtend (returning Vector<int> with Vector<int> or Vector<uint> offsets, and Vector<uint> versions)
• GatherVectorUInt32WithByteOffsetsZeroExtendFirstFaulting (same pattern)
• GatherVectorUInt32ZeroExtend (same pattern)
• GatherVectorUInt32ZeroExtendFirstFaulting (same pattern)

The 64-bit variants (Vector<long>, Vector<ulong>) are correctly retained — these make semantic sense for "zero-extend from 32-bit to 64-bit."

✅ JIT Intrinsic Table — Correctly Narrowed

The changes to hwintrinsiclistarm64sve.h correctly mark the int/uint type slots (positions 4-5 in the instruction array) as INS_invalid, leaving only long/ulong (positions 6-7) with valid instructions.

✅ Breaking Change Handling — Appropriate

All 16 suppressions are correctly added with:

• DiagnosticId: CP0002 (member exists in left but not right)
• Left: net10.0/System.Runtime.Intrinsics.dll
• Right: net11.0/System.Runtime.Intrinsics.dll

Since these APIs are marked [Experimental], formal breaking change documentation is not strictly required per maintainer feedback.

✅ Test Updates — Consistent

Test data for the removed overloads has been removed from SveTests.cs.

---

Multi-model review: GPT-5 concurred with LGTM verdict.

[dhartglassMSFT]

jit pieces lgtm
If it's unclear, I think the Unexpected HW intrinsic failures from spmi are expected when removing an intrinsic like this