Optimize Half comparison operators and CompareTo

[tannergooding]

Fixes #43117.

Half.CompareTo(Half) previously delegated to this < other, this > other, and this == other. Since those operators were rewritten to handle signed values and zeros, each is now a full managed method with its own NaN/sign/zero handling -- so the common non-NaN path ran roughly six IsNaN checks plus redundant sign/zero work before falling into a bit comparison it could have done directly.

This rewrites CompareTo to handle NaN once and then compare a monotonic sign-magnitude ordering key derived directly from the bits (both zeros collapse to the same key, preserving +0 == -0).

---

It also adds an AVX2 fast path to CompareTo, operator <, and operator <= that compares via (float)Half. On TARGET_XARCH + AVX2 the Half->float conversion lowers to a hardware vcvtph2ps, so comparing as float is materially cheaper. operator > / operator >= are thin right < left / right <= left wrappers and inherit the fast path automatically, and the existing Min/Max/*Magnitude* helpers already delegate through float/relational operators, so they benefit as well. Avx2.IsSupported is a JIT-time constant, so the non-taken branch is eliminated and there's no overhead on other targets; the bit-based logic remains as the hardware-independent fallback.

The equality operators (==, !=, Equals) are intentionally left on the bit logic -- converting to float there measured slower, since the bit equality check is nearly free and doesn't justify two conversions.

Half->float is lossless and order-preserving (NaN and signed-zero semantics preserved), so all fast paths match the existing bit semantics exactly.

---

Measurements (local, AVX2 machine; 4096 comparisons/op; in-process BenchmarkDotNet)

CompareTo, mixed data:

Impl	Time	ns/cmp
Old (operator-chained)	13.06 us	~3.19
Bit-key only	7.10 us	~1.73
Bit-key + AVX2 float path	3.47 us	~0.85

operator <: 4.76 us -> 2.41 us (~2x).

With AVX2 disabled the change is still a win over the old code (CompareTo 12.25 -> 7.06 us via the bit-key fallback), and the equality path measured slower as float (2.16 -> 3.20 us), which is why it was left alone.

Tests: extended HalfTests.CompareTo_TestData with signed-zero, subnormal-sign, and NaN/infinity ordering cases. Full HalfTests (1629 cases) pass both with AVX2 enabled and with DOTNET_EnableAVX2=0 exercising the fallback.

Note

This PR description and the code changes were drafted with the assistance of GitHub Copilot.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Pull request overview

This PR optimizes System.Half comparison behavior by (1) rewriting Half.CompareTo(Half) to avoid operator chaining and redundant NaN/sign/zero checks and (2) adding an AVX2-gated fast path for <, <=, and CompareTo that compares via lossless (float)Half conversion.

Changes:

• Update Half.CompareTo(Half) to handle NaN up-front and otherwise compare via a monotonic, bit-derived ordering key (collapsing +0/-0).
• Add AVX2 fast paths for operator <, operator <=, and CompareTo(Half) that compare through float on supported targets.
• Extend HalfTests CompareTo_TestData with additional signed-zero, subnormal-sign, and NaN/infinity ordering cases.

Show a summary per file

File	Description
src/libraries/System.Private.CoreLib/src/System/Half.cs	Reworks CompareTo(Half) to avoid operator chaining and adds AVX2-gated float-based fast paths for <, <=, and CompareTo.
src/libraries/System.Runtime/tests/System.Runtime.Tests/System/HalfTests.cs	Adds extra CompareTo test vectors covering signed zero, subnormal sign ordering, and NaN vs infinity ordering.

Copilot's findings

• Files reviewed: 2/2 changed files
• Comments generated: 0

[dotnet-policy-service]

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