JIT: Fold insertps chains that source a scalar extracted from another vector

[tannergooding]

Recognizes insertps chains where the value being inserted is itself a scalar extracted from another vector and folds them into a single insertps that selects the source element directly via count_s (bits 6-7 of the imm8).

That is, it rewrites:

Insert(vector, CreateScalarUnsafe(vector2.GetElement(idx1)), idx2)

into a single insertps reading lane idx1 of vector2, folding away the separate extraction (which otherwise requires its own movshdup/unpckhps/shufps).

Details

insertps operates purely on the 32-bit lanes of the low 128 bits of its source register, so this fold is valid for:

• Any 4-byte element type (float, int, uint) — count_s selects a 32-bit dword, which maps to the element index only when the element size is 4 bytes.
• A source vector of any width (128/256/512) — provided the extracted element is one of the first four lanes. That is all count_s can encode and all that resides in the low 128 bits, which is the only part insertps reads. Feeding a wider (ymm/zmm) source register to the 128-bit insertps is physically correct since the VEX.128 form reads the xmm view.

Both NI_Vector_ToScalar (source index 0) and constant-index NI_Vector_GetElement feeders are handled. When the extraction reads from a contained memory operand, the existing lowering already produces an optimal insertps xmm, m32 (which encodes the element offset in the address), so the fold keeps the register form only when the source is used from a register, avoiding a forced full-vector load.

Because the fold keys off the (already-narrowed, SIMD16) GetElement node rather than the vector feeding it, it naturally handles wide sources whose element access has been lowered through GetLower/GetLower128/GetUpper/ExtractVector128 — the correct 128-bit chunk is materialized and count_s selects the right lane within it.

Diffs

SuperPMI asmdiffs, windows x64, all collections: -141 bytes, 18 methods improved, 0 regressed (no PerfScore regressions).

Collection	Δ bytes	Methods	PerfScore (in diffs)
libraries.crossgen2	-29	5 improved	-8.31%
libraries.pmi	-8	2 improved	-7.68%
libraries_tests_no_tiered	-104	11 improved	-0.88%

Representative improvements: Matrix4x4.CreateScale, Matrix3x2.CreateScale, Vector3.Reflect, Vector2.Reflect, Quaternion.Lerp, and various Matrix4x4 billboard/shadow/decompose helpers.

A regression test covering register/memory sources, chained inserts, and wide (Vector256) sources across all element indices is included.

Note

This PR was authored with the assistance of GitHub Copilot.

[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 enhances x64/x86 JIT lowering for Sse41.Insert/insertps by folding a pattern where the inserted scalar is immediately sourced from another vector element, allowing the lowering to select the source lane directly via the count_s imm8 field and eliminate the separate extract+scalar-materialization sequence.

Changes:

• Add a lowering-time fold in LowerHWIntrinsic (xarch) to collapse Insert(..., CreateScalarUnsafe(GetElement(...)), ...) into a single insertps with updated count_s.
• Add a new JIT optimization test project validating result correctness across index combinations, chained inserts, and a Vector256<float>-sourced scenario.

Show a summary per file

File	Description
src/coreclr/jit/lowerxarch.cpp	Adds the CreateScalarUnsafe(GetElement(...))-source folding logic in the NI_X86Base_Insert lowering path.
src/tests/JIT/opt/Vectorization/InsertScalarFromVector.cs	New xUnit test covering combinations, chaining, and wide-source scenarios for the new fold.
src/tests/JIT/opt/Vectorization/InsertScalarFromVector.csproj	New test project file to build the added test.

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]

@MihuBot -nuget

[tannergooding]

CC. @EgorBo, @dotnet/jit-contrib.

Simple improvement to the insertps folding logic in lowering that removes a known pessimization. Gets -1218 bytes of codegen, namely in ImageSharp and BepuPhysics.