JIT: Track AVX state from register-producing widths

[tannergooding]

Follow up on #134410 by tracking AVX upper state from register-producing widths rather than vector types or memory-store widths. This removes unnecessary vzeroupper instructions while accounting for previously missed wide loads, copies, and integer-division intermediates.

• Account for allocator-inserted spills/reloads and actual local register copies.
• Narrow unsafe-widening/extraction copies to the defined source or required result width, and use XMM zeroing in integer-vector division.
• Preserve the existing call/prolog/epilog clearing policy. Native-boundary policy changes, including the pre-existing unmanaged calli gap, remain out of scope.

SuperPMI results

CI build 1610798, testing head b20f2c3d058:

Target	Compared contexts	Net code-size change	FullOpts contribution
Windows x64	2,814,515	+810 bytes	−1,008 bytes
Linux x64	3,256,661	+15,645 bytes	+11,970 bytes
Windows x86	2,576,563	+11,088 bytes	+8,784 bytes

ARM, ARM64, and WASM have no assembly diffs. Windows corpus coverage expanded substantially, so raw totals are not directly comparable with earlier runs. Remaining growth includes both corrected wide-producer accounting and conservative method-wide cleanup placement; it is not all individually necessary transition cleanup.

JIT executed-instruction counts increased approximately 0.011–0.031% on xarch; these are compiler-work measurements, not application-throughput results. The assembly runs report zero failing compilations, with symmetric missing-context counts between base and diff.

Validated with Checked x64/all-target JIT builds, focused execution and disassembly checks across default/AVX2/AVX/SSE/software configurations, targeted register-stress runs, and JIT formatting. Runtime execution was on Windows x64; cross-target replay is not runtime execution on those platforms.

Note

This PR description and implementation were prepared with GitHub Copilot assistance.

[azure-pipelines]

Azure Pipelines:
Successfully started running 6 pipeline(s).
10 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[Copilot]

Copilot review overview

🟡 Changes recommended

Local SIMD stores still need correction, and focused GetLower128 regression coverage is missing.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1

Open (1)

• Exclude SIMD local stores from AVX-width tracking · New

What changed in this PR

Updates x64 JIT AVX-state tracking to reflect actual register-producing widths and unmanaged call boundaries.

Changes:

• Refines SIMD width tracking, reload handling, and division temporary codegen.
• Recognizes unmanaged calli boundaries.
• Adds functional and disassembly regression coverage.

File	Reviewed changes
src/​tests/​JIT/​opt/​Vectorization/​AvxState.csproj	Configures the AVX-state regression test.
src/​tests/​JIT/​opt/​Vectorization/​AvxState.cs	Adds functional and disassembly coverage.
src/​coreclr/​jit/​lsraxarch.cpp	Tracks SIMD widths through LSRA.
src/​coreclr/​jit/​lsra.cpp	Accounts for allocator reloads and copies.
src/​coreclr/​jit/​hwintrinsiccodegenxarch.cpp	Narrows register moves and division zeroing.
src/​coreclr/​jit/​gentree.cpp	Handles unmanaged calls requiring AVX-state clearing.

[Copilot]

Copilot review overview

🟡 Changes recommended

Review findings identify a critical AVX/AVX2 assertion risk and gaps in tracking and regression coverage.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1

Open (1)

• Avoid invalid AVX width for Vector256 division · New

Resolved since last review (1)

• Exclude SIMD local stores from AVX-width tracking

Previously missed (1)

In code that hasn't changed since last review

Test does not verify unmanaged calli vzeroupper behavior

src/​tests/​JIT/​opt/​Vectorization/​AvxState.cs:84

This pattern does not actually verify the new unmanaged-calli behavior: CallUnmanaged necessarily emits an xorps/vxorps to materialize Vector128<float>.Zero, so the alternation can match that instruction even when NeedsVzeroupper still returns false for IsUnmanaged(). The regression would therefore pass against the old JIT; constrain the check to vzeroupper in AVX-enabled runs (and separately cover the no-AVX fallback), or otherwise make the call-site assertion unambiguous.

[tannergooding]

Regarding the unmanaged calli test finding in the review summary: the leading space in {{vzeroupper| xorps }} is intentional. The xorps alternative matches the legacy instruction used without AVX, but cannot match vxorps.

The preserved pre-fix AVX-enabled disassembly contains vxorps xmm0, xmm0, xmm0 and no vzeroupper; it fails this exact check. The fixed disassembly includes vzeroupper and passes. The test therefore distinguishes the missing unmanaged-call clearing from the no-AVX fallback as intended.

Note

This reply was drafted with GitHub Copilot assistance.

[tannergooding]

@EgorBo for this one it does some more cleanup and fixes other places we were incorrectly handling vzeroupper.

It still doesn't handle every possible case of unmanaged call (such as through calli, certain delegates, or fnptrs) because that caused far too many regressions and needs more complex handling to do properly.

The size regressions that are still present are generally correctness improvements. Ideally we'd have a bit more sophisticated tracked to avoid doing the operation unnecessarily, particularly when there's like only a single YMM usage late or optionally in the method, but that is a much more complex and involved fix.

This at least generally fixes the remaining cases that were trivially incorrectly tracking the state.

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved AVX-state accounting gaps remain for wide constants, spill paths without reloads, and regression coverage.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1

Open (1)

• Track wide SIMD constants in contained local stores · New

Resolved since last review (1)

• Avoid invalid AVX width for Vector256 division

[MichalPetryka]

certain delegates

Don't those just use IL stubs with calli? So they should be handled by calli handling cc @jkoritzinsky