JIT: omit redundant return-buffer initialization

[benaadams]

A non-GC struct return buffer can be cleared in the prolog even though its first defining call overwrites it completely. Recognize that call as explicit initialization using the existing return-buffer definition helper and reference counts. Retain initialization for GC-containing buffers, partial definitions, promoted fields, async methods, and potentially observable exception paths.

Split from #133833 at @jakobbotsch's request so this change can be reviewed independently.

Base: main at a652cdfed564.

Focused generated code

Windows x64 method sizes. Tier1 measurements use a repeated harness with tiering and dynamic PGO enabled; the harness waits for tier-up and captures Tier1 output. FullOpts measurements disable tiering for deterministic codegen checks.

Method	FullOpts bytes	Tier1 bytes
FullyDefined	112 -> 95	112 -> 95

Assembly example: FullyDefined

FullOpts, Windows x64. Instruction bytes are omitted; ... denotes unchanged assembly.

 ...
        push     rsi
        push     rbx
        sub      rsp, 56
-       vxorps   xmm4, xmm4, xmm4
-       vmovdqa  xmmword ptr [rsp+0x20], xmm4
-       xor      eax, eax
-       mov      qword ptr [rsp+0x30], rax
        mov      rbx, rcx
 G_M24874_IG02:
        lea      rcx, [rsp+0x20]
 ...
-; Total code size: 112 bytes
+; Total code size: 95 bytes

SuperPMI results

Windows x64 Checked JIT, eleven SuperPMI collections; identical settings on both sides and loop alignment disabled.

Comparable contexts	Before bytes	After bytes	Saved
983,234	238,457,491	238,455,292	2,199

131 contexts shrink; 7 grow by 36 bytes total; 5 equal-size assembly differences. Zero compilation failures. The existing recordings contain 983,424 contexts. Updated main lacks recorded answers in 190 contexts on both sides; those contexts are excluded from the table.

Seven contexts grow by 36 bytes in total through zeroing/layout changes. The two largest increases are 14 bytes each in C# LocalRewriter.MakeDecimalLiteral, with higher static PerfScore. These tradeoffs are included in the net figure.

Validation

• Windows x64 Checked and Release JIT builds; Linux x64 Release JIT build.
• Six existing physical-promotion regression files, run with the candidate JIT in normal, forced-promotion and JitStress=2 modes: 18 successful runs.
• Focused behavioral tests pass on base and candidate in optimized, forced-promotion, JitStress=2 and tiering/PGO modes.
• Exception, prior-read and partial-buffer tests retain .locals init while omitting explicit initialization stores. JIT dumps confirm the exception, reference-count and partial-definition guards are exercised. A GC-containing return-buffer test collects before returning; its caller retains GC-safe initialization.
• Focused tests also pass with Windows x64 GCStress=0xC; the Release test project builds successfully.
• Windows x64 codegen checks pass with the change and fail on its base.
• Linux x64 behavioral tests pass on both sides. The codegen assertions for this ABI-specific example are Windows-only.
• Changed-line clang-format 17.0.6 and git diff checks.

Release JIT throughput was measured with PIN instruction counts (candidate versus this PR's base):

Collection	Comparable contexts	Before JIT instructions	After JIT instructions	Change
benchmarks.run.windows.x64.checked.mch	51,617	87,397,498,731	87,366,103,140	-0.03592%
hashset-pgo-recollected.mch	1,307	558,875,294	558,775,175	-0.01791%

These small JIT instruction-count differences are not application-throughput measurements. Code size and static PerfScore also do not establish execution speed. No ARM64 or x86 execution is claimed.

[azure-pipelines]

Azure Pipelines:
Successfully started running 5 pipeline(s).
11 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]

🟡 Changes recommended

Regression coverage does not yet exercise the new safeguards, and the project reference ordering nit remains.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR removes redundant JIT prolog initialization for fully defined, non-GC struct return buffers while preserving safety cases.

Changes:

• Recognizes eligible return-buffer calls as explicit initialization.
• Preserves initialization for partial, GC-containing, promoted, asynchronous, and observable exception cases.
• Adds regression/codegen tests and registers the test project.

File summaries

File	Summary / review note
src/tests/JIT/Directed/physicalpromotion/ReturnBufferInitialization.csproj	Configures the return-buffer initialization tests.
src/tests/JIT/Directed/physicalpromotion/ReturnBufferInitialization.cs	Adds behavioral and codegen coverage; additional first-definition and GC-containing cases are needed to exercise the new safeguards.
src/tests/JIT/Directed/Directed_do.csproj	Registers the test project; its reference should be moved after readbackbeforeqmark to preserve ordering.
src/coreclr/jit/optimizer.cpp	Adds guarded recognition of eligible return-buffer initialization.

Review details

Suppressed comments (1)

src/tests/JIT/Directed/Directed_do.csproj:12

• The MergedWrapperProjectReference entries in this group are kept in path order (addressinterference through readbackbeforeqmark), but the new ReturnBufferInitialization entry is inserted before them. Please move it after readbackbeforeqmark so the list remains ordered.

    <MergedWrapperProjectReference Include="physicalpromotion\ReturnBufferInitialization.csproj" />

• Files reviewed: 4/4 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🔵 Needs a closer look

The change affects safety-sensitive JIT code generation and requires final human review.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 0 new
• Review effort level: Lite