JIT: Refine promotion costing for cold small parameter fields

[jakobbotsch]

Withhold bitwise parameter-register extraction credit for byte/short fields whose weighted access count is less than 10% of the entry-block weight. Preserve direct register mappings and wider-field costing.

Many byte/short fields can fit into each parameter register, so eagerly extracting them can introduce substantial work and register pressure even when most fields are rarely accessed. Larger fields naturally limit the number of values that fit in a register and therefore the number of eager extractions. Restrict the heuristic to small fields to target this pattern.

This intentionally imprecise heuristic targets cases where eager field extraction is not profitable, such as Guid.CompareTo returning before its later fields are accessed. It does not precisely model extraction costs, register pressure, or shared spill/copy costs, and is not a general profitability model. Initialization of promoted fields remains unchanged.

Fix #119770

PriorityQueue<Guid, Guid> — fix-119770 vs baseline

Environment: Linux/x64, AMD Ryzen 9 5950X, Release JITs, tiered PGO enabled, ReadyToRun disabled. BenchmarkDotNet: two launches, eight warmups, fifteen measured iterations.

Benchmark	Size	Baseline	fix-119770	Ratio	BDN classification
HeapSort	10	186.2 ns	142.1 ns	0.76	Faster
Dequeue_And_Enqueue	10	559.2 ns	496.0 ns	0.89	Faster
K_Max_Elements	10	182.9 ns	158.3 ns	0.87	Faster
HeapSort	100	4.191 µs	2.785 µs	0.66	Faster
Dequeue_And_Enqueue	100	11.279 µs	9.234 µs	0.82	Faster
K_Max_Elements	100	846.0 ns	867.7 ns	1.03	Same
HeapSort	1000	83.540 µs	66.290 µs	0.79	Faster
Dequeue_And_Enqueue	1000	237.735 µs	212.093 µs	0.89	Faster
K_Max_Elements	1000	5.577 µs	4.592 µs	0.82	Faster

[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.

[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.

[jakobbotsch]

cc @dotnet/jit-contrib PTAL @EgorBo

Minor diffs. This is really just targeting the regression we saw in benchmarks, which has to do with eagerly extracting lots of fields from Guid in Guid.CompareTo when PGO data shows that they won't be used (because the first check returns immediately).

I have a better fix in mind for .NET 12 where we do these extractions lazily. #133690 is the first step towards this.

[Copilot]

🟡 Changes recommended

Add before/after evidence covering the reported regression and wider-field control cases.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Refines RyuJIT physical-promotion costing to avoid eagerly extracting rarely accessed small parameter fields.

Changes:

• Adds extraction-credit control for parameter-register mapping.
• Applies a 10% weighted-access threshold to cold byte/short fields.
• Preserves direct mappings and wider-field behavior.

File summaries

File	Summary
src/coreclr/jit/promotion.h	Extends the parameter-register mapping declaration.
src/coreclr/jit/promotion.cpp	Implements threshold-based extraction costing.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Lite

[jakobbotsch]

/backport to release/11.0

[github-actions]

Started backporting to release/11.0 (link to workflow run)