JIT: Fix truncated histogram likelihoods

[AndyAyersMS]

Fixes #133973

Truncated class, method, and value histograms transferred all omitted
probability to the first returned entry. This artificially inflated its
likelihood and could trigger unjustified guarded devirtualization or constant
specialization.

Only normalize rounding error when all histogram entries are returned.

Validation:

• Checked JIT build and Release clr+libs build
• Issue repro and end-to-end value-profile validation
• Windows x64 SuperPMI: 251 diffs across 3 of 9 collections, no failing
  compiles, and 7,773 fewer bytes overall

Performance:

• Periodic uniform dispatch: 4.677 ns vs. 4.411 ns baseline (+6%)
• Shuffled uniform dispatch: 4.646 ns vs. 4.805 ns baseline (-3%)

The differing results reflect removal of an invalid speculative guard. This
change corrects profile probabilities rather than promising a uniform
throughput improvement.

Note

This pull request description was generated with GitHub Copilot.

[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

Add value-histogram regression coverage and correct the shared class/method documentation.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Fixes JIT PGO histogram likelihood normalization so truncated histograms no longer inflate the first result.

Changes:

• Normalize only complete class, method, and value histograms.
• Add class and method histogram regression tests.
• Update helper documentation and test configuration.

File summaries

File	Summary
src/tests/JIT/PGO/LikelyClass/LikelyClass.csproj	Configures the JIT regression test.
src/tests/JIT/PGO/LikelyClass/LikelyClass.cs	Tests truncated, complete, and unknown-handle class/method histograms.
src/coreclr/jit/likelyclass.cpp	Corrects normalization and updates documentation; value-histogram regression coverage and neutral class/method wording remain needed.

Review details

Suppressed comments (1)

src/coreclr/jit/likelyclass.cpp:328

• This shared return-value documentation is also used by getLikelyMethods (see the “See documentation on getLikelyClasses above” note below), so describing the result as class records written to pLikelyClasses is inaccurate for method histograms. Use neutral class/method wording so the contract is correct for both exported helpers.

//    Number of likely class records written to pLikelyClasses

• Files reviewed: 3/3 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🟡 Changes recommended

Add deterministic regression coverage for truncated class/method and value histograms.

Get a fresh assessment by requesting another Copilot review.

Review details

Suppressed comments (1)

src/coreclr/jit/likelyclass.cpp:453

• The value-histogram branch is also changed for the independently important case where more than eight constants are observed, but no regression test is included for profile-driven constant specialization with truncated data. Please add deterministic coverage (or a direct getLikelyValues test) that verifies the top value keeps its raw likelihood and does not trigger specialization after the truncation fix.

            if (numberOfLikelyConst == h.countHistogramElements)

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

[AndyAyersMS]

@EgorBo PTAL
fyi @dotnet/jit-contrib

A couple hundred SPMI diffs, mostly cases where we no longer act on profile data since it is no longer artificially inflated.