JIT: fix edge likelihoods for loop cloning

[AndyAyersMS]

We were over-estimating the edge likelihoods leading to the fast clone, causing profiles to inflate. The old code was not aware of the total number of checks that need to pass to reach the fast clone, so each edge's likelihood was slightly too high.

This shows up prominently when we clone a set of nested loops.

Fix by counting how many conditional branches we must pass through before reaching the fast clone, and use that to compute the proper likelihoods.

[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 adjusts how loop cloning assigns edge likelihoods for the fast-path gating condition chain so that the cumulative probability of reaching the fast clone matches the intended fastPathWeightScaleFactor, even when the condition chain is constructed across multiple CondToStmtInBlock calls (e.g., block/deref conditions plus cloning conditions).

Changes:

• Extend LoopCloneContext::CondToStmtInBlock to accept totalCondsInChain and use it to compute the per-conditional likelihood (Nth-root model) against the full chain length.
• In Compiler::optInsertLoopChoiceConditions, pre-count total condition blocks to be inserted and pass that count to each CondToStmtInBlock call.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/coreclr/jit/loopcloning.h	Updates the CondToStmtInBlock declaration and documents the new totalCondsInChain parameter.
src/coreclr/jit/loopcloning.cpp	Uses the full chain condition count to compute per-branch likelihoods; counts and threads the total through optInsertLoopChoiceConditions.

[AndyAyersMS]

@jakobbotsch PTAL
fyi @dotnet/jit-contrib

Definitely affects perf scores in some cases... still evaluating diffs, but what we were doing before was wrong.

benchmarks.run_pgo.windows has only ~200 code diffs locally, but −301,305,052 (−41.67% of base) perf score diff. Diffs look mostly to be layout changes.

[AndyAyersMS]

diffs