JIT: only skip loop inversion when the IV test is the bottom test

[AndyAyersMS]

The bail-out added in #128459 fires for any BBJ_COND latch (or BBJ_COND predecessor of a canonical BBJ_ALWAYS latch) that exits the loop. That block may be an unrelated early-exit and not the loop's continuation test — for example, Dictionary<TKey,TValue>+Enumerator.MoveNext's body has an early-exit BBJ_COND (if (entry.next >= -1) return true;) while the actual loop continuation is the top-tested _index < _count check in the header. The current predicate falsely classifies that shape as "already bottom-tested" and skips a beneficial inversion.

Restrict the bail-out to the case where the recognized IV test (from AnalyzeIteration) is at that latch / predecessor position. AnalyzeIteration is called lazily so only loops that match the structural shape pay the cost.

Fixes the Dictionary regression reported in #128910

System.Collections.Generic.Dictionary2+Enumerator[int,int]:MoveNext` codegen:

	code size	instructions	PerfScore
Pre-#128459 (target)	129 bytes	43	105.38
Post-#128459 (regressed)	114 bytes	38	114.88
With this fix	129 bytes	43	105.38 ✓

BDN IterateForEach<Int32>.Dictionary(Size: 512) on AMD Zen 3 (3 interleaved rounds, mean of 3):

• current main: 501.9 ns
• with fix: 490.1 ns (−2.4%)

The Viper lab is Zen 4 where the regression is +6%; the fix restores byte-identical pre-#128459 codegen, so the lab perf should fully recover.

SPMI impact

Collection	Code size	PerfScore overall	PerfScore on diffs	TP (release)
benchmarks.run_pgo	+0.18%	−0.04%	−1.07%	+0.46%
aspnet2.run	+0.09%	+0.07%	−0.77%	+0.28%
libraries.pmi	+0.07%	−0.02%	−0.79%	+0.11%

~1,600 methods get inversion re-enabled. TP cost mostly comes from running the actual inversion phase on those methods; the lazy AnalyzeIteration call itself is a small fraction.

Original target of #128459

The Perf_Encoding regression that #128459 fixed (Utf8Utility.GetPointerToFirstInvalidByte on arm64) has a true bottom-tested IV-controlled loop where AnalyzeIteration recognizes the latch as the IV test, so this refined predicate still bails out there.

Note

This PR description was drafted with assistance from GitHub Copilot CLI.

[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 refines Compiler::optTryInvertWhileLoop’s “already bottom-tested” bail-out so it only triggers when the back-edge test corresponds to the IV test recognized by FlowGraphNaturalLoop::AnalyzeIteration, avoiding false positives from unrelated early-exit conditionals that happen to feed the back-edge.

Changes:

• Tighten the “skip inversion for bottom-tested loops” check to require the exiting BBJ_COND at the latch (or predecessor of a canonical BBJ_ALWAYS latch) to match NaturalLoopIterInfo::TestBlock.
• Add a lazy, cached AnalyzeIteration invocation so only candidate loops pay the analysis cost.
• Update dump messages/comments to reflect the refined criteria.

[AndyAyersMS]

@jakobbotsch PTAL
@dotnet/jit-contrib

A bit more fussing with loop inversion, trying to refine the criteria for invertible multi-exit loops better, and fix some regressions.