JIT: Fix for IV opt to preserve updates to a counter that is live into an EH handler

[JulieLeeMSFT]

optLocalHasNonLoopUses relied on lvDoNotEnregister to detect locals live into exceptional exits. PR #127932 decoupled DNER from liveness (DNER is now set during LSRA/lowering, after optInductionVariables runs), so an induction variable live into a catch handler was no longer detected and its in-loop self-update was removed by optRemoveUnusedIVs, producing wrong results under full-opt codegen (crossgen2, jitstress, AggressiveOptimization).

Add an explicit lvTracked && IsLiveInOutOfHandler() check, since VisitRegularExitBlocks deliberately excludes handler blocks.

Fixes #128392.

[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 fixes a correctness hole in CoreCLR JIT induction-variable optimizations where a loop IV update could be removed even though the IV is live into an EH handler (e.g., used by a catch). The change makes optLocalHasNonLoopUses explicitly treat “live into handler” as a non-loop use, preventing optRemoveUnusedIVs from dropping the in-loop update.

Changes:

• Update optLocalHasNonLoopUses to conservatively return true when a tracked local is IsLiveInOutOfHandler().
• Adjust a comment in IV widening logic to reflect that exceptional-exit-live IVs are not profitable candidates (rather than relying on DNER being set).
• Add a JIT regression test (Runtime_128392) and include it in the regression csproj.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/coreclr/jit/inductionvariableopts.cpp	Adds an explicit EH-handler liveness check in optLocalHasNonLoopUses to avoid removing IV updates that are observable via exceptional control flow.
src/tests/JIT/Regression/JitBlue/Runtime_128392/Runtime_128392.cs	New xUnit regression covering an IV read in a catch after an exception thrown mid-loop.
src/tests/JIT/Regression/Regression_ro_2.csproj	Includes the new regression test source file in the build.

[AndyAyersMS]

Is this 11.0 or 10.0?

[jakobbotsch]

Is this 11.0 or 10.0?

This is a regression from #127932

[jakobbotsch]

I think this check should just be removed.

[JulieLeeMSFT]

@copilot, please remove this check if (varDsc->lvDoNotEnregister).

[jakobbotsch]

Can you introduce a similar assert as

runtime/src/coreclr/jit/inductionvariableopts.cpp

Lines 471 to 489 in 01cd5cc

	#ifdef DEBUG
	// We currently do not expect to ever widen IVs that are live into
	// exceptional exits. Such IVs are not currently register candidates (EH
	// write-thru is only for single def locals) which makes it unprofitable.
	// If this ever changes we need some more expansive handling here.
	loop->VisitLoopBlocks([=](BasicBlock* block) {
	block->VisitAllSuccs(this, [=](BasicBlock* succ) {
	if (!loop->ContainsBlock(succ) && bbIsHandlerBeg(succ))
	{
	assert(!optLocalIsLiveIntoBlock(lclNum, succ) &&
	"Candidate IV for widening is live into exceptional exit");
	}
	return BasicBlockVisit::Continue;
	});
	return BasicBlockVisit::Continue;
	});
	#endif

at the end of optLocalHasNonLoopUses?

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[JulieLeeMSFT]

#ifdef DEBUG
// We currently do not expect to ever widen IVs that are live into
// exceptional exits. Such IVs are not currently register candidates (EH
// write-thru is only for single def locals) which makes it unprofitable.
// If this ever changes we need some more expansive handling here.
loop->VisitLoopBlocks([=](BasicBlock* block) {
block->VisitAllSuccs(this, [=](BasicBlock* succ) {
if (!loop->ContainsBlock(succ) && bbIsHandlerBeg(succ))
{
assert(!optLocalIsLiveIntoBlock(lclNum, succ) &&
"Candidate IV for widening is live into exceptional exit");
}

         return BasicBlockVisit::Continue; 
     }); 

     return BasicBlockVisit::Continue; 
 }); 

#endif

@copilot, please add this code to the end of optLocalHasNonLoopUses before return false;.