JIT: Push initial defs and create initial def VNs for all tracked locals

[AndyAyersMS]

Have SSA create initial SSA defs for all locals instead of just the live subset (most of the time these will end up being unused). Have VN give these initial SSA defs suitable initial value numbers.

This fixes the "missing ssa def" issue seen in #105667 (where the only use of a local is an USEASG which does not make the use live) and also allows for a bit more VN based dead stores.

We could claw back some of the TP by only creating these unused things on demand.

Fixes #105667.

[AndyAyersMS]

FYI @dotnet/jit-contrib
@jakobbotsch PTAL

[AndyAyersMS]

I wonder if at some point VN dead stores could supplant the redundant zero inits phase....

[SingleAccretion]

This (presumably) obsoletes this workaround:

runtime/src/coreclr/jit/liveness.cpp

Lines 1194 to 1214 in ab03e0f

	if (isUse && !storeRemoved)
	{
	// SSA and VN treat "partial definitions" as true uses, so for this
	// front-end liveness pass we must add them to the live set in case
	// we failed to remove the dead store.
	if (varDsc->lvTracked)
	{
	VarSetOps::AddElemD(this, life, varDsc->lvVarIndex);
	}
	if (varDsc->lvPromoted)
	{
	for (unsigned fieldIndex = 0; fieldIndex < varDsc->lvFieldCnt; fieldIndex++)
	{
	LclVarDsc* fieldVarDsc = lvaGetDesc(varDsc->lvFieldLclStart + fieldIndex);
	if (fieldVarDsc->lvTracked)
	{
	VarSetOps::AddElemD(this, life, fieldVarDsc->lvVarIndex);
	}
	}
	}
	}

[AndyAyersMS]

/azp run runtime-coreclr jitstress

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[AndyAyersMS]

This (presumably) obsoletes this workaround:

Ah, that likely explains why the issue in #105667 wasn't more widespread -- partial defs by calls were bypassed in this special liveness update.

So we could extend that handling to cover calls too, as an alternative (and perhaps less risky) fix.

[jakobbotsch]

Ah, that likely explains why the issue in #105667 wasn't more widespread -- partial defs by calls were bypassed in this special liveness update.

So we could extend that handling to cover calls too, as an alternative (and perhaps less risky) fix.

Are you planning to change to handle it there?

[AndyAyersMS]

Ah, that likely explains why the issue in #105667 wasn't more widespread -- partial defs by calls were bypassed in this special liveness update.
So we could extend that handling to cover calls too, as an alternative (and perhaps less risky) fix.

Are you planning to change to handle it there?

Seems like a more surgical fix, so yeah let me do that.

We can consider this change for .NET 10.

[AndyAyersMS]

Going to hold off on this for now...

[AndyAyersMS]

Closing for now, may want to revisit in .NET 11.