Transform STRUCT-typed uses of primitives in local morph

[SingleAccretion]

And update global morph accordingly. This mainly involves not DNER-ing struct local fields right away, but waiting for global morph to do this, after it has had a chance to transform the field into an LCL_VAR.

Diffs are a bit mixed, due to requiring the substitution workaround but generally positive, and overall minor. The larger regressions in tests are due to us setting some DNERs later, thus expanding things into field-by-field copies more.

Note: this change is best reviewed with the "don't show whitespace diffs" setting.

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

And update global morph accordingly. This mainly involves giving not DNER-ing local fields right away, but waiting for global morph to do this, after it has had a chance to transform the field into an LCL_VAR.

Diffs are a bit mixed, due to requiring the substitution workaround but generally positive, and overall minor.

Author:	SingleAccretion
Assignees:	-
Labels:	area-CodeGen-coreclr, community-contribution
Milestone:	-

[SingleAccretion]

This now depends on #76491.

[SingleAccretion]

This is ready for review, but still depends on #76491.

@dotnet/jit-contrib

[BruceForstall]

@SingleAccretion Needs conflict resolution

[SingleAccretion]

Needs conflict resolution

Also done.

[jakobbotsch]

/azp run runtime-coreclr superpmi-diffs, runtime-coreclr jitstress, runtime-coreclr libraries-jitstress

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[SingleAccretion]

Libraries stress failures are #78909 and #78912.

[SingleAccretion]

Stress failures:

1. TypeGeneratorTests400-499: https://helixre107v0xdeko0k025g8.blob.core.windows.net/dotnet-runtime-refs-pull-78131-merge-7d3e4ae48c114120a0/TypeGeneratorTests400-499/1/console.0f672ed5.log?helixlogtype=result - looks to have passed actually?

12:46:58.844 Passed test: Loader/classloader/TypeGeneratorTests/TypeGeneratorTest499/Generated499/Generated499.dll
Expected: 100
Actual: 100
END EXECUTION - PASSED
+ export _commandExitCode=0
+ /usr/bin/python3 /datadisks/disk1/work/AF3D0989/p/reporter/run.py https://dev.azure.com/dnceng-public/ public 2032048 eyJ0eXAiOiJKV1QiLCJhbGciOiJSUzI1NiIsIng1dCI6Im9PdmN6NU1fN3AtSGpJS2xGWHo5M3VfVjBabyJ9.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.Sll8uONVwy9ZF9KTir7KbsJEDJ1UOe-vCIk4fL6oQfD52TnPQk082mPT1YKXwTrUuE0xSxNS7ehaYTSH5CGUN-RVDJDkgiJauEH6CLZvnNQrKJySShBowW7azSX7X_XYcpOT8-XafcEKFr_x1r-VUTgbl_GviHcFCTCv-KnHHOqIsj1Ma8DJRUzX3Lax12F2coGlh-ogqt8vm566VBsgTK6lUmJJuMXL3FceWGvVJUnpjdLYdtYcMIvU3s8o2_CE5zIWeVNH7GeqDEAW4afcydK9bQTmDxfffsWiT3zlhp0BO3BGlEZPrg5by-JVspcr30yakIn5YbrYLWmns4vmNg
Killed
['TypeGeneratorTests400-499' END OF WORK ITEM LOG: Command timed out, and was killed]

2. JIT\\Regression\\JitBlue\\Runtime_63354\\Runtime_63354\\Runtime_63354.cmd, x86: Test failure JIT\\Regression\\JitBlue\\Runtime_63354\\Runtime_63354\\Runtime_63354.cmd #78898.

[jakobbotsch]

Looks great to me, very nice to no longer have to think about these cases after local morph.
Are you going to optimize other uses (e.g. the liveness change we discussed) in a follow-up?

[SingleAccretion]

Are you going to optimize other uses (e.g. the liveness change we discussed) in a follow-up?

Yep.

[JulieLeeMSFT]

Hi @SingleAccretion, we are trying to clean up TODO comments. You touched this part of the code in this PR,

runtime/src/coreclr/jit/morph.cpp

Lines 11154 to 11167 in 7ff684a

	// TODO: support `genReturnBB != nullptr`, it requires #11413 to avoid `Incompatible types for
	// gtNewTempStore`.
	if (canFold && (genReturnBB == nullptr))
	{
	// Fold even if types do not match, lowering will handle it. This allows the local
	// to remain DNER-free and be enregistered.
	assert(lclFld->GetLclOffs() == 0);
	lclFld->ChangeType(varDsc->TypeGet());
	lclFld->SetOper(GT_LCL_VAR);
	}
	else if (!varDsc->lvDoNotEnregister)
	{
	lvaSetVarDoNotEnregister(lclNum DEBUGARG(DoNotEnregisterReason::BlockOpRet));
	}

. Our team members do not have much context on it, so I wanted to ask you if you know what to do to support genReturnBB != nullptr.
#11413 is now supported.

        // TODO: support `genReturnBB != nullptr`, it requires #11413 to avoid `Incompatible types for
        // gtNewTempStore`.
        if (canFold && (genReturnBB == nullptr))
        {
            // Fold even if types do not match, lowering will handle it. This allows the local
            // to remain DNER-free and be enregistered.
            assert(lclFld->GetLclOffs() == 0);
            lclFld->ChangeType(varDsc->TypeGet());
            lclFld->SetOper(GT_LCL_VAR);
        }
        else if (!varDsc->lvDoNotEnregister)
        {
            lvaSetVarDoNotEnregister(lclNum DEBUGARG(DoNotEnregisterReason::BlockOpRet));
        }

[JulieLeeMSFT]

BTW, the bitconverter instructions from #11413 does not hit this code path.

[SingleAccretion]

I wanted to ask you if you know what to do to support genReturnBB != nullptr

@JulieLeeMSFT it requires removing the property that allows "ABI-compatible" types under a GT_RETURN node in HIR. What the comment is talking about is that if you have something like:

struct Integer { int Value; }

Integer GetMyInt() {
  if (...)
     return new Integer(1); // A primitive - we can retype that
  else
     return GetAnotherInt(); // A call that return a primitive struct - we don't/can't retype that.
}

If this could be transformed by morph into something like:

if (...)
  GT_RETURN<struct>(CNS_INT<int>(1))
else
  GT_RETURN<struct>(CALL<struct>())

The common return local path would get mismatched types:

if (...)
  getReturnLcl<struct> = CNS_INT<int>(1); // Type mismatch!
else
  getReturnLcl<struct> = CALL<struct>();

GT_RETURN<struct>(getReturnLcl);

It is why this optimization is only done when we don't have a common return local - genReturnBB == nullptr.

[JulieLeeMSFT]

@JulieLeeMSFT it requires removing the property that allows "ABI-compatible" types under a GT_RETURN node in HIR. What the comment is talking about is that if you have something like:

Thanks @SingleAccretion for a quick response. I will take a look.

[jakobbotsch]

@JulieLeeMSFT This won't be a simple fix. It overlaps with the work I need to do to enable better ABI treatment with physical promotion, and hopefully I can get to that by .NET 10.
Basically, we can consider this TODO to be tracked by #86665. Feel free to remove the "TODO" part of the comment from the code base now that #86665 calls this scenario out explicitly.

[JulieLeeMSFT]

Sounds good. I will do that.