Check more patterns for retBuffer.

[sandreenko]

We force return buffers containing GC pointers to be on the stack and before we make a copy we check if the return buffer is already on the stack. Fix the check by adding LCL_VAR_ADDR, ADD(LCL_VAR_ADDR, CNST) patterns.

Fixes #39947, however, the issue happens because of the several factors:

1. EnableExtraSuperPmiQueries is a debug only variable that is set during collection;
2. debug only makeExtraStructQueries is called in checked and it could find a SIMD type when SIMD intrinsics are not used, it would call setUsesSIMDTypes(true) {_usesSIMDTypes = true; } in compiler;
3. based on the value of _usesSIMDTypes struct promotion could make different decisions, for example, promote struct A {simd16 a; };
4. lclmorph doesn't replace ADDR(LCL_VAR) as LCL_VAR_ADDR if the lclVar is promoted;
5. LCL_VAR_ADDR was not recognized by the function changed in this PR and it led to additional VM questions that were failing during release replay.

This PR fixes only the 5th. We could fix the first by:

1. disable EnableExtraSuperPmiQueries until we need it again;
2. don't record EnableExtraSuperPmiQueries value during collection (hack void MethodContext::recGetIntConfigValue(const WCHAR* name, int defaultValue, int result));
3. pass -jitoption force EnableExtraSuperPmiQueries=0 -jit2option force EnableExtraSuperPmiQueries=0 during SPMI replay;
4. do less in makeExtraStructQueries.

The rest SPMI chk/rel diffs are covered by #39908

[sandreenko]

PTAL @CarolEidt @BruceForstall @dotnet/jit-contrib

[CarolEidt]

LCL_VAR_ADDR was not recognized by the function changed in that PR

What PR are you referring to?

Regarding how to deal with EnableExtraSuperPmiQueries, I'd prefer option 3: pass -jitoption force EnableExtraSuperPmiQueries=0 -jit2option force EnableExtraSuperPmiQueries=0 during SPMI replay for checked/release diffs. But I'm flexible.

[CarolEidt]

LGTM

[sandreenko]

LCL_VAR_ADDR was not recognized by the function changed in that PR

What PR are you referring to?

This one, typo, sorry for the confusion.

Regarding how to deal with EnableExtraSuperPmiQueries, I'd prefer option 3: pass -jitoption force EnableExtraSuperPmiQueries=0 -jit2option force EnableExtraSuperPmiQueries=0 during SPMI replay for checked/release diffs. But I'm flexible.

Agree, I think it is safe and easy. Maybe we will lose some methods during SPMI cleaning phase but the result *.mch files should not show any diffs between chk/rel because of EnableExtraSuperPmiQueries. I will push the change to this PR.

[BruceForstall]

While forcing -jitoption force EnableExtraSuperPmiQueries=0 -jit2option force EnableExtraSuperPmiQueries=0 during replay seems expedient, it's also unsettling that enabling extra SuperPMI queries during collection can actually affect codegen. Presumably it could also cause the subsequent questions in the normal codegen to differ, leading to missing questions/answers on replay when NOT setting it.

I wonder if it would be possible to move the "extra" calls to the very end of JITing such that any potential side effects couldn't affect the already-generated code? That is, assuming these calls could even legally be made at that time.

[sandreenko]

Presumably it could also cause the subsequent questions in the normal codegen to differ, leading to missing questions/answers on replay when NOT setting it.

Yeah, as I said with this approach we could have some new failures during a cleaning stage and we will have fewer methods in the final mch.