Fix perf regressions with a recent change.

[sandreenko]

#54998 introduced a perf regression that this PR is fixing.

Thanks to @kunalspathak for catching and @EgorBo for repro help.

The regression was happening only when we switched from MinOpts to FullOpts and it was happening because of Tier0 QuickJitForLoops=disabled, so SPMI did not reproduce this scenario so it was not catched in my local testing.

We still want to have doNotEnreg flag to be known as soon as possible but with this change, we just add two additional iterations to set it for minopts in comparasing with the previous version that was trying to init it when we created a lclVar.

Why can't we init it when we create a lclVar? It is because at this time we don't know the final optimization level yet and we can't reset doNotEnreg flag after the optimization flag is finalized.

No SPMI diffs as expected but CoreRun with Tiering=1 shows improvements on methods with loops.

[sandreenko]

PTAL @kunalspathak @EgorBo @dotnet/jit-contrib

[kunalspathak]

Did you try running some BDN benchmarks that regressed and were able to see the gains come back?

[kunalspathak]

I was hoping to see MinOpts check here...why is it not there?

[sandreenko]

The comment was poorly written, what I was trying to explain here:

1. the main flag is (opts.compFlags & CLFLG_REGVAR), if it is not set we can't enregister variables;

2. right now it is set if we use CLFLG_MAXOPT and not set when using CLFLG_MINOPT, an example where we set it:
   

   runtime/src/coreclr/jit/compiler.cpp

   Lines 2615 to 2620 in 7f88911

   	// Don't optimize .cctors (except prejit) or if we're an inlinee
   	else if (!jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT) && ((info.compFlags & FLG_CCTOR) == FLG_CCTOR) &&
   	!compIsForInlining())
   	{
   	opts.compFlags = CLFLG_MINOPT;
   	}

3. in theory nothing prevents us to disable CLFLG_REGVAR when MinOpts is not set, or enable it when it is set, so I do not want to check isMinOpts instead of !compEnregLocals() here.

I have updated the comments to reflect that, please take a look.

[sandreenko]

Did you try running some BDN benchmarks that regressed and were able to see the gains come back?

Yes, they are back to the previous values.

[sandreenko]

the failures are unrelated.

[kunalspathak]

LGTM