Mark vars as do not enreg earlier in minopts.

[sandreenko]

and improve morph block init/copy a little bit.

There is a known problem that we start asking is it already known to be not on register? before the value is finalized during morph (block copy, should we do field by field) and lowering (do we need bitcast? can we mark the source as contained?) .
I have an markdown description of it in 6.0 planning.

For now, just mark them as doNotEnreg earlier when enregistration is disabled. Give us small improvements on minopts methods and prevent some regressions with future changes.

coreclr_tests.pmi.windows.x64.checked.mch
Total bytes of delta: -4
1 total methods with Code Size differences (1 improved, 0 regressed)

coreclr_tests.pmi.Linux.arm64.checked.mch
Total bytes of delta: -8
2 total methods with Code Size differences (2 improved, 0 regressed)

libraries_tests.pmi.Linux.arm64.checked.mch
Total bytes of delta: -28
5 total methods with Code Size differences (5 improved, 0 regressed)

benchmarks.run.Linux.x64.checked.mch
Total bytes of delta: -3
1 total methods with Code Size differences (1 improved, 0 regressed)

coreclr_tests.pmi.Linux.x64.checked.mch
Total bytes of delta: -4 
1 total methods with Code Size differences (1 improved, 0 regressed)

benchmarks.run.windows.arm64.checked.mch
Total bytes of delta: -8
1 total methods with Code Size differences (1 improved, 0 regressed)

coreclr_tests.pmi.windows.arm64.checked.mch
Total bytes of delta: -8
2 total methods with Code Size differences (2 improved, 0 regressed)

libraries_tests.pmi.windows.arm64.checked.mch
Total bytes of delta: -28
5 total methods with Code Size differences (5 improved, 0 regressed)

coreclr_tests.pmi.windows.x86.checked.mch
Total bytes of delta: -4
1 total methods with Code Size differences (1 improved, 0 regressed)

+no diffs improvements of morph block init/copy.

[sandreenko]

/azp run runtime-coreclr jitstress, runtime-coreclr outerloop

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[sandreenko]

after this change init/copy like:

           [000015] -ACXG-------              *  ASG       struct (copy)
           [000013] D----+-N----              +--*  LCL_VAR   struct<S, 8> V01 loc0
           [000012] --CXG+------              \--*  CALL      struct TestStructFields.Program.GetS1

don't require doNotEnreg flag. The condition was moved after we do m_comp->fgMorphBlockOperand to work after OBJ(ADDR(LCL)) folding.

[sandreenko]

the idea of if/else was that each block was setting lvaSetVarDoNotEnregister and there was no need to check the rest, now we have blocks that don't set it, so we need to check each.

[sandreenko]

PTAL @BruceForstall @dotnet/jit-contrib , another preparation for structs in registers, small changes and diffs.

[kunalspathak]

@sandreenko - We noticed several regressions possibly related to this change. Do you mind going through some of them to confirm if they are indeed from this change?

• [Perf] Changes at 6/29/2021 8:15:47 PM DrewScoggins/performance-2#7139
• [Perf] Changes at 7/1/2021 4:09:06 PM DrewScoggins/performance-2#7142
• [Perf] Changes at 7/1/2021 7:08:30 PM DrewScoggins/performance-2#7143
• [Perf] Changes at 7/1/2021 7:08:30 PM DrewScoggins/performance-2#7194

[EgorBo]

Simplest repro:

cd src\benchmarks\micro
dotnet run -c Release -f net5.0 -- --coreRun C:\prj\runtime\artifacts\tests\coreclr\windows.x64.Release\Tests\Core_Root\corerun.exe --filter "*IniArray.Test*" 

It prints around 179ms, then if I revert changes from this PR by hands and re-run - it prints 45ms. The benchmark itself seems to have the same codegen so I have no clue what's going on...

[sandreenko]

@sandreenko - We noticed several regressions possibly related to this change. Do you mind going through some of them to confirm if they are indeed from this change?

Sure, looking at them now, thanks for reporting it.

[sandreenko]

dotnet run -c Release -f net5.0 -- --coreRun C:\prj\runtime\artifacts\tests\coreclr\windows.x64.Release\Tests\Core_Root\corerun.exe --filter "*IniArray.Test*" 

Hm, don't see this:

D:\Sergey\git\runtime\.dotnet\dotnet.exe run -c Release -f net5.0 --  --cli D:\Sergey\git\runtime\.dotnet\dotnet.exe  --filter "*IniArray.Test*" --coreRun "D:\Sergey\git\runtime\artifacts\tests\coreclr\windows.x64.Release\Tests\Core_Root_base\corerun.exe"

|   Method |     Mean |    Error |   StdDev |   Median |      Min |      Max | Gen 0 | Gen 1 | Gen 2 | Allocated |
|--------- |---------:|---------:|---------:|---------:|---------:|---------:|------:|------:|------:|----------:|
| IniArray | 42.86 ms | 0.826 ms | 0.690 ms | 42.56 ms | 42.33 ms | 44.75 ms |     - |     - |     - |      88 B |

D:\Sergey\git\runtime\.dotnet\dotnet.exe run -c Release -f net5.0 --  --cli D:\Sergey\git\runtime\.dotnet\dotnet.exe  --filter "*IniArray.Test*" --coreRun "D:\Sergey\git\runtime\artifacts\tests\coreclr\windows.x64.Release\Tests\Core_Root_diff\corerun.exe"
|   Method |     Mean |    Error |   StdDev |   Median |      Min |      Max | Gen 0 | Gen 1 | Gen 2 | Allocated |
|--------- |---------:|---------:|---------:|---------:|---------:|---------:|------:|------:|------:|----------:|
| IniArray | 42.57 ms | 0.175 ms | 0.136 ms | 42.58 ms | 42.35 ms | 42.80 ms |     - |     - |     - |      88 B |

I will check other regressions.

[EgorBo]

@sandreenko I've just checked on a clean repo again - still reproduces for me

[EgorBo]

@sandreenko managed to get the diff between two coreruns: https://www.diffchecker.com/qh9EOTD4

[sandreenko]

@sandreenko managed to get the diff between two coreruns: https://www.diffchecker.com/qh9EOTD4

Thank you, now I see them, I had a system-wide tiering disabled that I forgot about.