Improvements for dead store removal.

[erozenfeld]

Improvements for dead store removal.

1. Don't mark fields of dependently promoted structs as untracked.

2. Remove some stores whose lhs local has a ref count of 1 when running
   late liveness. We can rely on ref counts since they are calculated right
   before the late liveness pass.

3. Remove dead GT_STORE_BLK in addition to GT_STOREIND in the late
   liveness pass.

4. Remove dead stores to untracked locals in the late liveness pass.

5. Allow optRemoveRedundantZeroInits to remove some redundant initializations
   of tracked locals. Move the phase to right after liveness so that SSA is correct
   after removing assignments to tracked locals.

[erozenfeld]

Framework x64 pmi diffs:

PMI CodeSize Diffs for System.Private.CoreLib.dll, framework assemblies for  default jit
Summary of Code Size diffs:
(Lower is better)
Total bytes of diff: -17783 (-0.04% of base)
    diff is an improvement.
Top file improvements (bytes):
       -9586 : Microsoft.Diagnostics.Tracing.TraceEvent.dasm (-0.31% of base)
       -2727 : Microsoft.CodeAnalysis.VisualBasic.dasm (-0.05% of base)
        -985 : Microsoft.CodeAnalysis.CSharp.dasm (-0.02% of base)
        -898 : System.Data.Common.dasm (-0.06% of base)
        -502 : Microsoft.CodeAnalysis.dasm (-0.03% of base)
        -457 : System.Private.CoreLib.dasm (-0.01% of base)
        -451 : System.Reflection.Metadata.dasm (-0.11% of base)
        -404 : Newtonsoft.Json.dasm (-0.05% of base)
        -359 : System.Threading.Tasks.Dataflow.dasm (-0.04% of base)
        -212 : System.Drawing.Primitives.dasm (-0.57% of base)
        -149 : System.Reflection.MetadataLoadContext.dasm (-0.08% of base)
        -140 : ILCompiler.Reflection.ReadyToRun.dasm (-0.10% of base)
        -102 : System.Linq.dasm (-0.01% of base)
         -71 : Newtonsoft.Json.Bson.dasm (-0.07% of base)
         -70 : System.Net.Http.dasm (-0.01% of base)
         -69 : System.Diagnostics.EventLog.dasm (-0.08% of base)
         -61 : System.Drawing.Common.dasm (-0.02% of base)
         -56 : System.Private.Xml.Linq.dasm (-0.03% of base)
         -45 : System.Linq.Parallel.dasm (-0.00% of base)
         -42 : System.Net.Security.dasm (-0.03% of base)
45 total files with Code Size differences (45 improved, 0 regressed), 219 unchanged.
Top method improvements (bytes):
       -2623 (-8.07% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - ClrPrivateTraceEventParser:EnumerateTemplates(Func`3,Action`1):this
       -2607 (-8.81% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - ClrTraceEventParser:EnumerateTemplates(Func`3,Action`1):this
       -2195 (-4.62% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - KernelTraceEventParser:EnumerateTemplates(Func`3,Action`1):this
       -2193 (-17.99% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - BoundTreeVisitor`2:VisitInternal(BoundNode,Vector`1):long:this
       -1761 (-2.67% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - CtfTraceEventSource:InitEventMap():Dictionary`2
        -279 (-0.25% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - ApplicationServerTraceEventParser:EnumerateTemplates(Func`3,Action`1):this
        -245 (-19.65% of base) : Microsoft.CodeAnalysis.CSharp.dasm - BinopEasyOut:TypeToIndex(TypeSymbol):Nullable`1
        -245 (-19.65% of base) : Microsoft.CodeAnalysis.CSharp.dasm - UnopEasyOut:TypeToIndex(TypeSymbol):Nullable`1
        -140 (-2.40% of base) : System.Data.Common.dasm - UnboxT`1:NullableField(Object):Nullable`1 (42 methods)
        -121 (-7.02% of base) : Microsoft.CodeAnalysis.dasm - SyntaxDiffer:GetNextAction():DiffAction:this
        -120 (-1.48% of base) : Newtonsoft.Json.dasm - JToken:op_Explicit(JToken):Nullable`1 (17 methods)
        -119 (-23.11% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - SyntaxRemover:AddEndOfLine():this
         -96 (-0.94% of base) : System.Threading.Tasks.Dataflow.dasm - BatchBlockTargetCore:RetrievePostponedItemsGreedyBounded(bool):this (7 methods)
         -93 (-0.83% of base) : System.Threading.Tasks.Dataflow.dasm - BatchBlockTargetCore:RetrievePostponedItemsNonGreedy(bool):this (7 methods)
         -85 (-1.19% of base) : System.Threading.Tasks.Dataflow.dasm - BatchBlockTargetCore:ConsumeReservedMessagesNonGreedy():this (7 methods)
         -85 (-1.45% of base) : System.Threading.Tasks.Dataflow.dasm - BatchBlockTargetCore:ConsumeReservedMessagesGreedyBounded():this (7 methods)
         -84 (-0.65% of base) : System.Linq.dasm - Enumerable:Sum(IEnumerable`1,Func`2):Nullable`1 (35 methods)
         -81 (-31.15% of base) : Microsoft.CodeAnalysis.CSharp.dasm - <>c:<HasEndOfLine>b__12_0(SyntaxTrivia):bool:this
         -81 (-31.64% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - _Closure$__:_Lambda$__12-0(SyntaxTrivia):bool:this
         -74 (-15.98% of base) : Microsoft.CodeAnalysis.CSharp.dasm - SyntaxRemover:AddEndOfLine():this
Top method improvements (percentages):
         -81 (-31.64% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - _Closure$__:_Lambda$__12-0(SyntaxTrivia):bool:this
         -81 (-31.15% of base) : Microsoft.CodeAnalysis.CSharp.dasm - <>c:<HasEndOfLine>b__12_0(SyntaxTrivia):bool:this
          -6 (-26.09% of base) : Microsoft.CodeAnalysis.dasm - SubsystemVersion:Create(int,int):SubsystemVersion
          -6 (-25.00% of base) : Microsoft.CodeAnalysis.dasm - Optional`1:op_Implicit(int):Optional`1
          -6 (-25.00% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - DimensionSize:ConstantSize(int):DimensionSize
          -6 (-25.00% of base) : System.Data.Common.dasm - SqlInt32:op_Implicit(int):SqlInt32
          -6 (-25.00% of base) : System.Private.CoreLib.dasm - Nullable`1:op_Implicit(int):Nullable`1
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - AssemblyDefinitionHandle:op_Implicit(AssemblyDefinitionHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - InterfaceImplementationHandle:op_Implicit(InterfaceImplementationHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - MethodDefinitionHandle:op_Implicit(MethodDefinitionHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - MethodImplementationHandle:op_Implicit(MethodImplementationHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - MethodSpecificationHandle:op_Implicit(MethodSpecificationHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - TypeDefinitionHandle:op_Implicit(TypeDefinitionHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - ExportedTypeHandle:op_Implicit(ExportedTypeHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - TypeReferenceHandle:op_Implicit(TypeReferenceHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - TypeSpecificationHandle:op_Implicit(TypeSpecificationHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - MemberReferenceHandle:op_Implicit(MemberReferenceHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - FieldDefinitionHandle:op_Implicit(FieldDefinitionHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - EventDefinitionHandle:op_Implicit(EventDefinitionHandle):Handle
          -6 (-25.00% of base) : System.Reflection.Metadata.dasm - PropertyDefinitionHandle:op_Implicit(PropertyDefinitionHandle):Handle
455 total methods with Code Size differences (455 improved, 0 regressed), 242842 unchanged.

Benchmarks x64 pmi diffs:

PMI CodeSize Diffs for benchstones and benchmarks game in f:\runtime1\artifacts\tests\coreclr\Windows_NT.x64.Release for  default jit
Summary of Code Size diffs:
(Lower is better)
Total bytes of diff: -52 (-0.01% of base)
    diff is an improvement.
Top file improvements (bytes):
         -40 : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm (-0.14% of base)
         -12 : BenchmarksGame\reverse-complement\reverse-complement-1\reverse-complement-1.dasm (-0.33% of base)
2 total files with Code Size differences (2 improved, 0 regressed), 80 unchanged.
Top method improvements (bytes):
         -14 (-3.23% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - ScalarFloatRenderer:RenderSingleThreadedWithADT(float,float,float,float,float):this
         -14 (-2.30% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - <>c__DisplayClass4_0:<RenderMultiThreadedWithADT>b__0(int):this (2 methods)
         -12 (-1.08% of base) : BenchmarksGame\reverse-complement\reverse-complement-1\reverse-complement-1.dasm - ReverseComplement_1:Bench(Stream,Stream)
          -6 (-10.17% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - ComplexFloat:square():ComplexFloat:this
          -6 (-9.38% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - ComplexFloat:op_Addition(ComplexFloat,ComplexFloat):ComplexFloat
Top method improvements (percentages):
          -6 (-10.17% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - ComplexFloat:square():ComplexFloat:this
          -6 (-9.38% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - ComplexFloat:op_Addition(ComplexFloat,ComplexFloat):ComplexFloat
         -14 (-3.23% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - ScalarFloatRenderer:RenderSingleThreadedWithADT(float,float,float,float,float):this
         -14 (-2.30% of base) : SIMD\ConsoleMandel\ConsoleMandel\ConsoleMandel.dasm - <>c__DisplayClass4_0:<RenderMultiThreadedWithADT>b__0(int):this (2 methods)
         -12 (-1.08% of base) : BenchmarksGame\reverse-complement\reverse-complement-1\reverse-complement-1.dasm - ReverseComplement_1:Bench(Stream,Stream)
5 total methods with Code Size differences (5 improved, 0 regressed), 1888 unchanged.

[erozenfeld]

@CarolEidt @sandreenko @AndyAyersMS PTAL cc: @dotnet/jit-contrib

[CarolEidt]

LGTM - I suspect that getting some of these conditions right was tricky! Have you measured the throughput or size impact of adding the defsInBlock map?

[CarolEidt]

FYI - the timeout on the CoreCLR Pri0 Runtime Tests Run Windows_NT x86 checked has been quite frequent lately. I've seen it on my last two PRs - one required 2 retries to get it to complete.
cc @dotnet/dnceng

[erozenfeld]

@CarolEidt

I suspect that getting some of these conditions right was tricky!

Yes, this took multiple iterations. Tracking down failures caused by removing stores to pinned locals with ref count 1 was painful.

Have you measured the throughput or size impact of adding the defsInBlock map?

I measured crossgen of SPC with pin-icount. Somewhat surprisingly, it shows a throughput improvement of 0.02%, which is close to noise level.

[sandreenko]

LGTM, do you want to run outerloop/stress tests before merge?

[briansull]

Looks Good

[erozenfeld]

LGTM, do you want to run outerloop/stress tests before merge?

Yes, will rebase and run outerloop and stress pipelines overnight.

[garath]

FYI - the timeout on the CoreCLR Pri0 Runtime Tests Run Windows_NT x86 checked has been quite frequent lately. I've seen it on my last two PRs - one required 2 retries to get it to complete.
cc @dotnet/dnceng

Noted, and thanks for highlighting it to us. I'll check some telemetry, but the likely cause is that this queue is very busy. We'll continue to monitor the pressure.

[erozenfeld]

Loader/binding/assemblies/assemblybugs/37910/Ii/Ii.sh failures in outerloop and jitstress2-jitstressregs pipelines are #38452.

JIT\jit64\mcc\interop\mcc_i7*\mcc_i7*.cmd failures in jitstress2-jitstressregs are #36199.

All failures in gcstress0x3-gcstress0xc pipeline are also seen in master except crossgen2smoke: https://dev.azure.com/dnceng/public/_build/results?buildId=705744&view=ms.vss-test-web.build-test-results-tab&runId=21861154&resultId=109360&paneView=debug
I'll try to repro it locally.

[erozenfeld]

Failure mode in crossgen2smoke is very similar to #34316.

[erozenfeld]

I was able to repro crossgen2smoke failure locally both with my changes (on the 6th run) and without my changes (on the 7th run). I opened #38482.

[MattGal]

FYI - the timeout on the CoreCLR Pri0 Runtime Tests Run Windows_NT x86 checked has been quite frequent lately. I've seen it on my last two PRs - one required 2 retries to get it to complete.
cc @dotnet/dnceng

Noted, and thanks for highlighting it to us. I'll check some telemetry, but the likely cause is that this queue is very busy. We'll continue to monitor the pressure.

@garath when you see logging like error : (NETCORE_ENGINEERING_TELEMETRY=Test) Work item ce3b1eca-cf51-4249-9c24-9a84e2fce15f/PayloadGroup0 in job ce3b1eca-cf51-4249-9c24-9a84e2fce15f has failed. Failure log: https://helix.dot.net/api/2019-06-17/jobs/ce3b1eca-cf51-4249-9c24-9a84e2fce15f/workitems/PayloadGroup0/console that's a work item timeout and not subject to the amount of work being sent in. Aside from helping to investigate the hangs there's not much we can do here to help the runtime team.