emit test for bounds checks against a 0 index

[nathan-moore]

Since array length is always non-negative, when we have a bounds check against a constant 0 index, we can emit a slightly cheaper/smaller test.

Found 156 files with textual diffs.

Summary of Code Size diffs:
(Lower is better)

Total bytes of diff: -3809 (-0.012% of base)
    diff is an improvement.

Top file improvements (bytes):
        -891 : Microsoft.Diagnostics.Tracing.TraceEvent.dasm (-0.027% of base)
        -597 : System.Private.CoreLib.dasm (-0.018% of base)
        -254 : System.Private.Xml.dasm (-0.008% of base)
        -121 : System.Data.Common.dasm (-0.011% of base)
        -101 : Microsoft.VisualBasic.Core.dasm (-0.024% of base)
         -77 : System.Net.Http.dasm (-0.013% of base)
         -68 : Microsoft.CodeAnalysis.dasm (-0.009% of base)
         -67 : Microsoft.CodeAnalysis.VisualBasic.dasm (-0.003% of base)
         -61 : System.Linq.Expressions.dasm (-0.002% of base)
         -57 : System.Runtime.Numerics.dasm (-0.083% of base)
         -51 : Microsoft.CodeAnalysis.CSharp.dasm (-0.002% of base)
         -51 : System.Private.DataContractSerialization.dasm (-0.007% of base)
         -50 : System.Drawing.Common.dasm (-0.018% of base)
         -44 : Newtonsoft.Json.dasm (-0.007% of base)
         -40 : System.CodeDom.dasm (-0.024% of base)
         -38 : System.Data.Odbc.dasm (-0.021% of base)
         -35 : FSharp.Core.dasm (-0.004% of base)
         -35 : System.Text.Json.dasm (-0.012% of base)
         -34 : System.Net.Primitives.dasm (-0.052% of base)
         -34 : System.ComponentModel.TypeConverter.dasm (-0.015% of base)

156 total files with Code Size differences (156 improved, 0 regressed), 111 unchanged.

Top method improvements (bytes):
         -38 (-1.332% of base) : System.Private.CoreLib.dasm - ArraySortHelper`2:HeapSort(Span`1,Span`1,IComparer`1) (11 methods)
         -30 (-1.162% of base) : System.Private.CoreLib.dasm - GenericArraySortHelper`2:HeapSort(Span`1,Span`1) (11 methods)
         -22 (-0.285% of base) : System.Private.CoreLib.dasm - Vector128`1:ToString():String:this (11 methods)
         -22 (-0.267% of base) : System.Private.CoreLib.dasm - Vector256`1:ToString():String:this (11 methods)
         -22 (-0.277% of base) : System.Private.CoreLib.dasm - Vector64`1:ToString():String:this (11 methods)
         -15 (-0.169% of base) : System.Private.Xml.dasm - XsltLoader:.ctor():this
         -14 (-0.536% of base) : System.Private.CoreLib.dasm - ArraySortHelper`1:HeapSort(Span`1,Comparison`1) (14 methods)
         -14 (-0.491% of base) : System.Net.Http.dasm - KnownHeaders:GetCandidate(StringAccessor):KnownHeader
         -13 (-0.569% of base) : System.Private.CoreLib.dasm - GenericArraySortHelper`1:HeapSort(Span`1) (13 methods)
         -12 (-0.259% of base) : System.Data.Common.dasm - FunctionNode:EvalFunction(int,ref,DataRow,int):Object:this
         -11 (-0.197% of base) : System.CodeDom.dasm - VBCodeGenerator:.cctor()
          -8 (-0.726% of base) : CommandLine.dasm - <>c:<get_FormatError>b__11_0(Error):String:this
          -7 (-2.160% of base) : Microsoft.CodeAnalysis.CSharp.dasm - CSharpSyntaxNode:CreateList(IEnumerable`1,bool):GreenNode:this
          -7 (-3.182% of base) : Newtonsoft.Json.Bson.dasm - BsonDataReader:BytesInSequence(ubyte):int:this
          -7 (-3.365% of base) : Newtonsoft.Json.Bson.dasm - BsonDataReader:.cctor()
          -7 (-3.125% of base) : Newtonsoft.Json.dasm - BsonReader:.cctor()
          -7 (-0.421% of base) : System.Data.Common.dasm - SqlDecimal:MpDiv(ReadOnlySpan`1,int,Span`1,int,Span`1,byref,Span`1,byref)
          -6 (-1.575% of base) : System.Private.CoreLib.dasm - Number:FormatExponent(byref,NumberFormatInfo,int,ushort,int,bool)
          -6 (-0.311% of base) : System.Private.CoreLib.dasm - Number:.cctor()
          -6 (-0.286% of base) : System.Private.CoreLib.dasm - CalendarData:CreateInvariant():CalendarData

Top method improvements (percentages):
          -7 (-3.365% of base) : Newtonsoft.Json.Bson.dasm - BsonDataReader:.cctor()
          -7 (-3.182% of base) : Newtonsoft.Json.Bson.dasm - BsonDataReader:BytesInSequence(ubyte):int:this
          -1 (-3.125% of base) : System.Collections.Immutable.dasm - ImmutableArrayExtensions:FirstOrDefault(ImmutableArray`1):__Canon
          -7 (-3.125% of base) : Newtonsoft.Json.dasm - BsonReader:.cctor()
          -1 (-2.703% of base) : Microsoft.VisualBasic.Core.dasm - CharType:FromString(String):ushort
          -1 (-2.703% of base) : Microsoft.VisualBasic.Core.dasm - Conversions:ToChar(String):ushort
          -1 (-2.564% of base) : System.Private.Xml.dasm - Compiler:IsPhantomNamespace(String):bool:this
          -1 (-2.564% of base) : xunit.execution.dotnet.dasm - Pack:UInt16_To_LE(ushort,ref)
          -1 (-2.500% of base) : System.Private.CoreLib.dasm - Statics:ShouldOverrideFieldName(String):bool
          -1 (-2.500% of base) : Microsoft.VisualBasic.Core.dasm - FileSystem:ChDrive(String)
          -1 (-2.500% of base) : System.Private.Xml.dasm - OutputScopeManager:Reset():this
          -1 (-2.500% of base) : xunit.execution.dotnet.dasm - Pack:UInt16_To_BE(ushort,ref)
          -1 (-2.439% of base) : System.Private.Xml.dasm - BigInteger:InitFromDigits(int,int,int):this
          -5 (-2.427% of base) : Microsoft.CodeAnalysis.dasm - UnionCollection`1:Create(ImmutableArray`1,Func`2):ICollection`1
          -1 (-2.381% of base) : Microsoft.CodeAnalysis.CSharp.dasm - CodeGenerator:IsMultidimensionalInitializer(ImmutableArray`1):bool
          -1 (-2.381% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - CodeGenerator:IsMultidimensionalInitializer(ImmutableArray`1):bool:this
          -1 (-2.381% of base) : xunit.execution.dotnet.dasm - Pack:BE_To_UInt16(ref):ushort
          -1 (-2.381% of base) : xunit.execution.dotnet.dasm - Pack:LE_To_UInt16(ref):ushort
          -1 (-2.326% of base) : Newtonsoft.Json.Bson.dasm - StringUtils:StartsWith(String,ushort):bool
          -1 (-2.326% of base) : Newtonsoft.Json.dasm - StringUtils:StartsWith(String,ushort):bool

3127 total methods with Code Size differences (3127 improved, 0 regressed), 188060 unchanged.

[EgorBo]

Nice!
Can you also remove dead variables in that function:

    GenTree* arrRef    = nullptr;
    int      lenOffset = 0;

[EgorBo]

(oops forgot my approves are "green" in this repo - treat it is a "like" 🙂)

[sandreenko]

LGTM, thank you, do you know if we can do something similar for arm64?

[sandreenko]

cc @dotnet/jit-contrib

[JulieLeeMSFT]

Thanks @nathan-moore for your contribution.
I am also interested to know if the same can be applied to Arm64.

[EgorBo]

@sandreenko @JulieLeeMSFT

static int Test(int[] arr) => arr[0];

Current codegen:

G_M50965_IG01:
        A9BF7BFD          stp     fp, lr, [sp,#-16]!
        910003FD          mov     fp, sp
                                                ;; bbWeight=1    PerfScore 1.50
G_M50965_IG02:
        B9400801          ldr     w1, [x0,#8]
        7100003F          cmp     w1, #0
        54000089          bls     G_M50965_IG04
        B9401000          ldr     w0, [x0,#16]
                                                ;; bbWeight=1    PerfScore 7.50
G_M50965_IG03:
        A8C17BFD          ldp     fp, lr, [sp],#16
        D65F03C0          ret     lr
                                                ;; bbWeight=1    PerfScore 2.00
G_M50965_IG04:
        97FF2D9A          bl      CORINFO_HELP_RNGCHKFAIL
        D4200000          brk     #0
                                                ;; bbWeight=0    PerfScore 0.00

this

    cmp     w1, #0
    bls     G_M50965_IG04

can be optimized to:

    cbz     w1, G_M50965_IG04  ;; or CBNZ

branch if (not)zero.

[nathan-moore]

@sandreenko @EgorBo I looked into what it will take to emit cbz here for arm. There isn't a way to emit it with genJumpToThrowHlpBlk, which is what everything currently uses. You can add a way, but since cbz is internally represented very differently from a normal jump, I don't see a good way of not duplicating the logic, which is what I did here: nathan-moore@8500e53. Does that seem like the right approach?

Also, I checked and there's already a peephole for this in the general case.

[sandreenko]

Thank you @nathan-moore and @EgorBo for the analysis. I think it is valuable to add this optimization for arm64 if it produces similar diffs there. However, it could be done in a separate PR and this one can be merged now.

Thanks @nathan-moore for this optimization.

[sandreenko]

Opened a follow-up for arm64: #42514.