JIT: fgLateDevirtualization should try and simplify type compares

[AndyAyersMS]

fgLateDevirtualization may be able to optimize type compares in cases where type information arrives late. For example a better type can be exposed by inlining or from a readonly static. For example:

class X 
{
    static object S() => "s";

    [MethodImpl(MethodImplOptions.NoInlining)]
    static bool IsString()
    {
        // This should be a jit-time constant
        return(S().GetType() == typeof(string));
    }

    public static int Main()
    {
        var b = IsString();
        return b ? 100 : -1;
    }
}

and once dotnet/coreclr#20886 is in place,

class X 
{
    static readonly object S;

    static X()
    {
        S = "s";
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    static bool IsString()
    {
        // This should be a jit-time constant
        return(S.GetType() == typeof(string));
    }

    public static int Main()
    {
        var p = S;
        var b = IsString();
        return b ? 100 : -1;
    }
}

category:cq
theme:devirtualization
skill-level:intermediate
cost:medium
impact:small

[AndyAyersMS]

Prototype changes: MoreTypeFolding.

Does not seem to fire very often... may need to audit all the various residual type compares to understand why.

PMI Diffs for System.Private.CoreLib.dll, framework assemblies [invoking .cctors] for x64 default jit
Summary:
(Lower is better)
Total bytes of diff: -105 (0.00% of base)
    diff is an improvement.
Top file improvements by size (bytes):
        -105 : Microsoft.CodeAnalysis.dasm (-0.01% of base)
1 total files with size differences (1 improved, 0 regressed), 128 unchanged.
Top method improvements by size (bytes):
        -105 (-16.67% of base) : Microsoft.CodeAnalysis.dasm - SmallDictionary`2:RightComplex(ref):ref (6 base, 5 diff methods)
Top method improvements by size (percentage):
        -105 (-16.67% of base) : Microsoft.CodeAnalysis.dasm - SmallDictionary`2:RightComplex(ref):ref (6 base, 5 diff methods)
1 total methods with size differences (1 improved, 0 regressed), 193468 unchanged.

am going to move this to Future for now.

[benaadams]

Re: Jit diffs, how easy would it be to include the ASP.NET Core assemblies as part of the common compares?

[AndyAyersMS]

Not too hard. We just need to figure out which versions to use and put the assemblies somewhere.

Feel free to open an issue over in jitutils.

[benaadams]

dotnet/jitutils#189

[AndyAyersMS]

Example 2 above is now optimized:

; Assembly listing for method X:IsString():bool
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; optimized code
; rsp based frame
; partially interruptible
; Final local variable assignments
;
;# V00 OutArgs      [V00    ] (  1,  1   )  lclBlk ( 0) [rsp+0x00]   "OutgoingArgSpace"
;
; Lcl frame size = 0

G_M36194_IG01:              ;; offset=0000H
                                                ;; bbWeight=1    PerfScore 0.00
G_M36194_IG02:              ;; offset=0000H
       B801000000           mov      eax, 1
                                                ;; bbWeight=1    PerfScore 0.25
G_M36194_IG03:              ;; offset=0005H
       C3                   ret
                                                ;; bbWeight=1    PerfScore 1.00

Example 1 above is not yet fully cleaned up.

; Assembly listing for method X:IsString():bool
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; optimized code
; rsp based frame
; partially interruptible
; Final local variable assignments
;
;# V00 OutArgs      [V00    ] (  1,  1   )  lclBlk ( 0) [rsp+0x00]   "OutgoingArgSpace"
;
; Lcl frame size = 0

G_M36194_IG01:              ;; offset=0000H
                                                ;; bbWeight=1    PerfScore 0.00
G_M36194_IG02:              ;; offset=0000H
       48B898317CB7C4010000 mov      rax, 0x1C4B77C3198
       488B00               mov      rax, gword ptr [rax]
       48BA4069BCAFF87F0000 mov      rdx, 0x7FF8AFBC6940
       483910               cmp      qword ptr [rax], rdx
       0F94C0               sete     al
       0FB6C0               movzx    rax, al
                                                ;; bbWeight=1    PerfScore 5.75
G_M36194_IG03:              ;; offset=0020H
       C3                   ret
                                                ;; bbWeight=1    PerfScore 1.00

Seems like case 1 should have been handled by #32790; need to drill in and see why it didn't get fully optimized (to null check plus constant).

[jakobbotsch]

Current codegen is:

; Method X:IsString():bool
G_M36194_IG01:
						;; size=0 bbWeight=1    PerfScore 0.00

G_M36194_IG02:
       mov      rax, 0xD1FFAB1E      ; System.String
       cmp      qword ptr [(reloc)], rax
       sete     al
       movzx    rax, al
						;; size=23 bbWeight=1    PerfScore 4.50

G_M36194_IG03:
       ret      
						;; size=1 bbWeight=1    PerfScore 1.00
; Total bytes of code: 24

Still not entirely folded, even with @EgorBo's recent changes

[EgorBo]

Bot examples are now return true;