[RyuJIT] Devirtualize Comparer<T>.Default

[EgorBo]

See #43796

Mostly a copy-paste from the existing EqualityComparer<T>.Default devirtualization.
Example:

private static int Compare(byte a, byte b) => Comparer<byte>.Default.Compare(a, b);

master:

; Method Program:Compare(ubyte,ubyte):int
G_M13367_IG01:              ;; offset=0000H
       57                   push     rdi
       56                   push     rsi
       4883EC28             sub      rsp, 40
       8BF1                 mov      esi, ecx
       8BFA                 mov      edi, edx
G_M13367_IG02:              ;; offset=000AH
       48B928002535FF7F0000 mov      rcx, 0x7FFF35250028
       BA08000000           mov      edx, 8
       E8F218755F           call     CORINFO_HELP_CLASSINIT_SHARED_DYNAMICCLASS
       48B9D02C0098F8010000 mov      rcx, 0x1F898002CD0
       488B09               mov      rcx, gword ptr [rcx]
       400FB6D6             movzx    rdx, sil
       440FB6C7             movzx    r8, dil
       488B01               mov      rax, qword ptr [rcx]
       488B4048             mov      rax, qword ptr [rax+72]
       488B4020             mov      rax, qword ptr [rax+32]
G_M13367_IG03:              ;; offset=003EH
       4883C428             add      rsp, 40
       5E                   pop      rsi
       5F                   pop      rdi
       48FFE0               rex.jmp  rax
; Total bytes of code: 71

PR:

; Method Program:Compare(ubyte,ubyte):int
G_M13367_IG01:              ;; offset=0000H
G_M13367_IG02:              ;; offset=0000H
       0FB6C1               movzx    rax, cl
       0FB6D2               movzx    rdx, dl
       2BC2                 sub      eax, edx
G_M13367_IG03:              ;; offset=0008H
       C3                   ret      
; Total bytes of code: 9

JIT diff is empty but it should allow to write generic versions of Max<T>/Min<T> functions (#43796).

[stephentoub]

Sweet!

[stephentoub]

cc: @eiriktsarpalis

[AndyAyersMS]

Seems like the enum comparer gets inlined and optimized w/o needing annotation, but can you double-check?

Did you try crossgenning your test and then running it? Would be good to get somebody more familiar with crossgen2 to sign off on the changes there.

Changes LGTM, just some comments on comments.

[EgorBo]

Seems like the enum comparer gets inlined and optimized w/o needing annotation, but can you double-check?

Did you try crossgenning your test and then running it? Would be good to get somebody more familiar with crossgen2 to sign off on the changes there.

Changes LGTM, just some comments on comments.

Yeah, the tests pass with crossgen2

[AndyAyersMS]

@dotnet/crossgen-contrib can somebody take a quick look at the crossgen2 changes here?

[trylek]

Crossgen2-related changes look good to me, thank you!

[eiriktsarpalis]

I ran my PriorityQueue benchmarks against the PR branch: this compares a heap implementation using Comparer<T>.Default.Compare calls versus an identical one using IComparable:

master branch

Method	Size	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
PriorityQueue_ComparerDefault	10	502.7 ns	9.83 ns	13.13 ns	-	-	-	-
PriorityQueue_Comparable	10	353.8 ns	6.86 ns	7.90 ns	-	-	-	-
PriorityQueue_ComparerDefault	50	4,171.1 ns	76.74 ns	68.02 ns	-	-	-	-
PriorityQueue_Comparable	50	2,380.8 ns	37.08 ns	32.87 ns	-	-	-	-
PriorityQueue_ComparerDefault	150	20,487.9 ns	364.63 ns	341.08 ns	-	-	-	-
PriorityQueue_Comparable	150	11,740.0 ns	138.82 ns	123.06 ns	-	-	-	-
PriorityQueue_ComparerDefault	500	92,828.3 ns	1,767.72 ns	1,567.04 ns	-	-	-	-
PriorityQueue_Comparable	500	64,291.4 ns	1,196.99 ns	1,378.46 ns	-	-	-	-
PriorityQueue_ComparerDefault	1000	205,028.1 ns	3,697.23 ns	4,257.74 ns	-	-	-	-
PriorityQueue_Comparable	1000	147,682.6 ns	2,176.21 ns	1,929.15 ns	-	-	-	-
PriorityQueue_ComparerDefault	10000	2,718,728.5 ns	54,288.34 ns	58,087.92 ns	-	-	-	1 B
PriorityQueue_Comparable	10000	1,935,928.9 ns	36,901.95 ns	34,518.11 ns	-	-	-	1 B
PriorityQueue_ComparerDefault	1000000	454,903,575.0 ns	8,828,957.20 ns	11,480,138.23 ns	-	-	-	144 B
PriorityQueue_Comparable	1000000	348,606,720.0 ns	6,908,734.31 ns	6,462,434.36 ns	-	-	-	2360 B

PR branch

Method	Size	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
PriorityQueue_ComparerDefault	10	349.2 ns	6.19 ns	7.83 ns	-	-	-	-
PriorityQueue_Comparable	10	327.2 ns	6.28 ns	5.87 ns	-	-	-	-
PriorityQueue_ComparerDefault	50	2,514.8 ns	18.71 ns	14.61 ns	-	-	-	-
PriorityQueue_Comparable	50	2,274.4 ns	37.11 ns	32.89 ns	-	-	-	-
PriorityQueue_ComparerDefault	150	12,616.7 ns	54.33 ns	45.37 ns	-	-	-	-
PriorityQueue_Comparable	150	11,299.3 ns	96.31 ns	90.09 ns	-	-	-	-
PriorityQueue_ComparerDefault	500	66,912.2 ns	541.71 ns	506.72 ns	-	-	-	-
PriorityQueue_Comparable	500	63,575.9 ns	375.87 ns	333.20 ns	-	-	-	-
PriorityQueue_ComparerDefault	1000	152,907.1 ns	554.02 ns	518.23 ns	-	-	-	-
PriorityQueue_Comparable	1000	147,177.2 ns	1,111.73 ns	1,039.92 ns	-	-	-	-
PriorityQueue_ComparerDefault	10000	2,060,862.3 ns	33,190.78 ns	29,422.76 ns	-	-	-	1 B
PriorityQueue_Comparable	10000	1,965,759.2 ns	39,021.23 ns	53,412.70 ns	-	-	-	1 B
PriorityQueue_ComparerDefault	1000000	372,963,264.3 ns	7,253,139.17 ns	6,429,718.95 ns	-	-	-	760 B
PriorityQueue_Comparable	1000000	362,177,334.8 ns	7,176,855.04 ns	9,076,423.69 ns	-	-	-	144 B

[AndyAyersMS]

Nice work @EgorBo!

I ran my PriorityQueue benchmarks ...

@adamsitnik Would be nice to fix the benchmark harness so you can pass multiple --corerun args and have it compare them, like it can do for --runtimes.

[stephentoub]

Would be nice to fix the benchmark harness so you can pass multiple --corerun args and have it compare them, like it can do for --runtimes.

You can, unless I'm misunderstanding what you're asking for. I do this all the time:

--corerun d:\coreclrtest\master\corerun.exe d:\coreclrtest\pr\corerun.exe

[AndyAyersMS]

I do this all the time ...

Guess I'm confused. Maybe what I was thinking was allowing a mixture of --runtimes and --corerun .

[akoeplinger]

@EgorBo this seems to cause a warning in CI, would you mind fixing it:

/__w/1/s/src/tests/JIT/opt/Devirtualization/Comparer_get_Default.cs(240,32): warning CS8632: The annotation for nullable reference types should only be used in code within a '#nullable' annotations context. [/__w/1/s/src/tests/JIT/opt/Devirtualization/Comparer_get_Default.csproj]