[Perf] Windows/x64: 2 Regressions on 5/31/2024 9:53:21 PM #35582

performanceautofiler · 2024-06-04T08:16:12Z

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22621
Queue	TigerWindows
Baseline	214cbbf3a7b60fd763a1401a853393f44c8cbc7d
Compare	9db548658b943f7b819428385cbd6304dadcb69a
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Memory.Span<Char>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
BinarySearch - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	10.62 ns	14.95 ns	1.41	0.05	True
BinarySearch - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	6.57 ns	8.74 ns	1.33	0.05	False

Test Report

Repro

General Docs link: https://github.com/dotnet/performance/blob/main/docs/benchmarking-workflow-dotnet-runtime.md

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Memory.Span&lt;Char&gt;*'

System.Memory.Span<Char>.BinarySearch(Size: 512)

ETL Files

Histogram

JIT Disasms

System.Memory.Span<Char>.BinarySearch(Size: 33)

ETL Files

Histogram

JIT Disasms

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

LoopedBard3 · 2024-06-04T16:59:34Z

Diff: dotnet/runtime@2280972...d8c59a4

jakobbotsch · 2024-06-24T13:53:59Z

Seems alignment related given that there is only one x64 config that seems to have regressed.
The diff seen below looks good as well -- we managed to get rid of a stack load and then hoisted an extension out from the loop.

`System.Memory.Span<Char>.BinarySearch(Size: 512)`

Hot functions:

(86.46%) SpanHelpers.BinarySearch (Tier-1)
- Has diffs
(6.13%) System.Memory.Span`1.BinarySearch (Tier-1)
- No diffs
(2.96%) Runnable_0.WorkloadActionUnroll (FullOpt)
- No diffs

Diffs

`[System.Private.CoreLib]SpanHelpers.BinarySearch(!!0&,int32,!!1)`

 ; Final local variable assignments
 ;
 ;  V00 arg0         [V00,T04] (  3,  9.20)   byref  ->  rbx         single-def
-;  V01 arg1         [V01,T05] (  3,  3   )     int  ->  rdx         single-def
-;  V02 arg2         [V02    ] (  3,  9.20)  ushort  ->  [rsp+0x60]  do-not-enreg[X] addr-exposed ld-addr-op single-def
+;  V01 arg1         [V01,T06] (  3,  3   )     int  ->  rdx         single-def
+;  V02 arg2         [V02,T07] (  3,  2.90)  ushort  ->   r8         ld-addr-op single-def
 ;  V03 loc0         [V03,T03] (  5, 16.21)     int  ->  rsi        
 ;  V04 loc1         [V04,T02] (  5, 16.51)     int  ->  rdi        
-;  V05 loc2         [V05,T00] (  5, 19.14)     int  ->  rbp        
-;  V06 loc3         [V06,T01] (  3, 18.66)     int  ->  rcx        
+;  V05 loc2         [V05,T00] (  5, 19.14)     int  ->  r14        
+;  V06 loc3         [V06,T01] (  3, 18.66)     int  ->  rax        
 ;* V07 loc4         [V07    ] (  0,  0   )  ushort  ->  zero-ref    ld-addr-op
 ;  V08 OutArgs      [V08    ] (  1,  1   )  struct (32) [rsp+0x00]  do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
 ;* V09 tmp1         [V09    ] (  0,  0   )   byref  ->  zero-ref   
 ;* V10 tmp2         [V10    ] (  0,  0   )  ushort  ->  zero-ref    "Inlining Arg"
+;  V11 cse0         [V11,T05] (  2,  8.10)     int  ->  rbp         hoist "CSE #01: aggressive"
 ;
-; Lcl frame size = 40
+; Lcl frame size = 32
 
 G_M30959_IG01:
+       push     r14
        push     rdi
        push     rsi
        push     rbp
        push     rbx
-       sub      rsp, 40
-       mov      dword ptr [rsp+0x60], r8d
+       sub      rsp, 32
        mov      rbx, rcx
-						;; size=16 bbWeight=1 PerfScore 5.50
+						;; size=13 bbWeight=1 PerfScore 5.50
 G_M30959_IG02:
        xor      esi, esi
        lea      edi, [rdx-0x01]
        test     edi, edi
-       jl       SHORT G_M30959_IG10
+       jl       SHORT G_M30959_IG11
 						;; size=9 bbWeight=1 PerfScore 2.00
 G_M30959_IG03:
-       lea      ebp, [rdi+rsi]
-       shr      ebp, 1
-       movzx    rcx, word  ptr [rsp+0x60]
-       movsxd   rax, ebp
-       movzx    rax, word  ptr [rbx+2*rax]
-       sub      ecx, eax
-       je       SHORT G_M30959_IG08
-						;; size=21 bbWeight=7.20 PerfScore 39.60
+       movzx    rbp, r8w
+						;; size=4 bbWeight=0.90 PerfScore 0.23
 G_M30959_IG04:
-       test     ecx, ecx
-       jle      SHORT G_M30959_IG06
-						;; size=4 bbWeight=4.26 PerfScore 5.33
+       lea      r14d, [rdi+rsi]
+       shr      r14d, 1
+       movsxd   rcx, r14d
+       movzx    rcx, word  ptr [rbx+2*rcx]
+       mov      eax, ebp
+       sub      eax, ecx
+       je       SHORT G_M30959_IG09
+						;; size=20 bbWeight=7.20 PerfScore 34.20
 G_M30959_IG05:
-       mov      rcx, 0xD1FFAB1E
-       call     CORINFO_HELP_COUNTPROFILE32
-       lea      esi, [rbp+0x01]
-       jmp      SHORT G_M30959_IG07
-						;; size=20 bbWeight=2.22 PerfScore 8.31
+       test     eax, eax
+       jle      SHORT G_M30959_IG07
+						;; size=4 bbWeight=4.26 PerfScore 5.33
 G_M30959_IG06:
        mov      rcx, 0xD1FFAB1E
        call     CORINFO_HELP_COUNTPROFILE32
-       lea      edi, [rbp-0x01]
-						;; size=18 bbWeight=2.05 PerfScore 3.58
+       lea      esi, [r14+0x01]
+       jmp      SHORT G_M30959_IG08
+						;; size=21 bbWeight=2.22 PerfScore 8.31
 G_M30959_IG07:
-       cmp      esi, edi
-       jle      SHORT G_M30959_IG03
-       jmp      SHORT G_M30959_IG10
-						;; size=6 bbWeight=5.26 PerfScore 17.11
-G_M30959_IG08:
        mov      rcx, 0xD1FFAB1E
        call     CORINFO_HELP_COUNTPROFILE32
-       mov      eax, ebp
-						;; size=17 bbWeight=0.47 PerfScore 0.71
+       lea      edi, [r14-0x01]
+						;; size=19 bbWeight=2.05 PerfScore 3.58
+G_M30959_IG08:
+       cmp      esi, edi
+       jle      SHORT G_M30959_IG04
+       jmp      SHORT G_M30959_IG11
+						;; size=6 bbWeight=5.26 PerfScore 17.11
 G_M30959_IG09:
-       add      rsp, 40
+       mov      rcx, 0xD1FFAB1E
+       call     CORINFO_HELP_COUNTPROFILE32
+       mov      eax, r14d
+						;; size=18 bbWeight=0.47 PerfScore 0.71
+G_M30959_IG10:
+       add      rsp, 32
        pop      rbx
        pop      rbp
        pop      rsi
        pop      rdi
+       pop      r14
        ret      
-						;; size=9 bbWeight=0.47 PerfScore 1.54
-G_M30959_IG10:
+						;; size=11 bbWeight=0.47 PerfScore 1.78
+G_M30959_IG11:
        mov      rcx, 0xD1FFAB1E
        call     CORINFO_HELP_COUNTPROFILE32
        mov      eax, esi
        not      eax
 						;; size=19 bbWeight=0.53 PerfScore 0.92
-G_M30959_IG11:
-       add      rsp, 40
+G_M30959_IG12:
+       add      rsp, 32
        pop      rbx
        pop      rbp
        pop      rsi
        pop      rdi
+       pop      r14
        ret      
-						;; size=9 bbWeight=0.53 PerfScore 1.71
+						;; size=11 bbWeight=0.53 PerfScore 1.97
 
-; Total bytes of code 148, prolog size 16, PerfScore 86.31, instruction count 49, allocated bytes for code 148 (MethodHash=971b8710) for method System.SpanHelpers:BinarySearch[ushort,ushort](byref,int,ushort):int (Instrumented Tier1)
+; Total bytes of code 155, prolog size 13, PerfScore 81.64, instruction count 52, allocated bytes for code 155 (MethodHash=971b8710) for method System.SpanHelpers:BinarySearch[ushort,ushort](byref,int,ushort):int (Instrumented Tier1)
 ; ============================================================
 
 ; Assembly listing for method System.SpanHelpers:BinarySearch[ushort,ushort](System.ReadOnlySpan`1[ushort],ushort):int (Instrumented Tier0)
@@ -214,52 +220,53 @@ G_M27467_IG05:
 ; optimized using Dynamic PGO
 ; rsp based frame
 ; fully interruptible
-; with Dynamic PGO: fgCalledCount is 31548
+; with Dynamic PGO: fgCalledCount is 30036
 ; 0 inlinees with PGO data; 1 single block inlinees; 0 inlinees without PGO data
 ; Final local variable assignments
 ;
-;  V00 arg0         [V00,T04] (  3, 12.17)   byref  ->  rcx         single-def
-;  V01 arg1         [V01,T05] (  3,  3   )     int  ->  rdx         single-def
-;  V02 arg2         [V02    ] (  3, 12.17)  ushort  ->  [rsp+0x18]  do-not-enreg[X] addr-exposed ld-addr-op single-def
-;  V03 loc0         [V03,T00] (  5, 30.51)     int  ->   r8        
-;  V04 loc1         [V04,T03] (  5, 22.34)     int  ->  rdx        
-;  V05 loc2         [V05,T01] (  5, 30.51)     int  ->  rax        
-;  V06 loc3         [V06,T02] (  3, 29.51)     int  ->  r10        
+;  V00 arg0         [V00,T04] (  3, 12.02)   byref  ->  rcx         single-def
+;  V01 arg1         [V01,T06] (  3,  3   )     int  ->  rdx         single-def
+;  V02 arg2         [V02,T07] (  3,  3   )  ushort  ->   r8         ld-addr-op single-def
+;  V03 loc0         [V03,T00] (  5, 30.06)     int  ->  r10        
+;  V04 loc1         [V04,T03] (  5, 22.04)     int  ->  rdx        
+;  V05 loc2         [V05,T01] (  5, 30.06)     int  ->  rax        
+;  V06 loc3         [V06,T02] (  3, 29.06)     int  ->  r11        
 ;* V07 loc4         [V07    ] (  0,  0   )  ushort  ->  zero-ref    ld-addr-op
 ;# V08 OutArgs      [V08    ] (  1,  1   )  struct ( 0) [rsp+0x00]  do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
 ;* V09 tmp1         [V09    ] (  0,  0   )   byref  ->  zero-ref   
 ;* V10 tmp2         [V10    ] (  0,  0   )  ushort  ->  zero-ref    "Inlining Arg"
+;  V11 cse0         [V11,T05] (  2, 11.02)     int  ->   r8         hoist "CSE #01: aggressive"
 ;
 ; Lcl frame size = 0
 
 G_M30959_IG01:
-       mov      dword ptr [rsp+0x18], r8d
-						;; size=5 bbWeight=1 PerfScore 1.00
+						;; size=0 bbWeight=1 PerfScore 0.00
 G_M30959_IG02:
-       xor      r8d, r8d
+       xor      r10d, r10d
        dec      edx
        js       SHORT G_M30959_IG08
+       movzx    r8, r8w
        align    [0 bytes for IG03]
-						;; size=7 bbWeight=1 PerfScore 1.50
+						;; size=11 bbWeight=1 PerfScore 1.75
 G_M30959_IG03:
-       lea      eax, [rdx+r8]
+       lea      eax, [rdx+r10]
        shr      eax, 1
-       movzx    r10, word  ptr [rsp+0x18]
        movsxd   r9, eax
        movzx    r9, word  ptr [rcx+2*r9]
-       sub      r10d, r9d
+       mov      r11d, r8d
+       sub      r11d, r9d
        je       SHORT G_M30959_IG06
-						;; size=25 bbWeight=10.17 PerfScore 55.93
+						;; size=22 bbWeight=10.02 PerfScore 47.59
 G_M30959_IG04:
-       test     r10d, r10d
+       test     r11d, r11d
        jle      SHORT G_M30959_IG07
-       lea      r8d, [rax+0x01]
-						;; size=9 bbWeight=9.17 PerfScore 16.05
+       lea      r10d, [rax+0x01]
+						;; size=9 bbWeight=9.02 PerfScore 15.78
 G_M30959_IG05:
-       cmp      r8d, edx
+       cmp      r10d, edx
        jle      SHORT G_M30959_IG03
        jmp      SHORT G_M30959_IG08
-						;; size=7 bbWeight=10.17 PerfScore 33.05
+						;; size=7 bbWeight=10.02 PerfScore 32.56
 G_M30959_IG06:
        ret      
 						;; size=1 bbWeight=1.00 PerfScore 1.00
@@ -268,14 +275,14 @@ G_M30959_IG07:
        jmp      SHORT G_M30959_IG05
 						;; size=5 bbWeight=0 PerfScore 0.00
 G_M30959_IG08:
-       mov      eax, r8d
+       mov      eax, r10d
        not      eax
 						;; size=5 bbWeight=0 PerfScore 0.00
 G_M30959_IG09:
        ret      
 						;; size=1 bbWeight=0 PerfScore 0.00
 
-; Total bytes of code 65, prolog size 5, PerfScore 108.53, instruction count 24, allocated bytes for code 65 (MethodHash=971b8710) for method System.SpanHelpers:BinarySearch[ushort,ushort](byref,int,ushort):int (Tier1)
+; Total bytes of code 61, prolog size 0, PerfScore 98.69, instruction count 24, allocated bytes for code 61 (MethodHash=971b8710) for method System.SpanHelpers:BinarySearch[ushort,ushort](byref,int,ushort):int (Tier1)
 ; ============================================================
 
 ; Assembly listing for method System.SpanHelpers:BinarySearch[ushort,ushort](System.ReadOnlySpan`1[ushort],ushort):int (Tier1)

performanceautofiler bot added arch-x64 branch-refs/heads/main kind-micro os-windows perf-regression runtime-coreclr untriaged labels Jun 4, 2024

performanceautofiler bot mentioned this issue Jun 4, 2024

[SENTINEL] Autofile run complete at 6/4/2024 8:18:38 AM. 16 issues filed. #35592

Closed

LoopedBard3 mentioned this issue Jun 4, 2024

JIT: Propagate LCL_ADDR nodes during local morph dotnet/runtime#102808

Merged

LoopedBard3 removed the untriaged label Jun 4, 2024

LoopedBard3 closed this as completed Jun 4, 2024

DrewScoggins mentioned this issue Aug 22, 2024

Performance Fundamental: Net 8 -> Net 9 Manual Comparison Report dotnet/runtime#106824

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Perf] Windows/x64: 2 Regressions on 5/31/2024 9:53:21 PM #35582

[Perf] Windows/x64: 2 Regressions on 5/31/2024 9:53:21 PM #35582

performanceautofiler bot commented Jun 4, 2024

System.Memory.Span<Char>.BinarySearch(Size: 512)

ETL Files

Histogram

JIT Disasms

System.Memory.Span<Char>.BinarySearch(Size: 33)

ETL Files

Histogram

JIT Disasms

Docs

LoopedBard3 commented Jun 4, 2024

jakobbotsch commented Jun 24, 2024 •

edited

Loading

`[System.Private.CoreLib]SpanHelpers.BinarySearch(!!0&,int32,!!1)`

[Perf] Windows/x64: 2 Regressions on 5/31/2024 9:53:21 PM #35582

[Perf] Windows/x64: 2 Regressions on 5/31/2024 9:53:21 PM #35582

Comments

performanceautofiler bot commented Jun 4, 2024

Run Information

Regressions in System.Memory.Span<Char>

Repro

System.Memory.Span<Char>.BinarySearch(Size: 512)

ETL Files

Histogram

JIT Disasms

System.Memory.Span<Char>.BinarySearch(Size: 33)

ETL Files

Histogram

JIT Disasms

Docs

LoopedBard3 commented Jun 4, 2024

jakobbotsch commented Jun 24, 2024 • edited Loading

System.Memory.Span<Char>.BinarySearch(Size: 512)

[System.Private.CoreLib]SpanHelpers.BinarySearch(!!0&,int32,!!1)

jakobbotsch commented Jun 24, 2024 •

edited

Loading

`System.Memory.Span<Char>.BinarySearch(Size: 512)`

`[System.Private.CoreLib]SpanHelpers.BinarySearch(!!0&,int32,!!1)`