Increase max loops optimized by RyuJIT from 16 to 64.

[BruceForstall]

16 seems remarkably small. 64 was not scientifically chosen, but does cover more
cases.

Note that this number is used to allocate the statically-sized loop table, as well as for memory allocation for value
numbering, so there is some overhead to increasing it.

A few microbenchmarks that have diffs show benefit, including 9% for MulMatrix

Method	Job	Toolchain	Mean	Error	StdDev	Median	Min	Max	Ratio
LLoops	Job-JXEMSM	\runtime2\artifacts\tests\coreclr\windows.x64.Checked\Tests\Core_Root\CoreRun.exe	559.9 ms	8.20 ms	7.67 ms	556.4 ms	550.6 ms	576.0 ms	1.00
LLoops	Job-MUOLTV	\runtime\artifacts\tests\coreclr\windows.x64.Checked\Tests\Core_Root\CoreRun.exe	552.3 ms	5.84 ms	5.46 ms	552.0 ms	542.4 ms	561.1 ms	0.99
MulMatrix	Job-JXEMSM	\runtime2\artifacts\tests\coreclr\windows.x64.Checked\Tests\Core_Root\CoreRun.exe	369.7 ms	4.01 ms	3.56 ms	369.9 ms	364.6 ms	376.9 ms	1.00
MulMatrix	Job-MUOLTV	\runtime\artifacts\tests\coreclr\windows.x64.Checked\Tests\Core_Root\CoreRun.exe	338.1 ms	2.69 ms	2.51 ms	337.7 ms	332.2 ms	341.9 ms	0.91
Puzzle	Job-JXEMSM	\runtime2\artifacts\tests\coreclr\windows.x64.Checked\Tests\Core_Root\CoreRun.exe	403.4 ms	6.93 ms	6.48 ms	402.3 ms	394.8 ms	412.5 ms	1.00
Puzzle	Job-MUOLTV	\runtime\artifacts\tests\coreclr\windows.x64.Checked\Tests\Core_Root\CoreRun.exe	394.9 ms	4.16 ms	3.68 ms	395.5 ms	388.2 ms	401.8 ms	0.98

spmi diffs:

Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 6264
Total bytes of diff: 6285
Total bytes of delta: 21 (0.34% of base)
Total relative delta: 0.00
diff is a regression.
relative diff is a regression.

Detail diffs

Top file regressions (bytes):
21 : 42451.dasm (0.34% of base)
1 total files with Code Size differences (0 improved, 1 regressed), 0 unchanged.
Top method regressions (bytes):
21 ( 0.34% of base) : 42451.dasm - RelationalModel:Create(IModel,IRelationalAnnotationProvider):IRelationalModel
Top method regressions (percentages):
21 ( 0.34% of base) : 42451.dasm - RelationalModel:Create(IModel,IRelationalAnnotationProvider):IRelationalModel
1 total methods with Code Size differences (0 improved, 1 regressed), 0 unchanged.

---

Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 39243
Total bytes of diff: 39940
Total bytes of delta: 697 (1.78% of base)
Total relative delta: 0.16
diff is a regression.
relative diff is a regression.

Detail diffs

Top file regressions (bytes):
536 : 25723.dasm (18.29% of base)
185 : 26877.dasm (2.21% of base)
66 : 16212.dasm (1.39% of base)
16 : 16196.dasm (0.36% of base)
13 : 13322.dasm (0.27% of base)
9 : 15596.dasm (0.16% of base)
Top file improvements (bytes):
-125 : 27270.dasm (-6.93% of base)
-3 : 13993.dasm (-0.05% of base)
8 total files with Code Size differences (2 improved, 6 regressed), 0 unchanged.
Top method regressions (bytes):
536 (18.29% of base) : 25723.dasm - Benchstone.BenchI.MulMatrix:Inner(System.Int32[][],System.Int32[][],System.Int32[][])
185 ( 2.21% of base) : 26877.dasm - Benchstone.BenchF.LLoops:Main1(int):this
66 ( 1.39% of base) : 16212.dasm - Jil.Deserialize.Methods:SkipWithLeadChar(System.IO.TextReader,int)
16 ( 0.36% of base) : 16196.dasm - DynamicClass:_DynamicMethod9(System.IO.TextReader,int):MicroBenchmarks.Serializers.MyEventsListerViewModel
13 ( 0.27% of base) : 13322.dasm - DynamicClass:Regex1_Go(System.Text.RegularExpressions.RegexRunner)
9 ( 0.16% of base) : 15596.dasm - DynamicClass:_DynamicMethod9(byref,int):MicroBenchmarks.Serializers.MyEventsListerViewModel
Top method improvements (bytes):
-125 (-6.93% of base) : 27270.dasm - Benchstone.BenchI.Puzzle:DoIt():bool:this
-3 (-0.05% of base) : 13993.dasm - Jil.Deserialize.Methods:SkipWithLeadCharThunkReader(byref,int)
Top method regressions (percentages):
536 (18.29% of base) : 25723.dasm - Benchstone.BenchI.MulMatrix:Inner(System.Int32[][],System.Int32[][],System.Int32[][])
185 ( 2.21% of base) : 26877.dasm - Benchstone.BenchF.LLoops:Main1(int):this
66 ( 1.39% of base) : 16212.dasm - Jil.Deserialize.Methods:SkipWithLeadChar(System.IO.TextReader,int)
16 ( 0.36% of base) : 16196.dasm - DynamicClass:_DynamicMethod9(System.IO.TextReader,int):MicroBenchmarks.Serializers.MyEventsListerViewModel
13 ( 0.27% of base) : 13322.dasm - DynamicClass:Regex1_Go(System.Text.RegularExpressions.RegexRunner)
9 ( 0.16% of base) : 15596.dasm - DynamicClass:_DynamicMethod9(byref,int):MicroBenchmarks.Serializers.MyEventsListerViewModel
Top method improvements (percentages):
-125 (-6.93% of base) : 27270.dasm - Benchstone.BenchI.Puzzle:DoIt():bool:this
-3 (-0.05% of base) : 13993.dasm - Jil.Deserialize.Methods:SkipWithLeadCharThunkReader(byref,int)
8 total methods with Code Size differences (2 improved, 6 regressed), 0 unchanged.

---

Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 142908
Total bytes of diff: 143387
Total bytes of delta: 479 (0.34% of base)
Total relative delta: 0.42
diff is a regression.
relative diff is a regression.

Detail diffs

Top file regressions (bytes):
536 : 248723.dasm (18.29% of base)
390 : 220441.dasm (14.45% of base)
390 : 220391.dasm (14.61% of base)
150 : 239253.dasm (1.78% of base)
71 : 234803.dasm (1.72% of base)
16 : 225588.dasm (1.00% of base)
16 : 225590.dasm (1.00% of base)
5 : 225285.dasm (0.26% of base)
Top file improvements (bytes):
-359 : 215690.dasm (-0.99% of base)
-320 : 215701.dasm (-1.16% of base)
-128 : 215723.dasm (-0.73% of base)
-128 : 215666.dasm (-0.62% of base)
-125 : 239280.dasm (-6.93% of base)
-29 : 216754.dasm (-0.33% of base)
-3 : 225316.dasm (-0.15% of base)
-3 : 225313.dasm (-0.15% of base)
16 total files with Code Size differences (8 improved, 8 regressed), 0 unchanged.
Top method regressions (bytes):
536 (18.29% of base) : 248723.dasm - Benchstone.BenchI.MulMatrix:Inner(System.Int32[][],System.Int32[][],System.Int32[][])
390 (14.45% of base) : 220441.dasm - VectorTest:Main():int
390 (14.61% of base) : 220391.dasm - VectorTest:Main():int
150 ( 1.78% of base) : 239253.dasm - Benchstone.BenchF.LLoops:Main1(int):this
71 ( 1.72% of base) : 234803.dasm - SmallLoop1:Main():int
16 ( 1.00% of base) : 225588.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
16 ( 1.00% of base) : 225590.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
5 ( 0.26% of base) : 225285.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
Top method improvements (bytes):
-359 (-0.99% of base) : 215690.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-320 (-1.16% of base) : 215701.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-128 (-0.73% of base) : 215723.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-128 (-0.62% of base) : 215666.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-125 (-6.93% of base) : 239280.dasm - Benchstone.BenchI.Puzzle:DoIt():bool:this
-29 (-0.33% of base) : 216754.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-3 (-0.15% of base) : 225316.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-3 (-0.15% of base) : 225313.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
Top method regressions (percentages):
536 (18.29% of base) : 248723.dasm - Benchstone.BenchI.MulMatrix:Inner(System.Int32[][],System.Int32[][],System.Int32[][])
390 (14.61% of base) : 220391.dasm - VectorTest:Main():int
390 (14.45% of base) : 220441.dasm - VectorTest:Main():int
150 ( 1.78% of base) : 239253.dasm - Benchstone.BenchF.LLoops:Main1(int):this
71 ( 1.72% of base) : 234803.dasm - SmallLoop1:Main():int
16 ( 1.00% of base) : 225588.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
16 ( 1.00% of base) : 225590.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
5 ( 0.26% of base) : 225285.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
Top method improvements (percentages):
-125 (-6.93% of base) : 239280.dasm - Benchstone.BenchI.Puzzle:DoIt():bool:this
-320 (-1.16% of base) : 215701.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-359 (-0.99% of base) : 215690.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-128 (-0.73% of base) : 215723.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-128 (-0.62% of base) : 215666.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-29 (-0.33% of base) : 216754.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-3 (-0.15% of base) : 225316.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
-3 (-0.15% of base) : 225313.dasm - IntelHardwareIntrinsicTest.Program:Main(System.String[]):int
16 total methods with Code Size differences (8 improved, 8 regressed), 0 unchanged.

---

Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 15351
Total bytes of diff: 15448
Total bytes of delta: 97 (0.63% of base)
Total relative delta: 0.08
diff is a regression.
relative diff is a regression.

Detail diffs

Top file regressions (bytes):
71 : 63973.dasm (7.63% of base)
26 : 106039.dasm (0.19% of base)
2 total files with Code Size differences (0 improved, 2 regressed), 1 unchanged.
Top method regressions (bytes):
71 ( 7.63% of base) : 63973.dasm - Microsoft.Diagnostics.Tracing.Parsers.Symbol.FileVersionTraceData:ToXml(System.Text.StringBuilder):System.Text.StringBuilder:this
26 ( 0.19% of base) : 106039.dasm - Microsoft.VisualBasic.CompilerServices.VBBinder:BindToMethod(int,System.Reflection.MethodBase[],byref,System.Reflection.ParameterModifier[],System.Globalization.CultureInfo,System.String[],byref):System.Reflection.MethodBase:this
Top method regressions (percentages):
71 ( 7.63% of base) : 63973.dasm - Microsoft.Diagnostics.Tracing.Parsers.Symbol.FileVersionTraceData:ToXml(System.Text.StringBuilder):System.Text.StringBuilder:this
26 ( 0.19% of base) : 106039.dasm - Microsoft.VisualBasic.CompilerServices.VBBinder:BindToMethod(int,System.Reflection.MethodBase[],byref,System.Reflection.ParameterModifier[],System.Globalization.CultureInfo,System.String[],byref):System.Reflection.MethodBase:this
2 total methods with Code Size differences (0 improved, 2 regressed), 1 unchanged.

---

Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 22540
Total bytes of diff: 22576
Total bytes of delta: 36 (0.16% of base)
Total relative delta: 0.01
diff is a regression.
relative diff is a regression.

Detail diffs

Top file regressions (bytes):
23 : 104600.dasm (0.15% of base)
14 : 43143.dasm (0.47% of base)
Top file improvements (bytes):
-1 : 35267.dasm (-0.03% of base)
3 total files with Code Size differences (1 improved, 2 regressed), 0 unchanged.
Top method regressions (bytes):
23 ( 0.15% of base) : 104600.dasm - Microsoft.VisualBasic.CompilerServices.VBBinder:BindToMethod(int,System.Reflection.MethodBase[],byref,System.Reflection.ParameterModifier[],System.Globalization.CultureInfo,System.String[],byref):System.Reflection.MethodBase:this
14 ( 0.47% of base) : 43143.dasm - Microsoft.CodeAnalysis.CSharp.Symbols.SourceMemberContainerTypeSymbol:ForceComplete(Microsoft.CodeAnalysis.SourceLocation,System.Threading.CancellationToken):this
Top method improvements (bytes):
-1 (-0.03% of base) : 35267.dasm - Microsoft.CodeAnalysis.CSharp.Syntax.InternalSyntax.LanguageParser:ParseNamespaceBody(byref,byref,byref,ushort):this
Top method regressions (percentages):
14 ( 0.47% of base) : 43143.dasm - Microsoft.CodeAnalysis.CSharp.Symbols.SourceMemberContainerTypeSymbol:ForceComplete(Microsoft.CodeAnalysis.SourceLocation,System.Threading.CancellationToken):this
23 ( 0.15% of base) : 104600.dasm - Microsoft.VisualBasic.CompilerServices.VBBinder:BindToMethod(int,System.Reflection.MethodBase[],byref,System.Reflection.ParameterModifier[],System.Globalization.CultureInfo,System.String[],byref):System.Reflection.MethodBase:this
Top method improvements (percentages):
-1 (-0.03% of base) : 35267.dasm - Microsoft.CodeAnalysis.CSharp.Syntax.InternalSyntax.LanguageParser:ParseNamespaceBody(byref,byref,byref,ushort):this
3 total methods with Code Size differences (1 improved, 2 regressed), 0 unchanged.

---

Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 53782
Total bytes of diff: 53837
Total bytes of delta: 55 (0.10% of base)
Total relative delta: 0.00
diff is a regression.
relative diff is a regression.

Detail diffs

Top file regressions (bytes):
28 : 28889.dasm (0.16% of base)
27 : 28887.dasm (0.15% of base)
2 total files with Code Size differences (0 improved, 2 regressed), 2 unchanged.
Top method regressions (bytes):
28 ( 0.16% of base) : 28889.dasm - ManagedTests.DynamicCSharp.Conformance.dynamic.statements.freach.freach003.freach003.Test:MainMethod():int
27 ( 0.15% of base) : 28887.dasm - ManagedTests.DynamicCSharp.Conformance.dynamic.statements.freach.freach004.freach004.Test:MainMethod():int
Top method regressions (percentages):
28 ( 0.16% of base) : 28889.dasm - ManagedTests.DynamicCSharp.Conformance.dynamic.statements.freach.freach003.freach003.Test:MainMethod():int
27 ( 0.15% of base) : 28887.dasm - ManagedTests.DynamicCSharp.Conformance.dynamic.statements.freach.freach004.freach004.Test:MainMethod():int
2 total methods with Code Size differences (0 improved, 2 regressed), 2 unchanged.

---

[BruceForstall]

/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress, runtime-coreclr outerloop

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[BruceForstall]

@AndyAyersMS @dotnet/jit-contrib PTAL

[BruceForstall]

If anyone has an opinion on how the "best" number should be chosen, I'd be happy to hear it. Making it dynamic perhaps would be an option as well: not have a maximum. Not sure if there are algorithms that would need to be reconsidered to avoid bad behavior for large numbers.

[kunalspathak]

Curious - does choosing 32 not showing enough wins? Can we expose a COMPlus_MaxLoops or something that will help us experiment more? Is there a TP impact?

[AndyAyersMS]

Does 64 cover all the SPMI cases...?

This might be a place where a more extensive SPMI collection would prove valuable.

[kunalspathak]

This might be a place where a more extensive SPMI collection would prove valuable.

Could you elaborate on that?

[AndyAyersMS]

This might be a place where a more extensive SPMI collection would prove valuable.

Could you elaborate on that?

I'm referring to some of the internal collections we had at one point, over much larger amounts of code.

[BruceForstall]

I enabled COUNT_LOOPS and ran it across a merged spmi mega-collection (and bumped the number of histogram buckets), and get:

---------------------------------------------------
Loop stats
---------------------------------------------------
Total number of methods with loops is 64366
Total number of loops is 92730
Maximum number of loops per method is 192
# of methods overflowing nat loop table is 49
Total number of 'unnatural' loops is 135600
# of methods overflowing unnat loop limit is 0
Total number of loops with an iterator is 32376
Total number of loops with a simple iterator is 32376
Total number of loops with a constant iterator is 5515
--------------------------------------------------
Loop count frequency table:
--------------------------------------------------
<= 0 ===> 11043 count ( 14% of total)
1 .. 1 ===> 48707 count ( 79% of total)
2 .. 2 ===> 10489 count ( 93% of total)
3 .. 3 ===> 2954 count ( 97% of total)
4 .. 4 ===> 932 count ( 98% of total)
5 .. 5 ===> 496 count ( 98% of total)
6 .. 6 ===> 268 count ( 99% of total)
7 .. 7 ===> 118 count ( 99% of total)
8 .. 8 ===> 95 count ( 99% of total)
9 .. 9 ===> 63 count ( 99% of total)
10 .. 10 ===> 50 count ( 99% of total)
11 .. 11 ===> 44 count ( 99% of total)
12 .. 12 ===> 19 count ( 99% of total)
13 .. 13 ===> 23 count ( 99% of total)
14 .. 14 ===> 34 count ( 99% of total)
15 .. 15 ===> 12 count ( 99% of total)
16 .. 16 ===> 13 count ( 99% of total)
17 .. 17 ===> 2 count ( 99% of total)
18 .. 18 ===> 9 count ( 99% of total)
19 .. 19 ===> 1 count ( 99% of total)
20 .. 20 ===> 6 count ( 99% of total)
21 .. 21 ===> 1 count ( 99% of total)
22 .. 22 ===> 1 count ( 99% of total)
23 .. 23 ===> 0 count ( 99% of total)
24 .. 24 ===> 1 count ( 99% of total)
25 .. 25 ===> 4 count ( 99% of total)
26 .. 26 ===> 1 count ( 99% of total)
27 .. 27 ===> 0 count ( 99% of total)
28 .. 27 ===> 0 count ( 99% of total)
28 .. 29 ===> 6 count ( 99% of total)
30 .. 30 ===> 2 count ( 99% of total)
31 .. 31 ===> 0 count ( 99% of total)
32 .. 32 ===> 0 count ( 99% of total)
33 .. 33 ===> 0 count ( 99% of total)
34 .. 34 ===> 0 count ( 99% of total)
35 .. 35 ===> 1 count ( 99% of total)
36 .. 36 ===> 2 count ( 99% of total)
37 .. 37 ===> 0 count ( 99% of total)
38 .. 38 ===> 0 count ( 99% of total)
39 .. 39 ===> 0 count ( 99% of total)
40 .. 40 ===> 0 count ( 99% of total)
41 .. 41 ===> 0 count ( 99% of total)
42 .. 42 ===> 0 count ( 99% of total)
43 .. 43 ===> 0 count ( 99% of total)
44 .. 44 ===> 0 count ( 99% of total)
45 .. 45 ===> 2 count ( 99% of total)
46 .. 46 ===> 2 count ( 99% of total)
47 .. 47 ===> 0 count ( 99% of total)
48 .. 48 ===> 1 count ( 99% of total)
49 .. 49 ===> 0 count ( 99% of total)
50 .. 50 ===> 0 count ( 99% of total)
51 .. 51 ===> 0 count ( 99% of total)
52 .. 52 ===> 0 count ( 99% of total)
53 .. 53 ===> 0 count ( 99% of total)
54 .. 54 ===> 0 count ( 99% of total)
55 .. 55 ===> 0 count ( 99% of total)
56 .. 56 ===> 0 count ( 99% of total)
57 .. 57 ===> 0 count ( 99% of total)
58 .. 58 ===> 0 count ( 99% of total)
59 .. 59 ===> 1 count (100% of total)
60 .. 60 ===> 0 count (100% of total)
> 60 ===> 6 count (100% of total)
--------------------------------------------------
Loop exit count frequency table:
--------------------------------------------------
<= 0 ===> 126 count ( 0% of total)
1 .. 1 ===> 57237 count ( 64% of total)
2 .. 2 ===> 21195 count ( 87% of total)
3 .. 3 ===> 5175 count ( 93% of total)
4 .. 4 ===> 2758 count ( 96% of total)
5 .. 5 ===> 2164 count ( 99% of total)
6 .. 6 ===> 892 count (100% of total)
> 6 ===> 3183 count (103% of total)
--------------------------------------------------

So, sticking with powers of 2, 32 only misses ~9 functions with > 32 loops. Even 16 max loops per function hits 99% of all loops (no surprise). Of course there's some crazy outlier: 192 loops in (at least) one function.

The stats for just the benchmarks is:

---------------------------------------------------
Loop stats
---------------------------------------------------
Total number of methods with loops is 2679
Total number of loops is 4326
Maximum number of loops per method is 61
# of methods overflowing nat loop table is 10
Total number of 'unnatural' loops is 4822
# of methods overflowing unnat loop limit is 0
Total number of loops with an iterator is 1534
Total number of loops with a simple iterator is 1534
Total number of loops with a constant iterator is 398
--------------------------------------------------
Loop count frequency table:
--------------------------------------------------
<= 0 ===> 213 count ( 7% of total)
1 .. 1 ===> 1984 count ( 75% of total)
2 .. 2 ===> 415 count ( 90% of total)
3 .. 3 ===> 144 count ( 95% of total)
4 .. 4 ===> 59 count ( 97% of total)
5 .. 5 ===> 29 count ( 98% of total)
6 .. 6 ===> 13 count ( 98% of total)
7 .. 7 ===> 5 count ( 98% of total)
8 .. 8 ===> 1 count ( 99% of total)
9 .. 9 ===> 2 count ( 99% of total)
10 .. 10 ===> 1 count ( 99% of total)
11 .. 11 ===> 1 count ( 99% of total)
12 .. 12 ===> 2 count ( 99% of total)
13 .. 13 ===> 5 count ( 99% of total)
14 .. 14 ===> 5 count ( 99% of total)
15 .. 15 ===> 3 count ( 99% of total)
16 .. 16 ===> 0 count ( 99% of total)
17 .. 17 ===> 0 count ( 99% of total)
18 .. 18 ===> 2 count ( 99% of total)
19 .. 19 ===> 1 count ( 99% of total)
20 .. 20 ===> 1 count ( 99% of total)
21 .. 21 ===> 0 count ( 99% of total)
22 .. 22 ===> 0 count ( 99% of total)
23 .. 23 ===> 0 count ( 99% of total)
24 .. 24 ===> 0 count ( 99% of total)
25 .. 25 ===> 0 count ( 99% of total)
26 .. 26 ===> 0 count ( 99% of total)
27 .. 27 ===> 0 count ( 99% of total)
28 .. 27 ===> 0 count ( 99% of total)
28 .. 29 ===> 1 count ( 99% of total)
30 .. 30 ===> 1 count ( 99% of total)
31 .. 31 ===> 0 count ( 99% of total)
32 .. 32 ===> 0 count ( 99% of total)
33 .. 33 ===> 0 count ( 99% of total)
34 .. 34 ===> 0 count ( 99% of total)
35 .. 35 ===> 0 count ( 99% of total)
36 .. 36 ===> 1 count ( 99% of total)
37 .. 37 ===> 0 count ( 99% of total)
38 .. 38 ===> 0 count ( 99% of total)
39 .. 39 ===> 0 count ( 99% of total)
40 .. 40 ===> 0 count ( 99% of total)
41 .. 41 ===> 0 count ( 99% of total)
42 .. 42 ===> 0 count ( 99% of total)
43 .. 43 ===> 0 count ( 99% of total)
44 .. 44 ===> 0 count ( 99% of total)
45 .. 45 ===> 1 count ( 99% of total)
46 .. 46 ===> 0 count ( 99% of total)
47 .. 47 ===> 0 count ( 99% of total)
48 .. 48 ===> 0 count ( 99% of total)
49 .. 49 ===> 0 count ( 99% of total)
50 .. 50 ===> 0 count ( 99% of total)
51 .. 51 ===> 0 count ( 99% of total)
52 .. 52 ===> 0 count ( 99% of total)
53 .. 53 ===> 0 count ( 99% of total)
54 .. 54 ===> 0 count ( 99% of total)
55 .. 55 ===> 0 count ( 99% of total)
56 .. 56 ===> 0 count ( 99% of total)
57 .. 57 ===> 0 count ( 99% of total)
58 .. 58 ===> 0 count ( 99% of total)
59 .. 59 ===> 1 count (100% of total)
60 .. 60 ===> 0 count (100% of total)
> 60 ===> 1 count (100% of total)
--------------------------------------------------
Loop exit count frequency table:
--------------------------------------------------
<= 0 ===> 0 count ( 0% of total)
1 .. 1 ===> 2426 count ( 58% of total)
2 .. 2 ===> 995 count ( 82% of total)
3 .. 3 ===> 330 count ( 90% of total)
4 .. 4 ===> 237 count ( 96% of total)
5 .. 5 ===> 96 count ( 98% of total)
6 .. 6 ===> 68 count (100% of total)
> 6 ===> 174 count (104% of total)
--------------------------------------------------

Unfortunately, the spmi collections, especially the benchmarks collection, have a lot of "MISSING" data currently, so it's not clear how that's skewing the data.

[BruceForstall]

Interestingly, switching from 64 to 32 max loops still gives MulMatrix 9% improvement, but puzzle regresses by 6%!

(puzzle has 45 loops in its main function)

[BruceForstall]

As for throughput: a PIN spmi run with max loops = 64 shows no statistically significant TP difference.

[BruceForstall]

Test failures are all infra or known issues

[AndyAyersMS]

LGTM