Unexpected performance difference with strange hack

[Lanayx]

Description

The following hack leads to 25% perf improvement

let mutable i = 0
while i < N do
    ...
    i <- i + 1

can be changed to

let mutable i = 0
while i < N do
    ...
    let mutable x = 1
    i <- i + x

// without hack:
// BenchmarkDotNet v0.14.0, Windows 11 (10.0.22631.4391/23H2/2023Update/SunValley3)
// AMD Ryzen 5 5600H with Radeon Graphics, 1 CPU, 12 logical and 6 physical cores
// .NET SDK 9.0.100
//   [Host]     : .NET 9.0.0 (9.0.24.52809), X64 RyuJIT AVX2 DEBUG
//   DefaultJob : .NET 9.0.0 (9.0.24.52809), X64 RyuJIT AVX2
//
//
// | Method     | Mean     | Error   | StdDev  | Gen0     | Gen1     | Gen2     | Allocated |
// |----------- |---------:|--------:|--------:|---------:|---------:|---------:|----------:|
// | WebUtility | 695.5 us | 4.14 us | 3.87 us | 332.0313 | 332.0313 | 332.0313 |   1370 KB |
// | Tools      | 668.1 us | 2.86 us | 2.39 us |  83.9844 |  68.3594 |        - |  691.6 KB |

// with hack:
// BenchmarkDotNet v0.14.0, Windows 11 (10.0.22631.4391/23H2/2023Update/SunValley3)
// AMD Ryzen 5 5600H with Radeon Graphics, 1 CPU, 12 logical and 6 physical cores
// .NET SDK 9.0.100
//   [Host]     : .NET 9.0.0 (9.0.24.52809), X64 RyuJIT AVX2 DEBUG
//   DefaultJob : .NET 9.0.0 (9.0.24.52809), X64 RyuJIT AVX2
//
//
// | Method     | Mean     | Error   | StdDev  | Gen0     | Gen1     | Gen2     | Allocated |
// |----------- |---------:|--------:|--------:|---------:|---------:|---------:|----------:|
// | WebUtility | 696.0 us | 2.37 us | 2.10 us | 332.0313 | 332.0313 | 332.0313 |   1370 KB |
// | Tools      | 498.0 us | 3.03 us | 2.83 us |  83.9844 |  68.3594 |        - |  691.6 KB |

Configuration

.NET SDK:
 Version:           9.0.100
 Commit:            59db016f11
 Workload version:  9.0.100-manifests.c6f19616
 MSBuild version:   17.12.7+5b8665660

Runtime Environment:
 OS Name:     Windows
 OS Version:  10.0.22631
 OS Platform: Windows
 RID:         win-x64
 Base Path:   C:\Program Files\dotnet\sdk\9.0.100\

.NET workloads installed:
There are no installed workloads to display.
Configured to use loose manifests when installing new manifests.

Host:
  Version:      9.0.0
  Architecture: x64
  Commit:       9d5a6a9aa4

Regression?

No, it's reproduced both with .NET8 and .NET9

Data

Repro is located here: Oxpecker, perfhack branch

1. Clone repository
2. Run tests/PerfTest project
3. Observe results
4. Remove dirty hack
5. Run test again and observe results

Analysis

I don't understand why difference occurs, some people say it's variable hoisting, but I couldn't find detailed explanation.

[etki]

Sharplab reports the following difference for a simple example:

  IL_001e: call int32 _::get_i()
  IL_0023: ldc.i4.1
  IL_0024: add
  IL_0025: call void _::set_i(int32)

vs.

  IL_0048: ldc.i4.1
> IL_0049: stloc.0
  IL_004a: call int32 _::get_i()
> IL_004f: ldloc.0
  IL_0050: add
  IL_0051: call void _::set_i(int32)

Given that only one of the examples above gets improvement, and that the JIT output is identical, i think this could be just a budgeting issue

Reference loop:

    L0017: cmp dword ptr [eax+4], 0xc9
    L001e: jge short L0039
    L0020: mov edx, [eax+8]
    L0023: add edx, [eax+4]
    L0026: mov [eax+8], edx
    L0029: mov edx, [eax+4]
    L002c: inc edx
    L002d: mov [eax+4], edx
    L0030: cmp dword ptr [eax+4], 0xc9
    L0037: jl short L0020

Only labels and limit are different in the hacked loop placed below the reference one:

    L0039: cmp dword ptr [eax+4], 0x192
    L0040: jge short L005b
    L0042: mov edx, [eax+8]
    L0045: add edx, [eax+4]
    L0048: mov [eax+8], edx
    L004b: mov edx, [eax+4]
    L004e: inc edx
    L004f: mov [eax+4], edx
    L0052: cmp dword ptr [eax+4], 0x192
    L0059: jl short L0042