JIT: PGO value-profiling for non-constant zero-init stackalloc

[EgorBo]

Revive my old PR that tried to do this, now that #127959 is merged

Instrument stackallocs of variable size for PGO + Generalize value probing so we can reuse it for other future ideas.

void Test(int len)
{
    Span<byte> buffer = stackalloc byte[len];
    Consume(buffer);
}

is optimized into additional fast path for most common length:

       cmp      rax, 100
       je       SHORT G_M44326_IG06
       ...
G_M44326_IG06:  ;; offset=0x004F
       test     dword ptr [rsp], esp
       sub      rsp, 112
       lea      rax, [rsp+0x20]
       vxorps   ymm0, ymm0, ymm0
       vmovdqu32 zmmword ptr [rax], zmm0
       vmovdqu32 zmmword ptr [rax+0x30], zmm0

when we call it mostly for the same size (100 in this example) so we zero it with just 3 AVX512 instructions instead of slow loop

Benchmark

using System.Runtime.CompilerServices;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(typeof(Bench).Assembly).Run(args);

public class Bench
{
    [Benchmark] public void Stackalloc32()     => Consume(stackalloc byte[GetVar(32)]);
    [Benchmark] public void Stackalloc40()     => Consume(stackalloc byte[GetVar(40)]);
    [Benchmark] public void Stackalloc50()     => Consume(stackalloc byte[GetVar(50)]);
    [Benchmark] public void Stackalloc64()     => Consume(stackalloc byte[GetVar(64)]);
    [Benchmark] public void Stackalloc128()    => Consume(stackalloc byte[GetVar(128)]);
    [Benchmark] public void Stackalloc200()    => Consume(stackalloc byte[GetVar(200)]);
    [Benchmark] public void Stackalloc256()    => Consume(stackalloc byte[GetVar(256)]);
    [Benchmark] public void Stackalloc512()    => Consume(stackalloc byte[GetVar(512)]);
    [Benchmark] public void Stackalloc1024()   => Consume(stackalloc byte[GetVar(1024)]);
    [Benchmark] public void Stackalloc16384()  => Consume(stackalloc byte[GetVar(16384)]);
    [Benchmark] public void Stackalloc524288() => Consume(stackalloc byte[GetVar(524288)]);

    [MethodImpl(MethodImplOptions.NoInlining)] static int GetVar(int a) => a;
    [MethodImpl(MethodImplOptions.NoInlining)] static void Consume(Span<byte> buffer) { }
}

Results

Speedup of PR vs main (>1.00× = PR is faster). Sizes <= 32 bytes are intentionally not specialized (variable-size loop is already cheap there).

Size	Linux x64 (EPYC 9V45)	Apple M4 arm64
32	1.00×	0.92×
40	1.25×	3.23×
50	1.37×	2.32×
64	1.34×	2.28×
128	1.77×	0.87×
200	2.26×	2.03×
256	2.46×	0.89×
512	3.76×	1.34×
1,024	2.43×	1.27×
16,384	2.47×	1.69×
524,288	2.54×	1.24×

Raw runs: EgorBot/Benchmarks#197.

Wins are large on Linux x64 across the board (up to 3.76× at 512 bytes). On Apple M4 the wins are also significant for non-power-of-two sizes (40 / 50 / 64 / 200 -> 2-3×), with two minor regressions at 128 and 256 (-13% / -11%) and parity-or-better elsewhere.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 4 comments.

[EgorBo]

Note

AI-generated benchmark (Copilot CLI). Re-run with larger size set after skipping <= 32 byte specialization.

@EgorBot -linux_amd -osx_arm64

using System.Runtime.CompilerServices;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(typeof(Bench).Assembly).Run(args);

public class Bench
{
    [Benchmark] public void Stackalloc32()     => Consume(stackalloc byte[GetVar(32)]);
    [Benchmark] public void Stackalloc40()     => Consume(stackalloc byte[GetVar(40)]);
    [Benchmark] public void Stackalloc50()     => Consume(stackalloc byte[GetVar(50)]);
    [Benchmark] public void Stackalloc64()     => Consume(stackalloc byte[GetVar(64)]);
    [Benchmark] public void Stackalloc128()    => Consume(stackalloc byte[GetVar(128)]);
    [Benchmark] public void Stackalloc200()    => Consume(stackalloc byte[GetVar(200)]);
    [Benchmark] public void Stackalloc256()    => Consume(stackalloc byte[GetVar(256)]);
    [Benchmark] public void Stackalloc512()    => Consume(stackalloc byte[GetVar(512)]);
    [Benchmark] public void Stackalloc1024()   => Consume(stackalloc byte[GetVar(1024)]);
    [Benchmark] public void Stackalloc16384()  => Consume(stackalloc byte[GetVar(16384)]);
    [Benchmark] public void Stackalloc524288() => Consume(stackalloc byte[GetVar(524288)]);

    [MethodImpl(MethodImplOptions.NoInlining)] static int GetVar(int a) => a;
    [MethodImpl(MethodImplOptions.NoInlining)] static void Consume(Span<byte> buffer) { }
}

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 7 comments.