Use SIMD for block inits with GC fields

[EgorBo]

Closes #83297

Currently, we conservatively give up on using SIMDs for structs (blocks) with GC references since SIMD (on x86/64 arch) has certain requirements to provide atomicity guarantees (data needs to be 16 byte aligned, CPU must support AVX2 and we must use aligned store instructions). Let's at least use SIMD for continuous non-GC parts of such structs, example:

struct MyStruct { string gc1; long a; long b; long c; long d; long e; long f; long g; long h; } MyStruct Test() => new();

Codegen diff:

; Method Bench:Test():Bench+MyStruct:this (FullOpts) xor eax, eax mov qword ptr [rdx], rax ;; <-- GC slot - mov qword ptr [rdx+0x08], rax - mov qword ptr [rdx+0x10], rax - mov qword ptr [rdx+0x18], rax - mov qword ptr [rdx+0x20], rax - mov qword ptr [rdx+0x28], rax - mov qword ptr [rdx+0x30], rax - mov qword ptr [rdx+0x38], rax - mov qword ptr [rdx+0x40], rax + vxorps xmm0, xmm0, xmm0 + vmovdqu32 zmmword ptr [rdx+0x08], zmm0 mov rax, rdx ret -; Total bytes of code: 41 +; Total bytes of code: 23

[EgorBo]

@EgorBot

using BenchmarkDotNet.Attributes; using BenchmarkDotNet.Running; BenchmarkRunner.Run<Bench>(args: args); public class Bench { public struct MyStructWithGC { object gc; long a; long b; long c; long d; long e; long f; long g; long h; } MyStructWithGC _fld1; MyStructWithGC _fld2; [Benchmark] public void Zeroing() { _fld1 = default; _fld2 = default; } }

[EgorBot]

@EgorBo

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish) AMD EPYC 7763, 1 CPU, 2 logical cores and 1 physical core .NET SDK 9.0.100-preview.3.24204.13 [Host] : .NET 9.0.0 (9.0.24.17209), X64 RyuJIT AVX2 Job-BWFEHP : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2 Job-NQQMIR : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2 

Method	Toolchain	Mean	Ratio
Zeroing	/Main/corerun	1.9355 ns	1.00
Zeroing	/PR/corerun	0.4288 ns	0.22

[EgorBo]

@jakobbotsch @kunalspathak @dotnet/jit-contrib PTAL, diffs aren't too big

[EgorBo]

Ping @jakobbotsch @kunalspathak @dotnet/jit-contrib 🙂 #102209 depends on it (for simpler diffs)

[jakobbotsch]

It would be nice to support it for copies as well (I think that would fix #90196 and #7469)

[kunalspathak]

LGTM