Add ref T overloads for hardware intrinsic functions which takes T*

[RamType0]

Currently,these are some hardware intrinsic functions which takes T*.
But requiring T* means requires pinning.
Obliviously,hardware intrinsic functions are used for performance,
but pinning will make lower performance.

[Dotnet-GitSync-Bot]

I couldn't figure out the best area label to add to this issue. Please help me learn by adding exactly one area label.

[john-h-k]

cc @tannergooding

These were, iirc, explicitly disregarded because they want intrinsics to have to be explicit in nature.

You can workaround it with reference casting, which generates pretty great asm

public Vector128<T> ReadVector128<T>(ref T start) where T : struct
    => Unsafe.As<T, Vector128<T>>(ref start);

You can see this approach being used in System.Buffers.Text.Base64

related convo:

Me:
Is there a way to load vectors through references rather than pointers?

Tanner:
No
by design
Primarily because reading more than a T from a ref is inherently unsafe

Tagging subscribers to this area: @tannergooding
Notify danmosemsft if you want to be subscribed.

[tannergooding]

I thought we had an existing issue for this, but I can't seem to find it right now...

but pinning will make lower performance.

Pinning is relatively low overhead, it effectively just involves storing a value to the stack and zeroing that location when done. While it can have overhead if you are pinning in a tight loop, it is no more expensive than stack spilling in general. Given the most likely scenario with HWIntrinsics is you are working with large amounts of data, pinning your data once upfront should be negligible in cost. Pinning when you have smaller amounts of data may have a larger relative impact, but it would ideally still not be noticeable due to being on small amounts of data. Instead, it would likely only be an issue if you are working with a large number of small inputs.

The downside to pinning is that it can inhibit the GC and indirectly impact performance if a compaction happens. However, ref arithmetic requires more verbose patterns that can impact readability/maintainability in already complex code. Exposing Sse.LoadVector128(ref float value) wouldn't help with that as you would still need to understand that:

1. ref float value loads 4 floats, not 1
2. Your inner loop needs to change to use Unsafe.Add<T>(ref T source, int elementOffset) to move to the next entry

This could somewhat be mitigated by taking a ref Vector128<float>, but it is still not ideal and you are more likely to have a float[] as input (otherwise you don't need to use LoadVector128 in the first place, you can just dereference the Vector128) and would require additional Unsafe.As<T, Vector128<T>>(ref T value) casts.

You could mitigate the ref arithmetic issue by exposing Sse.LoadVector128(ref T value, nint elementOffset), but that likely has its own drawbacks as well and doesn't necessarily address that the ref float really must be 4x float (or that ref byte must be 16x byte).

Last time I discussed this with @jkotas and @CarolEidt, it was determined that exposing overloads that took ref was not likely worthwhile at the time. I'm not aware of that stance changing.

[RamType0]

I also meant System.Runtime.Intrinsics.X86.Avx2.GatherXXX.

Me:
Is there a way to load vectors through references rather than pointers?

Tanner:
No
by design
Primarily because reading more than a T from a ref is inherently unsafe

I think hardware intrinsic is already inherently unsafe.

BTW,Unsafe.As<T,Vector> approach could cause DataMisalignedException on ARM platform.
You should use Unsafe.ReadUnaligned instead.
This approach is used for Span.IndexOf<byte>.
This code is trying to align vector manually.But we don't know when GC happens,and GC breaks alignment,we still need to use Unsafe.ReadUnaligned to load vector.

[GrabYourPitchforks]

I think hardware intrinsic is already inherently unsafe.

To the best of my knowledge, the only APIs in System.Runtime.Intrinsics that are unsafe are those that explicitly take pointers (and so you need to use an unsafe block to call them) or those that explicitly have Unsafe in the method name (see example). All other APIs should be safe unless I'm missing something.

This is the same reason we don't use ref byte to refer to a buffer in any of our public-facing APIs, with specific exceptions like Unsafe.ReadUnaligned<T>(ref byte).

[RamType0]

I think hardware intrinsic is already inherently unsafe.

To the best of my knowledge, the only APIs in System.Runtime.Intrinsics that are unsafe are those that explicitly take pointers (and so you need to use an unsafe block to call them) or those that explicitly have Unsafe in the method name (see example). All other APIs should be safe unless I'm missing something.

Yeah,ref T overload could be named Unsafe.

This is the same reason we don't use ref byte to refer to a buffer in any of our public-facing APIs, with specific exceptions like Unsafe.ReadUnaligned<T>(ref byte).

So,why not intrinsic functions become exception?
Most scenarios like using Unsafe class are starving for performance, even if it is unsafe.
These scenarios are very close to scenario where intrinsics are used.

Pinning is relatively low overhead, it effectively just involves storing a value to the stack and zeroing that location when done. While it can have overhead if you are pinning in a tight loop,

If we could avoid pinning in a tight loop,we could use SIMD operations effectively with extern methods.
This restriction will make lower advantages of intrinsics.

[GrabYourPitchforks]

Can you provide concrete examples (with pseudo-code) of scenarios that would benefit from a ref byte API being added to the intrinsics surface area? We're using the existing intrinsics APIs very effectively today and have not yet seen a need to move away from using pointers.

[tannergooding]

@GrabYourPitchforks, we use Unsafe.As and ref in a couple places in CoreCLR itself: https://source.dot.net/#System.Memory/System/Buffers/Text/Base64.cs,e2c2e333f97cbd35,references
IIRC, it was done to avoid GC overhead from long term or frequent pinning.

The use-case definitely exists, I think it's just a question of whether the existing support via Unsafe is fine or if exposing explicit overloads that take ref T, int elementIndex is worthwhile.
It's definitely not a lot of work to expose after all as it's basically just approving the APIs and adding them to a table.

[GrabYourPitchforks]

Right, I'm aware of our internal use cases. My main point was that the existing API surface (even if you have to mix + match API calls) appears to be sufficient for our needs. I can't think offhand of a scenario which is blocked because of missing APIs.

[john-h-k]

@GrabYourPitchforks the one thing i would say is that Unsafe methods don't have the same guarantees on other runtimes or whatever. Whereas you know you are getting the desired instruction with an actual intrinsic