JIT: Fixed containment of STOREIND of HW intrinsics ConvertTo*

[TIHan]

Resolves #92349 regarding the truncating CAST

Description

As an example, the HW Intrinsic SIMD scalar ConvertToInt32 will emit a vmovd instruction. This instruction can store a scalar value in memory or in a register, but only for 32 and 64 bits.

So, when we have something like this:

// pValue is a 'ushort*'
*pValue = (ushort)(uint)Sse2.ConvertToInt32(newMinZ16);

Acceptance Criteria

• Add test case

Note

We need to backport this to .NET 8.

[TIHan]

cc @tannergooding

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Resolves #92349 regarding the truncating CAST

Description

As an example, the HW Intrinsic SIMD scalar ConvertToInt32 will emit a vmovd instruction. This instruction can store a scalar value in memory or in a register, but only for 32 and 64 bits.

So, when we have something like this:

// pValue is a 'ushort*'
*pValue = (ushort)(uint)Sse2.ConvertToInt32(newMinZ16);

We need to keep the cast operation for int -> short, otherwise we will get an incorrect result which is what #92349 was telling us.

I tried to see if there was something we could do in codegen, but it is a lot easier to handle in morph where we get rid of these CASTs. The simplest thing to do is check to see if we have a SIMD scalar operation, and if we do, do not remove the CAST.

Acceptance Criteria

• Add test case

Author:	TIHan
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[jakobbotsch]

We need to keep the cast operation for int -> short, otherwise we will get an incorrect result which is what #92349 was telling us.

I tried to see if there was something we could do in codegen, but it is a lot easier to handle in morph where we get rid of these CASTs. The simplest thing to do is check to see if we have a SIMD scalar operation and we are storing the result for a small type, then do not remove the CAST.

We should fix this in the backend. STOREIND<ushort> should never be storing more than two bytes. With this PR the bug still exists, it's just harder to expose (but can definitely be done, e.g. you can write IL manually that doesn't create these casts in the first place).

It looks like a bug in ContainCheckStoreIndir. This line:

runtime/src/coreclr/jit/lowerxarch.cpp

Line 6507 in aba7c43

isContainable = varTypeIsIntegral(simdBaseType);

should be

isContainable = varTypeIsIntegral(simdBaseType) && (genTypeSize(src) == genTypeSize(node));

if vmovd has no 8/16 bit forms.

[TIHan]

We should fix this in the backend

Your fix works and is a lot simpler than what I had. Thank you.

[jakobbotsch]

Add a test?

[TIHan]

@dotnet/jit-contrib @jakobbotsch @tannergooding This is ready, pending CI.

[TIHan]

/backport to release/8.0

[github-actions]

Started backporting to release/8.0: https://github.com/dotnet/runtime/actions/runs/6280549647