[LoongArch64] add new ABI classifier.

[shushanhf]

Add new ABI classifier for LoongArch64. #100744

The first patch is only the LoongArch64Classifier.
The CodeGen::genHomeRegisterParams will be considered later whether to uniform within the codegencommon.cpp.

[dotnet-policy-service]

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

[jakobbotsch]

There are some formatting errors as well. Can you please fix those?

[jakobbotsch]

LGTM. Thank you!

[shushanhf]

Hi, @jakobbotsch
After LA64 using the CodeGen::genHomeRegisterParams within codegencommon.cpp, although I had some tests passed on LA64 and more tests are still in progressing.
I still am not sure here: the line-3335 and line-3041 within codegencommon.cpp.

Can you give some advices ?

Besides, althought the split args within the LoongArch64Classifier::Classify is ok, I also have to check the split args within CodeGen::genHomeRegisterParams.

[jakobbotsch]

Hi, @jakobbotsch After LA64 using the CodeGen::genHomeRegisterParams within codegencommon.cpp, although I had some tests passed on LA64 and more tests are still in progressing. I still am not sure here: the line-3335 and line-3041 within codegencommon.cpp.

Can you give some advices ?

Line 3335: This is handling to support SIMD types. The JIT can enregister a SIMD type parameter in a single vector register while it may be passed in multiple registers. For example, on arm64:

public static float Test(float x, Vector3 v)
{
    return v.Length();
}

Parameter V00 ABI info: [00..04) reg d0
Parameter V01 ABI info: 3 segments
  [0] [00..04) reg d1
  [1] [04..08) reg d2
  [2] [08..12) reg d3
...
*************** In genHomeRegisterParams()
4 registers in register parameter interference graph
  d1
  d0
    <- d3 (offset: 8)
    <- d2 (offset: 4)
    <- d1
  d2
  d3
IN000b:             fmov    s0, s1
IN000c:             mov     v0.s[2], v3.s[0]
IN000d:             mov     v0.s[1], v2.s[0]

I think you don't need this yet for LA64.

Line 3041: This is purely an optimization for arm64 to make stores to the stack always the same size, so that we can use stp more often. It doesn't affect correctness.

[shushanhf]

Thanks very much.

[shushanhf]

Besides, althought the split args within the LoongArch64Classifier::Classify is ok, I also have to check the split args within CodeGen::genHomeRegisterParams.

#101340 (comment)