Revert potential UB due to aliasing + more WB removals

[EgorBo]

Remove potential UB from #111576 (move it to importer) + enable more cases where we eliminate write barriers, e.g. this:

public ref struct TestS
{
    public MyStruct s1;
    public MyStruct s2;

    public void Test()
    {
        s1 = s2; // no barrier needed, but we generate (two)
    }
}

public struct MyStruct // Not byref-like
{
    public string A;
    public string B;
}

[dotnet-policy-service]

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

[EgorBo]

@MihuBot

[EgorBo]

Some regressions with barrier-free code e.g. https://www.diffchecker.com/Sz7TI32G/

(when we save a nonbyref-like struct with gc pointers to a field of byreflike struct)

[jakobbotsch]

Is there any benefit to adding something like

if (m_store->AsBlk()->GetLayout()->IsStackOnly())
  indirFlags |= GTF_IND_TGT_NOT_HEAP;

here (and in morphblock.cpp)? Or will m_store already have this flag set for this case?

[EgorBo]

@jakobbotsch I don't see diffs, but this PR already produces too many missing contexts 🙂
Presumably, this is not enough, because some GT_STORE_BLK have this flag without its layout being IsStackOnly. E.g. when we store a non-byreflike struct to byreflike-struct's field

[jakobbotsch]

Yeah, it makes sense to me to propagate the flags, but it also makes sense to me to set the flags anew here, since the decomposition is "losing" the information at this point. I will leave it up to you.

[EgorBo]

Added. Looks like MihuBot found a few diffs with it

[EgorBo]

@MihuBot

[EgorBo]

@MihuBot

[EgorBo]

PTAL @jakobbotsch Diffs: MihuBot/runtime-utils#932 (SPMI diffs report missing contexts: https://dev.azure.com/dnceng-public/public/_build/results?buildId=926588&view=ms.vss-build-web.run-extensions-tab)

There are a few regressions where WB is lighter than no-WB (e.g. https://www.diffchecker.com/Sz7TI32G/)
And there are some regressions where previous VN based code could detect more opportunities (mostly, mixed with inlining). I managed to create a minimal repro:

[InlineArray(4)]
internal struct ArgumentData<T>
{
    private T _arg0;
}

[StructLayout(LayoutKind.Sequential)]
internal ref struct Test
{
    [MethodImpl(MethodImplOptions.NoInlining)]
    public Test(object a)
    {
        _args[1] = a;
    }

    internal ArgumentData<object?> _args;
}

here we end up with:

               [000011] -ACXG------                         *  STOREIND  ref   
               [000009] --C--------                         +--*  RET_EXPR  byref (for [000007])
               [000010] -----------                         \--*  LCL_VAR   ref    V01 arg1         

and we can't extract class info easily (stind.ref opcode doesn't contain any class token nor we can extract it from ops).

I might look into this and more cases separately (overall it's a regression only for previously merged PR)