[Mono AOT] Fix error when returning zero sized struct

[jkurdek]

This PR allows Mono AOT to handle returns of zero sized structs in registers.

Example:

public struct Foo
{
}

The [StructLayout(LayoutKind.Sequential, Size = 1)] attribute will be added during compilation SharpLab.

[StructLayout(LayoutKind.Sequential)] // or [StructLayout(LayoutKind.Auto)] 
public struct Foo
{
}

However, adding an explicit [StructLayout(LayoutKind.Sequential] or [StructLayout(LayoutKind.Auto)] will not generate any explicit size metadata causing the struct to have size 0 SharpLab.

---

Mono Codegen:

[StructLayout(LayoutKind.Sequential)]
public struct Foo
{
}

public class Bar
{
   public static Foo GetFoo() => new Foo();
}

When processing the return of the GetFoo() method mono will evaluate size of struct Foo to be equal to zero. Then it tries to pass it in registers, calculating that it needs 0 registers to do so.

runtime/src/mono/mono/mini/mini-arm64.c

Line 1708 in bedf340

nregs = align_size / 8;

Mono JIT handles this by omitting codegen part when cinfo->ret.nregs is equal to zero.

runtime/src/mono/mono/mini/mini-arm64.c

Lines 3894 to 3911 in d606c60

	case ArgVtypeInIRegs: {
	MonoInst *loc = cfg->arch.vret_addr_loc;
	int i;
	/* Load the destination address */
	g_assert (loc && loc->opcode == OP_REGOFFSET);
	code = emit_ldrx (code, ARMREG_LR, loc->inst_basereg, GTMREG_TO_INT (loc->inst_offset));
	for (i = 0; i < cinfo->ret.nregs; ++i) {
	int offs = i * 8;
	// emit a pair of str as one stp
	if (i+1 < cinfo->ret.nregs && IS_VALID_STPX_OFFSET (offs)) {
	arm_stpx (code, cinfo->ret.reg + i, cinfo->ret.reg + i + 1, ARMREG_LR, offs);
	i++;
	} else {
	arm_strx (code, cinfo->ret.reg + i, ARMREG_LR, offs);
	}
	}
	break;

This PR proposes to fix the issue in Mono AOT by treating zero sized structs as if they have size 1.

---

Fixes #103628
Fixes #106259

[jkurdek]

/cc @vitek-karas

[lambdageek]

This PR proposes to fix the issue in Mono AOT by treating zero sized structs as if they have size 1.

Won't this break scenarios where we might call an AOTed method from the JIT or vice-versa? Or does this not leak out into callers?

What happens if you set the ret type to void? do we get ill-formed LLVM IR because we end up trying to return a value?

[jkurdek]

Won't this break scenarios where we might call an AOTed method from the JIT or vice-versa? Or does this not leak out into callers?

I'm not sure I get it. The change is made only on llvm codegen level. The size of the struct is not overridden for any other calls. Without this change we would just encounter compile time error. Am I missing something?

I don't think void return type would even enter this code path. Will verify.

[jkurdek]

I have disabled the new test on x64 as it results in a new issue #106259

[jkurdek]

/azp run runtime-extra-platforms

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkurdek]

/azp run runtime-extra-platforms

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkurdek]

/azp run runtime-extra-platforms

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkurdek]

Failures seem to be unrelated.

[derjabkin]

Is this PR also going to be merged into version 8.0?

[vitek-karas]

@jkurdek when you have some time, please open a backport, I'll try to get it through the triage. I think it should be doable.

[jkurdek]

/backport to release/8.0-staging

[github-actions]

Started backporting to release/8.0-staging: https://github.com/dotnet/runtime/actions/runs/10636307769