R2RDump Unwind code decoding logic missed the extra word or dword in the UWOP_ALLOC_LARGE case.

[cshung]

When we write the UnwindInfo for stack space allocation, depending on the size, we write an extra WORD/DWORD to the UnwindInfo stream as follow:

runtime/src/coreclr/src/jit/unwindamd64.cpp

Lines 278 to 316 in 0b16471

	void Compiler::unwindAllocStackWindows(unsigned size)
	{
	assert(compGeneratingProlog);
	FuncInfoDsc* func = funCurrentFunc();
	assert(func->unwindHeader.Version == 1); // Can't call this before unwindBegProlog
	assert(func->unwindHeader.CountOfUnwindCodes == 0); // Can't call this after unwindReserve
	assert(size % 8 == 0); // Stack size is *always* 8 byte aligned
	UNWIND_CODE* code;
	if (size <= 128)
	{
	assert(func->unwindCodeSlot > sizeof(UNWIND_CODE));
	code = (UNWIND_CODE*)&func->unwindCodes[func->unwindCodeSlot -= sizeof(UNWIND_CODE)];
	code->UnwindOp = UWOP_ALLOC_SMALL;
	code->OpInfo = (size - 8) / 8;
	}
	else if (size <= 0x7FFF8)
	{
	assert(func->unwindCodeSlot > (sizeof(UNWIND_CODE) + sizeof(USHORT)));
	USHORT* codedSize = (USHORT*)&func->unwindCodes[func->unwindCodeSlot -= sizeof(USHORT)];
	*codedSize = (USHORT)(size / 8);
	code = (UNWIND_CODE*)&func->unwindCodes[func->unwindCodeSlot -= sizeof(UNWIND_CODE)];
	code->UnwindOp = UWOP_ALLOC_LARGE;
	code->OpInfo = 0;
	}
	else
	{
	assert(func->unwindCodeSlot > (sizeof(UNWIND_CODE) + sizeof(ULONG)));
	ULONG* codedSize = (ULONG*)&func->unwindCodes[func->unwindCodeSlot -= sizeof(ULONG)];
	*codedSize = size;
	code = (UNWIND_CODE*)&func->unwindCodes[func->unwindCodeSlot -= sizeof(UNWIND_CODE)];
	code->UnwindOp = UWOP_ALLOC_LARGE;
	code->OpInfo = 1;
	}
	unsigned int cbProlog = unwindGetCurrentOffset(func);
	noway_assert((BYTE)cbProlog == cbProlog);
	code->CodeOffset = (BYTE)cbProlog;
	}

Notice how line 298 and line 307 moves the unwindCodeSlot further than just the usual sizeof(UNWIND_CODE).

However, the similar logic is missing from the decoder, here is the parsing code:

runtime/src/coreclr/src/tools/aot/ILCompiler.Reflection.ReadyToRun/Amd64/UnwindInfo.cs

Lines 63 to 79 in 0b16471

	public UnwindCode(byte[] image, int index, ref int offset)
	{
	Index = index;
	int off = offset;
	CodeOffset = NativeReader.ReadByte(image, ref off);
	byte op = NativeReader.ReadByte(image, ref off);
	UnwindOp = (UnwindOpCodes)(op & 15);
	OpInfo = (byte)(op >> 4);
	OffsetLow = CodeOffset;
	OffsetHigh = OpInfo;
	FrameOffset = NativeReader.ReadUInt16(image, ref offset);
	NextFrameOffset = -1;
	IsOpInfo = false;

The size of the extra word/dword happens to be a multiple of 2, therefore these word/dword are misinterpreted as extra unwind opcode. This is just wrong. This explains why we saw more than one unwind opcode having the same code offset in R2RDump.