Incorrect Behaviour: Arm64 CopyLdrLiteral rewrites Prefetch as Memory Read, which can Trash a Potentially Used Register

[Sewer56]

Describe the bug

This is a sister bug to:

• Incorrect Behaviour: Arm64 CopyLdrLiteral for non-SIMD inlines Pointer Read #305

In the following detours code

Detours/src/disasm.cpp

Lines 4226 to 4261 in 4b8c659

	BYTE CDetourDis::CopyLdrLiteral(BYTE* pSource, BYTE* pDest, ULONG instruction)
	{
	LdrLit19& decoded = (LdrLit19&)(instruction);
	PULONG pDstInst = (PULONG)(pDest);
	BYTE* pTarget = pSource + decoded.Imm();
	LONG64 delta = pTarget - pDest;
	// output as LDR
	if (delta >= -(1 << 21) && delta < (1 << 21))
	{
	EmitInstruction(pDstInst, LdrLit19::Assemble(decoded.s.Size, decoded.s.FpNeon, decoded.s.Rt, (LONG)delta));
	}
	// output as move immediate
	else if (decoded.s.FpNeon == 0)
	{
	UINT64 value = 0;
	switch (decoded.s.Size)
	{
	case 0: value = *(ULONG*)pTarget; break;
	case 1: value = *(UINT64*)pTarget; break;
	case 2: value = *(LONG*)pTarget; break;
	}
	EmitMovImmediate(pDstInst, decoded.s.Rt, value);
	}
	// FP/NEON register: compute address in x17 and load from there (BIG assumption that x17 isn't being used for anything!!)
	else
	{
	EmitMovImmediate(pDstInst, 17, (ULONG_PTR)pTarget);
	EmitInstruction(pDstInst, LdrFpNeonImm9::Assemble(2 + decoded.s.Size, decoded.s.Rt, 17, 0));
	}
	return (BYTE)((BYTE*)pDstInst - pDest);
	}

which rewrites LDR Literal, the PRFM (Prefetch) instruction is rewritten as a load.

This happens because LDR literal and PRFM use the same opcode, and thus the LDR code path would be used for PRFM

Detours/src/disasm.cpp

Lines 3967 to 3969 in 4b8c659

	} else if ((Instruction & 0x3b000000) == 0x18000000) {
	CopiedSize = CopyLdrLiteral(pSrc, pDst, Instruction);
	} else {

e.g. PRFM PLIL1KEEP, #0 is 0xD8000008. (0xD8000008 & 0x3b000000) == 0x18000000 is true, thus the LDR code path is executed for PRFM.

Expected behavior

Prefetch operation should not be rewritten as a load, as this will trash the existing value in the register, which may still be used by the function.

Additional context

I'm building a cross platform, multi architecture hooking library [big WIP]; and I found detours' source to be pretty invaluable as a reference for code rewriting.

This is just a small thing I noticed while working on code rewriting on my end while referencing detours.