Commit de4ad6b96434592b9da6a286d717794b5d5f5cf8 broke Dynarmic

[Tachi107]

Hi!

It seems that commit de4ad6b, "strictly check the size of general registers", broke the dynamic ARM recompiler Dynarmic.

In particular, the new check in opMR which throws ERR_BAD_MEM_SIZE is what broke it.

        void opMR(const Address& addr, const Reg& r, uint64_t type, int code, uint64_t type2 = 0, int code2 = NONE)
        {
                if (code2 == NONE) code2 = code;
                if (type2 && opROO(Reg(), addr, r, type2, code2)) return;
                if (addr.is64bitDisp()) XBYAK_THROW(ERR_CANT_USE_64BIT_DISP)
+               if (!(type & T_ALLOW_DIFF_SIZE) && r.getBit() <= BIT && addr.getBit() > 0 && addr.getBit() != r.getBit()) XBYAK_THROW(ERR_BAD_MEM_SIZE)
                bool rex2 = rex(addr, r, type);
                writeCode(type, r, code, rex2);
                opAddr(addr, r.getIdx());

I'm not really familiar with how Dynarmic uses Xbyak, but it'd be great to get this fixed nonetheless!

Bye :)

[herumi]

What is the dynamic ARM recompiler? Please show me the details of the error.

[Tachi107]

I'm talking about this project: https://salsa.debian.org/debian/dynarmic

As for the details of the error, I'm not really sure. I didn't write Dynarmic myself, I just maintain it as a dependency of other projects. I'm currently trying to figure out what it's causing the issue though.

[herumi]

That commit checks inconsistencies between register size and memory size, like "mov al, dword[rdx]" or "mov eax, byte[rdx]", and prevents unintended code generation. Please ask in dynamic repository whether they are writing such code.

[Tachi107]

Got it. I believe that yes, Dynarmic is writing such code extensively, since pretty much all of its tests fail with this new check. Still, Dynarmic works well even with such code. It has been used successfully in different programs, like a console emulator. Or maybe this is a false positive? Do you have suggestions as to how I could debug this? Is there a way to log code passed to Xbyak, to find out what where Dynarmic is generating bad code? Unfortunately, Dynarmic's original maintainer is currently unreachable, so I have to figure this out myself :/

[herumi]

Build your emulator with the compiler option "-O0 -g" and execute it with gdb:

gdb --args <executable file> <option>
(gdb) catch throw
(gdb) r
(gdb) bt

This will show you the location in the source code where Xbyak threw the exception.

However, if there are so many errors, it might be better to add a macro option to revert Xbyak to its original behavior.
Please wait a moment.

[herumi]

Will Dynarmic work if I just comment out line 2100 of xbyak.h?

[Tachi107]

Yes, that's what I did locally to work around the issue

[herumi]

I've added a macro, XBYAK_STRICT_CHECK_MEM_REG_SIZE, to control the size check.
8638978