asmjit/asmjit

Complete x86/x64 JIT and Remote Assembler for C++

muiloo1998

@muiloo1998

maybe it depends on the group of ARM processors

Petr Kobalicek

@kobalicek

Can you be a bit more specific? According to the manual it's not encodable with LDR

so I'm really wondering which opcode you tested

muiloo1998

@muiloo1998

Petr Kobalicek

@kobalicek

I'm still pretty new to aarch64, so maybe there is something I don't know :)

muiloo1998

@muiloo1998

here is offset range : https://www.keil.com/support/man/docs/armasm/armasm_dom1361289873425.htm

me too.

Petr Kobalicek

@kobalicek

but that's not AArch64, it's ARM and Thumb

muiloo1998

@muiloo1998

please see this document: https://developer.arm.com/documentation/ddi0487/latest/

Petr Kobalicek

@kobalicek

Well it has like 8000 pages

Can you just provide me the relevant info?

muiloo1998

@muiloo1998

"• Base plus a scaled 12-bit unsigned immediate offset or base plus an unscaled 9-bit signed immediate offset."

Petr Kobalicek

@kobalicek

Just to finalize this - according to AArch64 Manual LDR should use LDUR in case of signed or other immediate that is not encodable with LDR

I will fix this

of course this related to more instructions, like STR as well

Petr Kobalicek

@kobalicek

@muiloo1998 I have pushed a fix, let me know whether it fixes it

865

@ichi865_twitter

Hi @kobalicek
Could you give further hints on how i can solve that issue with asmjit?
asmjit/asmjit#201

After I added the bytecode to the emitter using embed(data, size), I indicated the base address.
But now how do I relocate the bytecode to an another base address so that relative calls and jumps are converted relative to the new base address.

Petr Kobalicek

@kobalicek

What do you mean by bytecode?

If you embed data with embed, asmjit has no knowledge about such data - no relocations, nothing, it's just a binary blob it appends to the current code buffer

865

@ichi865_twitter

@kobalicek I mean a sequence of bytes of some original function that has many jumps and calls. I copied the function to buffer, but it'll not be possible to call, since addresses lead to the wrong place.

then could you tell what can I do next with that blob in order to achieve the result I want?

Petr Kobalicek

@kobalicek

I don't know, maybe use some disassembler like zydis and reassemble it, but that is not something I would want to do. Or, if you control generating of that blob, add some metadata that you can use to properly relocate it

865

@ichi865_twitter

thanks, i'll try your 1st suggestion to start with.
but why do you think this is a not good way?

865

@ichi865_twitter

hi again @kobalicek
is it possible to skip some operands during parsing?

fld st0, dword ptr ds:[eax]

like skiping st0, so that there is no error

Petr Kobalicek

@kobalicek

but why do you think this is a not good way?

I think that if you generate the blob, you should provide the information to properly use it, guessing afterwards doesn't seem like the best idea

Well I don't know - the X87 instructions follow Intel X86 Arch Manual, so they match the manual. But I would also liked more if they were explicit

Maybe I would do another iteration to match LLVM in this case, but today I don't have time unfortunately

865

@ichi865_twitter

okay, fst/fstp also not supported
or I messed up with settings

Petr Kobalicek

@kobalicek

it's all supported

865

@ichi865_twitter

https://i.imgur.com/3Zwr15E.png
apparently I was wrong somewhere

Petr Kobalicek

@kobalicek

you can see all valid signatures here:

https://asmjit.com/asmgrid/

865

@ichi865_twitter

wau cool

J.V.

@JV81608177_twitter

Is asmjit suitable for a AOT compiler ? Or only for just in time compilation ?

Petr Kobalicek

@kobalicek

both, but with AOT you have to be extremely careful about relocations

in addition, asmjit doesn't produce binaries, like .so objects, this has to be done by users, but additions are welcome :)

John Högberg

@jhogberg

I think I've found a bug in the encoding of CMP Reg, Imm on ARM, the check to see if the immediate can be shifted seems inverted

        if (immValue > 0xFFFu) {
          if ((immValue & (0xFFFu << 12)) != 0)

Shouldn't it be ~(0xFFFul << 12)?

(I noticed it after not being able to encode cmp x0, 65536 which ought to be possible if I'm reading the docs right)

Petr Kobalicek

@kobalicek

I will check it out

Petr Kobalicek

@kobalicek

you are right, thanks, I will fix that

John Högberg

@jhogberg

Sweet, thanks :)

Petr Kobalicek

@kobalicek

@jhogberg Should be fixed, thanks!

Patrick Mackinlay

@pmackinlay

Hi @kobalicek, it's been a while but I'm back to working on my own jit in MAME. I'm having trouble understanding why an instruction is failing with an invalid address error.

The code is a.call(ptr(l, rax, 3, 8)), where l is a label.

I want call qword ptr[(l+8) + rax*8] to be emitted

Patrick Mackinlay

@pmackinlay

Aha, I see my error - rip-relative addressing can only use rip + offset; back to the books.

Petr Kobalicek

@kobalicek

Actually it would work in 32-bit mode - the label would be translated to absolute address, but in 64-bit mode this doesn't work

Where communities thrive

People

Repo info

Activity