Add more EIP syncs for x86 target memory hooks

[Evian-Zhang]

Inside memory hooks, value of RIP register may be wrongly retrieved, especially for SIMD instructions (since they do not use gen_op_ld_v and gen_op_st_v for memory access).

However, this behavior is random and I cannot provide an example for this bug. I came across this bug when unicorn is used as fuzzing executor with AFL++.

BTW, you can see many usages of tcg_gen_qemu_ld_xx and tcg_gen_qemu_st_xx across the whole translate.c. Maybe we should add this EIP sync before each callsite? However, there are too many usages (in some rarely-seen instructions), and I could not tell whether we should add this sync. So this PR will focus on the most common usages.

[wtdcode]

Could you add tests to illustrate the instructions?

[Evian-Zhang]

Sorry I could not provide a concrete test case, as this bug occurs randomly.

I could explain in theory. As the eip is already synced in normal memory accesses, this makes sure that the RIP we get in memory hooks is correct. However, the SIMD instructions do not use those functions to access memories. So in some rare cases where eip is not synced, we could only got the wrong eip that lastly synced.

As we already sync eip in normal memory accesses, there is no reason not adding those syncs for SIMD memory accesses.

[wtdcode]

Sorry I could not provide a concrete test case, as this bug occurs randomly.

I could explain in theory. As the eip is already synced in normal memory accesses, this makes sure that the RIP we get in memory hooks is correct. However, the SIMD instructions do not use those functions to access memories. So in some rare cases where eip is not synced, we could only got the wrong eip that lastly synced.

As we already sync eip in normal memory accesses, there is no reason not adding those syncs for SIMD memory accesses.

I understand this happens in a quite large codebase. I mean you can provide a unit case including only 1 SIMD instruction and 1 memory hook and the PC is not synced. You can find lots of samples in tests/unit/test_x86.c

[wtdcode]

This ensures our future releases won’t break your uses cases.

[Evian-Zhang]

I understand the necessity to add test case. However, by saying "random", I mean the bug is not reproducible. I do meet one SIMD instruction which does not sync EIP during one execution, however, in another execution, it is synced, where somewhere else is not synced. This is because I meet this bug when I use unicorn as fuzzing engine with AFL++, which uses a lot of low-level optimizations such as inter-process TB caches, which makes this bug hard to catch and hard to reproduce in a small testcase.

[wtdcode]

I understand the necessity to add test case. However, by saying "random", I mean the bug is not reproducible. I do meet one SIMD instruction which does not sync EIP during one execution, however, in another execution, it is synced, where somewhere else is not synced. This is because I meet this bug when I use unicorn as fuzzing engine with AFL++, which uses a lot of low-level optimizations such as inter-process TB caches, which makes this bug hard to catch and hard to reproduce in a small testcase.

Are you using unicornafl?

[Evian-Zhang]

Yes

[wtdcode]

Unicornafl bundled a quite old Unicorn at this moment, which I will update once next patch release lands. Have you updated the bundled unicorn version?

[Evian-Zhang]

I use the latest unicorn at dev branch.