Problem with encoding jmps to absolute addresses

[fotisl]

Jumps to absolute addresses are not always encoded correctly.
For example:

ks = Ks(KS_ARCH_ARM, KS_MODE_ARM)
ks.asm(b"b #0x0010", 0x0000)
ks.asm(b"b #0x1010", 0x1000)

Since the b instruction encodes a relative offset to the target address, the instruction encoding should be the same, however it isn't. You can verify the incorrect encoding using capstone.

The same happens with many architectures, such as KS_ARCH_ARM(other instructions such as bl), KS_ARCH_ARM64 (b, bl etc), KS_ARCH_MIPS(KS_MODE_MIPS32 with b, beq, etc), KS_ARCH_SPARC(b, etc).

The problem originates at the MCCodeEmitter (for example, for the SPARC architecture at llvm/lib/Target/Sparc/MCTargetDesc/SparcMCCodeEmitter.cpp) which returns the immediate when it encounters it as an argument at a branch, and not a relative offset.

[aquynh]

Can you send in a testcase? Thanks

[fotisl]

I have made a simple program that demonstrates this bug for ARM, ARM64, SPARC, MIPS, PPC and SystemZ.
bug.txt

[aquynh]

can you rewrite this into a regress test and send a PR to put it under directory suite/regress/? this is very simple, you can take https://github.com/keystone-engine/beta/blob/master/suite/regress/test1.py as example.

[fotisl]

Pull request #109 created.

[aquynh]

how far have you gone to attempt to fix this issue?

[fotisl]

I have found the source of the problem, and given enough time I can fix it. When do you plan to release keystone?

[aquynh]

fixed all now, thanks!