Set lr correctly in blx esil on arm

[bqv]

• [x] Mark this if you consider it ready to merge
• [ ] I've added tests (optional)
• [ ] I wrote some lines in the book (optional)

Description

see #22373

[trufae]

Uhm is your branch based on master?

[bqv]

oops nope. 5.8.8, will rebase

[trufae]

i think the problem is that the program counter results in an unaligned address,and the proper behaviour here would be to remove the lower bit and set the thumb mode flag instead

[bqv]

This and the issue may have been a misunderstanding, I've updated to a build from master and some things have changed or been solved already; issue I'm trying to solve now is - Emulating the bx lr instruction below results in a misalignment in disassembly:

[0x00002804]> aes; pd8@pc-8; aer=
│           0x00002828      4000           lsls r0, r0, 1
│           0x0000282a      a0f57050       sub.w r0, r0, 0x3c00        ; arg1
│           0x0000282e      0088           ldrh r0, [r0]               ; arg1
│           ;-- aav.0x00002830:
│           ;-- pc:
│           ;-- r15:
└           *0x002830*      7047           bx lr
            0x00002832      00bf           nop
            0x00002834      7a2c           cmp r4, 0x7a                ; 2'z'
            0x00002836      0000           movs r0, r0
            0x00002838      9222           movs r2, 0x92
     sb 0x00000000       sl 0x00000000       fp 0x20078000       ip 0x00007f14
     sp 0x20077ff8       lr 0x00001de6       pc 0x00002830       r0 0x00000000
     r1 0x00004aa2       r2 0x00000000       r3 0x00000000       r4 0x00000000
     r5 0x00000000       r6 0x00000000       r7 0x20077ff8       r8 0x00000000
     r9 0x00000000      r10 0x00000000      r11 0x20078000      r12 0x00007f14
    r13 0x20077ff8      r14 0x00001de6      r15 0x00002830     cpsr 0x80000000
[0x00002804]> aes; pd8@pc-8; aer=
┌ 12: sym.Java_kl_ime_oh_BinaryDictionary_toAccentLess (int16_t arg3);
│ rg: 1 (vars 0, args 1)
│ bp: 0 (vars 0, args 0)
│ sp: 0 (vars 0, args 0)
│           0x00001ddc      80b5           push {r7, lr}
│           0x00001dde      6f46           mov r7, sp
│           0x00001de0      90b2           uxth r0, r2                 ; arg3
│           0x00001de2  ~   fff77cee       blx fcn.latinime::Dictionary.toAccentLess_unsigned_short_ ; 0x1adc
│           ;-- pc:
│           ;-- r15:
└           *0x001de4*  ~   7cee80bd       cdp p13, 7, c11, c12, c0, 4
│           ;-- lr:
│           ;-- r14:
└           0x00001de6      80bd           pop {r7, pc}
┌ 244: sym.Java_kl_ime_oh_BinaryDictionary_getSuggestionsNative (int16_t arg1, int16_t arg3, int16_t arg4);
│ rg: 3 (vars 0, args 3)
│ bp: 0 (vars 0, args 0)
│ sp: 6 (vars 6, args 0)
│           0x00001de8      f0b5           push {r4, r5, r6, r7, lr}
│           0x00001dea      03af           add r7, sp, 0xc
     sb 0x00000000       sl 0x00000000       fp 0x20078000       ip 0x00007f14
     sp 0x20077ff8       lr 0x00001de6       pc 0x00001de4       r0 0x00000000
     r1 0x00004aa2       r2 0x00000000       r3 0x00000000       r4 0x00000000
     r5 0x00000000       r6 0x00000000       r7 0x20077ff8       r8 0x00000000
     r9 0x00000000      r10 0x00000000      r11 0x20078000      r12 0x00007f14
    r13 0x20077ff8      r14 0x00001de6      r15 0x00001de4     cpsr 0x80000000
[0x00001de4]> 

where the instruction's esil is:

[0x00001de4]> aoe @ 0x2830
0x2830 1,lr,-,pc,:=

Stepping over again, behaves as expected, though. Is this expected? Naturally I can make this go away with asm.flags.middle=1, but it seems strange that pc ends up misaligned

[trufae]

Sorry for the late response. kinda busy irl :( let me try to give you a proper response here. It's been a while since i really do some thumb/arm binary analysis, and i think it's been always a pending task to be addressed.. and maybe it's the time to do so if you are in the mood to discuss and test things.

I'm not sure the way it works right now, or how it should be working, but basically the way that ARM switches between thumb and arm modes is depending on the lower bit of the program counter which maps to the thumb bit of the cpsr register.

So, i see different things there that must be clarified before moving further.

• Some ELF symbols tell the disassembler if a function is thumb or arm (but sometimes this doesnt works well and results in bad anal hints. ahb-* is required to get those bins to work) <- there's a ticket for this already.
• On raw firmwares, the arm/thumb switching is usually spotted by doing esil emulation on the actual program, maybe we should do a collection of assembly snippets that do this, so we can use it for testing in a more precise way.
• What should ESIL do when PC&1 is set? this is sn arm-thumb-specific thing, and i think it should be clearing the lower bit and setting the thumb flag.
• Executing an unaligned instruction results in an exception, so I guess the ARM cpus just handle this exception to check the lower bit and do what's described above.
• But also, branch instructions (the ones where the esil expression modify the PC register) should check for the thumb bit before moving forward, and behave properly.