[libunwind] Fix ARM EHABI unwinding instruction calculation

[smeenai]

Summary

Section 9.3 of the ARM EHABI [1] states that the 01xxxxxx instruction
should be interpreted as follows:

vsp = vsp - (xxxxxx << 2) - 4

However, libunwind is adding 4 instead of subtracting 4. libgcc's
unwinder follows the spec here, so I'm inclined to believe this is a bug
in libunwind. (Perhaps the intent was to have parentheses around the
expression, such that it was a - (b + 4) instead of a - b + 4, but
it's hard to say at this point.) Interestingly, this appears to go all
the way back to the initial commit of ARM EHABI unwinding, back in June
2014: https://github.com/llvm/llvm-project/commit/97080e0c5eb4.

[1] https://developer.arm.com/docs/ihi0038/c/exception-handling-abi-for-the-arm-architecture-abi-2018q4-documentation#ehabi32-section9-3

[rprichard]

I think libunwind is already correct. It's doing:

vsp -= (xxxxxx << 2) + 4

The addition has higher precedence than the compound assignment, so it's equivalent to:

vsp = vsp - ((xxxxxx << 2) + 4)

i.e.

vsp = vsp - (xxxxxx << 2) - 4

[smeenai]

Ah, you're right. I mis-parsed that. Whoops.

libunwind/src/Unwind-EHABI.cpp

_Unwind_VRS_Interpret(_Unwind_Context *context, const uint32_t *data,
   bool wrotePC = false;
   bool finish = false;
   while (offset < len && !finish) {
     uint8_t byte = getByte(data, offset++);
     if ((byte & 0x80) == 0) {
       uint32_t sp;
       _Unwind_VRS_Get(context, _UVRSC_CORE, UNW_ARM_SP, _UVRSD_UINT32, &sp);
       if (byte & 0x40)
-        sp -= (((uint32_t)byte & 0x3f) << 2) + 4;
+        sp -= (((uint32_t)byte & 0x3f) << 2) - 4;
       else
         sp += ((uint32_t)byte << 2) + 4;
       _Unwind_VRS_Set(context, _UVRSC_CORE, UNW_ARM_SP, _UVRSD_UINT32, &sp);
     } else {
       switch (byte & 0xf0) {
         case 0x80: {
           if (offset >= len)
             return _URC_FAILURE;