[PATCH] D51432: [AArch64] Unwinding support for return address signing

[Oliver Stannard via Phabricator via cfe-commits]

olista01 added inline comments.


================
Comment at: src/DwarfInstructions.hpp:210
+        register unsigned long long x16 __asm("x16") = cfa;
+        asm("autia1716": "+r"(x17): "r"(x16));
+        returnAddress = x17;
----------------
I don't think this will work for cross-unwinding builds: for them, _LIBUNWIND_TARGET_AARCH64 is defined even when the compilation target is not AArch64, so this instruction won't exist.

Fully supporting cross-unwinding looks non-trivial: we'd need to either provide some way to ask the client to authenticate a pointer on the target, or strip the high bits of the pointer (which requires knowing the virtual address size of the target). For now, I think it's OK to not support cross-unwinding.


================
Comment at: src/Registers.hpp:1835
+  if (((regNum >= 0) && (regNum < 32)) || regNum == UNW_ARM64_RA_SIGN_STATE)
     return _registers.__x[regNum];
+
----------------
When regNum == UNW_ARM64_RA_SIGN_STATE, the index into __x is out of range. We'll need to add new storage to hold this value, I'd suggest replacing the current padding value in the GPRs struct, as that will avoid changing the layout expected by the context save/restore functions.


================
Comment at: src/Registers.hpp:1845
     _registers.__sp = value;
-  else if ((regNum >= 0) && (regNum < 32))
+  else if ((regNum >= 0) && (regNum < 32) || regNum == UNW_ARM64_RA_SIGN_STATE)
     _registers.__x[regNum] = value;
----------------
Ditto.


Repository:
  rUNW libunwind

https://reviews.llvm.org/D51432