[Lldb-commits] [PATCH] D12556: Introduce new address class eAddressClassDataIntermixedCode

[Greg Clayton via lldb-commits]

clayborg added a comment.

Changing all $d symbols to always say they are eAddressClassDataIntermixedCode is wrong because the symbols in the .data section now would be marked as eAddressClassDataIntermixedCode.

To clarify a few things, lets say we have the following code:

0x1000: bx <addr> Non-tail call in a no return function
0x1004: [data-pool] Marked with $d mapping symbol

Should just claim that 0x1000 is eAddressClassCode and that 0x1004 is eAddressClassData. We don't need a new eAddressClassDataIntermixedCode for this, it should just say eAddressClassData for 0x1004.

For return addresses we that are on the stack on in the LR, we should actually be sanitizing them before we start doing lookups. So the return address would be 0x1005 in this case you are talking about right? Maybe we always get the address class of the return address and check if it is eAddressClassData. If it is, we know something is wrong since the return address can't go to data and we work around this by recognizing that fact.

I would rather not just say that all data is eAddressClassDataIntermixedCode for ARM and Thumb, that seems like the wrong fix.


http://reviews.llvm.org/D12556