[PATCH] D36160: Liveness issues in tail merging, and the ARM::LR saved-but-not-restored

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

Fundamentally, the problem with LR on ARM is that it isn't really a callee-save register.  We just pretend it's callee-save to avoid explicitly tracking its lifetime throughout the function.  Actually, ARM isn't the only architecture that does this; the weird part about ARM is that it has the instruction "pop {pc}" (which loads the return address directly from the stack, without loading it back into LR).

If we are going to add a marking to CalleeSavedInfo, it would probably be better to add it in assignCalleeSavedSpillSlots, where we construct the CalleeSavedInfo, rather than waiting until restoreCalleeSavedRegisters.


Repository:
  rL LLVM

https://reviews.llvm.org/D36160