[PATCH] D114652: [AMDGPU] Treat call clobbered return address registers s[30:31] as callee saved

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp:630-633
+  // The return address register is call clobbered and the CFI needs to track
+  // its location. Hence it is handled specially.
+  if (!MFI->isEntryFunction())
+    reserveRegisterTuples(Reserved, getReturnAddressReg(MF));
----------------
RamNalamothu wrote:
> arsenm wrote:
> > I don't think it's necessary to reserve this
> But otherwise the register allocation thinks it is available for allocation as `s[30:31]` are scratch registers at the moment and its usage with return instruction is hidden with using SI_RETURN through the register allocation and we don't emit Live-ins anymore.
> 
> And we will have to reserve it here even after these registers actually get picked from **CSR range** because we want to emit save/restore exclusively so that the unwind info machinery knows it and can generate CFI as needed for return address. That's the reason why all of this got started right.
I think something special happens with the return address we are ignoring, but I haven't followed it all the way through.

AArch64 passes LR to the TargetRegisterInfo constructor, where we're passing the dummy PC_REG. If you replace that with the real return address, does that help?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D114652/new/

https://reviews.llvm.org/D114652