[PATCH] D26111: AArch64: Schedule DeadRegisterDefinitionsPass before regalloc.

[Geoff Berry via llvm-commits]

gberry added a comment.

This mostly looks good to me, though I have a few questions below.



================
Comment at: lib/Target/AArch64/AArch64DeadRegisterDefinitionsPass.cpp:92
+        continue;
+      // We should not have any relevant physreg defs that are replacable by
+      // zero before register allocation. So we just check for dead vreg defs.
----------------
Perhaps you could add an assert here checking this?


================
Comment at: lib/Target/AArch64/AArch64DeadRegisterDefinitionsPass.cpp:96
+      if (!TargetRegisterInfo::isVirtualRegister(Reg) ||
+          (!MO.isDead() && !MRI->use_nodbg_empty(Reg)))
         continue;
----------------
Is this use_nodbg_empty check necessary?  Can we end up with virtual registers marked dead with dbg uses?


================
Comment at: lib/Target/AArch64/AArch64DeadRegisterDefinitionsPass.cpp:148
   TRI = MF.getSubtarget().getRegisterInfo();
-  bool Changed = false;
   DEBUG(dbgs() << "***** AArch64DeadRegisterDefinitions *****\n");
----------------
Perhaps you should commit this bug fix (local Changed shadowing the member) separately?


https://reviews.llvm.org/D26111