[AMDGPU] Avoid RegScavenger::forward in copyPhysReg/indirectCopyToAGPR

RegScavenger::backward is preferred because it does not rely on accurate
kill flags.

Differential Revision: https://reviews.llvm.org/D150571

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 4f695a8..a38a835 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -618,8 +618,8 @@ static void indirectCopyToAGPR(const SIInstrInfo &TII,
     }
   }
 
-  RS.enterBasicBlock(MBB);
-  RS.forward(MI);
+  RS.enterBasicBlockEnd(MBB);
+  RS.backward(MI);
 
   // Ideally we want to have three registers for a long reg_sequence copy
   // to hide 2 waitstates between v_mov_b32 and accvgpr_write.