diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -614,7 +614,8 @@
   // Registers in the sequence are allocated contiguously so we can just
   // use register number to pick one of three round-robin temps.
   unsigned RegNo = DestReg % 3;
-  Register Tmp = AMDGPU::VGPR32;
+  Register Tmp =
+      MBB.getParent()->getInfo<SIMachineFunctionInfo>()->getVGPRForAGPRCopy();
   assert(MBB.getParent()->getRegInfo().isReserved(Tmp) &&
          "VGPR used for an intermediate copy should have been reserved.");
 
diff --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
@@ -493,6 +493,16 @@
   // frame, so save it here and add it to the RegScavenger later.
   Optional<int> ScavengeFI;
 
+private:
+  Register VGPRForAGPRCopy;
+
+public:
+  Register getVGPRForAGPRCopy() const {
+    assert(VGPRForAGPRCopy &&
+           "Valid VGPR for AGPR copy must have been identified by now");
+    return VGPRForAGPRCopy;
+  }
+
 public: // FIXME
   /// If this is set, an SGPR used for save/restore of the register used for the
   /// frame pointer.
diff --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
@@ -187,6 +187,12 @@
   S = F.getFnAttribute("amdgpu-gds-size").getValueAsString();
   if (!S.empty())
     S.consumeInteger(0, GDSSize);
+
+  // On GFX908, in order to guarantee copying between AGPRs, we need a scratch
+  // VGPR available at all times.
+  if (ST.hasMAIInsts() && !ST.hasGFX90AInsts()) {
+    VGPRForAGPRCopy = AMDGPU::VGPR_32RegClass.getRegister(32);
+  }
 }
 
 void SIMachineFunctionInfo::limitOccupancy(const MachineFunction &MF) {
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -698,7 +698,7 @@
   // On GFX908, in order to guarantee copying between AGPRs, we need a scratch
   // VGPR available at all times.
   if (ST.hasMAIInsts() && !ST.hasGFX90AInsts()) {
-    reserveRegisterTuples(Reserved, AMDGPU::VGPR32);
+    reserveRegisterTuples(Reserved, MFI->getVGPRForAGPRCopy());
   }
 
   for (auto Reg : MFI->WWMReservedRegs) {
@@ -1553,8 +1553,8 @@
       assert(EltSize == 4);
 
       if (!TmpIntermediateVGPR) {
-        assert(MF->getRegInfo().isReserved(AMDGPU::VGPR32));
-        TmpIntermediateVGPR = AMDGPU::VGPR32;
+        TmpIntermediateVGPR = FuncInfo->getVGPRForAGPRCopy();
+        assert(MF->getRegInfo().isReserved(TmpIntermediateVGPR));
       }
       if (IsStore) {
         auto AccRead = BuildMI(MBB, MI, DL,