Index: lib/Target/AMDGPU/SIInstrInfo.h
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.h
+++ lib/Target/AMDGPU/SIInstrInfo.h
@@ -434,6 +434,12 @@
     return getNamedOperand(const_cast<MachineInstr &>(MI), OpName);
   }
 
+  /// Get required immediate operand
+  int64_t getNamedImmOperand(const MachineInstr &MI, unsigned OpName) const {
+    int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName);
+    return MI.getOperand(Idx).getImm();
+  }
+
   uint64_t getDefaultRsrcDataFormat() const;
   uint64_t getScratchRsrcWords23() const;
 };
Index: lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.cpp
+++ lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1977,25 +1977,61 @@
     } else {
       // This instructions is the _OFFSET variant, so we need to convert it to
       // ADDR64.
+      assert(MBB.getParent()->getSubtarget<AMDGPUSubtarget>().getGeneration()
+             < AMDGPUSubtarget::VOLCANIC_ISLANDS &&
+             "FIXME: Need to emit flat atomics here");
+
       MachineOperand *VData = getNamedOperand(*MI, AMDGPU::OpName::vdata);
       MachineOperand *Offset = getNamedOperand(*MI, AMDGPU::OpName::offset);
       MachineOperand *SOffset = getNamedOperand(*MI, AMDGPU::OpName::soffset);
-
-      // Create the new instruction.
       unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI->getOpcode());
-      MachineInstr *Addr64 =
-        BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
-        .addOperand(*VData)
-        .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
-                                    // This will be replaced later
-                                    // with the new value of vaddr.
-        .addOperand(*SRsrc)
-        .addOperand(*SOffset)
-        .addOperand(*Offset)
-        .addImm(0) // glc
-        .addImm(0) // slc
-        .addImm(0) // tfe
-        .setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+
+      // Atomics rith return have have an additional tied operand and are
+      // missing some of the special bits.
+      MachineOperand *VDataIn = getNamedOperand(*MI, AMDGPU::OpName::vdata_in);
+      MachineInstr *Addr64;
+
+      if (!VDataIn) {
+        // Regular buffer load / store.
+        MachineInstrBuilder MIB
+          = BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
+          .addOperand(*VData)
+          .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
+          // This will be replaced later
+          // with the new value of vaddr.
+          .addOperand(*SRsrc)
+          .addOperand(*SOffset)
+          .addOperand(*Offset);
+
+        // Atomics do not have this operand.
+        if (const MachineOperand *GLC
+            = getNamedOperand(*MI, AMDGPU::OpName::glc)) {
+          MIB.addImm(GLC->getImm());
+        }
+
+        MIB.addImm(getNamedImmOperand(*MI, AMDGPU::OpName::slc));
+
+        if (const MachineOperand *TFE
+            = getNamedOperand(*MI, AMDGPU::OpName::tfe)) {
+          MIB.addImm(TFE->getImm());
+        }
+
+        MIB.setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+        Addr64 = MIB;
+      } else {
+        // Atomics with return.
+        Addr64 = BuildMI(MBB, MI, MI->getDebugLoc(), get(Addr64Opcode))
+          .addOperand(*VData)
+          .addOperand(*VDataIn)
+          .addReg(AMDGPU::NoRegister) // Dummy value for vaddr.
+          // This will be replaced later
+          // with the new value of vaddr.
+          .addOperand(*SRsrc)
+          .addOperand(*SOffset)
+          .addOperand(*Offset)
+          .addImm(getNamedImmOperand(*MI, AMDGPU::OpName::slc))
+          .setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+      }
 
       MI->removeFromParent();
       MI = Addr64;
Index: lib/Target/AMDGPU/SIInstrInfo.td
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.td
+++ lib/Target/AMDGPU/SIInstrInfo.td
@@ -2433,6 +2433,8 @@
   } // mayStore = 1, mayLoad = 1, hasPostISelHook = 1
 }
 
+// FIXME: tfe can't be an operand because it requires a separate
+// opcode because it needs an N+1 register class dest register.
 multiclass MUBUF_Load_Helper <mubuf op, string name, RegisterClass regClass,
                               ValueType load_vt = i32,
                               SDPatternOperator ld = null_frag> {
Index: test/CodeGen/AMDGPU/move-to-valu-atomicrmw.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/move-to-valu-atomicrmw.ll
@@ -0,0 +1,52 @@
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; XUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+; FIXME: broken on VI because flat instructions need to be emitted
+; instead of addr64 equivalent of the _OFFSET variants.
+
+; Check that moving the pointer out of the resource descriptor to
+; vaddr works for atomics.
+
+declare i32 @llvm.r600.read.tidig.x() #1
+
+; GCN-LABEL: {{^}}atomic_max_i32:
+; GCN: buffer_atomic_smax v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:400 glc{{$}}
+define void @atomic_max_i32(i32 addrspace(1)* %out, i32 addrspace(1)* addrspace(1)* %in, i32 addrspace(1)* %x, i32 %y) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x()
+  %tid.gep = getelementptr i32 addrspace(1)*, i32 addrspace(1)* addrspace(1)* %in, i32 %tid
+  %ptr = load volatile i32 addrspace(1)*, i32 addrspace(1)* addrspace(1)* %tid.gep
+  %xor = xor i32 %tid, 1
+  %cmp = icmp ne i32 %xor, 0
+  br i1 %cmp, label %atomic, label %exit
+
+atomic:
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i32 100
+  %ret = atomicrmw max i32 addrspace(1)* %gep, i32 %y seq_cst
+  store i32 %ret, i32 addrspace(1)* %out
+  br label %exit
+
+exit:
+  ret void
+}
+
+; GCN-LABEL: {{^}}atomic_max_i32_noret:
+; GCN: buffer_atomic_smax v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:400{{$}}
+define void @atomic_max_i32_noret(i32 addrspace(1)* %out, i32 addrspace(1)* addrspace(1)* %in, i32 addrspace(1)* %x, i32 %y) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x()
+  %tid.gep = getelementptr i32 addrspace(1)*, i32 addrspace(1)* addrspace(1)* %in, i32 %tid
+  %ptr = load volatile i32 addrspace(1)*, i32 addrspace(1)* addrspace(1)* %tid.gep
+  %xor = xor i32 %tid, 1
+  %cmp = icmp ne i32 %xor, 0
+  br i1 %cmp, label %atomic, label %exit
+
+atomic:
+  %gep = getelementptr i32, i32 addrspace(1)* %ptr, i32 100
+  %ret = atomicrmw max i32 addrspace(1)* %gep, i32 %y seq_cst
+  br label %exit
+
+exit:
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }