diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -1890,19 +1890,30 @@
 }
 
 // Match (32-bit SGPR base) + sext(imm offset)
-bool AMDGPUDAGToDAGISel::SelectScratchSAddr(SDNode *N,
-                                            SDValue Addr,
+bool AMDGPUDAGToDAGISel::SelectScratchSAddr(SDNode *Parent, SDValue Addr,
                                             SDValue &SAddr,
                                             SDValue &Offset) const {
   if (Addr->isDivergent())
     return false;
 
-  SAddr = Addr;
+  const MachinePointerInfo &PtrInfo = cast<MemSDNode>(Parent)->getPointerInfo();
+  MachineFunction &MF = CurDAG->getMachineFunction();
+  SDLoc DL(Addr);
+
   int64_t COffsetVal = 0;
 
-  if (CurDAG->isBaseWithConstantOffset(Addr)) {
+  ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr);
+  if (CAddr && isStackPtrRelative(PtrInfo)) {
+    COffsetVal = CAddr->getSExtValue();
+    // In a call sequence, stores to the argument stack area are relative to the
+    // stack pointer.
+    const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
+    SAddr = CurDAG->getRegister(Info->getStackPtrOffsetReg(), MVT::i32);
+  } else if (CurDAG->isBaseWithConstantOffset(Addr)) {
     COffsetVal = cast<ConstantSDNode>(Addr.getOperand(1))->getSExtValue();
     SAddr = Addr.getOperand(0);
+  } else {
+    SAddr = Addr;
   }
 
   SAddr = SelectSAddrFI(CurDAG, SAddr);
@@ -1917,14 +1928,15 @@
 
     COffsetVal = SplitImmOffset;
 
-    SDLoc DL(N);
     SDValue AddOffset =
-        getMaterializedScalarImm32(Lo_32(RemainderOffset), DL);
+        SAddr.getOpcode() == ISD::TargetFrameIndex
+            ? getMaterializedScalarImm32(Lo_32(RemainderOffset), DL)
+            : CurDAG->getTargetConstant(RemainderOffset, DL, MVT::i32);
     SAddr = SDValue(CurDAG->getMachineNode(AMDGPU::S_ADD_U32, DL, MVT::i32,
                                            SAddr, AddOffset), 0);
   }
 
-  Offset = CurDAG->getTargetConstant(COffsetVal, SDLoc(), MVT::i16);
+  Offset = CurDAG->getTargetConstant(COffsetVal, DL, MVT::i16);
 
   return true;
 }
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
@@ -160,14 +160,13 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 0
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, external_void_func_i1@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, external_void_func_i1@rel32@hi+12
-; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s32
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 0
@@ -257,7 +256,6 @@
 ; GFX10-SCRATCH-NEXT:    global_load_ubyte v0, v[0:1], off glc dlc
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 0
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
@@ -266,7 +264,7 @@
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-SCRATCH-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s2
+; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s32
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s1, v40, 1
@@ -356,7 +354,6 @@
 ; GFX10-SCRATCH-NEXT:    global_load_ubyte v0, v[0:1], off glc dlc
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 0
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
@@ -365,7 +362,7 @@
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-SCRATCH-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s2
+; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s32
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s1, v40, 1
@@ -4184,7 +4181,6 @@
 ; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v32, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 0
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
@@ -4205,7 +4201,7 @@
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, external_void_func_v32i32_i32@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, external_void_func_v32i32_i32@rel32@hi+12
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(8)
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v33, s2
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v33, s32
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s1, v40, 1
@@ -4967,14 +4963,13 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 0
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, external_void_func_i1_inreg@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, external_void_func_i1_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    scratch_store_byte off, v0, s32
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 0
@@ -8351,16 +8346,10 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 18
 ; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 20
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 16
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s20, 12
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s36, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s21, 8
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s22, 4
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s23, 0
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s37, 1
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s36, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s37, 1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s38, 2
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s39, 3
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s40, 4
@@ -8389,15 +8378,6 @@
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, s50
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, s49
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v3, s48
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, s47
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v5, s46
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s2
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s3
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s20
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v3, s21
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v4, s22
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v5, s23
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s20, s36
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s21, s37
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s22, s38
@@ -8408,6 +8388,14 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s27, s43
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s28, s44
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s29, s45
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, s47
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v5, s46
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:20
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s32 offset:16
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s32 offset:12
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v3, s32 offset:8
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v4, s32 offset:4
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v5, s32
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 16
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s1, v40, 17
@@ -8631,20 +8619,12 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 18
 ; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 24
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s20, 20
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s21, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s36, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s22, 12
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s23, 4
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s24, 0
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s37, 1
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s36, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s37, 1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s38, 2
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s39, 3
-; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s40, 4
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s41, 5
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s42, 6
@@ -8657,10 +8637,13 @@
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s49, 13
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s50, 14
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s51, 15
+; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-SCRATCH-NEXT:    s_load_dword s2, s[0:1], 0x0
 ; GFX10-SCRATCH-NEXT:    ; kill: killed $sgpr0_sgpr1
 ; GFX10-SCRATCH-NEXT:    ; kill: killed $sgpr0_sgpr1
+; GFX10-SCRATCH-NEXT:    s_clause 0x1
 ; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[36:51], s[0:1], 0x40
+; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x0
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, external_void_func_v32i32_i32_inreg@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, external_void_func_v32i32_i32_inreg@rel32@hi+12
@@ -8668,31 +8651,29 @@
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 17
 ; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, s2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 8
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, s50
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, s49
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s3
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:24
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, s51
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v3, s48
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, s47
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v5, s46
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s25, s41
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s20
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s21
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s22
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v3, s2
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v4, s23
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v5, s24
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s20, s36
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s21, s37
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s22, s38
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s23, s39
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s24, s40
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s25, s41
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s26, s42
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s27, s43
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s28, s44
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s29, s45
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, s47
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v5, s46
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:20
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s32 offset:16
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s32 offset:12
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v3, s32 offset:8
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v4, s32 offset:4
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v5, s32
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 16
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s1, v40, 17
@@ -8804,21 +8785,19 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 4
+; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    s_clause 0x1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v32, off, s33 offset:4
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v33, off, s33
-; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 0
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, stack_passed_f64_arg@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, stack_passed_f64_arg@rel32@hi+12
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v32, s2
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v32, s32 offset:4
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v33, s3
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v33, s32
 ; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s1, v40, 1
@@ -8977,21 +8956,18 @@
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s32 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 15
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 12
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 14
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 12
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 13
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 8
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s0
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 4
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 13
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
+; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v3, 12
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s1
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:12
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s32 offset:8
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s32 offset:4
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v3, s32
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 0
@@ -9024,8 +9000,6 @@
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v29, 9
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v30, 10
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v31, 11
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, external_void_func_12xv3i32@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, external_void_func_12xv3i32@rel32@hi+12
@@ -9218,32 +9192,24 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 15
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 14
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 28
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 24
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 11
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s1
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 13
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 12
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 20
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 16
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 13
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s1
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 10
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 12
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 9
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s0
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 8
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 8
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 4
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
+; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v3, 9
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:28
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s32 offset:24
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s32 offset:20
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 12
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 11
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 10
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, 8
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s1
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:16
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s32 offset:12
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s32 offset:8
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v3, s32 offset:4
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v4, s32
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 0
@@ -9276,8 +9242,6 @@
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v29, 5
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v30, 6
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v31, 7
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, external_void_func_8xv5i32@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, external_void_func_8xv5i32@rel32@hi+12
@@ -9466,32 +9430,24 @@
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 0x41700000
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 0x41600000
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 28
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 24
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 0x41300000
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s1
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 0x41500000
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 0x41400000
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 20
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 16
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 0x41500000
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s33, 2
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s1
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 0x41200000
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 12
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 0x41100000
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s0
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 8
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 0x41000000
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s0
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 4
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
+; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v3, 0x41100000
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:28
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s32 offset:24
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s32 offset:20
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 0x41400000
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 0x41300000
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 0x41200000
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, 0x41000000
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s0
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s1
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v0, s32 offset:16
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v1, s32 offset:12
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s32 offset:8
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v3, s32 offset:4
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v4, s32
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, 0
@@ -9524,8 +9480,6 @@
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v29, 0x40a00000
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v30, 0x40c00000
 ; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v31, 0x40e00000
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_add_u32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    s_getpc_b64 s[0:1]
 ; GFX10-SCRATCH-NEXT:    s_add_u32 s0, s0, external_void_func_8xv5f32@rel32@lo+4
 ; GFX10-SCRATCH-NEXT:    s_addc_u32 s1, s1, external_void_func_8xv5f32@rel32@hi+12
diff --git a/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll b/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll
--- a/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll
+++ b/llvm/test/CodeGen/AMDGPU/non-entry-alloca.ll
@@ -66,16 +66,15 @@
 ; FLATSCR-NEXT:    s_cbranch_scc1 BB0_3
 ; FLATSCR-NEXT:  ; %bb.2: ; %bb.1
 ; FLATSCR-NEXT:    s_mov_b32 s2, s32
-; FLATSCR-NEXT:    s_movk_i32 s3, 0x1000
-; FLATSCR-NEXT:    s_add_i32 s4, s2, s3
+; FLATSCR-NEXT:    s_add_i32 s3, s2, 0x1000
 ; FLATSCR-NEXT:    v_mov_b32_e32 v1, 0
+; FLATSCR-NEXT:    s_add_u32 s2, s2, 0x1000
 ; FLATSCR-NEXT:    v_mov_b32_e32 v2, 1
-; FLATSCR-NEXT:    s_add_u32 s2, s2, s3
 ; FLATSCR-NEXT:    scratch_store_dwordx2 off, v[1:2], s2
 ; FLATSCR-NEXT:    s_lshl_b32 s2, s6, 2
-; FLATSCR-NEXT:    s_mov_b32 s32, s4
-; FLATSCR-NEXT:    s_add_i32 s4, s4, s2
-; FLATSCR-NEXT:    scratch_load_dword v2, off, s4
+; FLATSCR-NEXT:    s_mov_b32 s32, s3
+; FLATSCR-NEXT:    s_add_i32 s3, s3, s2
+; FLATSCR-NEXT:    scratch_load_dword v2, off, s3
 ; FLATSCR-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
 ; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; FLATSCR-NEXT:    v_add_u32_e32 v0, v2, v0
@@ -255,7 +254,7 @@
 ; FLATSCR-LABEL: func_non_entry_block_static_alloca_align4:
 ; FLATSCR:       ; %bb.0: ; %entry
 ; FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; FLATSCR-NEXT:    s_mov_b32 s5, s33
+; FLATSCR-NEXT:    s_mov_b32 s4, s33
 ; FLATSCR-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
 ; FLATSCR-NEXT:    s_mov_b32 s33, s32
 ; FLATSCR-NEXT:    s_add_u32 s32, s32, 16
@@ -267,16 +266,15 @@
 ; FLATSCR-NEXT:    s_cbranch_execz BB2_3
 ; FLATSCR-NEXT:  ; %bb.2: ; %bb.1
 ; FLATSCR-NEXT:    s_mov_b32 s2, s32
-; FLATSCR-NEXT:    s_movk_i32 s3, 0x1000
-; FLATSCR-NEXT:    s_add_i32 s4, s2, s3
+; FLATSCR-NEXT:    s_add_i32 s3, s2, 0x1000
+; FLATSCR-NEXT:    s_add_u32 s2, s2, 0x1000
 ; FLATSCR-NEXT:    v_mov_b32_e32 v2, 0
 ; FLATSCR-NEXT:    v_mov_b32_e32 v3, 1
-; FLATSCR-NEXT:    s_add_u32 s2, s2, s3
 ; FLATSCR-NEXT:    scratch_store_dwordx2 off, v[2:3], s2
-; FLATSCR-NEXT:    v_lshl_add_u32 v2, v4, 2, s4
+; FLATSCR-NEXT:    v_lshl_add_u32 v2, v4, 2, s3
 ; FLATSCR-NEXT:    scratch_load_dword v2, v2, off
 ; FLATSCR-NEXT:    v_and_b32_e32 v3, 0x3ff, v5
-; FLATSCR-NEXT:    s_mov_b32 s32, s4
+; FLATSCR-NEXT:    s_mov_b32 s32, s3
 ; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; FLATSCR-NEXT:    v_add_u32_e32 v2, v2, v3
 ; FLATSCR-NEXT:    global_store_dword v[0:1], v2, off
@@ -286,7 +284,7 @@
 ; FLATSCR-NEXT:    global_store_dword v[0:1], v0, off
 ; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; FLATSCR-NEXT:    s_sub_u32 s32, s32, 16
-; FLATSCR-NEXT:    s_mov_b32 s33, s5
+; FLATSCR-NEXT:    s_mov_b32 s33, s4
 ; FLATSCR-NEXT:    s_setpc_b64 s[30:31]
 
 entry: