Index: lib/Target/R600/SIInstrInfo.td
===================================================================
--- lib/Target/R600/SIInstrInfo.td
+++ lib/Target/R600/SIInstrInfo.td
@@ -362,12 +362,12 @@
   : SOPC_Helper<op, SSrc_64, i64, opName, cond>;
 
 class SOPK_32 <bits<5> op, string opName, list<dag> pattern> : SOPK <
-  op, (outs SReg_32:$dst), (ins i16imm:$src0),
+  op, (outs SReg_32:$dst), (ins u16imm:$src0),
   opName#" $dst, $src0", pattern
 >;
 
 class SOPK_64 <bits<5> op, string opName, list<dag> pattern> : SOPK <
-  op, (outs SReg_64:$dst), (ins i16imm:$src0),
+  op, (outs SReg_64:$dst), (ins u16imm:$src0),
   opName#" $dst, $src0", pattern
 >;
 
Index: lib/Target/R600/SILowerControlFlow.cpp
===================================================================
--- lib/Target/R600/SILowerControlFlow.cpp
+++ lib/Target/R600/SILowerControlFlow.cpp
@@ -586,6 +586,8 @@
     MachineBasicBlock::iterator Start = MBB.getFirstNonPHI();
     const MCInstrDesc &SMovK = TII->get(AMDGPU::S_MOVK_I32);
 
+    assert(isInt<16>(StackOffset) && isInt<16>(StackSizeBytes));
+
     BuildMI(MBB, Start, NoDL, SMovK, AMDGPU::FLAT_SCR_LO)
       .addImm(StackOffset);
 
Index: lib/Target/R600/SIShrinkInstructions.cpp
===================================================================
--- lib/Target/R600/SIShrinkInstructions.cpp
+++ lib/Target/R600/SIShrinkInstructions.cpp
@@ -189,6 +189,19 @@
       Next = std::next(I);
       MachineInstr &MI = *I;
 
+      // Try to use S_MOVK_I32, which will save 4 bytes for small immediates.
+      if (MI.getOpcode() == AMDGPU::S_MOV_B32) {
+        const MachineOperand &Src = MI.getOperand(1);
+
+        // TODO: Handle FPImm?
+        if (Src.isImm()) {
+          if (isInt<16>(Src.getImm()) && !TII->isInlineConstant(Src)) {
+            MI.setDesc(TII->get(AMDGPU::S_MOVK_I32));
+            continue;
+          }
+        }
+      }
+
       if (!TII->hasVALU32BitEncoding(MI.getOpcode()))
         continue;
 
Index: test/CodeGen/R600/flat-address-space.ll
===================================================================
--- test/CodeGen/R600/flat-address-space.ll
+++ test/CodeGen/R600/flat-address-space.ll
@@ -156,8 +156,8 @@
 ; Check for prologue initializing special SGPRs pointing to scratch.
 ; CHECK-LABEL: {{^}}store_flat_scratch:
 ; CHECK: S_MOVK_I32 flat_scratch_lo, 0
-; CHECK-NO-PROMOTE: S_MOVK_I32 flat_scratch_hi, 40
-; CHECK-PROMOTE: S_MOVK_I32 flat_scratch_hi, 0
+; CHECK-NO-PROMOTE: S_MOVK_I32 flat_scratch_hi, 0x28{{$}}
+; CHECK-PROMOTE: S_MOVK_I32 flat_scratch_hi, 0x0{{$}}
 ; CHECK: FLAT_STORE_DWORD
 ; CHECK: S_BARRIER
 ; CHECK: FLAT_LOAD_DWORD
Index: test/CodeGen/R600/or.ll
===================================================================
--- test/CodeGen/R600/or.ll
+++ test/CodeGen/R600/or.ll
@@ -116,7 +116,7 @@
 
 ; SI-LABEL: {{^}}vector_or_i64_loadimm:
 ; SI-DAG: S_MOV_B32 [[LO_S_IMM:s[0-9]+]], 0xdf77987f
-; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 0x146f
+; SI-DAG: S_MOVK_I32 [[HI_S_IMM:s[0-9]+]], 0x146f
 ; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
 ; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
 ; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
Index: test/CodeGen/R600/s_movk_i32.ll
===================================================================
--- /dev/null
+++ test/CodeGen/R600/s_movk_i32.ll
@@ -0,0 +1,184 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+; SI-LABEL: {{^}}s_movk_i32_k0:
+; SI-DAG: S_MOV_B32 [[LO_S_IMM:s[0-9]+]], 0xffff{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 1{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k0(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 4295032831 ; ((1 << 16) - 1) | (1 << 32)
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k1:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x7fff{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 1{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k1(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 4295000063 ; ((1 << 15) - 1) | (1 << 32)
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k2:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x7fff{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 64{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k2(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 274877939711 ; ((1 << 15) - 1) | (64 << 32)
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k3:
+; SI-DAG: S_MOV_B32 [[LO_S_IMM:s[0-9]+]], 0x8000{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 1{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k3(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 4295000064 ; (1 << 15) | (1 << 32)
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k4:
+; SI-DAG: S_MOV_B32 [[LO_S_IMM:s[0-9]+]], 0x20000{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 1{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k4(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 4295098368 ; (1 << 17) | (1 << 32)
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k5:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0xffef{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 0xff00ffff{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k5(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 18374967954648334319 ; -17 & 0xff00ffffffffffff
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k6:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x41{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 63{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k6(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 270582939713 ; 65 | (63 << 32)
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k7:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x2000{{$}}
+; SI-DAG: S_MOVK_I32 [[HI_S_IMM:s[0-9]+]], 0x4000{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k7(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 70368744185856; ((1 << 13)) | ((1 << 14) << 32)
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+
+; SI-LABEL: {{^}}s_movk_i32_k8:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x8000{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 0x11111111{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k8(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 1229782942255906816 ; 0x11111111ffff8000
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k9:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x8001{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 0x11111111{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k9(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 1229782942255906817 ; 0x11111111ffff8001
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k10:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x8888{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 0x11111111{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k10(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 1229782942255909000 ; 0x11111111ffff8888
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k11:
+; SI-DAG: S_MOVK_I32 [[LO_S_IMM:s[0-9]+]], 0x8fff{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 0x11111111{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k11(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 1229782942255910911 ; 0x11111111ffff8fff
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; SI-LABEL: {{^}}s_movk_i32_k12:
+; SI-DAG: S_MOV_B32 [[LO_S_IMM:s[0-9]+]], 0xffff7001{{$}}
+; SI-DAG: S_MOV_B32 [[HI_S_IMM:s[0-9]+]], 0x11111111{{$}}
+; SI-DAG: BUFFER_LOAD_DWORDX2 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}},
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[LO_S_IMM]], v[[LO_VREG]]
+; SI-DAG: V_OR_B32_e32 {{v[0-9]+}}, [[HI_S_IMM]], v[[HI_VREG]]
+; SI: S_ENDPGM
+define void @s_movk_i32_k12(i64 addrspace(1)* %out, i64 addrspace(1)* %a, i64 addrspace(1)* %b) {
+  %loada = load i64 addrspace(1)* %a, align 4
+  %or = or i64 %loada, 1229782942255902721 ; 0x11111111ffff7001
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
Index: test/CodeGen/R600/smrd.ll
===================================================================
--- test/CodeGen/R600/smrd.ll
+++ test/CodeGen/R600/smrd.ll
@@ -24,7 +24,7 @@
 
 ; SMRD load with an offset greater than the largest possible immediate.
 ; CHECK-LABEL: {{^}}smrd2:
-; CHECK: S_MOV_B32 s[[OFFSET:[0-9]]], 0x400
+; CHECK: S_MOVK_I32 s[[OFFSET:[0-9]]], 0x400
 ; CHECK: S_LOAD_DWORD s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
 ; CHECK: S_ENDPGM
 define void @smrd2(i32 addrspace(1)* %out, i32 addrspace(2)* %ptr) {
@@ -81,7 +81,7 @@
 ; largets possible immediate.
 ; immediate offset.
 ; CHECK-LABEL: {{^}}smrd_load_const2:
-; CHECK: S_MOV_B32 s[[OFFSET:[0-9]]], 0x400
+; CHECK: S_MOVK_I32 s[[OFFSET:[0-9]]], 0x400
 ; CHECK: S_BUFFER_LOAD_DWORD s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
 define void @smrd_load_const2(<16 x i8> addrspace(2)* inreg, <16 x i8> addrspace(2)* inreg, <32 x i8> addrspace(2)* inreg, i32 inreg, <2 x i32>, <2 x i32>, <2 x i32>, <3 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, float, float, float, float, float, float, float, float, float) #0 {
 main_body: