Index: llvm/trunk/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
+++ llvm/trunk/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp
@@ -514,6 +514,7 @@
   DebugLoc DL = CI.I->getDebugLoc();
 
   unsigned BaseReg = AddrReg->getReg();
+  unsigned BaseSubReg = AddrReg->getSubReg();
   unsigned BaseRegFlags = 0;
   if (CI.BaseOff) {
     unsigned ImmReg = MRI->createVirtualRegister(&AMDGPU::SGPR_32RegClass);
@@ -525,15 +526,16 @@
 
     TII->getAddNoCarry(*MBB, CI.Paired, DL, BaseReg)
       .addReg(ImmReg)
-      .addReg(AddrReg->getReg());
+      .addReg(AddrReg->getReg(), 0, BaseSubReg);
+    BaseSubReg = 0;
   }
 
   MachineInstrBuilder Read2 = BuildMI(*MBB, CI.Paired, DL, Read2Desc, DestReg)
-                                  .addReg(BaseReg, BaseRegFlags) // addr
-                                  .addImm(NewOffset0)            // offset0
-                                  .addImm(NewOffset1)            // offset1
-                                  .addImm(0)                     // gds
-                                  .cloneMergedMemRefs({&*CI.I, &*CI.Paired});
+                        .addReg(BaseReg, BaseRegFlags, BaseSubReg) // addr
+                        .addImm(NewOffset0)                        // offset0
+                        .addImm(NewOffset1)                        // offset1
+                        .addImm(0)                                 // gds
+                        .cloneMergedMemRefs({&*CI.I, &*CI.Paired});
 
   (void)Read2;
 
@@ -601,6 +603,7 @@
   DebugLoc DL = CI.I->getDebugLoc();
 
   unsigned BaseReg = AddrReg->getReg();
+  unsigned BaseSubReg = AddrReg->getSubReg();
   unsigned BaseRegFlags = 0;
   if (CI.BaseOff) {
     unsigned ImmReg = MRI->createVirtualRegister(&AMDGPU::SGPR_32RegClass);
@@ -612,17 +615,18 @@
 
     TII->getAddNoCarry(*MBB, CI.Paired, DL, BaseReg)
       .addReg(ImmReg)
-      .addReg(AddrReg->getReg());
+      .addReg(AddrReg->getReg(), 0, BaseSubReg);
+    BaseSubReg = 0;
   }
 
   MachineInstrBuilder Write2 = BuildMI(*MBB, CI.Paired, DL, Write2Desc)
-                                   .addReg(BaseReg, BaseRegFlags) // addr
-                                   .add(*Data0)                   // data0
-                                   .add(*Data1)                   // data1
-                                   .addImm(NewOffset0)            // offset0
-                                   .addImm(NewOffset1)            // offset1
-                                   .addImm(0)                     // gds
-                                   .cloneMergedMemRefs({&*CI.I, &*CI.Paired});
+                        .addReg(BaseReg, BaseRegFlags, BaseSubReg) // addr
+                        .add(*Data0)                               // data0
+                        .add(*Data1)                               // data1
+                        .addImm(NewOffset0)                        // offset0
+                        .addImm(NewOffset1)                        // offset1
+                        .addImm(0)                                 // gds
+                        .cloneMergedMemRefs({&*CI.I, &*CI.Paired});
 
   moveInstsAfter(Write2, CI.InstsToMove);
 
Index: llvm/trunk/test/CodeGen/AMDGPU/merge-load-store-vreg.mir
===================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/merge-load-store-vreg.mir
+++ llvm/trunk/test/CodeGen/AMDGPU/merge-load-store-vreg.mir
@@ -6,7 +6,7 @@
 # $vcc, which is used in _e32); this ensures that $vcc is not inadvertently
 # clobbered.
 
-# GCN-LABEL: name: kernel
+# GCN-LABEL: name: ds_combine_base_offset{{$}}
 
 # VI: V_ADD_I32_e64 %6, %0,
 # VI-NEXT: DS_WRITE2_B32 killed %7, %0, %3, 0, 8,
@@ -21,7 +21,37 @@
 --- |
   @0 = internal unnamed_addr addrspace(3) global [256 x float] undef, align 4
 
-  define amdgpu_kernel void @kernel() {
+  define amdgpu_kernel void @ds_combine_base_offset() {
+    bb.0:
+      br label %bb2
+
+    bb1:
+      ret void
+
+    bb2:
+      %tmp = getelementptr inbounds [256 x float], [256 x float] addrspace(3)* @0, i32 0, i32 0
+      %tmp1 = getelementptr inbounds float, float addrspace(3)* %tmp, i32 8
+      %tmp2 = getelementptr inbounds float, float addrspace(3)* %tmp, i32 16
+      %tmp3 = getelementptr inbounds float, float addrspace(3)* %tmp, i32 24
+      br label %bb1
+    }
+
+  define amdgpu_kernel void @ds_combine_base_offset_subreg() {
+    bb.0:
+      br label %bb2
+
+    bb1:
+      ret void
+
+    bb2:
+      %tmp = getelementptr inbounds [256 x float], [256 x float] addrspace(3)* @0, i32 0, i32 0
+      %tmp1 = getelementptr inbounds float, float addrspace(3)* %tmp, i32 8
+      %tmp2 = getelementptr inbounds float, float addrspace(3)* %tmp, i32 16
+      %tmp3 = getelementptr inbounds float, float addrspace(3)* %tmp, i32 24
+      br label %bb1
+    }
+
+  define amdgpu_kernel void @ds_combine_subreg() {
     bb.0:
       br label %bb2
 
@@ -36,7 +66,7 @@
       br label %bb1
     }
 ---
-name:            kernel
+name:            ds_combine_base_offset
 body:             |
   bb.0:
     %0:vgpr_32 = IMPLICIT_DEF
@@ -58,3 +88,69 @@
     S_CBRANCH_VCCNZ %bb.1, implicit $vcc
     S_BRANCH %bb.1
 ...
+
+# GCN-LABEL: name: ds_combine_base_offset_subreg{{$}}
+
+# VI: V_ADD_I32_e64 %6, %0.sub0,
+# VI-NEXT: DS_WRITE2_B32 killed %7, %0.sub0, %3.sub0, 0, 8,
+# VI: V_ADD_I32_e64 %10, %3.sub0,
+# VI-NEXT: DS_READ2_B32 killed %11, 0, 8,
+
+# GFX9: V_ADD_U32_e64 %6, %0.sub0,
+# GFX9-NEXT: DS_WRITE2_B32_gfx9 killed %7, %0.sub0, %3.sub0, 0, 8,
+# GFX9: V_ADD_U32_e64 %9, %3.sub0,
+# GFX9-NEXT: DS_READ2_B32_gfx9 killed %10, 0, 8,
+---
+name:            ds_combine_base_offset_subreg
+body:             |
+  bb.0:
+    %0:vreg_64 = IMPLICIT_DEF
+    S_BRANCH %bb.2
+
+  bb.1:
+    S_ENDPGM
+
+  bb.2:
+    %1:sreg_64_xexec = V_CMP_NE_U32_e64 %0.sub0, 0, implicit $exec
+    %2:vgpr_32 = V_CNDMASK_B32_e64 0, 1, %1, implicit $exec
+    V_CMP_NE_U32_e32 1, %2, implicit-def $vcc, implicit $exec
+    DS_WRITE_B32 %0.sub0, %0.sub0, 1024, 0, implicit $m0, implicit $exec :: (store 4 into %ir.tmp)
+    undef %3.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec
+    DS_WRITE_B32 %0.sub0, %3.sub0, 1056, 0, implicit $m0, implicit $exec :: (store 4 into %ir.tmp1)
+    %4:vgpr_32 = DS_READ_B32 %3.sub0, 1088, 0, implicit $m0, implicit $exec :: (load 4 from %ir.tmp2)
+    %5:vgpr_32 = DS_READ_B32 %3.sub0, 1120, 0, implicit $m0, implicit $exec :: (load 4 from %ir.tmp3)
+    $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+    S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+    S_BRANCH %bb.1
+...
+
+# GCN-LABEL: name: ds_combine_subreg{{$}}
+
+# VI: DS_WRITE2_B32 %0.sub0, %0.sub0, %3.sub0, 0, 8,
+# VI: DS_READ2_B32 %3.sub0, 0, 8,
+
+# GFX9: DS_WRITE2_B32_gfx9 %0.sub0, %0.sub0, %3.sub0, 0, 8,
+# GFX9: DS_READ2_B32_gfx9 %3.sub0, 0, 8,
+---
+name:            ds_combine_subreg
+body:             |
+  bb.0:
+    %0:vreg_64 = IMPLICIT_DEF
+    S_BRANCH %bb.2
+
+  bb.1:
+    S_ENDPGM
+
+  bb.2:
+    %1:sreg_64_xexec = V_CMP_NE_U32_e64 %0.sub0, 0, implicit $exec
+    %2:vgpr_32 = V_CNDMASK_B32_e64 0, 1, %1, implicit $exec
+    V_CMP_NE_U32_e32 1, %2, implicit-def $vcc, implicit $exec
+    DS_WRITE_B32 %0.sub0, %0.sub0, 0, 0, implicit $m0, implicit $exec :: (store 4 into %ir.tmp)
+    undef %3.sub0:vreg_64 = V_MOV_B32_e32 0, implicit $exec
+    DS_WRITE_B32 %0.sub0, %3.sub0, 32, 0, implicit $m0, implicit $exec :: (store 4 into %ir.tmp1)
+    %4:vgpr_32 = DS_READ_B32 %3.sub0, 0, 0, implicit $m0, implicit $exec :: (load 4 from %ir.tmp2)
+    %5:vgpr_32 = DS_READ_B32 %3.sub0, 32, 0, implicit $m0, implicit $exec :: (load 4 from %ir.tmp3)
+    $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+    S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+    S_BRANCH %bb.1
+...