diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index ddf251f38bf..743bf6f7a92 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -273,6 +273,16 @@ bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
          (Dest->getPrimitiveSizeInBits() % 32 == 0);
 }
 
+bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
+  // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
+  // limited number of native 64-bit operations. Shrinking an operation to fit
+  // in a single 32-bit register should always be helpful. As currently used,
+  // this is much less general than the name suggests, and is only used in
+  // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
+  // not profitable, and may actually be harmful.
+  return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
+}
+
 //===---------------------------------------------------------------------===//
 // TargetLowering Callbacks
 //===---------------------------------------------------------------------===//
diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
index 2595c51d166..a2504ef1661 100644
--- a/lib/Target/R600/AMDGPUISelLowering.h
+++ b/lib/Target/R600/AMDGPUISelLowering.h
@@ -87,6 +87,8 @@ public:
   virtual bool isFNegFree(EVT VT) const override;
   virtual bool isTruncateFree(EVT Src, EVT Dest) const override;
   virtual bool isTruncateFree(Type *Src, Type *Dest) const override;
+  virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const override;
+
   virtual MVT getVectorIdxTy() const override;
   virtual bool isLoadBitCastBeneficial(EVT, EVT) const override;
   virtual SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
diff --git a/test/CodeGen/R600/trunc.ll b/test/CodeGen/R600/trunc.ll
index 6bbd7f7b510..e775fe7c989 100644
--- a/test/CodeGen/R600/trunc.ll
+++ b/test/CodeGen/R600/trunc.ll
@@ -16,14 +16,27 @@ define void @trunc_i64_to_i32_store(i32 addrspace(1)* %out, i64 %in) {
   ret void
 }
 
+; SI-LABEL: @trunc_load_shl_i64:
+; SI-DAG: S_LOAD_DWORDX2
+; SI-DAG: S_LOAD_DWORD [[SREG:s[0-9]+]],
+; SI: S_LSHL_B32 [[SHL:s[0-9]+]], [[SREG]], 2
+; SI: V_MOV_B32_e32 [[VSHL:v[0-9]+]], [[SHL]]
+; SI: BUFFER_STORE_DWORD [[VSHL]],
+define void @trunc_load_shl_i64(i32 addrspace(1)* %out, i64 %a) {
+  %b = shl i64 %a, 2
+  %result = trunc i64 %b to i32
+  store i32 %result, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
 ; SI-LABEL: @trunc_shl_i64:
-; SI: S_LOAD_DWORDX2
-; SI: S_LOAD_DWORDX2 [[SREG:s\[[0-9]+:[0-9]+\]]]
-; SI: S_LSHL_B64 s{{\[}}[[LO_SREG:[0-9]+]]:{{[0-9]+\]}}, [[SREG]], 2
-; SI: MOV_B32_e32 v[[LO_VREG:[0-9]+]], s[[LO_SREG]]
+; SI: S_LOAD_DWORDX2 s{{\[}}[[LO_SREG:[0-9]+]]:{{[0-9]+\]}},
+; SI: V_ADD_I32_e32 v[[LO_ADD:[0-9]+]], s[[LO_SREG]],
+; SI: V_LSHL_B64 v{{\[}}[[LO_VREG:[0-9]+]]:{{[0-9]+\]}}, v{{\[}}[[LO_ADD]]:{{[0-9]+\]}}, 2
 ; SI: BUFFER_STORE_DWORD v[[LO_VREG]],
 define void @trunc_shl_i64(i32 addrspace(1)* %out, i64 %a) {
-  %b = shl i64 %a, 2
+  %aa = add i64 %a, 234 ; Prevent shrinking store.
+  %b = shl i64 %aa, 2
   %result = trunc i64 %b to i32
   store i32 %result, i32 addrspace(1)* %out, align 4
   ret void