diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 2f5ee0c56c6..f98f0dd292a 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -848,6 +848,31 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op,
     }
     break;
   }
+  case ISD::BUILD_PAIR: {
+    EVT HalfVT = Op.getOperand(0).getValueType();
+    unsigned HalfBitWidth = HalfVT.getScalarSizeInBits();
+
+    APInt MaskLo = NewMask.getLoBits(HalfBitWidth).trunc(HalfBitWidth);
+    APInt MaskHi = NewMask.getHiBits(HalfBitWidth).trunc(HalfBitWidth);
+
+    APInt KnownZeroLo, KnownOneLo;
+    APInt KnownZeroHi, KnownOneHi;
+
+    if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownZeroLo,
+                             KnownOneLo, TLO, Depth + 1))
+      return true;
+
+    if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownZeroHi,
+                             KnownOneHi, TLO, Depth + 1))
+      return true;
+
+    KnownZero = KnownZeroLo.zext(BitWidth) |
+                KnownZeroHi.zext(BitWidth).shl(HalfBitWidth);
+
+    KnownOne = KnownOneLo.zext(BitWidth) |
+               KnownOneHi.zext(BitWidth).shl(HalfBitWidth);
+    break;
+  }
   case ISD::ZERO_EXTEND: {
     unsigned OperandBitWidth =
       Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
diff --git a/test/CodeGen/R600/simplify-demanded-bits-build-pair.ll b/test/CodeGen/R600/simplify-demanded-bits-build-pair.ll
new file mode 100644
index 00000000000..d9f60ea1a4d
--- /dev/null
+++ b/test/CodeGen/R600/simplify-demanded-bits-build-pair.ll
@@ -0,0 +1,36 @@
+; RUN: llc -verify-machineinstrs -march=r600 -mcpu=SI < %s | FileCheck -check-prefix=SI %s
+
+; 64-bit select was originally lowered with a build_pair, and this
+; could be simplified to 1 cndmask instead of 2, but that broken when
+; it started being implemented with a v2i32 build_vector and
+; bitcasting.
+define void @trunc_select_i64(i32 addrspace(1)* %out, i64 %a, i64 %b, i32 %c) {
+  %cmp = icmp eq i32 %c, 0
+  %select = select i1 %cmp, i64 %a, i64 %b
+  %trunc = trunc i64 %select to i32
+  store i32 %trunc, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @trunc_load_alloca_i64:
+; SI: V_MOVRELS_B32
+; SI-NOT: V_MOVRELS_B32
+; SI: S_ENDPGM
+define void @trunc_load_alloca_i64(i64 addrspace(1)* %out, i32 %a, i32 %b) {
+  %idx = add i32 %a, %b
+  %alloca = alloca i64, i32 4
+  %gep0 = getelementptr i64* %alloca, i64 0
+  %gep1 = getelementptr i64* %alloca, i64 1
+  %gep2 = getelementptr i64* %alloca, i64 2
+  %gep3 = getelementptr i64* %alloca, i64 3
+  store i64 24, i64* %gep0, align 8
+  store i64 9334, i64* %gep1, align 8
+  store i64 3935, i64* %gep2, align 8
+  store i64 9342, i64* %gep3, align 8
+  %gep = getelementptr i64* %alloca, i32 %idx
+  %load = load i64* %gep, align 8
+  %mask = and i64 %load, 4294967296
+  %add = add i64 %mask, -1
+  store i64 %add, i64 addrspace(1)* %out, align 4
+  ret void
+}