Index: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -4463,6 +4463,9 @@
                ? ByteProvider::getConstantZero()
                : calculateByteProvider(NarrowOp, Index, Depth + 1);
   }
+  case ISD::BSWAP:
+    return calculateByteProvider(Op->getOperand(0), ByteWidth - Index - 1,
+                                 Depth + 1);
   case ISD::LOAD: {
     auto L = cast<LoadSDNode>(Op.getNode());
 
Index: llvm/trunk/test/CodeGen/AArch64/load-combine-big-endian.ll
===================================================================
--- llvm/trunk/test/CodeGen/AArch64/load-combine-big-endian.ll
+++ llvm/trunk/test/CodeGen/AArch64/load-combine-big-endian.ll
@@ -331,3 +331,25 @@
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[0]) | (i32) bswap(p[1] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr   w8, [x0]
+; CHECK-NEXT: rev w0, w8
+; CHECK-NEXT: ret
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp5, 16
+  %tmp7 = or i32 %tmp6, %tmp2
+  ret i32 %tmp7
+}
Index: llvm/trunk/test/CodeGen/AArch64/load-combine.ll
===================================================================
--- llvm/trunk/test/CodeGen/AArch64/load-combine.ll
+++ llvm/trunk/test/CodeGen/AArch64/load-combine.ll
@@ -318,3 +318,26 @@
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[1]) | (i32) bswap(p[0] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr    w8, [x0]
+; CHECK-NEXT: rev w0, w8
+; CHECK-NEXT: ret
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp2, 16
+  %tmp7 = or i32 %tmp6, %tmp5
+  ret i32 %tmp7
+}
Index: llvm/trunk/test/CodeGen/ARM/load-combine-big-endian.ll
===================================================================
--- llvm/trunk/test/CodeGen/ARM/load-combine-big-endian.ll
+++ llvm/trunk/test/CodeGen/ARM/load-combine-big-endian.ll
@@ -449,3 +449,37 @@
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[0]) | (i32) bswap(p[1] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr r0, [r0]
+; CHECK-NEXT: mov r1, #65280
+; CHECK-NEXT: mov r2, #16711680
+; CHECK-NEXT: and r1, r1, r0, lsr #8
+; CHECK-NEXT: and r2, r2, r0, lsl #8
+; CHECK-NEXT: orr r1, r1, r0, lsr #24
+; CHECK-NEXT: orr r0, r2, r0, lsl #24
+; CHECK-NEXT: orr r0, r0, r1
+; CHECK-NEXT: mov pc, lr
+
+; CHECK-ARMv6-LABEL: load_i32_by_bswap_i16:
+; CHECK-ARMv6: ldr  r0, [r0]
+; CHECK-ARMv6-NEXT: rev r0, r0
+; CHECK-ARMv6-NEXT: bx  lr
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp5, 16
+  %tmp7 = or i32 %tmp6, %tmp2
+  ret i32 %tmp7
+}
Index: llvm/trunk/test/CodeGen/ARM/load-combine.ll
===================================================================
--- llvm/trunk/test/CodeGen/ARM/load-combine.ll
+++ llvm/trunk/test/CodeGen/ARM/load-combine.ll
@@ -407,3 +407,37 @@
   %tmp18 = or i32 %tmp13, %tmp17
   ret i32 %tmp18
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p; // p is 4 byte aligned
+; (i32) bswap(p[1]) | (i32) bswap(p[0] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK: ldr  r0, [r0]
+; CHECK-NEXT: mov r1, #65280
+; CHECK-NEXT: mov r2, #16711680
+; CHECK-NEXT: and r1, r1, r0, lsr #8
+; CHECK-NEXT: and r2, r2, r0, lsl #8
+; CHECK-NEXT: orr r1, r1, r0, lsr #24
+; CHECK-NEXT: orr r0, r2, r0, lsl #24
+; CHECK-NEXT: orr r0, r0, r1
+; CHECK-NEXT: mov pc, lr
+
+; CHECK-ARMv6-LABEL: load_i32_by_bswap_i16:
+; CHECK-ARMv6: ldr r0, [r0]
+; CHECK-ARMv6-NEXT: rev r0, r0
+; CHECK-ARMv6-NEXT: bx  lr
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp2, 16
+  %tmp7 = or i32 %tmp6, %tmp5
+  ret i32 %tmp7
+}
Index: llvm/trunk/test/CodeGen/X86/load-combine.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/load-combine.ll
+++ llvm/trunk/test/CodeGen/X86/load-combine.ll
@@ -869,3 +869,35 @@
   store i64 %conv75, i64* %dst, align 8
   ret void
 }
+
+declare i16 @llvm.bswap.i16(i16)
+
+; i16* p;
+; (i32) bswap(p[1]) | (i32) bswap(p[0] << 16)
+define i32 @load_i32_by_bswap_i16(i32* %arg) {
+; CHECK-LABEL: load_i32_by_bswap_i16:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; CHECK-NEXT:    movl (%eax), %eax
+; CHECK-NEXT:    bswapl %eax
+; CHECK-NEXT:    retl
+;
+; CHECK64-LABEL: load_i32_by_bswap_i16:
+; CHECK64:       # BB#0:
+; CHECK64-NEXT:    movl (%rdi), %eax
+; CHECK64-NEXT:    bswapl %eax
+; CHECK64-NEXT:    retq
+
+
+  %tmp = bitcast i32* %arg to i16*
+  %tmp1 = load i16, i16* %tmp, align 4
+  %tmp11 = call i16 @llvm.bswap.i16(i16 %tmp1)
+  %tmp2 = zext i16 %tmp11 to i32
+  %tmp3 = getelementptr inbounds i16, i16* %tmp, i32 1
+  %tmp4 = load i16, i16* %tmp3, align 1
+  %tmp41 = call i16 @llvm.bswap.i16(i16 %tmp4)
+  %tmp5 = zext i16 %tmp41 to i32
+  %tmp6 = shl nuw nsw i32 %tmp2, 16
+  %tmp7 = or i32 %tmp6, %tmp5
+  ret i32 %tmp7
+}