Index: lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
+++ lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
@@ -554,7 +554,7 @@
       BitOffset += SrcEltBits;
       if (BitOffset >= WideBits) {
         WideIdx++;
-        Offset -= WideBits;
+        Offset -= WideBits / 8;
         if (Offset > 0) {
           ShAmt = DAG.getConstant(SrcEltBits - Offset,
                                   TLI.getShiftAmountTy(WideVT));
Index: test/CodeGen/X86/pr15267.ll
===================================================================
--- test/CodeGen/X86/pr15267.ll
+++ test/CodeGen/X86/pr15267.ll
@@ -4,8 +4,7 @@
   %ret = load <4 x i3>* %in, align 1
   ret <4 x i3> %ret
 }
-
-; CHECK: test1
+; CHECK-LABEL: test1
 ; CHECK: movzwl
 ; CHECK: shrl $3
 ; CHECK: andl $7
@@ -25,7 +24,7 @@
   ret <4 x i1> %ret
 }
 
-; CHECK: test2
+; CHECK-LABEL: test2
 ; CHECK: movzbl
 ; CHECK: shrl
 ; CHECK: andl $1
@@ -46,7 +45,7 @@
   ret <4 x i64> %sext
 }
 
-; CHECK: test3
+; CHECK-LABEL: test3
 ; CHECK: movzbl
 ; CHECK: movq
 ; CHECK: shlq
@@ -67,3 +66,71 @@
 ; CHECK: vpunpcklqdq
 ; CHECK: vinsertf128
 ; CHECK: ret
+
+define <16 x i4> @test4(<16 x i4>* %in) nounwind {
+  %ret = load <16 x i4>* %in, align 1
+  ret <16 x i4> %ret
+}
+
+; CHECK-LABEL: test4
+; CHECK: movl
+; CHECK-NEXT: shrl
+; CHECK-NEXT: andl
+; CHECK-NEXT: movl
+; CHECK-NEXT: andl
+; CHECK-NEXT: vmovd
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movl
+; CHECK-NEXT: shrl
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movl
+; CHECK-NEXT: shrl
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movl
+; CHECK-NEXT: shrl
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movl
+; CHECK-NEXT: shrl
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movl
+; CHECK-NEXT: shrl
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movl
+; CHECK-NEXT: shrl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movq
+; CHECK-NEXT: shrq
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movq
+; CHECK-NEXT: shrq
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movq
+; CHECK-NEXT: shrq
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movq
+; CHECK-NEXT: shrq
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movq
+; CHECK-NEXT: shrq
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movq
+; CHECK-NEXT: shrq
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: movq
+; CHECK-NEXT: shrq
+; CHECK-NEXT: andl
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: shrq
+; CHECK-NEXT: vpinsrb
+; CHECK-NEXT: retq