Index: lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- lib/Target/X86/X86ISelLowering.cpp
+++ lib/Target/X86/X86ISelLowering.cpp
@@ -19325,8 +19325,12 @@
   // Merge this shuffle operation's mask into our accumulated mask. This is
   // a bit tricky as the shuffle may have a different size from the root.
   if (OpMask.size() == IncomingMask.size()) {
-    for (int M : IncomingMask)
-      Mask.push_back(OpMask[M]);
+    for (int M : IncomingMask) {
+      if (M == SM_SentinelZero)
+        Mask.push_back(SM_SentinelZero);
+      else
+        Mask.push_back(OpMask[M]);
+    }
   } else if (OpMask.size() < IncomingMask.size()) {
     assert(IncomingMask.size() % OpMask.size() == 0 &&
            "The smaller number of elements must divide the larger.");
Index: test/CodeGen/X86/pr20540.ll
===================================================================
--- /dev/null
+++ test/CodeGen/X86/pr20540.ll
@@ -0,0 +1,13 @@
+; RUN: llc  < %s -mtriple=x86_64-apple-darwin -march=x86  -mcpu=corei7 | FileCheck %s
+
+declare i32 @llvm.x86.sse2.pmovmskb.128(<16 x i8>)
+
+define void @f_fu_(float* noalias nocapture %RET, <8 x i8> %__mask) {
+  ; CHECK: pshufb
+  %mask8.i.i.i = shufflevector <8 x i8> %__mask, <8 x i8> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 8, i32 8, i32 8, i32 8, i32 8, i32 8, i32 8>
+  %m.i.i.i = call i32 @llvm.x86.sse2.pmovmskb.128(<16 x i8> %mask8.i.i.i)
+  %RET.i = bitcast float* %RET to i32*
+  store i32 %m.i.i.i, i32* %RET.i
+  ret void
+}
+