Index: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
@@ -7553,6 +7553,23 @@
 
   // Check if this is a broadcast of a scalar. We special case lowering
   // for scalars so that we can more effectively fold with loads.
+  // First, look through bitcast: if the original value has a larger element
+  // type than the shuffle, the broadcast element is in essence truncated.
+  // Make that explicit to ease folding.
+  if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
+    EVT EltVT = VT.getVectorElementType();
+    SDValue V0 = V.getOperand(0);
+    EVT V0VT = V0.getValueType();
+
+    if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
+        ((V0.getOpcode() == ISD::BUILD_VECTOR ||
+         (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
+      V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
+      BroadcastIdx = 0;
+    }
+  }
+
+  // Also check the simpler case, where we can directly reuse the scalar.
   if (V.getOpcode() == ISD::BUILD_VECTOR ||
       (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
     V = V.getOperand(BroadcastIdx);
Index: llvm/trunk/lib/Target/X86/X86InstrSSE.td
===================================================================
--- llvm/trunk/lib/Target/X86/X86InstrSSE.td
+++ llvm/trunk/lib/Target/X86/X86InstrSSE.td
@@ -8338,6 +8338,13 @@
 defm VPBROADCASTQ  : avx2_broadcast<0x59, "vpbroadcastq", i64mem, loadi64, v2i64, v4i64>;
 
 let Predicates = [HasAVX2] in {
+  // loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
+  // This means we'll encounter truncated i32 loads; match that here.
+  def : Pat<(v8i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
+            (VPBROADCASTWrm addr:$src)>;
+  def : Pat<(v16i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
+            (VPBROADCASTWYrm addr:$src)>;
+
   // Provide aliases for broadcast from the same register class that
   // automatically does the extract.
   def : Pat<(v32i8 (X86VBroadcast (v32i8 VR256:$src))),
Index: llvm/trunk/test/CodeGen/X86/vector-shuffle-128-v16.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/vector-shuffle-128-v16.ll
+++ llvm/trunk/test/CodeGen/X86/vector-shuffle-128-v16.ll
@@ -1404,8 +1404,7 @@
 ;
 ; AVX2-LABEL: insert_dup_mem_v16i8_i32:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpbroadcastb %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastb (%rdi), %xmm0
 ; AVX2-NEXT:    retq
   %tmp = load i32, i32* %ptr, align 4
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0
@@ -1451,9 +1450,7 @@
 ;
 ; AVX2-LABEL: insert_dup_mem_v16i8_sext_i8:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    movsbl (%rdi), %eax
-; AVX2-NEXT:    vmovd %eax, %xmm0
-; AVX2-NEXT:    vpbroadcastb %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastb (%rdi), %xmm0
 ; AVX2-NEXT:    retq
   %tmp = load i8, i8* %ptr, align 1
   %tmp1 = sext i8 %tmp to i32
Index: llvm/trunk/test/CodeGen/X86/vector-shuffle-128-v8.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/vector-shuffle-128-v8.ll
+++ llvm/trunk/test/CodeGen/X86/vector-shuffle-128-v8.ll
@@ -2175,8 +2175,7 @@
 ;
 ; AVX2-LABEL: insert_dup_mem_v8i16_i32:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpbroadcastw %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastw (%rdi), %xmm0
 ; AVX2-NEXT:    retq
   %tmp = load i32, i32* %ptr, align 4
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0
Index: llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v16.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v16.ll
+++ llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v16.ll
@@ -3301,8 +3301,7 @@
 ;
 ; AVX2-LABEL: insert_dup_mem_v16i16_i32:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpbroadcastw %xmm0, %ymm0
+; AVX2-NEXT:    vpbroadcastw (%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %tmp = load i32, i32* %ptr, align 4
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0
Index: llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v32.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v32.ll
+++ llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v32.ll
@@ -1986,8 +1986,7 @@
 ;
 ; AVX2-LABEL: insert_dup_mem_v32i8_i32:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
+; AVX2-NEXT:    vpbroadcastb (%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %tmp = load i32, i32* %ptr, align 4
   %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0
@@ -2008,9 +2007,7 @@
 ;
 ; AVX2-LABEL: insert_dup_mem_v32i8_sext_i8:
 ; AVX2:       # BB#0:
-; AVX2-NEXT:    movsbl (%rdi), %eax
-; AVX2-NEXT:    vmovd %eax, %xmm0
-; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
+; AVX2-NEXT:    vpbroadcastb (%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %tmp = load i8, i8* %ptr, align 1
   %tmp1 = sext i8 %tmp to i32