[x86] Unify the horizontal adding used for popcount lowering taking the

chandlerc · chandlerc · commit cb58910ce8a9 · 2015-05-30T10:35:03.000Z
best approach of each.

For vNi16, we use SHL + ADD + SRL pattern that seem easily the best.

For vNi32, we use the PUNPCK + PSADBW + PACKUSWB pattern. In some cases
there is a huge improvement with this in IACA's estimated throughput --
over 2x higher throughput!!!! -- but the measurements are too good to be
true. In one narrow case, the SHL + ADD + SHL + ADD + SRL pattern looks
slightly faster, but I'm not sure I believe any of the measurements at
this point. Both are the exact same uops though. Hard to be confident of
anything past that.

If anyone wants to collect very detailed (Agner-level) timings with the
result of this patch, or with the i32 case replaced with SHL + ADD + SHl
+ ADD + SRL, I'd be very interested. Note that you'll need to test it on
both Ivybridge and Haswell, with both SSE3, SSSE3, and AVX selected as
I saw unique behavior in each of these buckets with IACA all of which
should be checked against measured performance.

But this patch is still a useful improvement by dropping duplicate work
and getting the much nicer PSADBW lowering for v2i64.

I'd still like to rephrase this in terms of generic horizontal sum. It's
a bit lame to have a special case of that just for popcount.

llvm-svn: 238652
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -17344,63 +17344,19 @@ static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
     return DAG.getBitcast(VT, V);
   }
 
-  // To obtain pop count for each i16 element, shuffle the byte pop count to get
-  // even and odd elements into distinct vectors, add them and zero-extend each
-  // i8 elemento into i16, i.e.:
-  //
-  //  B -> pop count per i8
-  //  W -> pop count per i16
-  //
-  //  Y = shuffle B, undef <0, 2, ...>
-  //  Z = shuffle B, undef <1, 3, ...>
-  //  W = zext <... x i8> to <... x i16> (Y + Z)
-  //
-  // Use a byte shuffle mask that matches PSHUFB.
-  //
+  // The only element type left is i16.
   assert(EltVT == MVT::i16 && "Unknown how to handle type");
-  SDValue Undef = DAG.getUNDEF(ByteVecVT);
-  SmallVector<int, 32> MaskA, MaskB;
-
-  // We can't use PSHUFB across lanes, so do the shuffle and sum inside each
-  // 128-bit lane, and then collapse the result.
-  int NumLanes = VecSize / 128;
-  assert(VecSize % 128 == 0 && "Must have 16-byte multiple vectors!");
-  for (int i = 0; i < NumLanes; ++i) {
-    for (int j = 0; j < 8; ++j) {
-      MaskA.push_back(i * 16 + j * 2);
-      MaskB.push_back(i * 16 + (j * 2) + 1);
-    }
-    MaskA.append((size_t)8, -1);
-    MaskB.append((size_t)8, -1);
-  }
-
-  SDValue ShuffA = DAG.getVectorShuffle(ByteVecVT, DL, V, Undef, MaskA);
-  SDValue ShuffB = DAG.getVectorShuffle(ByteVecVT, DL, V, Undef, MaskB);
-  V = DAG.getNode(ISD::ADD, DL, ByteVecVT, ShuffA, ShuffB);
-
-  SmallVector<int, 4> Mask;
-  for (int i = 0; i < NumLanes; ++i)
-    Mask.push_back(2 * i);
-  Mask.append((size_t)NumLanes, -1);
-
-  int NumI64Elts = VecSize / 64;
-  MVT VecI64VT = MVT::getVectorVT(MVT::i64, NumI64Elts);
-
-  V = DAG.getBitcast(VecI64VT, V);
-  V = DAG.getVectorShuffle(VecI64VT, DL, V, DAG.getUNDEF(VecI64VT), Mask);
-  V = DAG.getBitcast(ByteVecVT, V);
-
-  // Zero extend i8s into i16 elts
-  SmallVector<int, 16> ZExtInRegMask;
-  for (int i = 0; i < NumElts; ++i) {
-    ZExtInRegMask.push_back(i);
-    ZExtInRegMask.push_back(2 * NumElts);
-  }
 
-  return DAG.getBitcast(
-      VT, DAG.getVectorShuffle(ByteVecVT, DL, V,
-                               getZeroVector(ByteVecVT, Subtarget, DAG, DL),
-                               ZExtInRegMask));
+  // To obtain pop count for each i16 element starting from the pop count for
+  // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
+  // right by 8. It is important to shift as i16s as i8 vector shift isn't
+  // directly supported.
+  SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
+  SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
+  SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
+  V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
+                  DAG.getBitcast(ByteVecVT, V));
+  return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
 }
 
 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
@@ -17526,28 +17482,12 @@ static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
   // At this point, V contains the byte-wise population count, and we are
   // merely doing a horizontal sum if necessary to get the wider element
   // counts.
-  //
-  // FIXME: There is a different lowering strategy above for the horizontal sum
-  // of byte-wise population counts. This one and that one should be merged,
-  // using the fastest of the two for each size.
-  MVT ByteVT = MVT::getVectorVT(MVT::i8, VecSize / 8);
-  MVT ShiftVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
-  V = DAG.getBitcast(ByteVT, V);
-  assert(Len <= 64 && "We don't support element sizes of more than 64 bits!");
-  assert(isPowerOf2_32(Len) && "Only power of two element sizes supported!");
-  for (int i = Len; i > 8; i /= 2) {
-    SDValue Shl = DAG.getBitcast(
-        ByteVT, GetShift(ISD::SHL, DAG.getBitcast(ShiftVT, V), i / 2));
-    V = DAG.getNode(ISD::ADD, DL, ByteVT, V, Shl);
-  }
-
-  // The high byte now contains the sum of the element bytes. Shift it right
-  // (if needed) to make it the low byte.
-  V = DAG.getBitcast(VT, V);
-  if (Len > 8)
-    V = GetShift(ISD::SRL, V, Len - 8);
+  if (EltVT == MVT::i8)
+    return V;
 
-  return V;
+  return LowerHorizontalByteSum(
+      DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
+      DAG);
 }
 
 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
diff --git a/llvm/test/CodeGen/X86/vector-popcnt-128.ll b/llvm/test/CodeGen/X86/vector-popcnt-128.ll
@@ -24,16 +24,9 @@ define <2 x i64> @testv2i64(<2 x i64> %in) {
 ; SSE2-NEXT:    psrlq $4, %xmm1
 ; SSE2-NEXT:    paddq %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
+; SSE2-NEXT:    pxor %xmm0, %xmm0
+; SSE2-NEXT:    psadbw %xmm0, %xmm1
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllq $32, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psllq $16, %xmm1
-; SSE2-NEXT:    paddb %xmm0, %xmm1
-; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllq $8, %xmm0
-; SSE2-NEXT:    paddb %xmm1, %xmm0
-; SSE2-NEXT:    psrlq $56, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: testv2i64:
@@ -52,16 +45,9 @@ define <2 x i64> @testv2i64(<2 x i64> %in) {
 ; SSE3-NEXT:    psrlq $4, %xmm1
 ; SSE3-NEXT:    paddq %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
+; SSE3-NEXT:    pxor %xmm0, %xmm0
+; SSE3-NEXT:    psadbw %xmm0, %xmm1
 ; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllq $32, %xmm0
-; SSE3-NEXT:    paddb %xmm1, %xmm0
-; SSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSE3-NEXT:    psllq $16, %xmm1
-; SSE3-NEXT:    paddb %xmm0, %xmm1
-; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllq $8, %xmm0
-; SSE3-NEXT:    paddb %xmm1, %xmm0
-; SSE3-NEXT:    psrlq $56, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: testv2i64:
@@ -130,13 +116,14 @@ define <4 x i32> @testv4i32(<4 x i32> %in) {
 ; SSE2-NEXT:    psrld $4, %xmm1
 ; SSE2-NEXT:    paddd %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
+; SSE2-NEXT:    pxor %xmm0, %xmm0
 ; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    psllq $16, %xmm2
-; SSE2-NEXT:    paddb %xmm1, %xmm2
-; SSE2-NEXT:    movdqa %xmm2, %xmm0
-; SSE2-NEXT:    psllq $8, %xmm0
-; SSE2-NEXT:    paddb %xmm2, %xmm0
-; SSE2-NEXT:    psrld $24, %xmm0
+; SSE2-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; SSE2-NEXT:    psadbw %xmm0, %xmm2
+; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE2-NEXT:    psadbw %xmm0, %xmm1
+; SSE2-NEXT:    packuswb %xmm2, %xmm1
+; SSE2-NEXT:    movdqa %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE3-LABEL: testv4i32:
@@ -155,13 +142,14 @@ define <4 x i32> @testv4i32(<4 x i32> %in) {
 ; SSE3-NEXT:    psrld $4, %xmm1
 ; SSE3-NEXT:    paddd %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
+; SSE3-NEXT:    pxor %xmm0, %xmm0
 ; SSE3-NEXT:    movdqa %xmm1, %xmm2
-; SSE3-NEXT:    psllq $16, %xmm2
-; SSE3-NEXT:    paddb %xmm1, %xmm2
-; SSE3-NEXT:    movdqa %xmm2, %xmm0
-; SSE3-NEXT:    psllq $8, %xmm0
-; SSE3-NEXT:    paddb %xmm2, %xmm0
-; SSE3-NEXT:    psrld $24, %xmm0
+; SSE3-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm0[2],xmm2[3],xmm0[3]
+; SSE3-NEXT:    psadbw %xmm0, %xmm2
+; SSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; SSE3-NEXT:    psadbw %xmm0, %xmm1
+; SSE3-NEXT:    packuswb %xmm2, %xmm1
+; SSE3-NEXT:    movdqa %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: testv4i32:
@@ -247,7 +235,7 @@ define <8 x i16> @testv8i16(<8 x i16> %in) {
 ; SSE2-NEXT:    paddw %xmm0, %xmm1
 ; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:    psllq $8, %xmm0
+; SSE2-NEXT:    psllw $8, %xmm0
 ; SSE2-NEXT:    paddb %xmm1, %xmm0
 ; SSE2-NEXT:    psrlw $8, %xmm0
 ; SSE2-NEXT:    retq
@@ -269,30 +257,27 @@ define <8 x i16> @testv8i16(<8 x i16> %in) {
 ; SSE3-NEXT:    paddw %xmm0, %xmm1
 ; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
 ; SSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSE3-NEXT:    psllq $8, %xmm0
+; SSE3-NEXT:    psllw $8, %xmm0
 ; SSE3-NEXT:    paddb %xmm1, %xmm0
 ; SSE3-NEXT:    psrlw $8, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; SSSE3-LABEL: testv8i16:
 ; SSSE3:       # BB#0:
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
-; SSSE3-NEXT:    movdqa %xmm0, %xmm3
-; SSSE3-NEXT:    pand %xmm2, %xmm3
-; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm4
-; SSSE3-NEXT:    pshufb %xmm3, %xmm4
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; SSSE3-NEXT:    movdqa %xmm0, %xmm2
+; SSSE3-NEXT:    pand %xmm1, %xmm2
+; SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
+; SSSE3-NEXT:    movdqa %xmm3, %xmm4
+; SSSE3-NEXT:    pshufb %xmm2, %xmm4
 ; SSSE3-NEXT:    psrlw $4, %xmm0
-; SSSE3-NEXT:    pand %xmm2, %xmm0
-; SSSE3-NEXT:    pshufb %xmm0, %xmm1
-; SSSE3-NEXT:    paddb %xmm4, %xmm1
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
-; SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u]
-; SSSE3-NEXT:    pshufb {{.*#+}} xmm1 = xmm1[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]
-; SSSE3-NEXT:    paddb %xmm0, %xmm1
-; SSSE3-NEXT:    pxor %xmm0, %xmm0
-; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; SSSE3-NEXT:    movdqa %xmm1, %xmm0
+; SSSE3-NEXT:    pand %xmm1, %xmm0
+; SSSE3-NEXT:    pshufb %xmm0, %xmm3
+; SSSE3-NEXT:    paddb %xmm4, %xmm3
+; SSSE3-NEXT:    movdqa %xmm3, %xmm0
+; SSSE3-NEXT:    psllw $8, %xmm0
+; SSSE3-NEXT:    paddb %xmm3, %xmm0
+; SSSE3-NEXT:    psrlw $8, %xmm0
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: testv8i16:
@@ -308,10 +293,9 @@ define <8 x i16> @testv8i16(<8 x i16> %in) {
 ; SSE41-NEXT:    pshufb %xmm0, %xmm3
 ; SSE41-NEXT:    paddb %xmm4, %xmm3
 ; SSE41-NEXT:    movdqa %xmm3, %xmm0
-; SSE41-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u]
-; SSE41-NEXT:    pshufb {{.*#+}} xmm3 = xmm3[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]
-; SSE41-NEXT:    paddb %xmm0, %xmm3
-; SSE41-NEXT:    pmovzxbw {{.*#+}} xmm0 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; SSE41-NEXT:    psllw $8, %xmm0
+; SSE41-NEXT:    paddb %xmm3, %xmm0
+; SSE41-NEXT:    psrlw $8, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: testv8i16:
@@ -324,10 +308,9 @@ define <8 x i16> @testv8i16(<8 x i16> %in) {
 ; AVX-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
 ; AVX-NEXT:    vpaddb %xmm2, %xmm0, %xmm0
-; AVX-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u]
-; AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]
-; AVX-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX-NEXT:    vpsllw $8, %xmm0, %xmm1
+; AVX-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vpsrlw $8, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %out = call <8 x i16> @llvm.ctpop.v8i16(<8 x i16> %in)
   ret <8 x i16> %out
diff --git a/llvm/test/CodeGen/X86/vector-popcnt-256.ll b/llvm/test/CodeGen/X86/vector-popcnt-256.ll
@@ -108,23 +108,19 @@ define <16 x i16> @testv16i16(<16 x i16> %in) {
 ; AVX1-NEXT:    vpand %xmm1, %xmm4, %xmm4
 ; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
 ; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm4 = <1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u>
-; AVX1-NEXT:    vpshufb %xmm4, %xmm2, %xmm5
-; AVX1-NEXT:    vmovdqa {{.*#+}} xmm6 = <0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u>
-; AVX1-NEXT:    vpshufb %xmm6, %xmm2, %xmm2
-; AVX1-NEXT:    vpaddb %xmm5, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxbw {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; AVX1-NEXT:    vpsllw $8, %xmm2, %xmm4
+; AVX1-NEXT:    vpaddb %xmm2, %xmm4, %xmm2
+; AVX1-NEXT:    vpsrlw $8, %xmm2, %xmm2
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm5
-; AVX1-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
+; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm4
+; AVX1-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
 ; AVX1-NEXT:    vpsrlw $4, %xmm0, %xmm0
 ; AVX1-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
-; AVX1-NEXT:    vpaddb %xmm5, %xmm0, %xmm0
-; AVX1-NEXT:    vpshufb %xmm4, %xmm0, %xmm1
-; AVX1-NEXT:    vpshufb %xmm6, %xmm0, %xmm0
-; AVX1-NEXT:    vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT:    vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX1-NEXT:    vpaddb %xmm4, %xmm0, %xmm0
+; AVX1-NEXT:    vpsllw $8, %xmm0, %xmm1
+; AVX1-NEXT:    vpaddb %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm0
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
 ; AVX1-NEXT:    retq
 ;
@@ -138,11 +134,9 @@ define <16 x i16> @testv16i16(<16 x i16> %in) {
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpshufb %ymm0, %ymm3, %ymm0
 ; AVX2-NEXT:    vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpshufb {{.*#+}} ymm1 = ymm0[1,3,5,7,9,11,13,15,u,u,u,u,u,u,u,u,17,19,21,23,25,27,29,31,u,u,u,u,u,u,u,u]
-; AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30,u,u,u,u,u,u,u,u]
-; AVX2-NEXT:    vpaddb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
-; AVX2-NEXT:    vpmovzxbw {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero,xmm0[8],zero,xmm0[9],zero,xmm0[10],zero,xmm0[11],zero,xmm0[12],zero,xmm0[13],zero,xmm0[14],zero,xmm0[15],zero
+; AVX2-NEXT:    vpsllw $8, %ymm0, %ymm1
+; AVX2-NEXT:    vpaddb %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vpsrlw $8, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
   %out = call <16 x i16> @llvm.ctpop.v16i16(<16 x i16> %in)
   ret <16 x i16> %out
