[x86] use more phadd for reductions

[spatel]

Summary

This is part of what is requested by PR42023:
https://bugs.llvm.org/show_bug.cgi?id=42023

There's an extension needed for FP add, but exactly how we would specify that using flags is not clear to me, so I left that as a TODO.
We're still missing patterns for partial reductions when the input vector is 256-bit or 512-bit, but I think that's a failure of vector narrowing. If we can reduce the widths, then this matching should work on those tests.

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

35618 ↗

(On Diff #209897)

(style) early-out if this is not the case

llvm/test/CodeGen/X86/vector-reduce-add.ll

680 ↗

(On Diff #209897)

ideally we'd have these cases combine as ll

[spatel]

llvm/test/CodeGen/X86/vector-reduce-add.ll

680 ↗

(On Diff #209897)

'll' - do you mean alter/override the ExpandReductions pass in IR?

[spatel]

Patch updated:
Early exit if wrong types or subtarget.

[RKSimon]

llvm/test/CodeGen/X86/vector-reduce-add.ll

680 ↗

(On Diff #209897)

sorry - "ideally we'd have these cases combine as well"

[spatel]

Patch updated:
Allow 256-bit reductions by extracting and using 1 more 128-bit hop.

[spatel]

Patch updated - no functional changes from the previous draft:

1. Move local variable for NumElts closer to uses.
2. Add TODO comment about handling bigger-than-256-bit types.

[RKSimon]

LGTM - cheers

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

case X86ISD::FRSQRT: {
     return DAG.getNode(Vec.getOpcode(), DL, VT, ExtOps);
   }
   default:
     return SDValue();
   }
   llvm_unreachable("All opcodes should return within switch");
 }
 
+/// Try to convert a vector reduction sequence composed of binops and shuffles
+/// into horizontal ops.
+static SDValue combineReductionToHorizontal(SDNode *ExtElt, SelectionDAG &DAG,
+                                            const X86Subtarget &Subtarget) {
+  assert(ExtElt->getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unexpected caller");
+  bool OptForSize = DAG.getMachineFunction().getFunction().hasOptSize();
+  if (!Subtarget.hasFastHorizontalOps() && !OptForSize)
+    return SDValue();
+  SDValue Index = ExtElt->getOperand(1);
+  if (!isNullConstant(Index))
+    return SDValue();
+
+  // TODO: Allow FADD with reduction and/or reassociation and no-signed-zeros.
+  ISD::NodeType Opc;
+  SDValue Rdx = DAG.matchBinOpReduction(ExtElt, Opc, {ISD::ADD});
+  if (!Rdx)
+    return SDValue();
+
+  EVT VT = ExtElt->getValueType(0);
+  EVT VecVT = ExtElt->getOperand(0).getValueType();
+  if (VecVT.getScalarType() != VT)
+    return SDValue();
+
+  unsigned HorizOpcode = Opc == ISD::ADD ? X86ISD::HADD : X86ISD::FHADD;
+  SDLoc DL(ExtElt);
+
+  // 256-bit horizontal instructions operate on 128-bit chunks rather than
+  // across the whole vector, so we need an extract + hop preliminary stage.
+  // This is the only step where the operands of the hop are not the same value.
+  // TODO: We could extend this to handle 512-bit or even longer vectors.
+  if (((VecVT == MVT::v16i16 || VecVT == MVT::v8i32) && Subtarget.hasSSSE3()) ||
+      ((VecVT == MVT::v8f32 || VecVT == MVT::v4f64) && Subtarget.hasSSE3())) {
+    unsigned NumElts = VecVT.getVectorNumElements();
+    SDValue Hi = extract128BitVector(Rdx, NumElts / 2, DAG, DL);
+    SDValue Lo = extract128BitVector(Rdx, 0, DAG, DL);
+    VecVT = EVT::getVectorVT(*DAG.getContext(), VT, NumElts / 2);
+    Rdx = DAG.getNode(HorizOpcode, DL, VecVT, Hi, Lo);
+  }
+  if (!((VecVT == MVT::v8i16 || VecVT == MVT::v4i32) && Subtarget.hasSSSE3()) &&
+      !((VecVT == MVT::v4f32 || VecVT == MVT::v2f64) && Subtarget.hasSSE3()))
+    return SDValue();
+
+  // extract (add (shuf X), X), 0 --> extract (hadd X, X), 0
+  assert(Rdx.getValueType() == VecVT && "Unexpected reduction match");
+  unsigned ReductionSteps = Log2_32(VecVT.getVectorNumElements());
+  for (unsigned i = 0; i != ReductionSteps; ++i)
+    Rdx = DAG.getNode(HorizOpcode, DL, VecVT, Rdx, Rdx);
+
+  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Rdx, Index);
+}
+
 /// Detect vector gather/scatter index generation and convert it from being a
 /// bunch of shuffles and extracts into a somewhat faster sequence.
 /// For i686, the best sequence is apparently storing the value and loading
 /// scalars back, while for x64 we should use 64-bit extracts and shifts.
 static SDValue combineExtractVectorElt(SDNode *N, SelectionDAG &DAG,
                                        TargetLowering::DAGCombinerInfo &DCI,
                                        const X86Subtarget &Subtarget) {
   if (SDValue NewOp = combineExtractWithShuffle(N, DAG, DCI, Subtarget))
static SDValue combineExtractVectorElt(SDNode *N, SelectionDAG &DAG,
   // Attempt to replace an all_of/any_of horizontal reduction with a MOVMSK.
   if (SDValue Cmp = combineHorizontalPredicateResult(N, DAG, Subtarget))
     return Cmp;
 
   // Attempt to replace min/max v8i16/v16i8 reductions with PHMINPOSUW.
   if (SDValue MinMax = combineHorizontalMinMaxResult(N, DAG, Subtarget))
     return MinMax;
 
+  if (SDValue V = combineReductionToHorizontal(N, DAG, Subtarget))
+    return V;
+
   if (SDValue V = scalarizeExtEltFP(N, DAG))
     return V;
 
   // Attempt to extract a i1 element by using MOVMSK to extract the signbits
   // and then testing the relevant element.
   if (CIdx && SrcVT.getScalarType() == MVT::i1) {
     SmallVector<SDNode *, 16> BoolExtracts;
     auto IsBoolExtract = [&BoolExtracts](SDNode *Use) {

llvm/trunk/test/CodeGen/X86/phaddsub-extract.ll

 ; SSE3-SLOW-NEXT:    psrld $16, %xmm1
 ; SSE3-SLOW-NEXT:    paddw %xmm0, %xmm1
 ; SSE3-SLOW-NEXT:    movd %xmm1, %eax
 ; SSE3-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE3-SLOW-NEXT:    retq
 ;
 ; SSE3-FAST-LABEL: hadd16_8:
 ; SSE3-FAST:       # %bb.0:
-; SSE3-FAST-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE3-FAST-NEXT:    paddw %xmm0, %xmm1
-; SSE3-FAST-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; SSE3-FAST-NEXT:    paddw %xmm1, %xmm0
+; SSE3-FAST-NEXT:    phaddw %xmm0, %xmm0
+; SSE3-FAST-NEXT:    phaddw %xmm0, %xmm0
 ; SSE3-FAST-NEXT:    phaddw %xmm0, %xmm0
 ; SSE3-FAST-NEXT:    movd %xmm0, %eax
 ; SSE3-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE3-FAST-NEXT:    retq
 ;
 ; AVX-SLOW-LABEL: hadd16_8:
 ; AVX-SLOW:       # %bb.0:
 ; AVX-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX-SLOW-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
 ; AVX-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
 ; AVX-SLOW-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
 ; AVX-SLOW-NEXT:    vpsrld $16, %xmm0, %xmm1
 ; AVX-SLOW-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
 ; AVX-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX-SLOW-NEXT:    retq
 ;
 ; AVX-FAST-LABEL: hadd16_8:
 ; AVX-FAST:       # %bb.0:
-; AVX-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX-FAST-NEXT:    retq
   %x224 = shufflevector <8 x i16> %x223, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %x225 = add <8 x i16> %x223, %x224
   %x226 = shufflevector <8 x i16> %x225, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %x227 = add <8 x i16> %x225, %x226

 ; SSE3-SLOW-NEXT:    paddd %xmm0, %xmm1
 ; SSE3-SLOW-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
 ; SSE3-SLOW-NEXT:    paddd %xmm1, %xmm0
 ; SSE3-SLOW-NEXT:    movd %xmm0, %eax
 ; SSE3-SLOW-NEXT:    retq
 ;
 ; SSE3-FAST-LABEL: hadd32_4:
 ; SSE3-FAST:       # %bb.0:
-; SSE3-FAST-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE3-FAST-NEXT:    paddd %xmm0, %xmm1
-; SSE3-FAST-NEXT:    phaddd %xmm1, %xmm1
-; SSE3-FAST-NEXT:    movd %xmm1, %eax
+; SSE3-FAST-NEXT:    phaddd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    phaddd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    movd %xmm0, %eax
 ; SSE3-FAST-NEXT:    retq
 ;
 ; AVX-SLOW-LABEL: hadd32_4:
 ; AVX-SLOW:       # %bb.0:
 ; AVX-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
 ; AVX-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX-SLOW-NEXT:    retq
 ;
 ; AVX-FAST-LABEL: hadd32_4:
 ; AVX-FAST:       # %bb.0:
-; AVX-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX-FAST-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX-FAST-NEXT:    retq
   %x226 = shufflevector <4 x i32> %x225, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
   %x227 = add <4 x i32> %x225, %x226
   %x228 = shufflevector <4 x i32> %x227, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
   %x229 = add <4 x i32> %x227, %x228
   %x230 = extractelement <4 x i32> %x229, i32 0
; AVX512-FAST-NEXT:    retq
   %x229 = add <16 x i32> %x227, %x228
   %x230 = extractelement <16 x i32> %x229, i32 0
   ret i32 %x230
 }
 
 define i16 @hadd16_8_optsize(<8 x i16> %x223) optsize {
 ; SSE3-LABEL: hadd16_8_optsize:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE3-NEXT:    paddw %xmm0, %xmm1
-; SSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,2,3]
-; SSE3-NEXT:    paddw %xmm1, %xmm0
+; SSE3-NEXT:    phaddw %xmm0, %xmm0
+; SSE3-NEXT:    phaddw %xmm0, %xmm0
 ; SSE3-NEXT:    phaddw %xmm0, %xmm0
 ; SSE3-NEXT:    movd %xmm0, %eax
 ; SSE3-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hadd16_8_optsize:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    vmovd %xmm0, %eax
 ; AVX-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX-NEXT:    retq
   %x224 = shufflevector <8 x i16> %x223, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %x225 = add <8 x i16> %x223, %x224
   %x226 = shufflevector <8 x i16> %x225, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %x227 = add <8 x i16> %x225, %x226
   %x228 = shufflevector <8 x i16> %x227, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %x229 = add <8 x i16> %x227, %x228
   %x230 = extractelement <8 x i16> %x229, i32 0
   ret i16 %x230
 }
 
 define i32 @hadd32_4_optsize(<4 x i32> %x225) optsize {
 ; SSE3-LABEL: hadd32_4_optsize:
 ; SSE3:       # %bb.0:
-; SSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; SSE3-NEXT:    paddd %xmm0, %xmm1
-; SSE3-NEXT:    phaddd %xmm1, %xmm1
-; SSE3-NEXT:    movd %xmm1, %eax
+; SSE3-NEXT:    phaddd %xmm0, %xmm0
+; SSE3-NEXT:    phaddd %xmm0, %xmm0
+; SSE3-NEXT:    movd %xmm0, %eax
 ; SSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hadd32_4_optsize:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    vmovd %xmm0, %eax
 ; AVX-NEXT:    retq
   %x226 = shufflevector <4 x i32> %x225, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
   %x227 = add <4 x i32> %x225, %x226
   %x228 = shufflevector <4 x i32> %x227, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
   %x229 = add <4 x i32> %x227, %x228
   %x230 = extractelement <4 x i32> %x229, i32 0

llvm/trunk/test/CodeGen/X86/vector-reduce-add-widen.ll

 ; AVX1-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
 ; AVX1-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v4i32:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v4i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm0

 ; AVX1-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    vzeroupper
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v8i32:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm1, %xmm0
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v8i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1

 ; AVX1-SLOW-NEXT:    vpsrld $16, %xmm0, %xmm1
 ; AVX1-SLOW-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v8i16:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]

 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-SLOW-NEXT:    vzeroupper
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v16i16:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm1, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v16i16:
 ; AVX2:       # %bb.0:

llvm/trunk/test/CodeGen/X86/vector-reduce-add.ll

 ; AVX1-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
 ; AVX1-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v4i32:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v4i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX2-NEXT:    vpaddd %xmm1, %xmm0, %xmm0

 ; AVX1-SLOW-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    vzeroupper
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v8i32:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm1, %xmm0
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v8i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1

 ; AVX1-SLOW-NEXT:    vpsrld $16, %xmm0, %xmm1
 ; AVX1-SLOW-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v8i16:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]

 ; AVX1-SLOW-NEXT:    vmovd %xmm0, %eax
 ; AVX1-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-SLOW-NEXT:    vzeroupper
 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test_v16i16:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
-; AVX1-FAST-NEXT:    vpaddw %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm1, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vmovd %xmm0, %eax
 ; AVX1-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq
 ;
 ; AVX2-LABEL: test_v16i16:
 ; AVX2:       # %bb.0: