[x86] lower extracted fadd/fsub to horizontal vector math

[spatel]

Summary

This would show up if we fix horizontal reductions to narrow as they go along, but it's an improvement for size and/or Jaguar (fast-hops) independent of that.

We need to do this late to not interfere with other pattern matching of larger horizontal sequences.

I'm guessing we would extend this to integer ops too, but I figured I better stop here and get feedback on the initial FP part in case I've missed something.

[spatel]

Ping.

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

18340

We already have shouldCombineToHorizontalOp

[spatel]

lib/Target/X86/X86ISelLowering.cpp

18340

Right - I’m moving/renaming it here because it’s used in both combining and lowering after this patch. I can make that an NFC pre-commit.

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

18340

Yes, please do the NFC pre-commit first

[spatel]

Patch updated:
No functional diff from the previous rev, but rebased after pre-committing the move/rename of the helper with rL350193.

One thing I'm not sure about - do we want to enable this for AVX512-sized vectors too? Ie, can we assert that the extract is from 128/256/512-bit instead of checking that explicitly?

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

18351

Do both ops have to be one use for this to be useful?

18370

Is it useful to support any suitable neighbouring pairs of extracted values?

18376

Is this important? I think its what you were asking about AVX512 (which doesn't have HADDPS). I'd start by adding avx512f/avx512vl tests to haddsub-undef.ll

[spatel]

lib/Target/X86/X86ISelLowering.cpp

18351

It's not as clearly a win, but I'll ease that constraint, and we can decide if it looks better.

18370

Yes, but arguably the other patterns are less likely (the 0/1 case is what I'm seeing from adjusting the expansion of horizontal reductions). Let me make that a TODO.

18376

Right - we would want to treat 512-bit extract identically to 256-bit. Ie, it doesn't really matter that there's no 512 version of haddps, we want to narrow it to a 128-bit vector first either way.

[spatel]

Patch updated:

1. Ease extra uses constraint - we might still be able to avoid a shuffle/extract.
2. Allow 512-bit vectors.
3. Add TODO to for other extract index pairs.

[spatel]

test/CodeGen/X86/haddsub.ll

1034

This wasn't clear from the patch update comment: we have 512-bit tests and AVX512 RUN lines. However, the output is identical to regular AVX1 because we're just extracting the low 128-bits.

[RKSimon]

LGTM - please can you add AVX512 tests to haddsub-undef.ll - we also need AVX-SLOW/AVX-FAST common prefixes as well

[spatel]

In D56011#1345388, @RKSimon wrote:

LGTM - please can you add AVX512 tests to haddsub-undef.ll - we also need AVX-SLOW/AVX-FAST common prefixes as well

Rearranged/added tests as preliminary commits:
rL350356
rL350357
rL350358
rL350362
rL350364

[spatel]

Reopening - I reverted with rL350373.
I didn't realize the custom legalization would affect the cost models used by the vectorizers; there are potentially IR test changes that go with this.

[spatel]

Patch updated:
Include cost model and SLP test changes. Should we be overriding the cost of scalar FADD/FSUB, so we don't have these diffs?

Without that, it looks like we fall back to this questionable logic:

// If the operation is custom lowered, then assume that the code is twice
// as expensive.

[RKSimon]

In D56011#1345978, @spatel wrote:

Patch updated:
Include cost model and SLP test changes. Should we be overriding the cost of scalar FADD/FSUB, so we don't have these diffs?

Yes - please add the relevant cost changes so the slp/cost test changes go away.

Without that, it looks like we fall back to this questionable logic:

// If the operation is custom lowered, then assume that the code is twice
// as expensive.

That is some VERY questionable logic.....

[spatel]

Patch updated:
Add cost model overrides for FADD/FSUB to preserve existing behavior.
I suspect that these should be adjusted, and we should include more opcodes to avoid the default trap, but that's another patch. This maintains the current state, so no test diffs outside of codegen.

[spatel]

lib/Target/X86/X86TargetTransformInfo.cpp

839–842

Looking harder at Agner's numbers - I didn't remember that P4 actually had 2-cycle throughput for addsd/subsd. So this is correct assuming the baseline for SSE2 is a P4 (willamette, prescott).

But the baseline chip for SSE1 below here is a P3, and that had 1-cycle throughput for addss/subss according to Agner.

I'll fix the comments.

[spatel]

Patch updated:
Updated cost model changes - for SSE2 (P4), the default 2-cycle throughput cost matches Agner.
For SSE1 (P3), the default 2-cycle throughput cost does not match a P3 implementation (but hopefully nobody cares at this point).

[spatel]

Patch updated:
Rebased after cost model enhancements in rL350403. So now this patch is once again independent of any IR diffs.

[RKSimon]

LGTM thanks

lib/Target/X86/X86ISelLowering.cpp

for (auto VT : { MVT::f32, MVT::f64 }) {
       setOperationAction(ISD::FABS, VT, Custom);
 
       // Use XORP to simulate FNEG.
       setOperationAction(ISD::FNEG, VT, Custom);
 
       // Use ANDPD and ORPD to simulate FCOPYSIGN.
       setOperationAction(ISD::FCOPYSIGN, VT, Custom);
 
+      // These might be better off as horizontal vector ops.
+      setOperationAction(ISD::FADD, VT, Custom);
+      setOperationAction(ISD::FSUB, VT, Custom);
+
       // We don't support sin/cos/fmod
       setOperationAction(ISD::FSIN   , VT, Expand);
       setOperationAction(ISD::FCOS   , VT, Expand);
       setOperationAction(ISD::FSINCOS, VT, Expand);
     }
 
     // Lower this to MOVMSK plus an AND.
     setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);

 /// Horizontal vector math instructions may be slower than normal math with
 /// shuffles. Limit horizontal op codegen based on size/speed trade-offs, uarch
 /// implementation, and likely shuffle complexity of the alternate sequence.
 static bool shouldUseHorizontalOp(bool IsSingleSource, SelectionDAG &DAG,
                                   const X86Subtarget &Subtarget) {
   bool IsOptimizingSize = DAG.getMachineFunction().getFunction().optForSize();
   bool HasFastHOps = Subtarget.hasFastHorizontalOps();
   return !IsSingleSource || IsOptimizingSize || HasFastHOps;
 }
 
+/// Depending on uarch and/or optimizing for size, we might prefer to use a
+/// vector operation in place of the typical scalar operation.
+static SDValue lowerFaddFsub(SDValue Op, SelectionDAG &DAG,
+                             const X86Subtarget &Subtarget) {
+  MVT VT = Op.getSimpleValueType();
+  assert(VT == MVT::f32 || VT == MVT::f64 && "Only expecting float/double");
+
+  // If both operands have other uses, this is probably not profitable.
+  // Horizontal FP add/sub were added with SSE3.
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  if ((!LHS.hasOneUse() && !RHS.hasOneUse()) || !Subtarget.hasSSE3())
+    return Op;
+
+  if (LHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+      RHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+      LHS.getOperand(0) != RHS.getOperand(0))
+    return Op;
+
+  if (!isa<ConstantSDNode>(LHS.getOperand(1)) ||
+      !isa<ConstantSDNode>(RHS.getOperand(1)) ||
+      !shouldUseHorizontalOp(true, DAG, Subtarget))
+    return Op;
+
+  // Allow commuted 'hadd' ops.
+  // TODO: Allow commuted fsub by negating the result of FHSUB?
+  // TODO: This can be extended to handle other adjacent extract pairs.
+  auto HOpcode = Op.getOpcode() == ISD::FADD ? X86ISD::FHADD : X86ISD::FHSUB;
+  unsigned LExtIndex = LHS.getConstantOperandVal(1);
+  unsigned RExtIndex = RHS.getConstantOperandVal(1);
+  if (LExtIndex == 1 && RExtIndex == 0 && HOpcode == X86ISD::FHADD)
+    std::swap(LExtIndex, RExtIndex);
+  if (LExtIndex != 0 || RExtIndex != 1)
+    return Op;
+
+  SDValue X = LHS.getOperand(0);
+  EVT VecVT = X.getValueType();
+  unsigned BitWidth = VecVT.getSizeInBits();
+  assert((BitWidth == 128 || BitWidth == 256 || BitWidth == 512) &&
+         "Not expecting illegal vector widths here");
+
+  // Creating a 256-bit horizontal op would be wasteful, and there is no 512-bit
+  // equivalent, so extract the 256/512-bit source op to 128-bit.
+  // This is free: ymm/zmm -> xmm.
+  SDLoc DL(Op);
+  if (BitWidth == 256 || BitWidth == 512)
+    X = extract128BitVector(X, 0, DAG, DL);
+
+  // fadd (extractelt (X, 0), extractelt (X, 1)) --> extractelt (hadd X, X), 0
+  // fadd (extractelt (X, 1), extractelt (X, 0)) --> extractelt (hadd X, X), 0
+  // fsub (extractelt (X, 0), extractelt (X, 1)) --> extractelt (hsub X, X), 0
+  // The extract of element 0 is free: the scalar result is element 0.
+  SDValue HOp = DAG.getNode(HOpcode, DL, X.getValueType(), X, X);
+  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, HOp,
+                     DAG.getIntPtrConstant(0, DL));
+}
+
 /// The only differences between FABS and FNEG are the mask and the logic op.
 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
   assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
          "Wrong opcode for lowering FABS or FNEG.");
 
   bool IsFABS = (Op.getOpcode() == ISD::FABS);
 
SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::ZERO_EXTEND_VECTOR_INREG:
   case ISD::SIGN_EXTEND_VECTOR_INREG:
     return LowerEXTEND_VECTOR_INREG(Op, Subtarget, DAG);
   case ISD::FP_TO_SINT:
   case ISD::FP_TO_UINT:         return LowerFP_TO_INT(Op, DAG);
   case ISD::FP_EXTEND:          return LowerFP_EXTEND(Op, DAG);
   case ISD::LOAD:               return LowerLoad(Op, Subtarget, DAG);
   case ISD::STORE:              return LowerStore(Op, Subtarget, DAG);
+  case ISD::FADD:
+  case ISD::FSUB:               return lowerFaddFsub(Op, DAG, Subtarget);
   case ISD::FABS:
   case ISD::FNEG:               return LowerFABSorFNEG(Op, DAG);
   case ISD::FCOPYSIGN:          return LowerFCOPYSIGN(Op, DAG);
   case ISD::FGETSIGN:           return LowerFGETSIGN(Op, DAG);
   case ISD::SETCC:              return LowerSETCC(Op, DAG);
   case ISD::SETCCCARRY:         return LowerSETCCCARRY(Op, DAG);
   case ISD::SELECT:             return LowerSELECT(Op, DAG);
   case ISD::BRCOND:             return LowerBRCOND(Op, DAG);

lib/Target/X86/X86TargetTransformInfo.cpp

static const CostTblEntry SSE2CostTable[] = {
     { ISD::MUL,  MVT::v8i16,       1 }, // pmullw
     { ISD::MUL,  MVT::v4i32,       6 }, // 3*pmuludq/4*shuffle
     { ISD::MUL,  MVT::v2i64,       8 }, // 3*pmuludq/3*shift/2*add
 
     { ISD::FDIV, MVT::f32,        23 }, // Pentium IV from http://www.agner.org/
     { ISD::FDIV, MVT::v4f32,      39 }, // Pentium IV from http://www.agner.org/
     { ISD::FDIV, MVT::f64,        38 }, // Pentium IV from http://www.agner.org/
     { ISD::FDIV, MVT::v2f64,      69 }, // Pentium IV from http://www.agner.org/
+
+    // The default cost model may consider these basic ops (addsd/subsd) more
+    // expensive just because we custom lower them. Explicitly set these so the
+    // cost is independent of our lowering implementation.
+    // TODO: The cost of "2" for FP ops is apparently due to the fact that P4
+    // era chips were running integer units twice as fast as FP units. But these
+    // costs should be relative to other FP costs above here, so they should be
+    // "1". Alternatively, other FP costs should be scaled up by a factor of 2.
+    { ISD::FADD, MVT::f64,         2 },
+    { ISD::FSUB, MVT::f64,         2 }
   };
 
   if (ST->hasSSE2())
     if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
       return LT.first * Entry->Cost;
 
   static const CostTblEntry SSE1CostTable[] = {
     { ISD::FDIV, MVT::f32,   17 }, // Pentium III from http://www.agner.org/
     { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
+
+    // The default cost model may consider these basic ops (addss/subss) more
+    // expensive just because we custom lower them. Explicitly set these so the
+    // cost is independent of our lowering implementation.
+    // TODO: The cost of "2" for FP ops is apparently due to the fact that P4
+    // era chips were running integer units twice as fast as FP units. But these
+    // costs should be relative to other FP costs above here, so they should be
+    // "1". Alternatively, other FP costs should be scaled up by a factor of 2.
+    { ISD::FADD, MVT::f32,    2 },
+    { ISD::FSUB, MVT::f32,    2 }
   };
 
   if (ST->hasSSE1())
     if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
       return LT.first * Entry->Cost;
 
   // It is not a good idea to vectorize division. We have to scalarize it and
   // in the process we will often end up having to spilling regular

test/CodeGen/X86/haddsub-undef.ll

; AVX-NEXT:    retq
   %vecext6 = extractelement <4 x float> %b, i32 0
   %vecext7 = extractelement <4 x float> %b, i32 1
   %add8 = fadd float %vecext6, %vecext7
   %vecinit9 = insertelement <4 x float> %vecinit5, float %add8, i32 2
   ret <4 x float> %vecinit9
 }
 
 define <4 x float> @test4_undef(<4 x float> %a, <4 x float> %b) {
-; SSE-LABEL: test4_undef:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE-NEXT:    addss %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE-SLOW-LABEL: test4_undef:
+; SSE-SLOW:       # %bb.0:
+; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: test4_undef:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE-FAST-LABEL: test4_undef:
+; SSE-FAST:       # %bb.0:
+; SSE-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: test4_undef:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: test4_undef:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %vecext = extractelement <4 x float> %a, i32 0
   %vecext1 = extractelement <4 x float> %a, i32 1
   %add = fadd float %vecext, %vecext1
   %vecinit = insertelement <4 x float> undef, float %add, i32 0
   ret <4 x float> %vecinit
 }
 
 define <2 x double> @test5_undef(<2 x double> %a, <2 x double> %b) {
-; SSE-LABEL: test5_undef:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movapd %xmm0, %xmm1
-; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE-NEXT:    addsd %xmm0, %xmm1
-; SSE-NEXT:    movapd %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE-SLOW-LABEL: test5_undef:
+; SSE-SLOW:       # %bb.0:
+; SSE-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE-SLOW-NEXT:    addsd %xmm0, %xmm1
+; SSE-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: test5_undef:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE-FAST-LABEL: test5_undef:
+; SSE-FAST:       # %bb.0:
+; SSE-FAST-NEXT:    haddpd %xmm0, %xmm0
+; SSE-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: test5_undef:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: test5_undef:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %vecext = extractelement <2 x double> %a, i32 0
   %vecext1 = extractelement <2 x double> %a, i32 1
   %add = fadd double %vecext, %vecext1
   %vecinit = insertelement <2 x double> undef, double %add, i32 0
   ret <2 x double> %vecinit
 }
 
 define <4 x float> @test6_undef(<4 x float> %a, <4 x float> %b) {
; AVX-NEXT:    retq
   %vecext2 = extractelement <4 x float> %b, i32 2
   %vecext3 = extractelement <4 x float> %b, i32 3
   %add4 = fadd float %vecext2, %vecext3
   %vecinit5 = insertelement <4 x float> %vecinit, float %add4, i32 3
   ret <4 x float> %vecinit5
 }
 
 define <4 x float> @test8_undef(<4 x float> %a, <4 x float> %b) {
-; SSE-LABEL: test8_undef:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE-NEXT:    addss %xmm0, %xmm1
-; SSE-NEXT:    movaps %xmm0, %xmm2
-; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE-NEXT:    addss %xmm2, %xmm0
-; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
-; SSE-NEXT:    movaps %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE-SLOW-LABEL: test8_undef:
+; SSE-SLOW:       # %bb.0:
+; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE-SLOW-NEXT:    addss %xmm0, %xmm1
+; SSE-SLOW-NEXT:    movaps %xmm0, %xmm2
+; SSE-SLOW-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
+; SSE-SLOW-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE-SLOW-NEXT:    addss %xmm2, %xmm0
+; SSE-SLOW-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE-SLOW-NEXT:    movaps %xmm1, %xmm0
+; SSE-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: test8_undef:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm1
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; AVX-NEXT:    vaddss %xmm0, %xmm2, %xmm0
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
-; AVX-NEXT:    retq
+; SSE-FAST-LABEL: test8_undef:
+; SSE-FAST:       # %bb.0:
+; SSE-FAST-NEXT:    movaps %xmm0, %xmm1
+; SSE-FAST-NEXT:    haddps %xmm0, %xmm1
+; SSE-FAST-NEXT:    movaps %xmm0, %xmm2
+; SSE-FAST-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
+; SSE-FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE-FAST-NEXT:    addss %xmm2, %xmm0
+; SSE-FAST-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE-FAST-NEXT:    movaps %xmm1, %xmm0
+; SSE-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: test8_undef:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm1
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX-SLOW-NEXT:    vaddss %xmm0, %xmm2, %xmm0
+; AVX-SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: test8_undef:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm1
+; AVX-FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX-FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX-FAST-NEXT:    vaddss %xmm0, %xmm2, %xmm0
+; AVX-FAST-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-FAST-NEXT:    retq
   %vecext = extractelement <4 x float> %a, i32 0
   %vecext1 = extractelement <4 x float> %a, i32 1
   %add = fadd float %vecext, %vecext1
   %vecinit = insertelement <4 x float> undef, float %add, i32 0
   %vecext2 = extractelement <4 x float> %a, i32 2
   %vecext3 = extractelement <4 x float> %a, i32 3
   %add4 = fadd float %vecext2, %vecext3
   %vecinit5 = insertelement <4 x float> %vecinit, float %add4, i32 2
; AVX-NEXT:    retq
   %vecext2 = extractelement <8 x float> %b, i32 2
   %vecext3 = extractelement <8 x float> %b, i32 3
   %add4 = fadd float %vecext2, %vecext3
   %vecinit5 = insertelement <8 x float> %vecinit, float %add4, i32 3
   ret <8 x float> %vecinit5
 }
 
 define <8 x float> @test11_undef(<8 x float> %a, <8 x float> %b) {
-; SSE-LABEL: test11_undef:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE-NEXT:    addss %xmm1, %xmm0
-; SSE-NEXT:    movshdup {{.*#+}} xmm1 = xmm3[1,1,3,3]
-; SSE-NEXT:    addss %xmm3, %xmm1
-; SSE-NEXT:    movddup {{.*#+}} xmm1 = xmm1[0,0]
-; SSE-NEXT:    retq
+; SSE-SLOW-LABEL: test11_undef:
+; SSE-SLOW:       # %bb.0:
+; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm3[1,1,3,3]
+; SSE-SLOW-NEXT:    addss %xmm3, %xmm1
+; SSE-SLOW-NEXT:    movddup {{.*#+}} xmm1 = xmm1[0,0]
+; SSE-SLOW-NEXT:    retq
+;
+; SSE-FAST-LABEL: test11_undef:
+; SSE-FAST:       # %bb.0:
+; SSE-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE-FAST-NEXT:    haddps %xmm3, %xmm3
+; SSE-FAST-NEXT:    movddup {{.*#+}} xmm1 = xmm3[0,0]
+; SSE-FAST-NEXT:    retq
 ;
 ; AVX-LABEL: test11_undef:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vhaddps %ymm0, %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %vecext = extractelement <8 x float> %a, i32 0
   %vecext1 = extractelement <8 x float> %a, i32 1
   %add = fadd float %vecext, %vecext1

 ; AVX1-SLOW-NEXT:    retq
 ;
 ; AVX1-FAST-LABEL: test13_v16f32_undef:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    retq
 ;
-; AVX512-LABEL: test13_v16f32_undef:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX512-NEXT:    vaddss %xmm1, %xmm0, %xmm1
-; AVX512-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX512-NEXT:    vpermilps {{.*#+}} xmm3 = xmm0[3,1,2,3]
-; AVX512-NEXT:    vaddss %xmm3, %xmm2, %xmm2
-; AVX512-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
-; AVX512-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX512-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX512-NEXT:    vaddss %xmm2, %xmm0, %xmm2
-; AVX512-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm2[0],xmm1[3]
-; AVX512-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX512-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; AVX512-NEXT:    vaddss %xmm0, %xmm2, %xmm0
-; AVX512-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
-; AVX512-NEXT:    retq
+; AVX512-SLOW-LABEL: test13_v16f32_undef:
+; AVX512-SLOW:       # %bb.0:
+; AVX512-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm1
+; AVX512-SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX512-SLOW-NEXT:    vpermilps {{.*#+}} xmm3 = xmm0[3,1,2,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm3, %xmm2, %xmm2
+; AVX512-SLOW-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
+; AVX512-SLOW-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX512-SLOW-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm2, %xmm0, %xmm2
+; AVX512-SLOW-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm2[0],xmm1[3]
+; AVX512-SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX512-SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm0, %xmm2, %xmm0
+; AVX512-SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX512-SLOW-NEXT:    retq
+;
+; AVX512-FAST-LABEL: test13_v16f32_undef:
+; AVX512-FAST:       # %bb.0:
+; AVX512-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm1
+; AVX512-FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX512-FAST-NEXT:    vpermilps {{.*#+}} xmm3 = xmm0[3,1,2,3]
+; AVX512-FAST-NEXT:    vaddss %xmm3, %xmm2, %xmm2
+; AVX512-FAST-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
+; AVX512-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX512-FAST-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
+; AVX512-FAST-NEXT:    vaddss %xmm2, %xmm0, %xmm2
+; AVX512-FAST-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm2[0],xmm1[3]
+; AVX512-FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX512-FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX512-FAST-NEXT:    vaddss %xmm0, %xmm2, %xmm0
+; AVX512-FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX512-FAST-NEXT:    retq
   %vecext = extractelement <16 x float> %a, i32 0
   %vecext1 = extractelement <16 x float> %a, i32 1
   %add1 = fadd float %vecext, %vecext1
   %vecinit1 = insertelement <16 x float> undef, float %add1, i32 0
   %vecext2 = extractelement <16 x float> %a, i32 2
   %vecext3 = extractelement <16 x float> %a, i32 3
   %add2 = fadd float %vecext2, %vecext3
   %vecinit2 = insertelement <16 x float> %vecinit1, float %add2, i32 1

test/CodeGen/X86/haddsub.ll

; AVX-NEXT:    retq
   %res0 = insertelement <2 x float> undef, float %op0, i32 0
   %res1 = insertelement <2 x float> %res0, float %op1, i32 1
   ret <2 x float> %res1
 }
 
 ; 128-bit vectors, float/double, fadd/fsub
 
 define float @extract_extract_v4f32_fadd_f32(<4 x float> %x) {
-; SSE3-LABEL: extract_extract_v4f32_fadd_f32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f32_fadd_f32:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f32_fadd_f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f32_fadd_f32:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f32_fadd_f32:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f32_fadd_f32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x float> %x, i32 0
   %x1 = extractelement <4 x float> %x, i32 1
   %x01 = fadd float %x0, %x1
   ret float %x01
 }
 
 define float @extract_extract_v4f32_fadd_f32_commute(<4 x float> %x) {
-; SSE3-LABEL: extract_extract_v4f32_fadd_f32_commute:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f32_fadd_f32_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f32_fadd_f32_commute:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f32_fadd_f32_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f32_fadd_f32_commute:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm0, %xmm1, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f32_fadd_f32_commute:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x float> %x, i32 0
   %x1 = extractelement <4 x float> %x, i32 1
   %x01 = fadd float %x1, %x0
   ret float %x01
 }
 
 define double @extract_extract_v2f64_fadd_f64(<2 x double> %x) {
-; SSE3-LABEL: extract_extract_v2f64_fadd_f64:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    addsd %xmm0, %xmm1
-; SSE3-NEXT:    movapd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v2f64_fadd_f64:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    addsd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v2f64_fadd_f64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v2f64_fadd_f64:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v2f64_fadd_f64:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v2f64_fadd_f64:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <2 x double> %x, i32 0
   %x1 = extractelement <2 x double> %x, i32 1
   %x01 = fadd double %x0, %x1
   ret double %x01
 }
 
 define double @extract_extract_v2f64_fadd_f64_commute(<2 x double> %x) {
-; SSE3-LABEL: extract_extract_v2f64_fadd_f64_commute:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    addsd %xmm0, %xmm1
-; SSE3-NEXT:    movapd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v2f64_fadd_f64_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    addsd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v2f64_fadd_f64_commute:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddsd %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v2f64_fadd_f64_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v2f64_fadd_f64_commute:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vaddsd %xmm0, %xmm1, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v2f64_fadd_f64_commute:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <2 x double> %x, i32 0
   %x1 = extractelement <2 x double> %x, i32 1
   %x01 = fadd double %x1, %x0
   ret double %x01
 }
 
 define float @extract_extract_v4f32_fsub_f32(<4 x float> %x) {
-; SSE3-LABEL: extract_extract_v4f32_fsub_f32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    subss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f32_fsub_f32:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    subss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f32_fsub_f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vsubss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f32_fsub_f32:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    hsubps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f32_fsub_f32:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vsubss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f32_fsub_f32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhsubps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x float> %x, i32 0
   %x1 = extractelement <4 x float> %x, i32 1
   %x01 = fsub float %x0, %x1
   ret float %x01
 }
 
 define float @extract_extract_v4f32_fsub_f32_commute(<4 x float> %x) {
 ; SSE3-LABEL: extract_extract_v4f32_fsub_f32_commute:

 ; AVX-NEXT:    retq
   %x0 = extractelement <4 x float> %x, i32 0
   %x1 = extractelement <4 x float> %x, i32 1
   %x01 = fsub float %x1, %x0
   ret float %x01
 }
 
 define double @extract_extract_v2f64_fsub_f64(<2 x double> %x) {
-; SSE3-LABEL: extract_extract_v2f64_fsub_f64:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    subsd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v2f64_fsub_f64:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    subsd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v2f64_fsub_f64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v2f64_fsub_f64:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    hsubpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v2f64_fsub_f64:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v2f64_fsub_f64:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhsubpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <2 x double> %x, i32 0
   %x1 = extractelement <2 x double> %x, i32 1
   %x01 = fsub double %x0, %x1
   ret double %x01
 }
 
 define double @extract_extract_v2f64_fsub_f64_commute(<2 x double> %x) {
 ; SSE3-LABEL: extract_extract_v2f64_fsub_f64_commute:
; AVX-NEXT:    retq
   %x1 = extractelement <2 x double> %x, i32 1
   %x01 = fsub double %x1, %x0
   ret double %x01
 }
 
 ; 256-bit vectors, float/double, fadd/fsub
 
 define float @extract_extract_v8f32_fadd_f32(<8 x float> %x) {
-; SSE3-LABEL: extract_extract_v8f32_fadd_f32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v8f32_fadd_f32:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v8f32_fadd_f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v8f32_fadd_f32:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v8f32_fadd_f32:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v8f32_fadd_f32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <8 x float> %x, i32 0
   %x1 = extractelement <8 x float> %x, i32 1
   %x01 = fadd float %x0, %x1
   ret float %x01
 }
 
 define float @extract_extract_v8f32_fadd_f32_commute(<8 x float> %x) {
-; SSE3-LABEL: extract_extract_v8f32_fadd_f32_commute:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v8f32_fadd_f32_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v8f32_fadd_f32_commute:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v8f32_fadd_f32_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v8f32_fadd_f32_commute:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm0, %xmm1, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v8f32_fadd_f32_commute:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <8 x float> %x, i32 0
   %x1 = extractelement <8 x float> %x, i32 1
   %x01 = fadd float %x1, %x0
   ret float %x01
 }
 
 define double @extract_extract_v4f64_fadd_f64(<4 x double> %x) {
-; SSE3-LABEL: extract_extract_v4f64_fadd_f64:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    addsd %xmm0, %xmm1
-; SSE3-NEXT:    movapd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f64_fadd_f64:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    addsd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f64_fadd_f64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f64_fadd_f64:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f64_fadd_f64:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f64_fadd_f64:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x double> %x, i32 0
   %x1 = extractelement <4 x double> %x, i32 1
   %x01 = fadd double %x0, %x1
   ret double %x01
 }
 
 define double @extract_extract_v4f64_fadd_f64_commute(<4 x double> %x) {
-; SSE3-LABEL: extract_extract_v4f64_fadd_f64_commute:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    addsd %xmm0, %xmm1
-; SSE3-NEXT:    movapd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f64_fadd_f64_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    addsd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f64_fadd_f64_commute:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddsd %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f64_fadd_f64_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f64_fadd_f64_commute:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vaddsd %xmm0, %xmm1, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f64_fadd_f64_commute:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x double> %x, i32 0
   %x1 = extractelement <4 x double> %x, i32 1
   %x01 = fadd double %x1, %x0
   ret double %x01
 }
 
 define float @extract_extract_v8f32_fsub_f32(<8 x float> %x) {
-; SSE3-LABEL: extract_extract_v8f32_fsub_f32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    subss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v8f32_fsub_f32:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    subss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v8f32_fsub_f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vsubss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v8f32_fsub_f32:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    hsubps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v8f32_fsub_f32:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vsubss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v8f32_fsub_f32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhsubps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <8 x float> %x, i32 0
   %x1 = extractelement <8 x float> %x, i32 1
   %x01 = fsub float %x0, %x1
   ret float %x01
 }
 
+; Negative test...or get hoppy and negate?
+
 define float @extract_extract_v8f32_fsub_f32_commute(<8 x float> %x) {
 ; SSE3-LABEL: extract_extract_v8f32_fsub_f32_commute:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
 ; SSE3-NEXT:    subss %xmm0, %xmm1
 ; SSE3-NEXT:    movaps %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; AVX-LABEL: extract_extract_v8f32_fsub_f32_commute:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
 ; AVX-NEXT:    vsubss %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq
   %x0 = extractelement <8 x float> %x, i32 0
   %x1 = extractelement <8 x float> %x, i32 1
   %x01 = fsub float %x1, %x0
   ret float %x01
 }
 
 define double @extract_extract_v4f64_fsub_f64(<4 x double> %x) {
-; SSE3-LABEL: extract_extract_v4f64_fsub_f64:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    subsd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f64_fsub_f64:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    subsd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f64_fsub_f64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f64_fsub_f64:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    hsubpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f64_fsub_f64:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f64_fsub_f64:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhsubpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x double> %x, i32 0
   %x1 = extractelement <4 x double> %x, i32 1
   %x01 = fsub double %x0, %x1
   ret double %x01
 }
 
+; Negative test...or get hoppy and negate?
+
 define double @extract_extract_v4f64_fsub_f64_commute(<4 x double> %x) {
 ; SSE3-LABEL: extract_extract_v4f64_fsub_f64_commute:
 ; SSE3:       # %bb.0:
 ; SSE3-NEXT:    movapd %xmm0, %xmm1
 ; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
 ; SSE3-NEXT:    subsd %xmm0, %xmm1
 ; SSE3-NEXT:    movapd %xmm1, %xmm0
 ; SSE3-NEXT:    retq
 ;
 ; AVX-LABEL: extract_extract_v4f64_fsub_f64_commute:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
 ; AVX-NEXT:    vsubsd %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq
   %x0 = extractelement <4 x double> %x, i32 0
   %x1 = extractelement <4 x double> %x, i32 1
   %x01 = fsub double %x1, %x0
   ret double %x01
 }
 
 ; 512-bit vectors, float/double, fadd/fsub
 
 define float @extract_extract_v16f32_fadd_f32(<16 x float> %x) {
-; SSE3-LABEL: extract_extract_v16f32_fadd_f32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v16f32_fadd_f32:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v16f32_fadd_f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v16f32_fadd_f32:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v16f32_fadd_f32:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v16f32_fadd_f32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <16 x float> %x, i32 0
   %x1 = extractelement <16 x float> %x, i32 1
   %x01 = fadd float %x0, %x1
   ret float %x01
 }
 
 define float @extract_extract_v16f32_fadd_f32_commute(<16 x float> %x) {
-; SSE3-LABEL: extract_extract_v16f32_fadd_f32_commute:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v16f32_fadd_f32_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v16f32_fadd_f32_commute:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v16f32_fadd_f32_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v16f32_fadd_f32_commute:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm0, %xmm1, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v16f32_fadd_f32_commute:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <16 x float> %x, i32 0
   %x1 = extractelement <16 x float> %x, i32 1
   %x01 = fadd float %x1, %x0
   ret float %x01
 }
 
 define double @extract_extract_v8f64_fadd_f64(<8 x double> %x) {
-; SSE3-LABEL: extract_extract_v8f64_fadd_f64:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    addsd %xmm0, %xmm1
-; SSE3-NEXT:    movapd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v8f64_fadd_f64:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    addsd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v8f64_fadd_f64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v8f64_fadd_f64:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v8f64_fadd_f64:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v8f64_fadd_f64:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <8 x double> %x, i32 0
   %x1 = extractelement <8 x double> %x, i32 1
   %x01 = fadd double %x0, %x1
   ret double %x01
 }
 
 define double @extract_extract_v8f64_fadd_f64_commute(<8 x double> %x) {
-; SSE3-LABEL: extract_extract_v8f64_fadd_f64_commute:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    addsd %xmm0, %xmm1
-; SSE3-NEXT:    movapd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v8f64_fadd_f64_commute:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    addsd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    movapd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v8f64_fadd_f64_commute:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vaddsd %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v8f64_fadd_f64_commute:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    haddpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v8f64_fadd_f64_commute:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vaddsd %xmm0, %xmm1, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v8f64_fadd_f64_commute:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <8 x double> %x, i32 0
   %x1 = extractelement <8 x double> %x, i32 1
   %x01 = fadd double %x1, %x0
   ret double %x01
 }
 
 define float @extract_extract_v16f32_fsub_f32(<16 x float> %x) {
-; SSE3-LABEL: extract_extract_v16f32_fsub_f32:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    subss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v16f32_fsub_f32:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    subss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v16f32_fsub_f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vsubss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v16f32_fsub_f32:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    hsubps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v16f32_fsub_f32:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vsubss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v16f32_fsub_f32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhsubps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <16 x float> %x, i32 0
   %x1 = extractelement <16 x float> %x, i32 1
   %x01 = fsub float %x0, %x1
   ret float %x01
 }
 
 define float @extract_extract_v16f32_fsub_f32_commute(<16 x float> %x) {
 ; SSE3-LABEL: extract_extract_v16f32_fsub_f32_commute:

 ; AVX-NEXT:    retq
   %x0 = extractelement <16 x float> %x, i32 0
   %x1 = extractelement <16 x float> %x, i32 1
   %x01 = fsub float %x1, %x0
   ret float %x01
 }
 
 define double @extract_extract_v8f64_fsub_f64(<8 x double> %x) {
-; SSE3-LABEL: extract_extract_v8f64_fsub_f64:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movapd %xmm0, %xmm1
-; SSE3-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
-; SSE3-NEXT:    subsd %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v8f64_fsub_f64:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movapd %xmm0, %xmm1
+; SSE3-SLOW-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
+; SSE3-SLOW-NEXT:    subsd %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v8f64_fsub_f64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v8f64_fsub_f64:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    hsubpd %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v8f64_fsub_f64:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    vzeroupper
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v8f64_fsub_f64:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhsubpd %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vzeroupper
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <8 x double> %x, i32 0
   %x1 = extractelement <8 x double> %x, i32 1
   %x01 = fsub double %x0, %x1
   ret double %x01
 }
 
 define double @extract_extract_v8f64_fsub_f64_commute(<8 x double> %x) {
 ; SSE3-LABEL: extract_extract_v8f64_fsub_f64_commute:
; AVX-NEXT:    retq
   %x1 = extractelement <8 x double> %x, i32 1
   %x01 = fsub double %x1, %x0
   ret double %x01
 }
 
 ; Check output when 1 or both extracts have extra uses.
 
 define float @extract_extract_v4f32_fadd_f32_uses1(<4 x float> %x, float* %p) {
-; SSE3-LABEL: extract_extract_v4f32_fadd_f32_uses1:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movss %xmm0, (%rdi)
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f32_fadd_f32_uses1:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movss %xmm0, (%rdi)
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f32_fadd_f32_uses1:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovss %xmm0, (%rdi)
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f32_fadd_f32_uses1:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    movss %xmm0, (%rdi)
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f32_fadd_f32_uses1:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovss %xmm0, (%rdi)
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f32_fadd_f32_uses1:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vmovss %xmm0, (%rdi)
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x float> %x, i32 0
   store float %x0, float* %p
   %x1 = extractelement <4 x float> %x, i32 1
   %x01 = fadd float %x0, %x1
   ret float %x01
 }
 
 define float @extract_extract_v4f32_fadd_f32_uses2(<4 x float> %x, float* %p) {
-; SSE3-LABEL: extract_extract_v4f32_fadd_f32_uses2:
-; SSE3:       # %bb.0:
-; SSE3-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE3-NEXT:    movss %xmm1, (%rdi)
-; SSE3-NEXT:    addss %xmm1, %xmm0
-; SSE3-NEXT:    retq
+; SSE3-SLOW-LABEL: extract_extract_v4f32_fadd_f32_uses2:
+; SSE3-SLOW:       # %bb.0:
+; SSE3-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-SLOW-NEXT:    movss %xmm1, (%rdi)
+; SSE3-SLOW-NEXT:    addss %xmm1, %xmm0
+; SSE3-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: extract_extract_v4f32_fadd_f32_uses2:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vmovss %xmm1, (%rdi)
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; SSE3-FAST-LABEL: extract_extract_v4f32_fadd_f32_uses2:
+; SSE3-FAST:       # %bb.0:
+; SSE3-FAST-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE3-FAST-NEXT:    movss %xmm1, (%rdi)
+; SSE3-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE3-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: extract_extract_v4f32_fadd_f32_uses2:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vmovss %xmm1, (%rdi)
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: extract_extract_v4f32_fadd_f32_uses2:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vextractps $1, %xmm0, (%rdi)
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
   %x0 = extractelement <4 x float> %x, i32 0
   %x1 = extractelement <4 x float> %x, i32 1
   store float %x1, float* %p
   %x01 = fadd float %x0, %x1
   ret float %x01
 }
 
 define float @extract_extract_v4f32_fadd_f32_uses3(<4 x float> %x, float* %p1, float* %p2) {