[TargetLowering][AMDGPU][X86] Improve SimplifyDemandedBits bitcast handling

[RKSimon]

Summary

This patch adds support for BigBitWidth -> SmallBitWidth bitcasts, splitting the DemandedBits/Elts accordingly.

Re: the AMDGPU regression - @arsenm it looks there isn't much that generates BFE U32/S32 nodes in the DAG its mostly done in ISEL - in this case we need to match srl(and(shl(x,c1),c2),c1) - is this something that needs fixing first or are you OK with this change for now?

The X86 changes are all definite wins.

[RKSimon]

test/CodeGen/AMDGPU/store-weird-sizes.ll

49

@arsenm Any suggestions on what to do here - is it better to add another isel pattern or start creating bfe_user dag nodes?

[nhaehnle]

test/CodeGen/AMDGPU/store-weird-sizes.ll

49

Why doesn't the DAGCombiner simplify this to a single AND node?

[RKSimon]

rebase

[RKSimon]

test/CodeGen/AMDGPU/store-weird-sizes.ll

49

The AND appears too late (from a bitcasted i64 -> v2i32) for visitShiftByConstant to be called from the SRL node - I'm working on a generic fix at the moment.

[RKSimon]

Add AMDGPU srl(and(x,m),c) -> and(srl(x,c),srl(m,c)) canonicalization to improve BFE recognition

[RKSimon]

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

3170

@nhaehnle @arsenm I think this handles the BFE issue - I investigated putting this in DAGCombine but it caused a lot of noise on other targets - some improvements, some regressions.

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

3162

Braecs around this

lib/CodeGen/SelectionDAG/TargetLowering.cpp

if (SrcVT.isVector() && NumSrcEltBits > 1 &&
       if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef,
                                      KnownSrcZero, TLO, Depth + 1))
         return true;
 
       KnownBits KnownSrcBits;
       if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts,
                                KnownSrcBits, TLO, Depth + 1))
         return true;
+    } else if ((NumSrcEltBits % BitWidth) == 0 &&
+               TLO.DAG.getDataLayout().isLittleEndian()) {
+      unsigned Scale = NumSrcEltBits / BitWidth;
+      unsigned NumSrcElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1;
+      APInt DemandedSrcBits = APInt::getNullValue(NumSrcEltBits);
+      APInt DemandedSrcElts = APInt::getNullValue(NumSrcElts);
+      for (unsigned i = 0; i != NumElts; ++i)
+        if (DemandedElts[i]) {
+          unsigned Offset = (i % Scale) * BitWidth;
+          DemandedSrcBits.insertBits(DemandedBits, Offset);
+          DemandedSrcElts.setBit(i / Scale);
+        }
+
+      if (SrcVT.isVector()) {
+        APInt KnownSrcUndef, KnownSrcZero;
+        if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef,
+                                       KnownSrcZero, TLO, Depth + 1))
+          return true;
+      }
+
+      KnownBits KnownSrcBits;
+      if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts,
+                               KnownSrcBits, TLO, Depth + 1))
+        return true;
     }
 
     // If this is a bitcast, let computeKnownBits handle it.  Only do this on a
     // recursive call where Known may be useful to the caller.
     if (Depth > 0) {
-      Known = TLO.DAG.computeKnownBits(Op, Depth);
+      Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
       return false;
     }
     break;
   }
   case ISD::ADD:
   case ISD::MUL:
   case ISD::SUB: {
     // Add, Sub, and Mul don't demand any bits in positions beyond that

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

if (RHSVal == 63) {
     return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
   }
 
   return SDValue();
 }
 
 SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
-  if (N->getValueType(0) != MVT::i64)
-    return SDValue();
-
-  const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
+  auto *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
   if (!RHS)
     return SDValue();
 
+  EVT VT = N->getValueType(0);
+  SDValue LHS = N->getOperand(0);
   unsigned ShiftAmt = RHS->getZExtValue();
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc SL(N);
+
+  // fold (srl (and x, c1 << c2), c2) -> (and (srl(x, c2), c1)
+  // this improves the ability to match BFE patterns in isel.
+  if (LHS.getOpcode() == ISD::AND)
+    if (auto *Mask = dyn_cast<ConstantSDNode>(LHS.getOperand(1)))
+      if (Mask->getAPIntValue().isShiftedMask() &&
+          Mask->getAPIntValue().countTrailingZeros() == ShiftAmt) {
+        return DAG.getNode(
+            ISD::AND, SL, VT,
+            DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(0), N->getOperand(1)),
+            DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(1), N->getOperand(1)));
+      }
+
+  if (VT != MVT::i64)
+    return SDValue();
+
   if (ShiftAmt < 32)
     return SDValue();
 
   // srl i64:x, C for C >= 32
   // =>
   //   build_pair (srl hi_32(x), C - 32), 0
-
-  SelectionDAG &DAG = DCI.DAG;
-  SDLoc SL(N);
-
   SDValue One = DAG.getConstant(1, SL, MVT::i32);
   SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
 
-  SDValue VecOp = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, N->getOperand(0));
-  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32,
-                           VecOp, One);
+  SDValue VecOp = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, LHS);
+  SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecOp, One);
 
   SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32);
   SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst);
 
   SDValue BuildPair = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, Zero});
 
   return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildPair);
 }

test/CodeGen/AMDGPU/store-weird-sizes.ll

 ; HAWAII-NEXT:    s_mov_b32 m0, -1
 ; HAWAII-NEXT:    s_waitcnt lgkmcnt(0)
 ; HAWAII-NEXT:    v_mov_b32_e32 v1, s0
 ; HAWAII-NEXT:    v_mov_b32_e32 v3, s1
 ; HAWAII-NEXT:    v_mov_b32_e32 v2, s2
 ; HAWAII-NEXT:    ds_write_b16 v1, v2 offset:4
 ; HAWAII-NEXT:    s_waitcnt vmcnt(0)
 ; HAWAII-NEXT:    v_and_b32_e32 v0, 0x7f, v0
 ; HAWAII-NEXT:    ds_write_b8 v1, v0 offset:6
 ; HAWAII-NEXT:    ds_write_b32 v1, v3
 ; HAWAII-NEXT:    s_endpgm
 ;
 ; FIJI-LABEL: local_store_i55:
 ; FIJI:       ; %bb.0:
 ; FIJI-NEXT:    s_add_u32 s0, s4, 14
 ; FIJI-NEXT:    s_addc_u32 s1, s5, 0
 ; FIJI-NEXT:    v_mov_b32_e32 v0, s0

 ; GFX9-NEXT:    s_load_dword s1, s[4:5], 0x8
 ; GFX9-NEXT:    s_load_dword s2, s[4:5], 0xc
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v3, s1
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX9-NEXT:    ds_write_b16 v1, v2 offset:4
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_and_b32_e32 v0, 0x7f0000, v0
-; GFX9-NEXT:    ds_write_b8_d16_hi v1, v0 offset:6
+; GFX9-NEXT:    v_bfe_u32 v0, v0, 16, 7
+; GFX9-NEXT:    ds_write_b8 v1, v0 offset:6
 ; GFX9-NEXT:    ds_write_b32 v1, v3
 ; GFX9-NEXT:    s_endpgm
   store i55 %arg, i55 addrspace(3)* %ptr, align 8
   ret void
 }
 
 define amdgpu_kernel void @local_store_i48(i48 addrspace(3)* %ptr, i48 %arg) #0 {
 ; HAWAII-LABEL: local_store_i48:

test/CodeGen/X86/bitcast-setcc-256.ll

; AVX512BW-NEXT:    retq
   %a2 = bitcast <8 x i1> %a1 to i8
   store i8 %a2, i8* %p
   ret void
 }
 
 define void @bitcast_4i64_store(i4* %p, <4 x i64> %a0) {
 ; SSE2-SSSE3-LABEL: bitcast_4i64_store:
 ; SSE2-SSSE3:       # %bb.0:
-; SSE2-SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
-; SSE2-SSSE3-NEXT:    pxor %xmm2, %xmm1
-; SSE2-SSSE3-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-SSSE3-NEXT:    pcmpeqd %xmm1, %xmm3
-; SSE2-SSSE3-NEXT:    movdqa %xmm2, %xmm4
-; SSE2-SSSE3-NEXT:    pcmpgtd %xmm1, %xmm4
-; SSE2-SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,2]
-; SSE2-SSSE3-NEXT:    pand %xmm3, %xmm1
-; SSE2-SSSE3-NEXT:    por %xmm4, %xmm1
-; SSE2-SSSE3-NEXT:    pxor %xmm2, %xmm0
-; SSE2-SSSE3-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-SSSE3-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE2-SSSE3-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
-; SSE2-SSSE3-NEXT:    pand %xmm3, %xmm0
-; SSE2-SSSE3-NEXT:    por %xmm2, %xmm0
 ; SSE2-SSSE3-NEXT:    packssdw %xmm1, %xmm0
 ; SSE2-SSSE3-NEXT:    movmskps %xmm0, %eax
 ; SSE2-SSSE3-NEXT:    movb %al, (%rdi)
 ; SSE2-SSSE3-NEXT:    retq
 ;
 ; AVX12-LABEL: bitcast_4i64_store:
 ; AVX12:       # %bb.0:
 ; AVX12-NEXT:    vmovmskpd %ymm0, %eax

test/CodeGen/X86/bitcast-setcc-512.ll

 ; SSE-NEXT:    packssdw %xmm3, %xmm4
 ; SSE-NEXT:    packsswb %xmm0, %xmm4
 ; SSE-NEXT:    pmovmskb %xmm4, %eax
 ; SSE-NEXT:    movb %al, (%rdi)
 ; SSE-NEXT:    retq
 ;
 ; AVX1-LABEL: bitcast_8i64_store:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovmskps %ymm0, %eax
 ; AVX1-NEXT:    movb %al, (%rdi)
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: bitcast_8i64_store:
 ; AVX2:       # %bb.0:

test/CodeGen/X86/dagcombine-cse.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i386-apple-darwin -mattr=+sse2 | FileCheck %s --check-prefix=X32
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse2 | FileCheck %s --check-prefix=X64
 
 define i32 @t(i8* %ref_frame_ptr, i32 %ref_frame_stride, i32 %idxX, i32 %idxY) nounwind  {
 ; X32-LABEL: t:
 ; X32:       ## %bb.0: ## %entry
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X32-NEXT:    imull {{[0-9]+}}(%esp), %ecx
 ; X32-NEXT:    addl {{[0-9]+}}(%esp), %ecx
 ; X32-NEXT:    movl (%eax,%ecx), %eax
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: t:
 ; X64:       ## %bb.0: ## %entry
-; X64-NEXT:    ## kill: def $edx killed $edx def $rdx
-; X64-NEXT:    ## kill: def $esi killed $esi def $rsi
 ; X64-NEXT:    imull %ecx, %esi
-; X64-NEXT:    leal (%rsi,%rdx), %eax
-; X64-NEXT:    cltq
+; X64-NEXT:    addl %edx, %esi
+; X64-NEXT:    movslq %esi, %rax
 ; X64-NEXT:    movl (%rdi,%rax), %eax
-; X64-NEXT:    leal 4(%rsi,%rdx), %ecx
-; X64-NEXT:    movslq %ecx, %rcx
-; X64-NEXT:    movzwl (%rdi,%rcx), %ecx
-; X64-NEXT:    shlq $32, %rcx
-; X64-NEXT:    orq %rax, %rcx
-; X64-NEXT:    movq %rcx, %xmm0
+; X64-NEXT:    movq %rax, %xmm0
 ; X64-NEXT:    movd %xmm0, %eax
 ; X64-NEXT:    retq
 entry:
   %tmp7 = mul i32 %idxY, %ref_frame_stride    ; <i32> [#uses=2]
   %tmp9 = add i32 %tmp7, %idxX    ; <i32> [#uses=1]
   %tmp11 = getelementptr i8, i8* %ref_frame_ptr, i32 %tmp9    ; <i8*> [#uses=1]
   %tmp1112 = bitcast i8* %tmp11 to i32*    ; <i32*> [#uses=1]
   %tmp13 = load i32, i32* %tmp1112, align 4    ; <i32> [#uses=1]

test/CodeGen/X86/masked_store.ll

; AVX-NEXT:    retq
   %mask = icmp slt <1 x i64> %trigger, zeroinitializer
   call void @llvm.masked.store.v1f64.p0v1f64(<1 x double> %val, <1 x double>* %addr, i32 4, <1 x i1> %mask)
   ret void
 }
 
 define void @store_v2f64_v2i64(<2 x i64> %trigger, <2 x double>* %addr, <2 x double> %val) {
 ; SSE2-LABEL: store_v2f64_v2i64:
 ; SSE2:       ## %bb.0:
-; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
-; SSE2-NEXT:    pxor %xmm3, %xmm0
-; SSE2-NEXT:    movdqa %xmm3, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm4
-; SSE2-NEXT:    pand %xmm2, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    movd %xmm4, %eax
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE2-NEXT:    pxor %xmm2, %xmm0
+; SSE2-NEXT:    movdqa %xmm2, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm3, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB1_2
 ; SSE2-NEXT:  ## %bb.1: ## %cond.store
 ; SSE2-NEXT:    movlpd %xmm1, (%rdi)
 ; SSE2-NEXT:  LBB1_2: ## %else
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    por %xmm3, %xmm0
 ; SSE2-NEXT:    pextrw $4, %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB1_4
 ; SSE2-NEXT:  ## %bb.3: ## %cond.store1
 ; SSE2-NEXT:    movhpd %xmm1, 8(%rdi)
 ; SSE2-NEXT:  LBB1_4: ## %else2
 ; SSE2-NEXT:    retq
 ;

 define void @store_v4f64_v4i64(<4 x i64> %trigger, <4 x double>* %addr, <4 x double> %val) {
 ; SSE2-LABEL: store_v4f64_v4i64:
 ; SSE2:       ## %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648]
 ; SSE2-NEXT:    pxor %xmm4, %xmm0
 ; SSE2-NEXT:    movdqa %xmm4, %xmm5
 ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm5
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm7
-; SSE2-NEXT:    pand %xmm5, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    movd %xmm7, %eax
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm6, %xmm0
+; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB2_2
 ; SSE2-NEXT:  ## %bb.1: ## %cond.store
 ; SSE2-NEXT:    movlpd %xmm2, (%rdi)
 ; SSE2-NEXT:  LBB2_2: ## %else
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm0
-; SSE2-NEXT:    por %xmm6, %xmm0
 ; SSE2-NEXT:    pextrw $4, %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB2_4
 ; SSE2-NEXT:  ## %bb.3: ## %cond.store1
 ; SSE2-NEXT:    movhpd %xmm2, 8(%rdi)
 ; SSE2-NEXT:  LBB2_4: ## %else2
 ; SSE2-NEXT:    pxor %xmm4, %xmm1
 ; SSE2-NEXT:    movdqa %xmm4, %xmm0
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pand %xmm2, %xmm1
+; SSE2-NEXT:    pand %xmm0, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    pextrw $0, %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB2_6
 ; SSE2-NEXT:  ## %bb.5: ## %cond.store3
 ; SSE2-NEXT:    movlpd %xmm3, 16(%rdi)
 ; SSE2-NEXT:  LBB2_6: ## %else4

 
 ;
 ; vXi64
 ;
 
 define void @store_v2i64_v2i64(<2 x i64> %trigger, <2 x i64>* %addr, <2 x i64> %val) {
 ; SSE2-LABEL: store_v2i64_v2i64:
 ; SSE2:       ## %bb.0:
-; SSE2-NEXT:    movdqa {{.*#+}} xmm3 = [2147483648,2147483648]
-; SSE2-NEXT:    pxor %xmm3, %xmm0
-; SSE2-NEXT:    movdqa %xmm3, %xmm2
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pcmpeqd %xmm3, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm4
-; SSE2-NEXT:    pand %xmm2, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
-; SSE2-NEXT:    por %xmm3, %xmm4
-; SSE2-NEXT:    movd %xmm4, %eax
+; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
+; SSE2-NEXT:    pxor %xmm2, %xmm0
+; SSE2-NEXT:    movdqa %xmm2, %xmm3
+; SSE2-NEXT:    pcmpgtd %xmm0, %xmm3
+; SSE2-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm3, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
+; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB7_2
 ; SSE2-NEXT:  ## %bb.1: ## %cond.store
 ; SSE2-NEXT:    movq %xmm1, (%rdi)
 ; SSE2-NEXT:  LBB7_2: ## %else
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pand %xmm2, %xmm0
-; SSE2-NEXT:    por %xmm3, %xmm0
 ; SSE2-NEXT:    pextrw $4, %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB7_4
 ; SSE2-NEXT:  ## %bb.3: ## %cond.store1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
 ; SSE2-NEXT:    movq %xmm0, 8(%rdi)
 ; SSE2-NEXT:  LBB7_4: ## %else2
 ; SSE2-NEXT:    retq

 define void @store_v4i64_v4i64(<4 x i64> %trigger, <4 x i64>* %addr, <4 x i64> %val) {
 ; SSE2-LABEL: store_v4i64_v4i64:
 ; SSE2:       ## %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648]
 ; SSE2-NEXT:    pxor %xmm4, %xmm0
 ; SSE2-NEXT:    movdqa %xmm4, %xmm5
 ; SSE2-NEXT:    pcmpgtd %xmm0, %xmm5
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT:    movdqa %xmm0, %xmm7
-; SSE2-NEXT:    pand %xmm5, %xmm7
-; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,1,3,3]
-; SSE2-NEXT:    por %xmm6, %xmm7
-; SSE2-NEXT:    movd %xmm7, %eax
+; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
+; SSE2-NEXT:    pand %xmm5, %xmm6
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
+; SSE2-NEXT:    por %xmm6, %xmm0
+; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB8_2
 ; SSE2-NEXT:  ## %bb.1: ## %cond.store
 ; SSE2-NEXT:    movq %xmm2, (%rdi)
 ; SSE2-NEXT:  LBB8_2: ## %else
-; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[0,0,2,2]
-; SSE2-NEXT:    pand %xmm5, %xmm0
-; SSE2-NEXT:    por %xmm6, %xmm0
 ; SSE2-NEXT:    pextrw $4, %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB8_4
 ; SSE2-NEXT:  ## %bb.3: ## %cond.store1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,0,1]
 ; SSE2-NEXT:    movq %xmm0, 8(%rdi)
 ; SSE2-NEXT:  LBB8_4: ## %else2
 ; SSE2-NEXT:    pxor %xmm4, %xmm1
 ; SSE2-NEXT:    movdqa %xmm4, %xmm0
 ; SSE2-NEXT:    pcmpgtd %xmm1, %xmm0
-; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
 ; SSE2-NEXT:    pcmpeqd %xmm4, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
-; SSE2-NEXT:    pand %xmm2, %xmm1
+; SSE2-NEXT:    pand %xmm0, %xmm1
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
 ; SSE2-NEXT:    por %xmm1, %xmm0
 ; SSE2-NEXT:    pextrw $0, %xmm0, %eax
 ; SSE2-NEXT:    testb $1, %al
 ; SSE2-NEXT:    je LBB8_6
 ; SSE2-NEXT:  ## %bb.5: ## %cond.store3
 ; SSE2-NEXT:    movq %xmm3, 16(%rdi)
 ; SSE2-NEXT:  LBB8_6: ## %else4

test/CodeGen/X86/movmsk-cmp.ll

; SKX-NEXT:    retq
   %tmp1 = bitcast <16 x i1> %tmp to i16
   %tmp2 = icmp eq i16 %tmp1, 0
   ret i1 %tmp2
 }
 
 define i1 @allones_v4i64_sign(<4 x i64> %arg) {
 ; SSE2-LABEL: allones_v4i64_sign:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
-; SSE2-NEXT:    pxor %xmm2, %xmm1
-; SSE2-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm3
-; SSE2-NEXT:    movdqa %xmm2, %xmm4
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,2]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    por %xmm4, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pand %xmm3, %xmm0
-; SSE2-NEXT:    por %xmm2, %xmm0
 ; SSE2-NEXT:    packssdw %xmm1, %xmm0
 ; SSE2-NEXT:    movmskps %xmm0, %eax
 ; SSE2-NEXT:    cmpb $15, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX-LABEL: allones_v4i64_sign:
 ; AVX:       # %bb.0:
; SKX-NEXT:    retq
   %tmp1 = bitcast <4 x i1> %tmp to i4
   %tmp2 = icmp eq i4 %tmp1, -1
   ret i1 %tmp2
 }
 
 define i1 @allzeros_v4i64_sign(<4 x i64> %arg) {
 ; SSE2-LABEL: allzeros_v4i64_sign:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
-; SSE2-NEXT:    pxor %xmm2, %xmm1
-; SSE2-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm3
-; SSE2-NEXT:    movdqa %xmm2, %xmm4
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,2]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    por %xmm4, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pand %xmm3, %xmm0
-; SSE2-NEXT:    por %xmm2, %xmm0
 ; SSE2-NEXT:    packssdw %xmm1, %xmm0
 ; SSE2-NEXT:    movmskps %xmm0, %eax
 ; SSE2-NEXT:    testb %al, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX-LABEL: allzeros_v4i64_sign:
 ; AVX:       # %bb.0:

 ; SSE2-NEXT:    packsswb %xmm0, %xmm1
 ; SSE2-NEXT:    pmovmskb %xmm1, %eax
 ; SSE2-NEXT:    cmpb $-1, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: allones_v8i64_sign:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovmskps %ymm0, %eax
 ; AVX1-NEXT:    cmpb $-1, %al
 ; AVX1-NEXT:    sete %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: allones_v8i64_sign:

 ; SSE2-NEXT:    packsswb %xmm0, %xmm1
 ; SSE2-NEXT:    pmovmskb %xmm1, %eax
 ; SSE2-NEXT:    testb %al, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: allzeros_v8i64_sign:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovmskps %ymm0, %eax
 ; AVX1-NEXT:    testb %al, %al
 ; AVX1-NEXT:    sete %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: allzeros_v8i64_sign:

 ; SSE2-NEXT:    packsswb %xmm0, %xmm0
 ; SSE2-NEXT:    pmovmskb %xmm0, %eax
 ; SSE2-NEXT:    cmpb $-1, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: allones_v8i64_and1:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpsllq $63, %xmm2, %xmm2
-; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsllq $63, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
 ; AVX1-NEXT:    vpsllq $63, %xmm2, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpsllq $63, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpsllq $63, %xmm2, %xmm2
+; AVX1-NEXT:    vpsllq $63, %xmm1, %xmm1
+; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovmskps %ymm0, %eax
 ; AVX1-NEXT:    cmpb $-1, %al
 ; AVX1-NEXT:    sete %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: allones_v8i64_and1:

 ; SSE2-NEXT:    packsswb %xmm0, %xmm0
 ; SSE2-NEXT:    pmovmskb %xmm0, %eax
 ; SSE2-NEXT:    testb %al, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: allzeros_v8i64_and1:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpsllq $63, %xmm2, %xmm2
-; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsllq $63, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
 ; AVX1-NEXT:    vpsllq $63, %xmm2, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpsllq $63, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpsllq $63, %xmm2, %xmm2
+; AVX1-NEXT:    vpsllq $63, %xmm1, %xmm1
+; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovmskps %ymm0, %eax
 ; AVX1-NEXT:    testb %al, %al
 ; AVX1-NEXT:    sete %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: allzeros_v8i64_and1:

 ; SSE2-NEXT:    packsswb %xmm0, %xmm0
 ; SSE2-NEXT:    pmovmskb %xmm0, %eax
 ; SSE2-NEXT:    cmpb $-1, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: allones_v8i64_and4:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpsllq $61, %xmm2, %xmm2
-; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsllq $61, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
 ; AVX1-NEXT:    vpsllq $61, %xmm2, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpsllq $61, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpsllq $61, %xmm2, %xmm2
+; AVX1-NEXT:    vpsllq $61, %xmm1, %xmm1
+; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovmskps %ymm0, %eax
 ; AVX1-NEXT:    cmpb $-1, %al
 ; AVX1-NEXT:    sete %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: allones_v8i64_and4:

 ; SSE2-NEXT:    packsswb %xmm0, %xmm0
 ; SSE2-NEXT:    pmovmskb %xmm0, %eax
 ; SSE2-NEXT:    testb %al, %al
 ; SSE2-NEXT:    sete %al
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: allzeros_v8i64_and4:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT:    vpsllq $61, %xmm2, %xmm2
-; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpsllq $61, %xmm1, %xmm1
-; AVX1-NEXT:    vpcmpgtq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
 ; AVX1-NEXT:    vpsllq $61, %xmm2, %xmm2
+; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
 ; AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vpsllq $61, %xmm0, %xmm0
 ; AVX1-NEXT:    vpcmpgtq %xmm0, %xmm3, %xmm0
 ; AVX1-NEXT:    vpackssdw %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vpsllq $61, %xmm2, %xmm2
+; AVX1-NEXT:    vpsllq $61, %xmm1, %xmm1
+; AVX1-NEXT:    vpackssdw %xmm2, %xmm1, %xmm1
 ; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovmskps %ymm0, %eax
 ; AVX1-NEXT:    testb %al, %al
 ; AVX1-NEXT:    sete %al
 ; AVX1-NEXT:    vzeroupper
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: allzeros_v8i64_and4:
; SKX-NEXT:    retq
   %c = bitcast <4 x i1> %b to i4
   %d = zext i4 %c to i32
   ret i32 %d
 }
 
 define i32 @movmskpd256(<4 x double> %x) {
 ; SSE2-LABEL: movmskpd256:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
-; SSE2-NEXT:    pxor %xmm2, %xmm1
-; SSE2-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-NEXT:    pcmpeqd %xmm1, %xmm3
-; SSE2-NEXT:    movdqa %xmm2, %xmm4
-; SSE2-NEXT:    pcmpgtd %xmm1, %xmm4
-; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,2]
-; SSE2-NEXT:    pand %xmm3, %xmm1
-; SSE2-NEXT:    por %xmm4, %xmm1
-; SSE2-NEXT:    pxor %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm2, %xmm3
-; SSE2-NEXT:    pcmpeqd %xmm0, %xmm3
-; SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
-; SSE2-NEXT:    pand %xmm3, %xmm0
-; SSE2-NEXT:    por %xmm2, %xmm0
 ; SSE2-NEXT:    packssdw %xmm1, %xmm0
 ; SSE2-NEXT:    movmskps %xmm0, %eax
 ; SSE2-NEXT:    retq
 ;
 ; AVX-LABEL: movmskpd256:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovmskpd %ymm0, %eax
 ; AVX-NEXT:    vzeroupper