[X86] Add initial support for unfolding broadcast loads from arithmetic instructions to enable LICM hoisting of the load

[craig.topper]

Summary

MachineLICM can hoist an invariant load, but if that load is folded it needs to be unfolded. On AVX512 sometimes this load is an broadcast load which we were previously unable to unfold. This patch adds initial support for that with a very basic list of supported instructions as a starting point.

[craig.topper]

llvm/lib/Target/X86/X86InstrFoldTables.h

28

This was done to get a free bit for the TB_FOLDED_BCAST flag

44–46

I re-encoded these to get room for the broadcast type. I can pre-commit this.

llvm/lib/Target/X86/X86InstrInfo.cpp

4808

Part of the re-encoding of the alignment field.

[RKSimon]

Does this break the X86FoldTablesEmitter in anyway?

In the future do you think we could we add something similar to support scalar ops and ops where memory size != register size?

llvm/lib/Target/X86/X86InstrFoldTables.cpp

5248

Do masked ops work? Possibly add a few to this initial test?

llvm/lib/Target/X86/X86InstrFoldTables.h

44–46

I'm happy for the bit adjustments to go in straight away (without the new broadcast bits - these need to stay in this patch).

[craig.topper]

In D67017#1653603, @RKSimon wrote:

Does this break the X86FoldTablesEmitter in anyway?

In the future do you think we could we add something similar to support scalar ops and ops where memory size != register size?

I don't think it will break the X86FoldTablesEmitter. The emitter just won't generate the new table.

Yes I think we could support scalar ops in the future.

llvm/lib/Target/X86/X86InstrFoldTables.cpp

5248

I think if the mask was dynamically all 0 then any memory fault would be masked. So I think if we unfold it, we would need to generate a masked broadcast as well to maintain the fault suppression. Which wouldn't be eligible for hoisting. Though we don't fold masked operations in the first place so unfolding without applying the mask would be fine today, but I wouldn't want it to break in the future.

[craig.topper]

Rebase

[RKSimon]

Are there any AVX512 ISAs that generate b/w broadcasts folded instructions?

llvm/lib/Target/X86/X86InstrInfo.cpp

5325

Test coverage?

5334

Test coverage?

[craig.topper]

Remove byte and word broadcasts

[RKSimon]

LGTM with 2 minors, cheers.

llvm/lib/Target/X86/X86InstrFoldTables.h

54

Unused bits 14-15

55

12 - 13 ?

llvm/lib/Target/X86/X86InstrFoldTables.h

 enum {
   // Select which memory operand is being unfolded.
   // (stored in bits 0 - 2)
   TB_INDEX_0    = 0,
   TB_INDEX_1    = 1,
   TB_INDEX_2    = 2,
   TB_INDEX_3    = 3,
   TB_INDEX_4    = 4,
   TB_INDEX_MASK = 0x7,
 
   // Do not insert the reverse map (MemOp -> RegOp) into the table.
   // This may be needed because there is a many -> one mapping.
   TB_NO_REVERSE   = 1 << 3,
 
   // Do not insert the forward map (RegOp -> MemOp) into the table.
   // This is needed for Native Client, which prohibits branch
   // instructions from using a memory operand.
   TB_NO_FORWARD   = 1 << 4,
 
   TB_FOLDED_LOAD  = 1 << 5,
   TB_FOLDED_STORE = 1 << 6,
-  // Unused bit 7
+  TB_FOLDED_BCAST = 1 << 7,
 
   // Minimum alignment required for load/store.
   // Used for RegOp->MemOp conversion. Encoded as Log2(Align) + 1 to allow 0
   // to mean align of 0.
   // (stored in bits 8 - 11)
   TB_ALIGN_SHIFT = 8,
   TB_ALIGN_NONE  =   0 << TB_ALIGN_SHIFT,
   TB_ALIGN_16    =   5 << TB_ALIGN_SHIFT,
   TB_ALIGN_32    =   6 << TB_ALIGN_SHIFT,
   TB_ALIGN_64    =   7 << TB_ALIGN_SHIFT,
   TB_ALIGN_MASK  = 0xf << TB_ALIGN_SHIFT,
 
-  // Unused bits 12-15
+  // Broadcast type.
+  // (stored in bits 12 - 15)
+  TB_BCAST_TYPE_SHIFT = 12,
+  TB_BCAST_D    =   0 << TB_BCAST_TYPE_SHIFT,
+  TB_BCAST_Q    =   1 << TB_BCAST_TYPE_SHIFT,
+  TB_BCAST_SS   =   2 << TB_BCAST_TYPE_SHIFT,
+  TB_BCAST_SD   =   3 << TB_BCAST_TYPE_SHIFT,
+  TB_BCAST_MASK = 0x3 << TB_BCAST_TYPE_SHIFT,
 };
 
 // This struct is used for both the folding and unfold tables. They KeyOp
 // is used to determine the sorting order.
 struct X86MemoryFoldTableEntry {
   uint16_t KeyOp;
   uint16_t DstOp;
   uint16_t Flags;

llvm/lib/Target/X86/X86InstrFoldTables.cpp

static const X86MemoryFoldTableEntry MemoryFoldTable4[] = {
   { X86::VXORPDZ128rrk,             X86::VXORPDZ128rmk,             0 },
   { X86::VXORPDZ256rrk,             X86::VXORPDZ256rmk,             0 },
   { X86::VXORPDZrrk,                X86::VXORPDZrmk,                0 },
   { X86::VXORPSZ128rrk,             X86::VXORPSZ128rmk,             0 },
   { X86::VXORPSZ256rrk,             X86::VXORPSZ256rmk,             0 },
   { X86::VXORPSZrrk,                X86::VXORPSZrmk,                0 },
 };
 
+static const X86MemoryFoldTableEntry BroadcastFoldTable2[] = {
+  { X86::VADDPDZ128rr,   X86::VADDPDZ128rmb,   TB_BCAST_SD },
+  { X86::VADDPDZ256rr,   X86::VADDPDZ256rmb,   TB_BCAST_SD },
+  { X86::VADDPDZrr,      X86::VADDPDZrmb,      TB_BCAST_SD },
+  { X86::VADDPSZ128rr,   X86::VADDPSZ128rmb,   TB_BCAST_SS },
+  { X86::VADDPSZ256rr,   X86::VADDPSZ256rmb,   TB_BCAST_SS },
+  { X86::VADDPSZrr,      X86::VADDPSZrmb,      TB_BCAST_SS },
+  { X86::VDIVPDZ128rr,   X86::VDIVPDZ128rmb,   TB_BCAST_SD },
+  { X86::VDIVPDZ256rr,   X86::VDIVPDZ256rmb,   TB_BCAST_SD },
+  { X86::VDIVPDZrr,      X86::VDIVPDZrmb,      TB_BCAST_SD },
+  { X86::VDIVPSZ128rr,   X86::VDIVPSZ128rmb,   TB_BCAST_SS },
+  { X86::VDIVPSZ256rr,   X86::VDIVPSZ256rmb,   TB_BCAST_SS },
+  { X86::VDIVPSZrr,      X86::VDIVPSZrmb,      TB_BCAST_SS },
+  { X86::VMULPDZ128rr,   X86::VMULPDZ128rmb,   TB_BCAST_SD },
+  { X86::VMULPDZ256rr,   X86::VMULPDZ256rmb,   TB_BCAST_SD },
+  { X86::VMULPDZrr,      X86::VMULPDZrmb,      TB_BCAST_SD },
+  { X86::VMULPSZ128rr,   X86::VMULPSZ128rmb,   TB_BCAST_SS },
+  { X86::VMULPSZ256rr,   X86::VMULPSZ256rmb,   TB_BCAST_SS },
+  { X86::VMULPSZrr,      X86::VMULPSZrmb,      TB_BCAST_SS },
+  { X86::VPADDDZ128rr,   X86::VPADDDZ128rmb,   TB_BCAST_D },
+  { X86::VPADDDZ256rr,   X86::VPADDDZ256rmb,   TB_BCAST_D },
+  { X86::VPADDDZrr,      X86::VPADDDZrmb,      TB_BCAST_D },
+  { X86::VPADDQZ128rr,   X86::VPADDQZ128rmb,   TB_BCAST_Q },
+  { X86::VPADDQZ256rr,   X86::VPADDQZ256rmb,   TB_BCAST_Q },
+  { X86::VPADDQZrr,      X86::VPADDQZrmb,      TB_BCAST_Q },
+  { X86::VPANDDZ128rr,   X86::VPANDDZ128rmb,   TB_BCAST_D },
+  { X86::VPANDDZ256rr,   X86::VPANDDZ256rmb,   TB_BCAST_D },
+  { X86::VPANDDZrr,      X86::VPANDDZrmb,      TB_BCAST_D },
+  { X86::VPANDNDZ128rr,  X86::VPANDNDZ128rmb,  TB_BCAST_D },
+  { X86::VPANDNDZ256rr,  X86::VPANDNDZ256rmb,  TB_BCAST_D },
+  { X86::VPANDNDZrr,     X86::VPANDNDZrmb,     TB_BCAST_D },
+  { X86::VPANDNQZ128rr,  X86::VPANDNQZ128rmb,  TB_BCAST_Q },
+  { X86::VPANDNQZ256rr,  X86::VPANDNQZ256rmb,  TB_BCAST_Q },
+  { X86::VPANDNQZrr,     X86::VPANDNQZrmb,     TB_BCAST_Q },
+  { X86::VPANDQZ128rr,   X86::VPANDQZ128rmb,   TB_BCAST_Q },
+  { X86::VPANDQZ256rr,   X86::VPANDQZ256rmb,   TB_BCAST_Q },
+  { X86::VPANDQZrr,      X86::VPANDQZrmb,      TB_BCAST_Q },
+  { X86::VPMULLDZ128rr,  X86::VPMULLDZ128rmb,  TB_BCAST_D },
+  { X86::VPMULLDZ256rr,  X86::VPMULLDZ256rmb,  TB_BCAST_D },
+  { X86::VPMULLDZrr,     X86::VPMULLDZrmb,     TB_BCAST_D },
+  { X86::VPMULLQZ128rr,  X86::VPMULLQZ128rmb,  TB_BCAST_Q },
+  { X86::VPMULLQZ256rr,  X86::VPMULLQZ256rmb,  TB_BCAST_Q },
+  { X86::VPMULLQZrr,     X86::VPMULLQZrmb,     TB_BCAST_Q },
+  { X86::VPORDZ128rr,    X86::VPORDZ128rmb,    TB_BCAST_D },
+  { X86::VPORDZ256rr,    X86::VPORDZ256rmb,    TB_BCAST_D },
+  { X86::VPORDZrr,       X86::VPORDZrmb,       TB_BCAST_D },
+  { X86::VPORQZ128rr,    X86::VPORQZ128rmb,    TB_BCAST_Q },
+  { X86::VPORQZ256rr,    X86::VPORQZ256rmb,    TB_BCAST_Q },
+  { X86::VPORQZrr,       X86::VPORQZrmb,       TB_BCAST_Q },
+  { X86::VPXORDZ128rr,   X86::VPXORDZ128rmb,   TB_BCAST_D },
+  { X86::VPXORDZ256rr,   X86::VPXORDZ256rmb,   TB_BCAST_D },
+  { X86::VPXORDZrr,      X86::VPXORDZrmb,      TB_BCAST_D },
+  { X86::VPXORQZ128rr,   X86::VPXORQZ128rmb,   TB_BCAST_Q },
+  { X86::VPXORQZ256rr,   X86::VPXORQZ256rmb,   TB_BCAST_Q },
+  { X86::VPXORQZrr,      X86::VPXORQZrmb,      TB_BCAST_Q },
+  { X86::VSUBPDZ128rr,   X86::VSUBPDZ128rmb,   TB_BCAST_SD },
+  { X86::VSUBPDZ256rr,   X86::VSUBPDZ256rmb,   TB_BCAST_SD },
+  { X86::VSUBPDZrr,      X86::VSUBPDZrmb,      TB_BCAST_SD },
+  { X86::VSUBPSZ128rr,   X86::VSUBPSZ128rmb,   TB_BCAST_SS },
+  { X86::VSUBPSZ256rr,   X86::VSUBPSZ256rmb,   TB_BCAST_SS },
+  { X86::VSUBPSZrr,      X86::VSUBPSZrmb,      TB_BCAST_SS },
+};
+
 static const X86MemoryFoldTableEntry *
 lookupFoldTableImpl(ArrayRef<X86MemoryFoldTableEntry> Table, unsigned RegOp) {
 #ifndef NDEBUG
   // Make sure the tables are sorted.
   static std::atomic<bool> FoldTablesChecked(false);
   if (!FoldTablesChecked.load(std::memory_order_relaxed)) {
     assert(std::is_sorted(std::begin(MemoryFoldTable2Addr),
                           std::end(MemoryFoldTable2Addr)) &&
assert(std::is_sorted(std::begin(MemoryFoldTable3),
            std::end(MemoryFoldTable3) &&
            "MemoryFoldTable3 is not sorted and unique!");
     assert(std::is_sorted(std::begin(MemoryFoldTable4),
                           std::end(MemoryFoldTable4)) &&
            std::adjacent_find(std::begin(MemoryFoldTable4),
                               std::end(MemoryFoldTable4)) ==
            std::end(MemoryFoldTable4) &&
            "MemoryFoldTable4 is not sorted and unique!");
+    assert(std::is_sorted(std::begin(BroadcastFoldTable2),
+                          std::end(BroadcastFoldTable2)) &&
+           std::adjacent_find(std::begin(BroadcastFoldTable2),
+                              std::end(BroadcastFoldTable2)) ==
+           std::end(BroadcastFoldTable2) &&
+           "BroadcastFoldTable2 is not sorted and unique!");
     FoldTablesChecked.store(true, std::memory_order_relaxed);
   }
 #endif
 
   const X86MemoryFoldTableEntry *Data = llvm::lower_bound(Table, RegOp);
   if (Data != Table.end() && Data->KeyOp == RegOp &&
       !(Data->Flags & TB_NO_FORWARD))
     return Data;
X86MemUnfoldTable() {
     for (const X86MemoryFoldTableEntry &Entry : MemoryFoldTable3)
       // Index 3, folded load
       addTableEntry(Entry, TB_INDEX_3 | TB_FOLDED_LOAD);
 
     for (const X86MemoryFoldTableEntry &Entry : MemoryFoldTable4)
       // Index 4, folded load
       addTableEntry(Entry, TB_INDEX_4 | TB_FOLDED_LOAD);
 
+    // Broadcast tables.
+    for (const X86MemoryFoldTableEntry &Entry : BroadcastFoldTable2)
+      // Index 2, folded broadcast
+      addTableEntry(Entry, TB_INDEX_2 | TB_FOLDED_LOAD | TB_FOLDED_BCAST);
+
     // Sort the memory->reg unfold table.
     array_pod_sort(Table.begin(), Table.end());
 
     // Now that it's sorted, ensure its unique.
     assert(std::adjacent_find(Table.begin(), Table.end()) == Table.end() &&
            "Memory unfolding table is not unique!");
   }
 

llvm/lib/Target/X86/X86InstrInfo.cpp

if (isTwoAddr && NumOps >= 2 && OpNum < 2 && MI.getOperand(0).isReg() &&
     }
 
     I = lookupFoldTable(MI.getOpcode(), OpNum);
   }
 
   if (I != nullptr) {
     unsigned Opcode = I->DstOp;
     unsigned MinAlign = (I->Flags & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT;
     MinAlign = MinAlign ? 1 << (MinAlign - 1) : 0;
     if (Align < MinAlign)
       return nullptr;
     bool NarrowToMOV32rm = false;
     if (Size) {
       const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
       const TargetRegisterClass *RC = getRegClass(MI.getDesc(), OpNum,
                                                   &RI, MF);
       unsigned RCSize = TRI.getRegSizeInBits(*RC) / 8;
if (!MMO->isLoad()) {
       StoreMMOs.push_back(MF.getMachineMemOperand(
           MMO, MMO->getFlags() & ~MachineMemOperand::MOLoad));
     }
   }
 
   return StoreMMOs;
 }
 
+static unsigned getBroadcastOpcode(const X86MemoryFoldTableEntry *I,
+                                   const TargetRegisterClass *RC,
+                                   const X86Subtarget &STI) {
+  assert(STI.hasAVX512() && "Expected at least AVX512!");
+  unsigned SpillSize = STI.getRegisterInfo()->getSpillSize(*RC);
+  assert((SpillSize == 64 || STI.hasVLX()) &&
+         "Can't broadcast less than 64 bytes without AVX512VL!");
+
+  switch (I->Flags & TB_BCAST_MASK) {
+  default: llvm_unreachable("Unexpected broadcast type!");
+  case TB_BCAST_D:
+    switch (SpillSize) {
+    default: llvm_unreachable("Unknown spill size");
+    case 16: return X86::VPBROADCASTDZ128m;
+    case 32: return X86::VPBROADCASTDZ256m;
+    case 64: return X86::VPBROADCASTDZm;
+    }
+    break;
+  case TB_BCAST_Q:
+    switch (SpillSize) {
+    default: llvm_unreachable("Unknown spill size");
+    case 16: return X86::VPBROADCASTQZ128m;
+    case 32: return X86::VPBROADCASTQZ256m;
+    case 64: return X86::VPBROADCASTQZm;
+    }
+    break;
+  case TB_BCAST_SS:
+    switch (SpillSize) {
+    default: llvm_unreachable("Unknown spill size");
+    case 16: return X86::VBROADCASTSSZ128m;
+    case 32: return X86::VBROADCASTSSZ256m;
+    case 64: return X86::VBROADCASTSSZm;
+    }
+    break;
+  case TB_BCAST_SD:
+    switch (SpillSize) {
+    default: llvm_unreachable("Unknown spill size");
+    case 16: return X86::VMOVDDUPZ128rm;
+    case 32: return X86::VBROADCASTSDZ256m;
+    case 64: return X86::VBROADCASTSDZm;
+    }
+    break;
+  }
+}
+
 bool X86InstrInfo::unfoldMemoryOperand(
     MachineFunction &MF, MachineInstr &MI, unsigned Reg, bool UnfoldLoad,
     bool UnfoldStore, SmallVectorImpl<MachineInstr *> &NewMIs) const {
   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(MI.getOpcode());
   if (I == nullptr)
     return false;
   unsigned Opc = I->DstOp;
   unsigned Index = I->Flags & TB_INDEX_MASK;
   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
+  bool FoldedBCast = I->Flags & TB_FOLDED_BCAST;
   if (UnfoldLoad && !FoldedLoad)
     return false;
   UnfoldLoad &= FoldedLoad;
   if (UnfoldStore && !FoldedStore)
     return false;
   UnfoldStore &= FoldedStore;
 
   const MCInstrDesc &MCID = get(Opc);
+
   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
   // TODO: Check if 32-byte or greater accesses are slow too?
   if (!MI.hasOneMemOperand() && RC == &X86::VR128RegClass &&
       Subtarget.isUnalignedMem16Slow())
     // Without memoperands, loadRegFromAddr and storeRegToStackSlot will
     // conservatively assume the address is unaligned. That's bad for
     // performance.
for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
     else if (Op.isReg() && Op.isImplicit())
       ImpOps.push_back(Op);
     else if (i < Index)
       BeforeOps.push_back(Op);
     else if (i > Index)
       AfterOps.push_back(Op);
   }
 
-  // Emit the load instruction.
+  // Emit the load or broadcast instruction.
   if (UnfoldLoad) {
     auto MMOs = extractLoadMMOs(MI.memoperands(), MF);
+
+    unsigned Opc;
+    if (FoldedBCast) {
+      Opc = getBroadcastOpcode(I, RC, Subtarget);
+    } else {
       unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
       bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
-    unsigned Opc = getLoadRegOpcode(Reg, RC, isAligned, Subtarget);
+      Opc = getLoadRegOpcode(Reg, RC, isAligned, Subtarget);
+    }
+
     DebugLoc DL;
     MachineInstrBuilder MIB = BuildMI(MF, DL, get(Opc), Reg);
     for (unsigned i = 0, e = AddrOps.size(); i != e; ++i)
       MIB.add(AddrOps[i]);
     MIB.setMemRefs(MMOs);
     NewMIs.push_back(MIB);
 
     if (UnfoldStore) {
X86InstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
 
   const X86MemoryFoldTableEntry *I = lookupUnfoldTable(N->getMachineOpcode());
   if (I == nullptr)
     return false;
   unsigned Opc = I->DstOp;
   unsigned Index = I->Flags & TB_INDEX_MASK;
   bool FoldedLoad = I->Flags & TB_FOLDED_LOAD;
   bool FoldedStore = I->Flags & TB_FOLDED_STORE;
+  bool FoldedBCast = I->Flags & TB_FOLDED_BCAST;
   const MCInstrDesc &MCID = get(Opc);
   MachineFunction &MF = DAG.getMachineFunction();
   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
   const TargetRegisterClass *RC = getRegClass(MCID, Index, &RI, MF);
   unsigned NumDefs = MCID.NumDefs;
   std::vector<SDValue> AddrOps;
   std::vector<SDValue> BeforeOps;
   std::vector<SDValue> AfterOps;
if (FoldedLoad) {
     EVT VT = *TRI.legalclasstypes_begin(*RC);
     auto MMOs = extractLoadMMOs(cast<MachineSDNode>(N)->memoperands(), MF);
     if (MMOs.empty() && RC == &X86::VR128RegClass &&
         Subtarget.isUnalignedMem16Slow())
       // Do not introduce a slow unaligned load.
       return false;
     // FIXME: If a VR128 can have size 32, we should be checking if a 32-byte
     // memory access is slow above.
+
+    unsigned Opc;
+    if (FoldedBCast) {
+      Opc = getBroadcastOpcode(I, RC, Subtarget);
+    } else {
       unsigned Alignment = std::max<uint32_t>(TRI.getSpillSize(*RC), 16);
       bool isAligned = !MMOs.empty() && MMOs.front()->getAlignment() >= Alignment;
-    Load = DAG.getMachineNode(getLoadRegOpcode(0, RC, isAligned, Subtarget), dl,
-                              VT, MVT::Other, AddrOps);
+      Opc = getLoadRegOpcode(0, RC, isAligned, Subtarget);
+    }
+
+    Load = DAG.getMachineNode(Opc, dl, VT, MVT::Other, AddrOps);
     NewNodes.push_back(Load);
 
     // Preserve memory reference information.
     DAG.setNodeMemRefs(cast<MachineSDNode>(Load), MMOs);
   }
 
   // Emit the data processing instruction.
   std::vector<EVT> VTs;

llvm/test/CodeGen/X86/avx512-broadcast-unfold.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-pc-linux -mattr=avx512vl | FileCheck %s
 
 ; Test that we can unfold constant pool loads when we're using avx512's
 ; ability to fold a broadcast load into an operation.
 
 define void @bcast_unfold_add_v16i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_add_v16i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB0_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpaddd {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpaddd 4096(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB0_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_add_v8i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_add_v8i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [2,2,2,2,2,2,2,2]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB1_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpaddd {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpaddd 4096(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovdqu %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB1_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_add_v4i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_add_v4i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [2,2,2,2]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB2_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vpaddd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovdqu %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpaddd 4096(%rdi,%rax), %xmm0, %xmm1
+; CHECK-NEXT:    vmovdqu %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB2_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 
 bb2:                                              ; preds = %bb2, %bb

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_add_v8i64(i64* %arg) {
 ; CHECK-LABEL: bcast_unfold_add_v8i64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vpbroadcastq {{.*#+}} zmm0 = [2,2,2,2,2,2,2,2]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB3_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpaddq {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vpaddq 8192(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB3_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_add_v4i64(i64* %arg) {
 ; CHECK-LABEL: bcast_unfold_add_v4i64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vpbroadcastq {{.*#+}} ymm0 = [2,2,2,2]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB4_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpaddq {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vpaddq 8192(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovdqu %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB4_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_mul_v16i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_mul_v16i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB6_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpmulld {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpmulld 4096(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB6_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_mul_v8i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_mul_v8i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [3,3,3,3,3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB7_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpmulld {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpmulld 4096(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovdqu %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB7_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_mul_v4i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_mul_v4i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB8_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vpmulld {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovdqu %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpmulld 4096(%rdi,%rax), %xmm0, %xmm1
+; CHECK-NEXT:    vmovdqu %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB8_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 
 bb2:                                              ; preds = %bb2, %bb

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_or_v16i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_or_v16i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB12_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpord {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpord 4096(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB12_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_or_v8i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_or_v8i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [3,3,3,3,3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB13_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpord {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vorps 4096(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB13_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_or_v4i32(i32* %arg) {
 ; CHECK-LABEL: bcast_unfold_or_v4i32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB14_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vpord {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovdqu %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vorps 4096(%rdi,%rax), %xmm0, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB14_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 
 bb2:                                              ; preds = %bb2, %bb

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_or_v8i64(i64* %arg) {
 ; CHECK-LABEL: bcast_unfold_or_v8i64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vpbroadcastq {{.*#+}} zmm0 = [3,3,3,3,3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB15_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vporq {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vporq 8192(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB15_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_or_v4i64(i64* %arg) {
 ; CHECK-LABEL: bcast_unfold_or_v4i64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [3,3,3,3]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB16_1: # %bb2
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vporq {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vorps 8192(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB16_1
 ; CHECK-NEXT:  # %bb.2: # %bb10
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb2
 

 bb10:                                             ; preds = %bb2
   ret void
 }
 
 define void @bcast_unfold_fneg_v16f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fneg_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB18_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpxord {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpxord 4096(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB18_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fneg_v8f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fneg_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB19_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpxord {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vxorps 4096(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB19_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fneg_v4f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fneg_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB20_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vpxord {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovdqu %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vxorps 4096(%rdi,%rax), %xmm0, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB20_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 
 bb1:                                              ; preds = %bb1, %bb

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fneg_v8f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fneg_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vpbroadcastq {{.*#+}} zmm0 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB21_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpxorq {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vpxorq 8192(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB21_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fneg_v4f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fneg_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB22_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpxorq {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vxorps 8192(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB22_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fabs_v16f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fabs_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB24_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpandd {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vpandd 4096(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB24_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 
 ; Function Attrs: nounwind readnone speculatable willreturn
 declare <16 x float> @llvm.fabs.v16f32(<16 x float>) #0
 
 define void @bcast_unfold_fabs_v8f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fabs_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB25_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpandd {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vandps 4096(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB25_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 
 ; Function Attrs: nounwind readnone speculatable willreturn
 declare <8 x float> @llvm.fabs.v8f32(<8 x float>) #0
 
 define void @bcast_unfold_fabs_v4f32(float* %arg) {
 ; CHECK-LABEL: bcast_unfold_fabs_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [NaN,NaN,NaN,NaN]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB26_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vpandd {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovdqu %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vandps 4096(%rdi,%rax), %xmm0, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB26_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 
 bb1:                                              ; preds = %bb1, %bb

 
 ; Function Attrs: nounwind readnone speculatable willreturn
 declare <4 x float> @llvm.fabs.v4f32(<4 x float>) #0
 
 define void @bcast_unfold_fabs_v8f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fabs_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vpbroadcastq {{.*#+}} zmm0 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB27_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu64 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vpandq {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovdqu64 %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vpandq 8192(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovdqu64 %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB27_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 
 ; Function Attrs: nounwind readnone speculatable willreturn
 declare <8 x double> @llvm.fabs.v8f64(<8 x double>) #0
 
 define void @bcast_unfold_fabs_v4f64(double* %arg) {
 ; CHECK-LABEL: bcast_unfold_fabs_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [NaN,NaN,NaN,NaN]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB28_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovdqu 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vpandq {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovdqu %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vandps 8192(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB28_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 
 ; Function Attrs: nounwind readnone speculatable willreturn
 declare <2 x double> @llvm.fabs.v2f64(<2 x double>) #0
 
 define void @bcast_unfold_fadd_v16f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fadd_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} zmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB30_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vaddps {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovups %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vaddps 4096(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovups %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB30_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fadd_v8f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fadd_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB31_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vaddps {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovups %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vaddps 4096(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB31_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fadd_v4f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fadd_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB32_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vaddps {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovups %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vaddps 4096(%rdi,%rax), %xmm0, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB32_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 
 bb1:                                              ; preds = %bb1, %bb

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fadd_v8f64(double* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fadd_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} zmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB33_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vaddpd {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovupd %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vaddpd 8192(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovupd %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB33_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fadd_v4f64(double* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fadd_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB34_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vaddpd {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovupd %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vaddpd 8192(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovupd %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB34_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fmul_v16f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fmul_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} zmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB36_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vmulps {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovups %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmulps 4096(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovups %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB36_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fmul_v8f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fmul_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB37_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vmulps {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovups %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmulps 4096(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB37_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fmul_v4f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fmul_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB38_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vmulps {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovups %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmulps 4096(%rdi,%rax), %xmm0, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB38_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 
 bb1:                                              ; preds = %bb1, %bb

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fmul_v8f64(double* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fmul_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} zmm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB39_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vmulpd {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovupd %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmulpd 8192(%rdi,%rax), %zmm0, %zmm1
+; CHECK-NEXT:    vmovupd %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB39_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fmul_v4f64(double* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fmul_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [3.0E+0,3.0E+0,3.0E+0,3.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB40_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vmulpd {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovupd %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmulpd 8192(%rdi,%rax), %ymm0, %ymm1
+; CHECK-NEXT:    vmovupd %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB40_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fdiv_v16f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fdiv_v16f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} zmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB42_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vdivps {{.*}}(%rip){1to16}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovups %zmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %zmm1
+; CHECK-NEXT:    vdivps %zmm0, %zmm1, %zmm1
+; CHECK-NEXT:    vmovups %zmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB42_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fdiv_v8f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fdiv_v8f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB43_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vdivps {{.*}}(%rip){1to8}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovups %ymm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %ymm1
+; CHECK-NEXT:    vdivps %ymm0, %ymm1, %ymm1
+; CHECK-NEXT:    vmovups %ymm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB43_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fdiv_v4f32(float* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fdiv_v4f32:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-4096, %rax # imm = 0xF000
+; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB44_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm0
-; CHECK-NEXT:    vdivps {{.*}}(%rip){1to4}, %xmm0, %xmm0
-; CHECK-NEXT:    vmovups %xmm0, 4096(%rdi,%rax)
+; CHECK-NEXT:    vmovups 4096(%rdi,%rax), %xmm1
+; CHECK-NEXT:    vdivps %xmm0, %xmm1, %xmm1
+; CHECK-NEXT:    vmovups %xmm1, 4096(%rdi,%rax)
 ; CHECK-NEXT:    addq $16, %rax
 ; CHECK-NEXT:    jne .LBB44_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 
 bb1:                                              ; preds = %bb1, %bb

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fdiv_v8f64(double* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fdiv_v8f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} zmm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB45_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm0
-; CHECK-NEXT:    vdivpd {{.*}}(%rip){1to8}, %zmm0, %zmm0
-; CHECK-NEXT:    vmovupd %zmm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %zmm1
+; CHECK-NEXT:    vdivpd %zmm0, %zmm1, %zmm1
+; CHECK-NEXT:    vmovupd %zmm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $64, %rax
 ; CHECK-NEXT:    jne .LBB45_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1
 

 bb9:                                              ; preds = %bb1
   ret void
 }
 
 define void @bcast_unfold_fdiv_v4f64(double* nocapture %arg) {
 ; CHECK-LABEL: bcast_unfold_fdiv_v4f64:
 ; CHECK:       # %bb.0: # %bb
 ; CHECK-NEXT:    movq $-8192, %rax # imm = 0xE000
+; CHECK-NEXT:    vbroadcastsd {{.*#+}} ymm0 = [2.0E+0,2.0E+0,2.0E+0,2.0E+0]
 ; CHECK-NEXT:    .p2align 4, 0x90
 ; CHECK-NEXT:  .LBB46_1: # %bb1
 ; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm0
-; CHECK-NEXT:    vdivpd {{.*}}(%rip){1to4}, %ymm0, %ymm0
-; CHECK-NEXT:    vmovupd %ymm0, 8192(%rdi,%rax)
+; CHECK-NEXT:    vmovupd 8192(%rdi,%rax), %ymm1
+; CHECK-NEXT:    vdivpd %ymm0, %ymm1, %ymm1
+; CHECK-NEXT:    vmovupd %ymm1, 8192(%rdi,%rax)
 ; CHECK-NEXT:    addq $32, %rax
 ; CHECK-NEXT:    jne .LBB46_1
 ; CHECK-NEXT:  # %bb.2: # %bb9
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 bb:
   br label %bb1