[ARM][RegAlloc] Add t2LoopEndDec

[dmgreen]

Summary

We currently have problems with the way that low overhead loops are specified, with LR sometimes being spilled between the t2LoopDec and the t2LoopEnd forcing the entire loop to be reverted late in the backend. As they will eventually become a single instruction, this patch introduces a t2LoopEndDec which is the combination of the two, combined before registry allocation to make sure this does not fail.

Unfortunately this instruction is a terminator that produces a value (and also branches - it only produces the value around the branching edge). So this needs some adjustment to phi elimination and the register allocator to make sure that we do not spill this LR def around the loop (needing to put a spill after the terminator). We treat the loop very carefully, making sure that there is nothing else like calls that would break it's ability to use LR. For that, this adds a isUnspillableTerminator to opt in the new behaviour.

This might obviously be a little contentious, so I would like to get opinions from people who know what they are talking about. There is a chance that this could cause problems, and so I have added an escape option incase. But I have not seen any problems in the testing that I've tried, and not reverting Low overhead loops is important for our performance. If this does work then we can hopefully do the same for t2WhileLoopStart and t2DoLoopStart instructions.

This patch also contains the code needed to convert or revert the t2LoopEndDec in the backend (which just needs a subs; bne) and the code pre-ra to create them.

[SjoerdMeijer]

This might obviously be a little contentious, so I would like to get opinions from people who know what they are talking about.

The ARM changes look very reasonable to me. To be a bit more precise, yes, this is definitely something we want, so agreed on the motivation and the (ARM) implementation, showing nice codegen changes.

But value producing terminators in regalloc is not my forte, so indeed best if someone approves that part.

Find inline some/mostly nits.

llvm/include/llvm/CodeGen/TargetInstrInfo.h

351

typo: trerminator

llvm/lib/CodeGen/CalcSpillWeights.cpp

227

Nit: guess you don't really need to check isTerminator as that will be checked in isUnspillableTerminator?

llvm/lib/Target/ARM/ARMInstrThumb2.td

5451

bikeshedding names: how about t2LoopDecEnd to more keep the semantics this is a decrement + end, in that order?

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

1302–1308

It wasn't immediately clear to me why we need this now.

1365

nit: perhaps an assert that MI is a loop end?

1632

nit: how about RevertLoopEnd(LoLoop.End, RevertLoopDec(LoLoop.Dec)); to get rid of the curly brackets around the if-else here?

llvm/lib/Target/ARM/MVETPAndVPTOptimisationsPass.cpp

39

nit: arm-enable-mergeenddec is a bit difficult to read. Since it is relative long already, how about making it even a bit longer, something like: arm-enable-merge-loopenddec?

40

typo? mergeing?

243–249

nit:

Replace the loop dec into the loop end as a single instruction

->

Replace the loop dec and loop end as a single instruction

or something similar...

[dmgreen]

Thanks for taking a look.

llvm/lib/CodeGen/CalcSpillWeights.cpp

227

I was trying to make it explicit, that it needed to be a terminator that defined the value and this shouldn't be used for anything else.

I can change that if you think it's better to remove it. It is currently just trying to be conservative.

llvm/lib/Target/ARM/ARMInstrThumb2.td

5451

Yeah, It does both together. I have no strong opinion on the order in the name. I guess I chose t2LoopEndDec as I felt that the major thing it did was behave as a loop end (like an LE instruction). The Dec somewhat of a side effect of that.

Can change it if you think it's better the other way. Let me know.

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

1302–1308

I updated the comment.

llvm/lib/Target/ARM/MVETPAndVPTOptimisationsPass.cpp

39

Thanks. This did feel like a particularly ugly option name.

[SjoerdMeijer]

Adding @mtrofin, @bjope, @arsenm for the unspillable terminator change.

llvm/lib/CodeGen/CalcSpillWeights.cpp

227

Not a big deal, whatever you prefer.

llvm/lib/Target/ARM/ARMInstrThumb2.td

5451

No strong opinion on this, I was bikeshedding anyway.

[mtrofin]

llvm/lib/CodeGen/CalcSpillWeights.cpp

227

from a readability perspective, isUnspillableTerminator suggests "isTerminator && isUnspillable".

How about having isUnspillableTerminator not virtual and doing this:

return isTerminator() && isUnspillableTerminatorImpl()

where isUnspillableTerminatorImpl is protected and virtual?

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

1375

Nit: addReg(MCRegister::NoRegister) is probably more readable.

[bjope]

llvm/lib/CodeGen/PHIElimination.cpp

452

This is only valid if there are no other uses of SrcReg (no anywhere else, and not in the current MBB that we are eliminating PHI nodes inside, such as another PHI node in MBB).

Haven't analyzed it more closely than that. But I don't know if it can be generally guaranteed that the def in an UnspillableTerminator only has a single use. Is such a rule guaranteed here (or even feasible)?

[dmgreen]

Thanks for the comments. Added a isUnspillableTerminatorImpl.

llvm/lib/CodeGen/CalcSpillWeights.cpp

227

Sure, sounds good to me.

llvm/lib/CodeGen/PHIElimination.cpp

452

Yeah. We are being very careful! We are only using these for hardware loop instructions that we control the insertion of and are careful to only do the transform to produce the instruction (t2LoopEndDec in this case) when it has a single remaining user. CheckUsers in MVETPAndVPTOptimisations::MergeLoopEnd does that.

That's why this is very opt-in and it's trying to be careful to say "Use with care".

[bjope]

llvm/lib/CodeGen/PHIElimination.cpp

452

My concern was that this is not inside the target specific part of the backend. So if some other target wants to use the concept of "unspillable terminators" then there must be some clear rules what to expect.

I think that having an assert that there is only one use would be appropriate here (to at least catch any violations to such a rule at this point where things would go wrong otherwise).

But maybe the MachineVerifier should help out detect problems if an unspillable terminator def has more than one use as well (to catch such problems already in a pass breaking the rule)?

And maybe we need to document what is expected related to "unspillable terminators" related to this restriction somewhere (not sure exactly where).

I guess there could be lots of generic passes, but also target specific passes, that must adhere to such a rule if it is allowed to insert an unspillable terminator already at ISel. So I still don't see how to ensure that all passes between ISel and PHI-elmination are aware about that they aren't allowed to introduce new uses of the unspillable terminator def? Or how close to PHI-elimination must a pass introducing "unspillable terminators" be?

[dmgreen]

Add a machine verifier check and an extra assert.

llvm/lib/CodeGen/PHIElimination.cpp

452

I agree. I have added a check to the verifier and an assert in here. In our case we are adding these for hardware loop instructions soon before registry allocation, after all the other mir level optimization have happened. There's also some details about the instruction in isUnspillableTerminatorImpl.

[dmgreen]

Ping

[dmgreen]

Rebase

[bjope]

llvm/lib/CodeGen/MachineVerifier.cpp

1553 ↗

(On Diff #309942)

typ: /Checl/Check/

1559 ↗

(On Diff #309942)

Use lists in MRI also contain dbg-uses. So I think we want to use use_instr_nodbg here to only count non-dbg-uses.

Also a bit unclear if zero uses should be valid or not (code is currently checking "at most one user", but code comments and the report says "should have a single user"). There is a hasOneDBGUser helper in MRI that could be used if there has to be one user.

[dmgreen]

Thanks. I think hasOneDBGUse would have probably worked in our use case, but checking use_instr_nodbg is a little more general.

[bjope]

@SjoerdMeijer : I've reviewed the non-target-specific changes now. And that part looks good to me now.

I don't know if the target-specific parts already has been reviewed (or if anything has changed lately). Can someone with ARM/Thumb2 knowledge can acknowledge that part and set "accept revision".

[SjoerdMeijer]

In D91358#2442675, @bjope wrote:

@SjoerdMeijer : I've reviewed the non-target-specific changes now. And that part looks good to me now.

I don't know if the target-specific parts already has been reviewed (or if anything has changed lately). Can someone with ARM/Thumb2 knowledge can acknowledge that part and set "accept revision".

Thanks for helping and reviewing!

I've commented on the ARM specific part, and that has been addressed. With your comments addressed too, this then LGTM.

[dmgreen]

Thanks folks.

llvm/include/llvm/CodeGen/TargetInstrInfo.h

public:
   /// subregisters has no single offset.
   ///
   /// Targets with nontrivial bigendian implementations may need to override
   /// this, particularly to support spilled vector registers.
   virtual bool getStackSlotRange(const TargetRegisterClass *RC, unsigned SubIdx,
                                  unsigned &Size, unsigned &Offset,
                                  const MachineFunction &MF) const;
 
+  /// Return true if the given instruction is terminator that is unspillable,
+  /// according to isUnspillableTerminatorImpl.
+  bool isUnspillableTerminator(const MachineInstr *MI) const {
+    return MI->isTerminator() && isUnspillableTerminatorImpl(MI);
+  }
+
   /// Returns the size in bytes of the specified MachineInstr, or ~0U
   /// when this function is not implemented by a target.
   virtual unsigned getInstSizeInBytes(const MachineInstr &MI) const {
     return ~0U;
   }
 
   /// Return true if the instruction is as cheap as a move instruction.
   ///
protected:
   /// If the specific machine instruction is a instruction that moves/copies
   /// value from one register to another register return destination and source
   /// registers as machine operands.
   virtual Optional<DestSourcePair>
   isCopyInstrImpl(const MachineInstr &MI) const {
     return None;
   }
 
+  /// Return true if the given terminator MI is not expected to spill. This
+  /// sets the live interval as not spillable and adjusts phi node lowering to
+  /// not introduce copies after the terminator. Use with care, as it may run
+  /// out of registers if not careful and should only have single phi users.
+  virtual bool isUnspillableTerminatorImpl(const MachineInstr *MI) const {
+    return false;
+  }
+
 public:
   /// If the specific machine instruction is a instruction that moves/copies
   /// value from one register to another register return destination and source
   /// registers as machine operands.
   /// For COPY-instruction the method naturally returns destination and source
   /// registers as machine operands, for all other instructions the method calls
   /// target-dependent implementation.
   Optional<DestSourcePair> isCopyInstr(const MachineInstr &MI) const {

llvm/lib/CodeGen/CalcSpillWeights.cpp

float VirtRegAuxInfo::futureWeight(LiveInterval &LI, SlotIndex Start,
                                    SlotIndex End) {
   return weightCalcHelper(LI, &Start, &End);
 }
 
 float VirtRegAuxInfo::weightCalcHelper(LiveInterval &LI, SlotIndex *Start,
                                        SlotIndex *End) {
   MachineRegisterInfo &MRI = MF.getRegInfo();
   const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
+  const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
   MachineBasicBlock *MBB = nullptr;
   MachineLoop *Loop = nullptr;
   bool IsExiting = false;
   float TotalWeight = 0;
   unsigned NumInstr = 0; // Number of instructions using LI
   SmallPtrSet<MachineInstr *, 8> Visited;
 
   std::pair<Register, Register> TargetHint = MRI.getRegAllocationHint(LI.reg());
if (IsLocalSplitArtifact && ((SI < *Start) || (SI > *End)))
       continue;
 
     NumInstr++;
     if (MI->isIdentityCopy() || MI->isImplicitDef())
       continue;
     if (!Visited.insert(MI).second)
       continue;
 
+    // For terminators that produce values, ask the backend if the register is
+    // not spillable.
+    if (TII.isUnspillableTerminator(MI) && MI->definesRegister(LI.reg())) {
+      LI.markNotSpillable();
+      return -1.0f;
+    }
+
     float Weight = 1.0f;
     if (IsSpillable) {
       // Get loop info for mi.
       if (MI->getParent() != MBB) {
         MBB = MI->getParent();
         Loop = Loops.getLoopFor(MBB);
         IsExiting = Loop ? Loop->isLoopExiting(MBB) : false;
       }

llvm/lib/CodeGen/PHIElimination.cpp

for (int i = NumSrcs - 1; i >= 0; --i) {
     MachineBasicBlock &opBlock = *MPhi->getOperand(i*2+2).getMBB();
 
     // Check to make sure we haven't already emitted the copy for this block.
     // This can happen because PHI nodes may have multiple entries for the same
     // basic block.
     if (!MBBsInsertedInto.insert(&opBlock).second)
       continue;  // If the copy has already been emitted, we're done.
 
+    MachineInstr *SrcRegDef = MRI->getVRegDef(SrcReg);
+    if (SrcRegDef && TII->isUnspillableTerminator(SrcRegDef)) {
+      assert(SrcRegDef->getOperand(0).isReg() &&
+             SrcRegDef->getOperand(0).isDef() &&
+             "Expected operand 0 to be a reg def!");
+      SrcRegDef->getOperand(0).setReg(IncomingReg);
+      continue;
+    }
+
     // Find a safe location to insert the copy, this may be the first terminator
     // in the block (or end()).
     MachineBasicBlock::iterator InsertPos =
       findPHICopyInsertPoint(&opBlock, &MBB, SrcReg);
 
     // Insert the copy.
     MachineInstr *NewSrcInstr = nullptr;
     if (!reusedIncoming && IncomingReg) {

llvm/lib/Target/ARM/ARMBaseInstrInfo.h

public:
   MachineBasicBlock::iterator
   insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
                      MachineBasicBlock::iterator &It, MachineFunction &MF,
                      const outliner::Candidate &C) const override;
 
   /// Enable outlining by default at -Oz.
   bool shouldOutlineFromFunctionByDefault(MachineFunction &MF) const override;
 
+  bool isUnspillableTerminatorImpl(const MachineInstr *MI) const override {
+    return MI->getOpcode() == ARM::t2LoopEndDec;
+  }
+
 private:
   /// Returns an unused general-purpose register which can be used for
   /// constructing an outlined call if one exists. Returns 0 otherwise.
   unsigned findRegisterToSaveLRTo(const outliner::Candidate &C) const;
 
   // Adds an instruction which saves the link register on top of the stack into
   /// the MachineBasicBlock \p MBB at position \p It.
   void saveLROnStack(MachineBasicBlock &MBB,

llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp

if (Opc == ARM::tPICADD || Opc == ARM::PICADD || Opc == ARM::PICSTR ||
       Opc == ARM::PICLDRSH || Opc == ARM::t2LDRpci_pic ||
       Opc == ARM::t2MOVi16_ga_pcrel || Opc == ARM::t2MOVTi16_ga_pcrel ||
       Opc == ARM::t2MOV_ga_pcrel)
     return outliner::InstrType::Illegal;
 
   // Be conservative with ARMv8.1 MVE instructions.
   if (Opc == ARM::t2BF_LabelPseudo || Opc == ARM::t2DoLoopStart ||
       Opc == ARM::t2DoLoopStartTP || Opc == ARM::t2WhileLoopStart ||
-      Opc == ARM::t2LoopDec || Opc == ARM::t2LoopEnd)
+      Opc == ARM::t2LoopDec || Opc == ARM::t2LoopEnd ||
+      Opc == ARM::t2LoopEndDec)
     return outliner::InstrType::Illegal;
 
   const MCInstrDesc &MCID = MI.getDesc();
   uint64_t MIFlags = MCID.TSFlags;
   if ((MIFlags & ARMII::DomainMask) == ARMII::DomainMVE)
     return outliner::InstrType::Illegal;
 
   // Is this a terminator for a basic block?

llvm/lib/Target/ARM/ARMInstrThumb2.td

t2PseudoInst<(outs),
                  (ins rGPR:$elts, brtarget:$target),
                  8, IIC_Br, []>,
                  Sched<[WriteBr]>;
 
 def t2LoopEnd :
   t2PseudoInst<(outs), (ins GPRlr:$elts, brtarget:$target),
   8, IIC_Br, []>, Sched<[WriteBr]>;
 
+def t2LoopEndDec :
+  t2PseudoInst<(outs GPRlr:$Rm), (ins GPRlr:$elts, brtarget:$target),
+  8, IIC_Br, []>, Sched<[WriteBr]>;
+
 } // end isBranch, isTerminator, hasSideEffects
 
 }
 
 } // end isNotDuplicable
 
 class CS<string iname, bits<4> opcode, list<dag> pattern=[]>
   : V8_1MI<(outs rGPR:$Rd), (ins GPRwithZRnosp:$Rn, GPRwithZRnosp:$Rm, pred_noal:$fcond),

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

private:
 
     void RevertWhile(MachineInstr *MI) const;
     void RevertDo(MachineInstr *MI) const;
 
     bool RevertLoopDec(MachineInstr *MI) const;
 
     void RevertLoopEnd(MachineInstr *MI, bool SkipCmp = false) const;
 
+    void RevertLoopEndDec(MachineInstr *MI) const;
+
     void ConvertVPTBlocks(LowOverheadLoop &LoLoop);
 
     MachineInstr *ExpandLoopStart(LowOverheadLoop &LoLoop);
 
     void Expand(LowOverheadLoop &LoLoop);
 
     void IterationCountDCE(LowOverheadLoop &LoLoop);
   };

 void LowOverheadLoop::Validate(ARMBasicBlockUtils *BBUtils) {
   if (Revert)
     return;
 
   // Check branch target ranges: WLS[TP] can only branch forwards and LE[TP]
   // can only jump back.
   auto ValidateRanges = [](MachineInstr *Start, MachineInstr *End,
                            ARMBasicBlockUtils *BBUtils, MachineLoop &ML) {
-    assert(End->getOperand(1).isMBB() &&
-           "Expected LoopEnd to target basic block!");
-
+    MachineBasicBlock *TgtBB = End->getOpcode() == ARM::t2LoopEnd
+                                   ? End->getOperand(1).getMBB()
+                                   : End->getOperand(2).getMBB();
     // TODO Maybe there's cases where the target doesn't have to be the header,
     // but for now be safe and revert.
-    if (End->getOperand(1).getMBB() != ML.getHeader()) {
+    if (TgtBB != ML.getHeader()) {
       LLVM_DEBUG(dbgs() << "ARM Loops: LoopEnd is not targeting header.\n");
       return false;
     }
 
     // The WLS and LE instructions have 12-bits for the label offset. WLS
     // requires a positive offset, while LE uses negative.
     if (BBUtils->getOffsetOf(End) < BBUtils->getOffsetOf(ML.getHeader()) ||
         !BBUtils->isBBInRange(End, ML.getHeader(), 4094)) {
bool ARMLowOverheadLoops::ProcessLoop(MachineLoop *ML) {
   for (auto *MBB : reverse(ML->getBlocks())) {
     for (auto &MI : *MBB) {
       if (MI.isDebugValue())
         continue;
       else if (MI.getOpcode() == ARM::t2LoopDec)
         LoLoop.Dec = &MI;
       else if (MI.getOpcode() == ARM::t2LoopEnd)
         LoLoop.End = &MI;
+      else if (MI.getOpcode() == ARM::t2LoopEndDec)
+        LoLoop.End = LoLoop.Dec = &MI;
       else if (isLoopStart(MI))
         LoLoop.Start = &MI;
       else if (MI.getDesc().isCall()) {
         // TODO: Though the call will require LE to execute again, does this
         // mean we should revert? Always executing LE hopefully should be
         // faster than performing a sub,cmp,br or even subs,br.
         LoLoop.Revert = true;
         LLVM_DEBUG(dbgs() << "ARM Loops: Found call.\n");
       } else {
         // Record VPR defs and build up their corresponding vpt blocks.
         // Check we know how to tail predicate any mve instructions.
         LoLoop.AnalyseMVEInst(&MI);
       }
     }
   }
 
   LLVM_DEBUG(LoLoop.dump());
   if (!LoLoop.FoundAllComponents()) {
     LLVM_DEBUG(dbgs() << "ARM Loops: Didn't find loop start, update, end\n");
     return false;
   }
 
-  // Check that the only instruction using LoopDec is LoopEnd.
+  // Check that the only instruction using LoopDec is LoopEnd. This can only
+  // happen when the Dec and End are separate, not a single t2LoopEndDec.
   // TODO: Check for copy chains that really have no effect.
-  SmallPtrSet<MachineInstr*, 2> Uses;
+  if (LoLoop.Dec != LoLoop.End) {
+    SmallPtrSet<MachineInstr *, 2> Uses;
     RDA->getReachingLocalUses(LoLoop.Dec, MCRegister::from(ARM::LR), Uses);
     if (Uses.size() > 1 || !Uses.count(LoLoop.End)) {
       LLVM_DEBUG(dbgs() << "ARM Loops: Unable to remove LoopDec.\n");
       LoLoop.Revert = true;
     }
+  }
   LoLoop.Validate(BBUtils.get());
   Expand(LoLoop);
   return true;
 }
 
 // WhileLoopStart holds the exit block, so produce a cmp lr, 0 and then a
 // beq that branches to the exit branch.
 // TODO: We could also try to generate a cbz if the value in LR is also in
void ARMLowOverheadLoops::RevertLoopEnd(MachineInstr *MI, bool SkipCmp) const {
 
   MachineBasicBlock *DestBB = MI->getOperand(1).getMBB();
   unsigned BrOpc = BBUtils->isBBInRange(MI, DestBB, 254) ?
     ARM::tBcc : ARM::t2Bcc;
 
   llvm::RevertLoopEnd(MI, TII, BrOpc, SkipCmp);
 }
 
+// Generate a subs, or sub and cmp, and a branch instead of an LE.
+void ARMLowOverheadLoops::RevertLoopEndDec(MachineInstr *MI) const {
+  LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to subs, br: " << *MI);
+  assert(MI->getOpcode() == ARM::t2LoopEndDec && "Expected a t2LoopEndDec!");
+  MachineBasicBlock *MBB = MI->getParent();
+
+  MachineInstrBuilder MIB =
+      BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::t2SUBri));
+  MIB.addDef(ARM::LR);
+  MIB.add(MI->getOperand(1));
+  MIB.addImm(1);
+  MIB.addImm(ARMCC::AL);
+  MIB.addReg(ARM::NoRegister);
+  MIB.addReg(ARM::CPSR);
+  MIB->getOperand(5).setIsDef(true);
+
+  MachineBasicBlock *DestBB = MI->getOperand(2).getMBB();
+  unsigned BrOpc =
+      BBUtils->isBBInRange(MI, DestBB, 254) ? ARM::tBcc : ARM::t2Bcc;
+
+  // Create bne
+  MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(BrOpc));
+  MIB.add(MI->getOperand(2)); // branch target
+  MIB.addImm(ARMCC::NE);      // condition code
+  MIB.addReg(ARM::CPSR);
+
+  MI->eraseFromParent();
+}
+
 // Perform dead code elimation on the loop iteration count setup expression.
 // If we are tail-predicating, the number of elements to be processed is the
 // operand of the VCTP instruction in the vector body, see getCount(), which is
 // register $r3 in this example:
 //
 //   $lr = big-itercount-expression
 //   ..
 //   $lr = t2DoLoopStart renamable $lr
void ARMLowOverheadLoops::Expand(LowOverheadLoop &LoLoop) {
   auto ExpandLoopEnd = [this](LowOverheadLoop &LoLoop) {
     MachineInstr *End = LoLoop.End;
     MachineBasicBlock *MBB = End->getParent();
     unsigned Opc = LoLoop.IsTailPredicationLegal() ?
       ARM::MVE_LETP : ARM::t2LEUpdate;
     MachineInstrBuilder MIB = BuildMI(*MBB, End, End->getDebugLoc(),
                                       TII->get(Opc));
     MIB.addDef(ARM::LR);
-    MIB.add(End->getOperand(0));
-    MIB.add(End->getOperand(1));
+    unsigned Off = LoLoop.Dec == LoLoop.End ? 1 : 0;
+    MIB.add(End->getOperand(Off + 0));
+    MIB.add(End->getOperand(Off + 1));
     LLVM_DEBUG(dbgs() << "ARM Loops: Inserted LE: " << *MIB);
     LoLoop.ToRemove.insert(LoLoop.Dec);
     LoLoop.ToRemove.insert(End);
     return &*MIB;
   };
 
   // TODO: We should be able to automatically remove these branches before we
   // get here - probably by teaching analyzeBranch about the pseudo
auto RemoveDeadBranch = [](MachineInstr *I) {
     }
   };
 
   if (LoLoop.Revert) {
     if (LoLoop.Start->getOpcode() == ARM::t2WhileLoopStart)
       RevertWhile(LoLoop.Start);
     else
       RevertDo(LoLoop.Start);
-    bool FlagsAlreadySet = RevertLoopDec(LoLoop.Dec);
-    RevertLoopEnd(LoLoop.End, FlagsAlreadySet);
+    if (LoLoop.Dec == LoLoop.End)
+      RevertLoopEndDec(LoLoop.End);
+    else
+      RevertLoopEnd(LoLoop.End, RevertLoopDec(LoLoop.Dec));
   } else {
     LoLoop.Start = ExpandLoopStart(LoLoop);
     RemoveDeadBranch(LoLoop.Start);
     LoLoop.End = ExpandLoopEnd(LoLoop);
     RemoveDeadBranch(LoLoop.End);
     if (LoLoop.IsTailPredicationLegal())
       ConvertVPTBlocks(LoLoop);
     for (auto *I : LoLoop.ToRemove) {
       LLVM_DEBUG(dbgs() << "ARM Loops: Erasing " << *I);

 bool ARMLowOverheadLoops::RevertNonLoops() {
   LLVM_DEBUG(dbgs() << "ARM Loops: Reverting any remaining pseudos...\n");
   bool Changed = false;
 
   for (auto &MBB : *MF) {
     SmallVector<MachineInstr*, 4> Starts;
     SmallVector<MachineInstr*, 4> Decs;
     SmallVector<MachineInstr*, 4> Ends;
+    SmallVector<MachineInstr *, 4> EndDecs;
 
     for (auto &I : MBB) {
       if (isLoopStart(I))
         Starts.push_back(&I);
       else if (I.getOpcode() == ARM::t2LoopDec)
         Decs.push_back(&I);
       else if (I.getOpcode() == ARM::t2LoopEnd)
         Ends.push_back(&I);
+      else if (I.getOpcode() == ARM::t2LoopEndDec)
+        EndDecs.push_back(&I);
     }
 
-    if (Starts.empty() && Decs.empty() && Ends.empty())
+    if (Starts.empty() && Decs.empty() && Ends.empty() && EndDecs.empty())
       continue;
 
     Changed = true;
 
     for (auto *Start : Starts) {
       if (Start->getOpcode() == ARM::t2WhileLoopStart)
         RevertWhile(Start);
       else
         RevertDo(Start);
     }
     for (auto *Dec : Decs)
       RevertLoopDec(Dec);
 
     for (auto *End : Ends)
       RevertLoopEnd(End);
+    for (auto *End : EndDecs)
+      RevertLoopEndDec(End);
   }
   return Changed;
 }
 
 FunctionPass *llvm::createARMLowOverheadLoopsPass() {
   return new ARMLowOverheadLoops();
 }

llvm/lib/Target/ARM/MVETPAndVPTOptimisationsPass.cpp

 #include "llvm/InitializePasses.h"
 #include "llvm/Support/Debug.h"
 #include <cassert>
 
 using namespace llvm;
 
 #define DEBUG_TYPE "arm-mve-vpt-opts"
 
+static cl::opt<bool>
+MergeEndDec("arm-enable-merge-loopenddec", cl::Hidden,
+    cl::desc("Enable merging Loop End and Dec instructions."),
+    cl::init(true));
+
 namespace {
 class MVETPAndVPTOptimisations : public MachineFunctionPass {
 public:
   static char ID;
   const Thumb2InstrInfo *TII;
   MachineRegisterInfo *MRI;
 
   MVETPAndVPTOptimisations() : MachineFunctionPass(ID) {}
static bool findLoopComponents(MachineLoop *ML, MachineRegisterInfo *MRI,
 
   // Find the loop end from the terminators.
   LoopEnd = nullptr;
   for (auto &T : Latch->terminators()) {
     if (T.getOpcode() == ARM::t2LoopEnd && T.getOperand(1).getMBB() == Header) {
       LoopEnd = &T;
       break;
     }
+    if (T.getOpcode() == ARM::t2LoopEndDec &&
+        T.getOperand(2).getMBB() == Header) {
+      LoopEnd = &T;
+      break;
+    }
   }
   if (!LoopEnd) {
     LLVM_DEBUG(dbgs() << "  no LoopEnd\n");
     return false;
   }
   LLVM_DEBUG(dbgs() << "  found loop end: " << *LoopEnd);
 
   // Find the dec from the use of the end. There may be copies between
   // instructions. We expect the loop to loop like:
   //   $vs = t2DoLoopStart ...
   // loop:
   //   $vp = phi [ $vs ], [ $vd ]
   //   ...
   //   $vd = t2LoopDec $vp
   //   ...
   //   t2LoopEnd $vd, loop
+  if (LoopEnd->getOpcode() == ARM::t2LoopEndDec)
+    LoopDec = LoopEnd;
+  else {
     LoopDec =
         LookThroughCOPY(MRI->getVRegDef(LoopEnd->getOperand(0).getReg()), MRI);
     if (!LoopDec || LoopDec->getOpcode() != ARM::t2LoopDec) {
       LLVM_DEBUG(dbgs() << "  didn't find LoopDec where we expected!\n");
       return false;
     }
+  }
   LLVM_DEBUG(dbgs() << "  found loop dec: " << *LoopDec);
 
   LoopPhi =
       LookThroughCOPY(MRI->getVRegDef(LoopDec->getOperand(1).getReg()), MRI);
   if (!LoopPhi || LoopPhi->getOpcode() != TargetOpcode::PHI ||
       LoopPhi->getNumOperands() != 5 ||
       (LoopPhi->getOperand(2).getMBB() != Latch &&
        LoopPhi->getOperand(4).getMBB() != Latch)) {
if (!LoopStart || LoopStart->getOpcode() != ARM::t2DoLoopStart) {
     return false;
   }
   LLVM_DEBUG(dbgs() << "  found loop start: " << *LoopStart);
 
   return true;
 }
 
 bool MVETPAndVPTOptimisations::MergeLoopEnd(MachineLoop *ML) {
+  if (!MergeEndDec)
+    return false;
+
   LLVM_DEBUG(dbgs() << "MergeLoopEnd on loop " << ML->getHeader()->getName()
                     << "\n");
 
   MachineInstr *LoopEnd, *LoopPhi, *LoopStart, *LoopDec;
   if (!findLoopComponents(ML, MRI, LoopStart, LoopPhi, LoopDec, LoopEnd))
     return false;
 
   // Check if there is an illegal instruction (a call) in the low overhead loop
bool MVETPAndVPTOptimisations::MergeLoopEnd(MachineLoop *ML) {
   if (LoopPhi->getOperand(2).getMBB() == ML->getLoopLatch()) {
     LoopPhi->getOperand(3).setReg(StartReg);
     LoopPhi->getOperand(1).setReg(DecReg);
   } else {
     LoopPhi->getOperand(1).setReg(StartReg);
     LoopPhi->getOperand(3).setReg(DecReg);
   }
 
-  LoopDec->getOperand(1).setReg(PhiReg);
-  LoopEnd->getOperand(0).setReg(DecReg);
+  // Replace the loop dec and loop end as a single instruction.
+  MachineInstrBuilder MI =
+      BuildMI(*LoopEnd->getParent(), *LoopEnd, LoopEnd->getDebugLoc(),
+              TII->get(ARM::t2LoopEndDec), DecReg)
+          .addReg(PhiReg)
+          .add(LoopEnd->getOperand(1));
+  LLVM_DEBUG(dbgs() << "Merged LoopDec and End into: " << *MI.getInstr());
 
+  LoopDec->eraseFromParent();
+  LoopEnd->eraseFromParent();
   for (auto *MI : Copies)
     MI->eraseFromParent();
   return true;
 }
 
 // Convert t2DoLoopStart to t2DoLoopStartTP if the loop contains VCTP
 // instructions. This keeps the VCTP count reg operand on the t2DoLoopStartTP
 // instruction, making the backend ARMLowOverheadLoops passes job of finding the
 // VCTP operand much simpler.
 bool MVETPAndVPTOptimisations::ConvertTailPredLoop(MachineLoop *ML,
                                               MachineDominatorTree *DT) {
   LLVM_DEBUG(dbgs() << "ConvertTailPredLoop on loop "
                     << ML->getHeader()->getName() << "\n");
 
   // Find some loop components including the LoopEnd/Dec/Start, and any VCTP's
   // in the loop.
   MachineInstr *LoopEnd, *LoopPhi, *LoopStart, *LoopDec;
   if (!findLoopComponents(ML, MRI, LoopStart, LoopPhi, LoopDec, LoopEnd))
     return false;
+  if (LoopDec != LoopEnd)
+    return false;
 
   SmallVector<MachineInstr *, 4> VCTPs;
   for (MachineBasicBlock *BB : ML->blocks())
     for (MachineInstr &MI : *BB)
       if (isVCTP(&MI))
         VCTPs.push_back(&MI);
 
   if (VCTPs.empty()) {

llvm/test/CodeGen/Thumb2/LowOverheadLoops/count_dominates_start.mir

body:             |
   ; CHECK:   [[t2SUBri1:%[0-9]+]]:rgpr = t2SUBri [[PHI4]], 8, 14 /* CC::al */, $noreg, $noreg
   ; CHECK:   [[COPY7:%[0-9]+]]:gpr = COPY [[t2SUBri1]]
   ; CHECK:   [[MVE_VLDRHU16_post:%[0-9]+]]:rgpr, [[MVE_VLDRHU16_post1:%[0-9]+]]:mqpr = MVE_VLDRHU16_post [[PHI]], 16, 1, [[MVE_VCTP16_]] :: (load 16 from %ir.lsr.iv35, align 2)
   ; CHECK:   [[MVE_VLDRHU16_post2:%[0-9]+]]:rgpr, [[MVE_VLDRHU16_post3:%[0-9]+]]:mqpr = MVE_VLDRHU16_post [[PHI1]], 16, 1, [[MVE_VCTP16_]] :: (load 16 from %ir.lsr.iv12, align 2)
   ; CHECK:   [[MVE_VMLADAVas16_:%[0-9]+]]:tgpreven = MVE_VMLADAVas16 [[PHI2]], killed [[MVE_VLDRHU16_post3]], killed [[MVE_VLDRHU16_post1]], 1, [[MVE_VCTP16_]]
   ; CHECK:   [[COPY8:%[0-9]+]]:gpr = COPY [[MVE_VMLADAVas16_]]
   ; CHECK:   [[COPY9:%[0-9]+]]:gpr = COPY [[MVE_VLDRHU16_post2]]
   ; CHECK:   [[COPY10:%[0-9]+]]:gpr = COPY [[MVE_VLDRHU16_post]]
-  ; CHECK:   [[t2LoopDec:%[0-9]+]]:gprlr = t2LoopDec [[PHI3]], 1
-  ; CHECK:   t2LoopEnd [[t2LoopDec]], %bb.3, implicit-def dead $cpsr
+  ; CHECK:   [[t2LoopEndDec:%[0-9]+]]:gprlr = t2LoopEndDec [[PHI3]], %bb.3, implicit-def $cpsr
   ; CHECK:   t2B %bb.4, 14 /* CC::al */, $noreg
   ; CHECK: bb.4.for.cond.cleanup:
   ; CHECK:   [[PHI5:%[0-9]+]]:gpr = PHI [[COPY3]], %bb.1, [[COPY8]], %bb.3
   ; CHECK:   $r0 = COPY [[PHI5]]
   ; CHECK:   tBX_RET 14 /* CC::al */, $noreg, implicit $r0
   bb.0.entry:
     successors: %bb.1(0x50000000), %bb.4(0x30000000)
     liveins: $r0, $r1, $r2

llvm/test/CodeGen/Thumb2/LowOverheadLoops/minloop.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve -verify-machineinstrs %s -o - | FileCheck %s
 
 define void @arm_min_q31(i32* nocapture readonly %pSrc, i32 %blockSize, i32* nocapture %pResult, i32* nocapture %pIndex) {
 ; CHECK-LABEL: arm_min_q31:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
 ; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}
-; CHECK-NEXT:    .pad #4
-; CHECK-NEXT:    sub sp, #4
 ; CHECK-NEXT:    ldr.w r12, [r0]
 ; CHECK-NEXT:    subs.w r9, r1, #1
 ; CHECK-NEXT:    beq .LBB0_3
 ; CHECK-NEXT:  @ %bb.1: @ %while.body.preheader
-; CHECK-NEXT:    subs r6, r1, #2
-; CHECK-NEXT:    and r7, r9, #3
-; CHECK-NEXT:    cmp r6, #3
-; CHECK-NEXT:    str r7, [sp] @ 4-byte Spill
+; CHECK-NEXT:    subs r7, r1, #2
+; CHECK-NEXT:    and r8, r9, #3
+; CHECK-NEXT:    cmp r7, #3
 ; CHECK-NEXT:    bhs .LBB0_4
 ; CHECK-NEXT:  @ %bb.2:
-; CHECK-NEXT:    mov.w r8, #0
+; CHECK-NEXT:    movs r6, #0
 ; CHECK-NEXT:    b .LBB0_6
 ; CHECK-NEXT:  .LBB0_3:
-; CHECK-NEXT:    mov.w r8, #0
+; CHECK-NEXT:    movs r6, #0
 ; CHECK-NEXT:    b .LBB0_10
 ; CHECK-NEXT:  .LBB0_4: @ %while.body.preheader.new
-; CHECK-NEXT:    bic r6, r9, #3
-; CHECK-NEXT:    movs r4, #1
-; CHECK-NEXT:    subs r6, #4
-; CHECK-NEXT:    mov.w r8, #0
-; CHECK-NEXT:    add.w lr, r4, r6, lsr #2
-; CHECK-NEXT:    movs r6, #4
-; CHECK-NEXT:    mov lr, lr
-; CHECK-NEXT:    mov r11, lr
+; CHECK-NEXT:    bic r7, r9, #3
+; CHECK-NEXT:    movs r6, #1
+; CHECK-NEXT:    subs r7, #4
+; CHECK-NEXT:    add.w lr, r6, r7, lsr #2
+; CHECK-NEXT:    movs r6, #0
+; CHECK-NEXT:    dls lr, lr
+; CHECK-NEXT:    movs r7, #4
 ; CHECK-NEXT:  .LBB0_5: @ %while.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    ldr r10, [r0, #16]!
-; CHECK-NEXT:    mov lr, r11
-; CHECK-NEXT:    sub.w lr, lr, #1
 ; CHECK-NEXT:    sub.w r9, r9, #4
-; CHECK-NEXT:    ldrd r7, r5, [r0, #-12]
-; CHECK-NEXT:    mov r11, lr
-; CHECK-NEXT:    ldr r4, [r0, #-4]
-; CHECK-NEXT:    cmp r12, r7
+; CHECK-NEXT:    ldrd r5, r4, [r0, #-12]
+; CHECK-NEXT:    ldr r11, [r0, #-4]
+; CHECK-NEXT:    cmp r12, r5
 ; CHECK-NEXT:    it gt
-; CHECK-NEXT:    subgt.w r8, r6, #3
-; CHECK-NEXT:    csel r7, r7, r12, gt
-; CHECK-NEXT:    cmp r7, r5
+; CHECK-NEXT:    subgt r6, r7, #3
+; CHECK-NEXT:    csel r5, r5, r12, gt
+; CHECK-NEXT:    cmp r5, r4
 ; CHECK-NEXT:    it gt
-; CHECK-NEXT:    subgt.w r8, r6, #2
-; CHECK-NEXT:    csel r7, r5, r7, gt
-; CHECK-NEXT:    cmp r7, r4
+; CHECK-NEXT:    subgt r6, r7, #2
+; CHECK-NEXT:    csel r5, r4, r5, gt
+; CHECK-NEXT:    cmp r5, r11
 ; CHECK-NEXT:    it gt
-; CHECK-NEXT:    subgt.w r8, r6, #1
-; CHECK-NEXT:    csel r7, r4, r7, gt
-; CHECK-NEXT:    cmp r7, r10
-; CHECK-NEXT:    csel r8, r6, r8, gt
-; CHECK-NEXT:    add.w r6, r6, #4
-; CHECK-NEXT:    csel r12, r10, r7, gt
-; CHECK-NEXT:    cmp.w lr, #0
-; CHECK-NEXT:    bne .LBB0_5
-; CHECK-NEXT:    b .LBB0_6
+; CHECK-NEXT:    subgt r6, r7, #1
+; CHECK-NEXT:    csel r5, r11, r5, gt
+; CHECK-NEXT:    cmp r5, r10
+; CHECK-NEXT:    csel r6, r7, r6, gt
+; CHECK-NEXT:    add.w r7, r7, #4
+; CHECK-NEXT:    csel r12, r10, r5, gt
+; CHECK-NEXT:    le lr, .LBB0_5
 ; CHECK-NEXT:  .LBB0_6: @ %while.end.loopexit.unr-lcssa
-; CHECK-NEXT:    ldr r7, [sp] @ 4-byte Reload
-; CHECK-NEXT:    cbz r7, .LBB0_10
+; CHECK-NEXT:    cmp.w r8, #0
+; CHECK-NEXT:    beq .LBB0_10
 ; CHECK-NEXT:  @ %bb.7: @ %while.body.epil
-; CHECK-NEXT:    ldr r4, [r0, #4]
+; CHECK-NEXT:    ldr r7, [r0, #4]
 ; CHECK-NEXT:    sub.w r1, r1, r9
-; CHECK-NEXT:    cmp r12, r4
-; CHECK-NEXT:    csel r8, r1, r8, gt
-; CHECK-NEXT:    csel r12, r4, r12, gt
-; CHECK-NEXT:    cmp r7, #1
+; CHECK-NEXT:    cmp r12, r7
+; CHECK-NEXT:    csel r6, r1, r6, gt
+; CHECK-NEXT:    csel r12, r7, r12, gt
+; CHECK-NEXT:    cmp.w r8, #1
 ; CHECK-NEXT:    beq .LBB0_10
 ; CHECK-NEXT:  @ %bb.8: @ %while.body.epil.1
-; CHECK-NEXT:    ldr r4, [r0, #8]
-; CHECK-NEXT:    cmp r12, r4
-; CHECK-NEXT:    csinc r8, r8, r1, le
-; CHECK-NEXT:    csel r12, r4, r12, gt
-; CHECK-NEXT:    cmp r7, #2
+; CHECK-NEXT:    ldr r7, [r0, #8]
+; CHECK-NEXT:    cmp r12, r7
+; CHECK-NEXT:    csinc r6, r6, r1, le
+; CHECK-NEXT:    csel r12, r7, r12, gt
+; CHECK-NEXT:    cmp.w r8, #2
 ; CHECK-NEXT:    beq .LBB0_10
 ; CHECK-NEXT:  @ %bb.9: @ %while.body.epil.2
 ; CHECK-NEXT:    ldr r0, [r0, #12]
 ; CHECK-NEXT:    cmp r12, r0
 ; CHECK-NEXT:    it gt
-; CHECK-NEXT:    addgt.w r8, r1, #2
+; CHECK-NEXT:    addgt r6, r1, #2
 ; CHECK-NEXT:    csel r12, r0, r12, gt
 ; CHECK-NEXT:  .LBB0_10: @ %while.end
 ; CHECK-NEXT:    str.w r12, [r2]
-; CHECK-NEXT:    str.w r8, [r3]
-; CHECK-NEXT:    add sp, #4
+; CHECK-NEXT:    str r6, [r3]
 ; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
 entry:
   %0 = load i32, i32* %pSrc, align 4
   %blkCnt.015 = add i32 %blockSize, -1
   %cmp.not17 = icmp eq i32 %blkCnt.015, 0
   br i1 %cmp.not17, label %while.end, label %while.body.preheader
 
 while.body.preheader:                             ; preds = %entry

llvm/test/CodeGen/Thumb2/LowOverheadLoops/mve-float-loops.ll

for.body:                                         ; preds = %for.body.prol.loopexit, %for.body
   %exitcond.3 = icmp eq i32 %inc.3, %N
   br i1 %exitcond.3, label %for.cond.cleanup, label %for.body
 }
 
 define arm_aapcs_vfpcc void @float_int_int_mul(i32* nocapture readonly %a, i32* nocapture readonly %b, float* nocapture %c, i32 %N) {
 ; CHECK-LABEL: float_int_int_mul:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    push {r4, r5, r6, lr}
-; CHECK-NEXT:    cmp r3, #0
-; CHECK-NEXT:    beq .LBB4_8
+; CHECK-NEXT:    cbz r3, .LBB4_8
 ; CHECK-NEXT:  @ %bb.1: @ %for.body.preheader
 ; CHECK-NEXT:    cmp r3, #3
 ; CHECK-NEXT:    bhi .LBB4_3
 ; CHECK-NEXT:  @ %bb.2:
 ; CHECK-NEXT:    mov.w r12, #0
 ; CHECK-NEXT:    b .LBB4_6
 ; CHECK-NEXT:  .LBB4_3: @ %vector.ph
 ; CHECK-NEXT:    bic r12, r3, #3

llvm/test/CodeGen/Thumb2/LowOverheadLoops/regalloc.ll

 ; CHECK-NEXT:    sxtb.w r12, r5
 ; CHECK-NEXT:    cmp r12, r11
 ; CHECK-NEXT:    csel r12, r11, r5, lt
 ; CHECK-NEXT:    ldrsb r5, [r7, #-3]
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    addlt r4, r6, #2
 ; CHECK-NEXT:    sxtb.w r10, r12
 ; CHECK-NEXT:    cmp r10, r5
-; CHECK-NEXT:    csel r10, r5, r12, lt
+; CHECK-NEXT:    csel r12, r5, r12, lt
 ; CHECK-NEXT:    ldrsb r5, [r7, #-2]
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    addlt r4, r6, #3
 ; CHECK-NEXT:    adds r6, #4
-; CHECK-NEXT:    sxtb.w r12, r10
-; CHECK-NEXT:    cmp r12, r5
-; CHECK-NEXT:    csel r12, r5, r10, lt
+; CHECK-NEXT:    sxtb.w r10, r12
+; CHECK-NEXT:    cmp r10, r5
+; CHECK-NEXT:    csel r12, r5, r12, lt
 ; CHECK-NEXT:    csel r10, r6, r4, lt
 ; CHECK-NEXT:    le lr, .LBB0_5
 ; CHECK-NEXT:  @ %bb.6: @ %while.end.loopexit.unr-lcssa.loopexit
 ; CHECK-NEXT:    add r0, r6
 ; CHECK-NEXT:    sub.w r9, r9, r6
 ; CHECK-NEXT:    cmp.w r8, #0
 ; CHECK-NEXT:    beq .LBB0_10
 ; CHECK-NEXT:  .LBB0_7: @ %while.body.epil

 ; CHECK-NEXT:    itt hi
 ; CHECK-NEXT:    addhi.w r3, r1, r2, lsl #1
 ; CHECK-NEXT:    cmphi r3, r0
 ; CHECK-NEXT:    bhi .LBB1_9
 ; CHECK-NEXT:  @ %bb.3: @ %vector.ph
 ; CHECK-NEXT:    bic r4, r2, #7
 ; CHECK-NEXT:    movs r5, #1
 ; CHECK-NEXT:    sub.w r3, r4, #8
-; CHECK-NEXT:    add.w r12, r1, r4, lsl #1
+; CHECK-NEXT:    add.w r12, r0, r4, lsl #1
 ; CHECK-NEXT:    add.w lr, r5, r3, lsr #3
 ; CHECK-NEXT:    and r5, r2, #7
-; CHECK-NEXT:    add.w r3, r0, r4, lsl #1
+; CHECK-NEXT:    add.w r3, r1, r4, lsl #1
 ; CHECK-NEXT:    dls lr, lr
 ; CHECK-NEXT:  .LBB1_4: @ %vector.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    vldrh.u16 q0, [r0], #16
 ; CHECK-NEXT:    vabs.f16 q0, q0
 ; CHECK-NEXT:    vstrb.8 q0, [r1], #16
 ; CHECK-NEXT:    le lr, .LBB1_4
 ; CHECK-NEXT:  @ %bb.5: @ %middle.block
 ; CHECK-NEXT:    cmp r4, r2
 ; CHECK-NEXT:    it eq
 ; CHECK-NEXT:    popeq {r4, r5, r7, pc}
 ; CHECK-NEXT:  .LBB1_6: @ %while.body.preheader18
 ; CHECK-NEXT:    dls lr, r5
 ; CHECK-NEXT:  .LBB1_7: @ %while.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    vldr.16 s0, [r3]
-; CHECK-NEXT:    add.w r0, r12, #2
-; CHECK-NEXT:    adds r3, #2
+; CHECK-NEXT:    vldr.16 s0, [r12]
+; CHECK-NEXT:    adds r0, r3, #2
+; CHECK-NEXT:    add.w r12, r12, #2
 ; CHECK-NEXT:    vabs.f16 s0, s0
-; CHECK-NEXT:    vstr.16 s0, [r12]
-; CHECK-NEXT:    mov r12, r0
+; CHECK-NEXT:    vstr.16 s0, [r3]
+; CHECK-NEXT:    mov r3, r0
 ; CHECK-NEXT:    le lr, .LBB1_7
 ; CHECK-NEXT:  .LBB1_8: @ %while.end
 ; CHECK-NEXT:    pop {r4, r5, r7, pc}
 ; CHECK-NEXT:  .LBB1_9:
 ; CHECK-NEXT:    mov r5, r2
-; CHECK-NEXT:    mov r3, r0
-; CHECK-NEXT:    mov r12, r1
+; CHECK-NEXT:    mov r12, r0
+; CHECK-NEXT:    mov r3, r1
 ; CHECK-NEXT:    b .LBB1_6
 entry:
   %cmp4 = icmp eq i32 %blockSize, 0
   br i1 %cmp4, label %while.end, label %while.body.preheader
 
 while.body.preheader:                             ; preds = %entry
   %min.iters.check = icmp ult i32 %blockSize, 8
   br i1 %min.iters.check, label %while.body.preheader18, label %vector.memcheck

llvm/test/CodeGen/Thumb2/mve-float32regloops.ll

 ; CHECK-NEXT:    b .LBB16_4
 ; CHECK-NEXT:  .LBB16_3: @ %while.end
 ; CHECK-NEXT:    @ in Loop: Header=BB16_4 Depth=1
 ; CHECK-NEXT:    ldr r0, [sp, #16] @ 4-byte Reload
 ; CHECK-NEXT:    subs.w r10, r10, #1
 ; CHECK-NEXT:    vstrb.8 q0, [r2], #16
 ; CHECK-NEXT:    add.w r0, r5, r0, lsl #2
 ; CHECK-NEXT:    add.w r5, r0, #16
-; CHECK-NEXT:    beq.w .LBB16_12
+; CHECK-NEXT:    beq .LBB16_12
 ; CHECK-NEXT:  .LBB16_4: @ %while.body
 ; CHECK-NEXT:    @ =>This Loop Header: Depth=1
 ; CHECK-NEXT:    @ Child Loop BB16_6 Depth 2
 ; CHECK-NEXT:    @ Child Loop BB16_10 Depth 2
 ; CHECK-NEXT:    vldrw.u32 q0, [r1], #16
 ; CHECK-NEXT:    ldrd r7, r6, [r12]
 ; CHECK-NEXT:    ldrd r0, r4, [r12, #8]
 ; CHECK-NEXT:    ldrd r3, lr, [r12, #16]

llvm/test/CodeGen/Thumb2/mve-postinc-dct.ll

 ; CHECK-NEXT:    @ Parent Loop BB7_2 Depth=1
 ; CHECK-NEXT:    @ => This Inner Loop Header: Depth=2
 ; CHECK-NEXT:    add.w r11, r3, r5
 ; CHECK-NEXT:    vctp.32 r10
 ; CHECK-NEXT:    vpsttt
 ; CHECK-NEXT:    vldrwt.u32 q0, [r12], #16
 ; CHECK-NEXT:    vldrwt.u32 q1, [r3], #16
 ; CHECK-NEXT:    vfmat.f32 q6, q1, q0
-; CHECK-NEXT:    add.w r6, r11, r5
+; CHECK-NEXT:    vstrw.32 q6, [sp, #48] @ 16-byte Spill
 ; CHECK-NEXT:    vpstt
 ; CHECK-NEXT:    vldrwt.u32 q1, [r11]
 ; CHECK-NEXT:    vfmat.f32 q7, q1, q0
-; CHECK-NEXT:    vstrw.32 q7, [sp, #48] @ 16-byte Spill
-; CHECK-NEXT:    vmov q7, q6
+; CHECK-NEXT:    add.w r6, r11, r5
 ; CHECK-NEXT:    vmov q6, q5
 ; CHECK-NEXT:    vmov q5, q3
 ; CHECK-NEXT:    vmov q3, q4
 ; CHECK-NEXT:    vpst
 ; CHECK-NEXT:    vldrwt.u32 q1, [r6]
 ; CHECK-NEXT:    vmov q4, q2
 ; CHECK-NEXT:    vldrw.u32 q2, [sp, #64] @ 16-byte Reload
-; CHECK-NEXT:    adds r7, r6, r5
 ; CHECK-NEXT:    vpst
 ; CHECK-NEXT:    vfmat.f32 q2, q1, q0
 ; CHECK-NEXT:    vstrw.32 q2, [sp, #64] @ 16-byte Spill
+; CHECK-NEXT:    adds r7, r6, r5
 ; CHECK-NEXT:    vpst
 ; CHECK-NEXT:    vldrwt.u32 q1, [r7]
 ; CHECK-NEXT:    vldrw.u32 q2, [sp, #80] @ 16-byte Reload
+; CHECK-NEXT:    adds r6, r7, r5
 ; CHECK-NEXT:    vpst
 ; CHECK-NEXT:    vfmat.f32 q2, q1, q0
-; CHECK-NEXT:    adds r6, r7, r5
 ; CHECK-NEXT:    vstrw.32 q2, [sp, #80] @ 16-byte Spill
 ; CHECK-NEXT:    vmov q2, q4
 ; CHECK-NEXT:    vmov q4, q3
 ; CHECK-NEXT:    vmov q3, q5
 ; CHECK-NEXT:    vmov q5, q6
-; CHECK-NEXT:    vmov q6, q7
-; CHECK-NEXT:    vldrw.u32 q7, [sp, #48] @ 16-byte Reload
+; CHECK-NEXT:    vldrw.u32 q6, [sp, #48] @ 16-byte Reload
 ; CHECK-NEXT:    adds r7, r6, r5
 ; CHECK-NEXT:    vpstt
 ; CHECK-NEXT:    vldrwt.u32 q1, [r6]
 ; CHECK-NEXT:    vfmat.f32 q2, q1, q0
 ; CHECK-NEXT:    sub.w r10, r10, #4
 ; CHECK-NEXT:    adds r6, r7, r5
 ; CHECK-NEXT:    vpstttt
 ; CHECK-NEXT:    vldrwt.u32 q1, [r7]

llvm/test/CodeGen/Thumb2/mve-postinc-lsr.ll

 ; CHECK-NEXT:    vpush {d8, d9, d10, d11, d12, d13, d14, d15}
 ; CHECK-NEXT:    .pad #56
 ; CHECK-NEXT:    sub sp, #56
 ; CHECK-NEXT:    cmp r2, #8
 ; CHECK-NEXT:    str r1, [sp, #20] @ 4-byte Spill
 ; CHECK-NEXT:    vstr s0, [sp, #4] @ 4-byte Spill
 ; CHECK-NEXT:    mov r1, r2
 ; CHECK-NEXT:    str r2, [sp, #8] @ 4-byte Spill
-; CHECK-NEXT:    blo.w .LBB7_9
+; CHECK-NEXT:    blo .LBB7_9
 ; CHECK-NEXT:  @ %bb.1:
 ; CHECK-NEXT:    ldr r2, [sp, #8] @ 4-byte Reload
 ; CHECK-NEXT:    movs r3, #1
 ; CHECK-NEXT:    mov.w r10, #0
 ; CHECK-NEXT:    str r2, [sp, #16] @ 4-byte Spill
 ; CHECK-NEXT:    lsrs r1, r2, #2
 ; CHECK-NEXT:    b .LBB7_3
 ; CHECK-NEXT:  .LBB7_2: @ in Loop: Header=BB7_3 Depth=1

llvm/test/CodeGen/Thumb2/mve-satmul-loops.ll

 ; CHECK-NEXT:    cmp r3, #0
 ; CHECK-NEXT:    beq.w .LBB0_8
 ; CHECK-NEXT:  @ %bb.1: @ %entry
 ; CHECK-NEXT:    cmp r3, #1
 ; CHECK-NEXT:    bne .LBB0_3
 ; CHECK-NEXT:  @ %bb.2:
 ; CHECK-NEXT:    movs r7, #0
 ; CHECK-NEXT:    mov r12, r0
-; CHECK-NEXT:    mov r9, r1
+; CHECK-NEXT:    mov r6, r1
 ; CHECK-NEXT:    mov r11, r2
 ; CHECK-NEXT:    b .LBB0_6
 ; CHECK-NEXT:  .LBB0_3: @ %vector.ph
 ; CHECK-NEXT:    str r3, [sp, #4] @ 4-byte Spill
 ; CHECK-NEXT:    bic r3, r3, #1
 ; CHECK-NEXT:    subs r7, r3, #2
 ; CHECK-NEXT:    movs r6, #1
 ; CHECK-NEXT:    adr r4, .LCPI0_0
 ; CHECK-NEXT:    str r3, [sp] @ 4-byte Spill
 ; CHECK-NEXT:    add.w lr, r6, r7, lsr #1
 ; CHECK-NEXT:    add.w r11, r2, r3, lsl #2
-; CHECK-NEXT:    add.w r9, r1, r3, lsl #2
+; CHECK-NEXT:    add.w r6, r1, r3, lsl #2
 ; CHECK-NEXT:    add.w r12, r0, r3, lsl #2
 ; CHECK-NEXT:    dls lr, lr
 ; CHECK-NEXT:    vldrw.u32 q0, [r4]
+; CHECK-NEXT:    mvn r10, #-2147483648
 ; CHECK-NEXT:    vmvn.i32 q1, #0x80000000
-; CHECK-NEXT:    mov.w r10, #-1
 ; CHECK-NEXT:  .LBB0_4: @ %vector.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    ldrd r4, r5, [r0]
+; CHECK-NEXT:    ldrd r4, r8, [r0]
 ; CHECK-NEXT:    adds r0, #8
-; CHECK-NEXT:    ldrd r7, r6, [r1]
+; CHECK-NEXT:    ldrd r7, r5, [r1]
 ; CHECK-NEXT:    adds r1, #8
-; CHECK-NEXT:    smull r8, r5, r6, r5
+; CHECK-NEXT:    smull r8, r5, r5, r8
 ; CHECK-NEXT:    smull r4, r7, r7, r4
 ; CHECK-NEXT:    asrl r8, r5, #31
 ; CHECK-NEXT:    asrl r4, r7, #31
-; CHECK-NEXT:    rsbs.w r3, r4, #-2147483648
+; CHECK-NEXT:    rsbs.w r9, r4, #-2147483648
 ; CHECK-NEXT:    vmov.32 q4[0], r4
-; CHECK-NEXT:    sbcs.w r3, r10, r7
-; CHECK-NEXT:    vmov.32 q4[1], r7
+; CHECK-NEXT:    mov.w r9, #-1
+; CHECK-NEXT:    sbcs.w r3, r9, r7
 ; CHECK-NEXT:    mov.w r3, #0
-; CHECK-NEXT:    vmov.32 q4[2], r8
+; CHECK-NEXT:    vmov.32 q4[1], r7
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    movlt r3, #1
 ; CHECK-NEXT:    cmp r3, #0
 ; CHECK-NEXT:    csetm r3, ne
-; CHECK-NEXT:    vmov.32 q4[3], r5
+; CHECK-NEXT:    vmov.32 q4[2], r8
 ; CHECK-NEXT:    vmov.32 q2[0], r3
+; CHECK-NEXT:    vmov.32 q4[3], r5
 ; CHECK-NEXT:    vmov.32 q2[1], r3
 ; CHECK-NEXT:    rsbs.w r3, r8, #-2147483648
-; CHECK-NEXT:    sbcs.w r3, r10, r5
-; CHECK-NEXT:    mvn r5, #-2147483648
+; CHECK-NEXT:    sbcs.w r3, r9, r5
 ; CHECK-NEXT:    mov.w r3, #0
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    movlt r3, #1
 ; CHECK-NEXT:    cmp r3, #0
 ; CHECK-NEXT:    csetm r3, ne
 ; CHECK-NEXT:    vmov.32 q2[2], r3
 ; CHECK-NEXT:    vmov.32 q2[3], r3
 ; CHECK-NEXT:    vbic q3, q0, q2
 ; CHECK-NEXT:    vand q2, q4, q2
 ; CHECK-NEXT:    vorr q2, q2, q3
 ; CHECK-NEXT:    vmov r4, s8
 ; CHECK-NEXT:    vmov r3, s9
-; CHECK-NEXT:    subs r4, r4, r5
+; CHECK-NEXT:    subs.w r4, r4, r10
 ; CHECK-NEXT:    sbcs r3, r3, #0
 ; CHECK-NEXT:    vmov r4, s10
 ; CHECK-NEXT:    mov.w r3, #0
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    movlt r3, #1
 ; CHECK-NEXT:    cmp r3, #0
 ; CHECK-NEXT:    csetm r3, ne
 ; CHECK-NEXT:    vmov.32 q3[0], r3
 ; CHECK-NEXT:    vmov.32 q3[1], r3
 ; CHECK-NEXT:    vmov r3, s11
-; CHECK-NEXT:    subs r4, r4, r5
+; CHECK-NEXT:    subs.w r4, r4, r10
 ; CHECK-NEXT:    sbcs r3, r3, #0
 ; CHECK-NEXT:    mov.w r3, #0
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    movlt r3, #1
 ; CHECK-NEXT:    cmp r3, #0
 ; CHECK-NEXT:    csetm r3, ne
 ; CHECK-NEXT:    vmov.32 q3[2], r3
 ; CHECK-NEXT:    vbic q4, q1, q3

 ; CHECK-NEXT:    sub.w lr, r3, r7
 ; CHECK-NEXT:    mov.w r0, #-1
 ; CHECK-NEXT:    dls lr, lr
 ; CHECK-NEXT:    mov.w r1, #-2147483648
 ; CHECK-NEXT:    mvn r2, #-2147483648
 ; CHECK-NEXT:  .LBB0_7: @ %for.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    ldr r3, [r12], #4
-; CHECK-NEXT:    ldr r4, [r9], #4
+; CHECK-NEXT:    ldr r4, [r6], #4
 ; CHECK-NEXT:    smull r4, r3, r4, r3
 ; CHECK-NEXT:    asrl r4, r3, #31
 ; CHECK-NEXT:    subs r5, r1, r4
 ; CHECK-NEXT:    sbcs.w r5, r0, r3
 ; CHECK-NEXT:    mov.w r5, #0
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    movlt r5, #1
 ; CHECK-NEXT:    cmp r5, #0

 ; CHECK-NEXT:    sub sp, #4
 ; CHECK-NEXT:    .vsave {d8, d9, d10, d11, d12, d13}
 ; CHECK-NEXT:    vpush {d8, d9, d10, d11, d12, d13}
 ; CHECK-NEXT:    .pad #8
 ; CHECK-NEXT:    sub sp, #8
 ; CHECK-NEXT:    cmp r3, #0
 ; CHECK-NEXT:    beq.w .LBB1_8
 ; CHECK-NEXT:  @ %bb.1: @ %for.body.preheader
-; CHECK-NEXT:    mov r8, r2
-; CHECK-NEXT:    movs r2, #0
+; CHECK-NEXT:    mov r9, r1
+; CHECK-NEXT:    movs r7, #0
 ; CHECK-NEXT:    cmp r3, #3
 ; CHECK-NEXT:    bhi .LBB1_3
 ; CHECK-NEXT:  @ %bb.2:
 ; CHECK-NEXT:    mov r12, r0
-; CHECK-NEXT:    mov r10, r1
-; CHECK-NEXT:    mov r11, r8
+; CHECK-NEXT:    mov r1, r9
+; CHECK-NEXT:    mov r11, r2
 ; CHECK-NEXT:    b .LBB1_6
 ; CHECK-NEXT:  .LBB1_3: @ %vector.ph
+; CHECK-NEXT:    bic r7, r3, #3
+; CHECK-NEXT:    adr r4, .LCPI1_0
+; CHECK-NEXT:    subs r1, r7, #4
 ; CHECK-NEXT:    str r3, [sp, #4] @ 4-byte Spill
-; CHECK-NEXT:    bic r3, r3, #3
-; CHECK-NEXT:    subs r2, r3, #4
-; CHECK-NEXT:    movs r7, #1
-; CHECK-NEXT:    str r3, [sp] @ 4-byte Spill
-; CHECK-NEXT:    add.w r11, r8, r3, lsl #2
-; CHECK-NEXT:    add.w lr, r7, r2, lsr #2
-; CHECK-NEXT:    adr r7, .LCPI1_0
-; CHECK-NEXT:    vldrw.u32 q0, [r7]
-; CHECK-NEXT:    adr r7, .LCPI1_1
-; CHECK-NEXT:    add.w r10, r1, r3, lsl #2
-; CHECK-NEXT:    add.w r12, r0, r3, lsl #2
+; CHECK-NEXT:    movs r3, #1
+; CHECK-NEXT:    vldrw.u32 q0, [r4]
+; CHECK-NEXT:    add.w lr, r3, r1, lsr #2
+; CHECK-NEXT:    str r7, [sp] @ 4-byte Spill
+; CHECK-NEXT:    adr r4, .LCPI1_1
+; CHECK-NEXT:    add.w r11, r2, r7, lsl #2
+; CHECK-NEXT:    add.w r1, r9, r7, lsl #2
+; CHECK-NEXT:    add.w r12, r0, r7, lsl #2
 ; CHECK-NEXT:    dls lr, lr
-; CHECK-NEXT:    vldrw.u32 q1, [r7]
-; CHECK-NEXT:    mov.w r3, #-1
-; CHECK-NEXT:    mvn r9, #-2147483648
+; CHECK-NEXT:    vldrw.u32 q1, [r4]
+; CHECK-NEXT:    mov.w r10, #-1
 ; CHECK-NEXT:  .LBB1_4: @ %vector.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    vldrw.u32 q2, [r0], #16
-; CHECK-NEXT:    vldrw.u32 q3, [r1], #16
-; CHECK-NEXT:    mov r2, lr
+; CHECK-NEXT:    vldrw.u32 q3, [r9], #16
 ; CHECK-NEXT:    vmov.f32 s16, s10
 ; CHECK-NEXT:    vmov.f32 s20, s14
 ; CHECK-NEXT:    vmov.f32 s18, s11
 ; CHECK-NEXT:    vmov.f32 s22, s15
 ; CHECK-NEXT:    vmullb.s32 q6, q5, q4
 ; CHECK-NEXT:    vmov.f32 s10, s9
 ; CHECK-NEXT:    vmov r7, s25
-; CHECK-NEXT:    vmov r6, s24
-; CHECK-NEXT:    asrl r6, r7, #31
-; CHECK-NEXT:    vmov lr, s26
-; CHECK-NEXT:    rsbs.w r5, r6, #-2147483648
+; CHECK-NEXT:    vmov r4, s24
+; CHECK-NEXT:    asrl r4, r7, #31
+; CHECK-NEXT:    vmov r8, s26
+; CHECK-NEXT:    rsbs.w r5, r4, #-2147483648
 ; CHECK-NEXT:    vmov.f32 s14, s13
-; CHECK-NEXT:    sbcs.w r5, r3, r7
+; CHECK-NEXT:    sbcs.w r5, r10, r7
 ; CHECK-NEXT:    mov.w r5, #0
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    movlt r5, #1
 ; CHECK-NEXT:    cmp r5, #0
 ; CHECK-NEXT:    csetm r5, ne
 ; CHECK-NEXT:    vmov.32 q4[0], r5
 ; CHECK-NEXT:    vmov.32 q4[1], r5
 ; CHECK-NEXT:    vmov r5, s27
-; CHECK-NEXT:    asrl lr, r5, #31
-; CHECK-NEXT:    vmov.32 q6[0], r6
-; CHECK-NEXT:    rsbs.w r4, lr, #-2147483648
+; CHECK-NEXT:    asrl r8, r5, #31
+; CHECK-NEXT:    vmov.32 q6[0], r4
+; CHECK-NEXT:    rsbs.w r6, r8, #-2147483648
 ; CHECK-NEXT:    vmov.32 q6[1], r7
-; CHECK-NEXT:    sbcs.w r4, r3, r5
-; CHECK-NEXT:    vmov.32 q6[2], lr
-; CHECK-NEXT:    mov.w r4, #0
+; CHECK-NEXT:    sbcs.w r6, r10, r5
+; CHECK-NEXT:    vmov.32 q6[2], r8
+; CHECK-NEXT:    mov.w r6, #0
 ; CHECK-NEXT:    vmov.32 q6[3], r5
 ; CHECK-NEXT:    it lt
-; CHECK-NEXT:    movlt r4, #1
-; CHECK-NEXT:    cmp r4, #0
-; CHECK-NEXT:    csetm r4, ne
-; CHECK-NEXT:    mov lr, r2
-; CHECK-NEXT:    vmov.32 q4[2], r4
-; CHECK-NEXT:    vmov.32 q4[3], r4
-; CHECK-NEXT:    vmov r4, s14
+; CHECK-NEXT:    movlt r6, #1
+; CHECK-NEXT:    cmp r6, #0
+; CHECK-NEXT:    csetm r6, ne
+; CHECK-NEXT:    mvn r8, #-2147483648
+; CHECK-NEXT:    vmov.32 q4[2], r6
+; CHECK-NEXT:    vmov.32 q4[3], r6
+; CHECK-NEXT:    vmov r6, s14
 ; CHECK-NEXT:    vbic q5, q0, q4
 ; CHECK-NEXT:    vand q4, q6, q4
 ; CHECK-NEXT:    vorr q4, q4, q5
-; CHECK-NEXT:    vmov r6, s16
-; CHECK-NEXT:    vmov r7, s17
-; CHECK-NEXT:    subs.w r6, r6, r9
-; CHECK-NEXT:    sbcs r7, r7, #0
-; CHECK-NEXT:    vmov r6, s18
-; CHECK-NEXT:    mov.w r7, #0
+; CHECK-NEXT:    vmov r5, s16
+; CHECK-NEXT:    vmov r4, s17
+; CHECK-NEXT:    subs.w r5, r5, r8
+; CHECK-NEXT:    sbcs r4, r4, #0
+; CHECK-NEXT:    vmov r5, s18
+; CHECK-NEXT:    mov.w r4, #0
+; CHECK-NEXT:    it lt
+; CHECK-NEXT:    movlt r4, #1
+; CHECK-NEXT:    cmp r4, #0
+; CHECK-NEXT:    csetm r4, ne
+; CHECK-NEXT:    vmov.32 q5[0], r4
+; CHECK-NEXT:    vmov.32 q5[1], r4
+; CHECK-NEXT:    vmov r4, s19
+; CHECK-NEXT:    subs.w r5, r5, r8
+; CHECK-NEXT:    vmov r5, s12
+; CHECK-NEXT:    sbcs r4, r4, #0
+; CHECK-NEXT:    mov.w r4, #0
 ; CHECK-NEXT:    it lt
-; CHECK-NEXT:    movlt r7, #1
-; CHECK-NEXT:    cmp r7, #0
-; CHECK-NEXT:    csetm r7, ne
-; CHECK-NEXT:    vmov.32 q5[0], r7
-; CHECK-NEXT:    vmov.32 q5[1], r7
-; CHECK-NEXT:    vmov r7, s19
-; CHECK-NEXT:    subs.w r6, r6, r9
-; CHECK-NEXT:    vmov r6, s12
-; CHECK-NEXT:    sbcs r7, r7, #0
-; CHECK-NEXT:    mov.w r7, #0
-; CHECK-NEXT:    it lt
-; CHECK-NEXT:    movlt r7, #1
-; CHECK-NEXT:    cmp r7, #0
-; CHECK-NEXT:    csetm r7, ne
-; CHECK-NEXT:    vmov.32 q5[2], r7
-; CHECK-NEXT:    vmov r7, s8
+; CHECK-NEXT:    movlt r4, #1
+; CHECK-NEXT:    cmp r4, #0
+; CHECK-NEXT:    csetm r4, ne
+; CHECK-NEXT:    vmov.32 q5[2], r4
+; CHECK-NEXT:    vmov r4, s8
 ; CHECK-NEXT:    vbic q6, q1, q5
 ; CHECK-NEXT:    vand q4, q4, q5
 ; CHECK-NEXT:    vorr q4, q4, q6
-; CHECK-NEXT:    smull r6, r7, r6, r7
-; CHECK-NEXT:    asrl r6, r7, #31
-; CHECK-NEXT:    rsbs.w r5, r6, #-2147483648
-; CHECK-NEXT:    vmov.32 q3[0], r6
-; CHECK-NEXT:    sbcs.w r5, r3, r7
+; CHECK-NEXT:    smull r4, r7, r5, r4
+; CHECK-NEXT:    asrl r4, r7, #31
+; CHECK-NEXT:    rsbs.w r5, r4, #-2147483648
+; CHECK-NEXT:    vmov.32 q3[0], r4
+; CHECK-NEXT:    sbcs.w r5, r10, r7
 ; CHECK-NEXT:    vmov.32 q3[1], r7
 ; CHECK-NEXT:    mov.w r5, #0
 ; CHECK-NEXT:    it lt
 ; CHECK-NEXT:    movlt r5, #1
 ; CHECK-NEXT:    cmp r5, #0
 ; CHECK-NEXT:    csetm r5, ne
 ; CHECK-NEXT:    vmov.32 q5[0], r5
 ; CHECK-NEXT:    vmov.32 q5[1], r5
 ; CHECK-NEXT:    vmov r5, s10
-; CHECK-NEXT:    smull r4, r5, r4, r5
-; CHECK-NEXT:    asrl r4, r5, #31
-; CHECK-NEXT:    rsbs.w r2, r4, #-2147483648
-; CHECK-NEXT:    vmov.32 q3[2], r4
-; CHECK-NEXT:    sbcs.w r2, r3, r5
+; CHECK-NEXT:    smull r6, r5, r6, r5
+; CHECK-NEXT:    asrl r6, r5, #31
+; CHECK-NEXT:    rsbs.w r3, r6, #-2147483648
+; CHECK-NEXT:    vmov.32 q3[2], r6
+; CHECK-NEXT:    sbcs.w r3, r10, r5
 ; CHECK-NEXT:    vmov.32 q3[3], r5
-; CHECK-NEXT:    mov.w r2, #0
+; CHECK-NEXT:    mov.w r3, #0
 ; CHECK-NEXT:    it lt
-; CHECK-NEXT:    movlt r2, #1
-; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    csetm r2, ne
-; CHECK-NEXT:    vmov.32 q5[2], r2
-; CHECK-NEXT:    vmov.32 q5[3], r2
+; CHECK-NEXT:    movlt r3, #1
+; CHECK-NEXT:    cmp r3, #0
+; CHECK-NEXT:    csetm r3, ne
+; CHECK-NEXT:    vmov.32 q5[2], r3
+; CHECK-NEXT:    vmov.32 q5[3], r3
 ; CHECK-NEXT:    vbic q2, q0, q5
 ; CHECK-NEXT:    vand q3, q3, q5
 ; CHECK-NEXT:    vorr q2, q3, q2
-; CHECK-NEXT:    vmov r7, s8
-; CHECK-NEXT:    vmov r2, s9
-; CHECK-NEXT:    subs.w r7, r7, r9
-; CHECK-NEXT:    sbcs r2, r2, #0
-; CHECK-NEXT:    vmov r7, s10
-; CHECK-NEXT:    mov.w r2, #0
-; CHECK-NEXT:    it lt
-; CHECK-NEXT:    movlt r2, #1
-; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    csetm r2, ne
-; CHECK-NEXT:    vmov.32 q3[0], r2
-; CHECK-NEXT:    vmov.32 q3[1], r2
-; CHECK-NEXT:    vmov r2, s11
-; CHECK-NEXT:    subs.w r7, r7, r9
-; CHECK-NEXT:    sbcs r2, r2, #0
-; CHECK-NEXT:    mov.w r2, #0
-; CHECK-NEXT:    it lt
-; CHECK-NEXT:    movlt r2, #1
-; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    csetm r2, ne
-; CHECK-NEXT:    vmov.32 q3[2], r2
+; CHECK-NEXT:    vmov r4, s8
+; CHECK-NEXT:    vmov r3, s9
+; CHECK-NEXT:    subs.w r4, r4, r8
+; CHECK-NEXT:    sbcs r3, r3, #0
+; CHECK-NEXT:    vmov r4, s10
+; CHECK-NEXT:    mov.w r3, #0
+; CHECK-NEXT:    it lt
+; CHECK-NEXT:    movlt r3, #1
+; CHECK-NEXT:    cmp r3, #0
+; CHECK-NEXT:    csetm r3, ne
+; CHECK-NEXT:    vmov.32 q3[0], r3
+; CHECK-NEXT:    vmov.32 q3[1], r3
+; CHECK-NEXT:    vmov r3, s11
+; CHECK-NEXT:    subs.w r4, r4, r8
+; CHECK-NEXT:    sbcs r3, r3, #0
+; CHECK-NEXT:    mov.w r3, #0
+; CHECK-NEXT:    it lt
+; CHECK-NEXT:    movlt r3, #1
+; CHECK-NEXT:    cmp r3, #0
+; CHECK-NEXT:    csetm r3, ne
+; CHECK-NEXT:    vmov.32 q3[2], r3
 ; CHECK-NEXT:    vbic q5, q1, q3
 ; CHECK-NEXT:    vand q2, q2, q3
 ; CHECK-NEXT:    vorr q2, q2, q5
 ; CHECK-NEXT:    vmov.f32 s9, s10
 ; CHECK-NEXT:    vmov.f32 s10, s16
 ; CHECK-NEXT:    vmov.f32 s11, s18
-; CHECK-NEXT:    vstrb.8 q2, [r8], #16
+; CHECK-NEXT:    vstrb.8 q2, [r2], #16
 ; CHECK-NEXT:    le lr, .LBB1_4
 ; CHECK-NEXT:  @ %bb.5: @ %middle.block
-; CHECK-NEXT:    ldrd r2, r3, [sp] @ 8-byte Folded Reload
-; CHECK-NEXT:    cmp r2, r3
+; CHECK-NEXT:    ldrd r7, r3, [sp] @ 8-byte Folded Reload
+; CHECK-NEXT:    cmp r7, r3
 ; CHECK-NEXT:    beq .LBB1_8
 ; CHECK-NEXT:  .LBB1_6: @ %for.body.preheader21
-; CHECK-NEXT:    sub.w lr, r3, r2
+; CHECK-NEXT:    sub.w lr, r3, r7
 ; CHECK-NEXT:    mov.w r0, #-1
 ; CHECK-NEXT:    dls lr, lr
-; CHECK-NEXT:    mov.w r1, #-2147483648
-; CHECK-NEXT:    mvn r3, #-2147483648
+; CHECK-NEXT:    mov.w r3, #-2147483648
+; CHECK-NEXT:    mvn r2, #-2147483648
 ; CHECK-NEXT:  .LBB1_7: @ %for.body
 ; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    ldr r2, [r12], #4
-; CHECK-NEXT:    ldr r4, [r10], #4
-; CHECK-NEXT:    smull r2, r5, r4, r2
-; CHECK-NEXT:    asrl r2, r5, #31
-; CHECK-NEXT:    subs r4, r1, r2
-; CHECK-NEXT:    sbcs.w r4, r0, r5
-; CHECK-NEXT:    mov.w r4, #0
-; CHECK-NEXT:    it lt
-; CHECK-NEXT:    movlt r4, #1
-; CHECK-NEXT:    cmp r4, #0
-; CHECK-NEXT:    csel r2, r2, r1, ne
-; CHECK-NEXT:    csel r4, r5, r0, ne
-; CHECK-NEXT:    subs r5, r2, r3
-; CHECK-NEXT:    sbcs r4, r4, #0
-; CHECK-NEXT:    csel r2, r2, r3, lt
-; CHECK-NEXT:    str r2, [r11], #4
+; CHECK-NEXT:    ldr r4, [r12], #4
+; CHECK-NEXT:    ldr r5, [r1], #4
+; CHECK-NEXT:    smull r4, r5, r5, r4
+; CHECK-NEXT:    asrl r4, r5, #31
+; CHECK-NEXT:    subs r6, r3, r4
+; CHECK-NEXT:    sbcs.w r6, r0, r5
+; CHECK-NEXT:    mov.w r6, #0
+; CHECK-NEXT:    it lt
+; CHECK-NEXT:    movlt r6, #1
+; CHECK-NEXT:    cmp r6, #0
+; CHECK-NEXT:    csel r4, r4, r3, ne
+; CHECK-NEXT:    csel r5, r5, r0, ne
+; CHECK-NEXT:    subs r6, r4, r2
+; CHECK-NEXT:    sbcs r5, r5, #0
+; CHECK-NEXT:    csel r4, r4, r2, lt
+; CHECK-NEXT:    str r4, [r11], #4
 ; CHECK-NEXT:    le lr, .LBB1_7
 ; CHECK-NEXT:  .LBB1_8: @ %for.cond.cleanup
 ; CHECK-NEXT:    add sp, #8
 ; CHECK-NEXT:    vpop {d8, d9, d10, d11, d12, d13}
 ; CHECK-NEXT:    add sp, #4
 ; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
 ; CHECK-NEXT:    .p2align 4
 ; CHECK-NEXT:  @ %bb.9:
for.body:                                         ; preds = %for.body.preheader21, %for.body
   br i1 %exitcond, label %for.cond.cleanup, label %for.body
 }
 
 define arm_aapcs_vfpcc void @ssatmul_8i_q15(i16* nocapture readonly %pSrcA, i16* nocapture readonly %pSrcB, i16* noalias nocapture %pDst, i32 %N) {
 ; CHECK-LABEL: ssatmul_8i_q15:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r4, r5, r6, lr}
 ; CHECK-NEXT:    push {r4, r5, r6, lr}
-; CHECK-NEXT:    cmp r3, #0
-; CHECK-NEXT:    beq .LBB7_8
+; CHECK-NEXT:    cbz r3, .LBB7_8
 ; CHECK-NEXT:  @ %bb.1: @ %for.body.preheader
 ; CHECK-NEXT:    cmp r3, #7
 ; CHECK-NEXT:    bhi .LBB7_3
 ; CHECK-NEXT:  @ %bb.2:
 ; CHECK-NEXT:    movs r5, #0
 ; CHECK-NEXT:    mov r12, r0
 ; CHECK-NEXT:    mov r6, r1
 ; CHECK-NEXT:    mov r4, r2
for.body:                                         ; preds = %for.body.preheader23, %for.body
   br i1 %exitcond, label %for.cond.cleanup, label %for.body
 }
 
 define arm_aapcs_vfpcc void @ssatmul_16i_q7(i8* nocapture readonly %pSrcA, i8* nocapture readonly %pSrcB, i8* noalias nocapture %pDst, i32 %N) {
 ; CHECK-LABEL: ssatmul_16i_q7:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r4, r5, r6, lr}
 ; CHECK-NEXT:    push {r4, r5, r6, lr}
-; CHECK-NEXT:    cmp r3, #0
-; CHECK-NEXT:    beq .LBB16_8
+; CHECK-NEXT:    cbz r3, .LBB16_8
 ; CHECK-NEXT:  @ %bb.1: @ %for.body.preheader
 ; CHECK-NEXT:    cmp r3, #15
 ; CHECK-NEXT:    bhi .LBB16_3
 ; CHECK-NEXT:  @ %bb.2:
 ; CHECK-NEXT:    movs r5, #0
 ; CHECK-NEXT:    mov r12, r0
 ; CHECK-NEXT:    mov r6, r1
 ; CHECK-NEXT:    mov r4, r2
for.cond.cleanup:                                 ; preds = %vector.body, %entry
   ret void
 }
 
 define arm_aapcs_vfpcc void @usatmul_8_q7(i8* nocapture readonly %pSrcA, i8* nocapture readonly %pSrcB, i8* noalias nocapture %pDst, i32 %N) {
 ; CHECK-LABEL: usatmul_8_q7:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r4, r5, r6, lr}
 ; CHECK-NEXT:    push {r4, r5, r6, lr}
-; CHECK-NEXT:    cmp r3, #0
-; CHECK-NEXT:    beq .LBB20_8
+; CHECK-NEXT:    cbz r3, .LBB20_8
 ; CHECK-NEXT:  @ %bb.1: @ %for.body.preheader
 ; CHECK-NEXT:    cmp r3, #7
 ; CHECK-NEXT:    bhi .LBB20_3
 ; CHECK-NEXT:  @ %bb.2:
 ; CHECK-NEXT:    movs r5, #0
 ; CHECK-NEXT:    mov r12, r0
 ; CHECK-NEXT:    mov r6, r1
 ; CHECK-NEXT:    mov r4, r2

llvm/test/CodeGen/Thumb2/mve-vldshuffle.ll

while.end:                                        ; preds = %while.body, %middle.block, %entry
   ret void
 }
 
 define void @arm_cmplx_mag_squared_f32(float* nocapture readonly %pSrc, float* nocapture %pDst, i32 %numSamples) {
 ; CHECK-LABEL: arm_cmplx_mag_squared_f32:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    .save {r4, r5, r7, lr}
 ; CHECK-NEXT:    push {r4, r5, r7, lr}
-; CHECK-NEXT:    cmp r2, #0
-; CHECK-NEXT:    beq .LBB1_8
+; CHECK-NEXT:    cbz r2, .LBB1_8
 ; CHECK-NEXT:  @ %bb.1: @ %while.body.preheader
 ; CHECK-NEXT:    cmp r2, #4
 ; CHECK-NEXT:    blo .LBB1_9
 ; CHECK-NEXT:  @ %bb.2: @ %vector.memcheck
 ; CHECK-NEXT:    add.w r3, r0, r2, lsl #3
 ; CHECK-NEXT:    cmp r3, r1
 ; CHECK-NEXT:    itt hi
 ; CHECK-NEXT:    addhi.w r3, r1, r2, lsl #2