[ARM][MVE] Validate tail predication values

[samparker]

Summary

Iterate through the loop and check that the observable values produced are the same whether tail predication happens or not.
We want to find out if the tail-predicated version of this loop will produce the same values as the loop in its original form. For this to be true, the newly inserted implicit predication must not change the the results. All MVE loads and stores have to be predicated, so we know that any load operands, or stored results are equivalent already.
Other explicitly predicated instructions will perform the same operation in the original loop and the tail-predicated form. We call predicated instructions 'Known' and their users are also Known if it overwrites an Known operand. This is because tail predication will cause the false lanes to not be updated, but we know the output because we know the input.
For any 'Unknown' instructions, we can check that they're only consumed by Known instructions because it means that the unknown false lane(s) are replaced with known lane(s).
What this should result in is that all instructions are dependent upon predicated inputs. It also means that for an iteration of the loop, a register should hold the same value whether tail predication has happened or not, with the exception when the differences are masked away by their user(s) and not observable elsewhere.

[samparker]

Rewrite.

[samparker]

• Renamed the sets.
• Change comment describing what I'm trying to achieve.

[SjoerdMeijer]

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

555

Would it be good to spell out here that we have these 2 use cases?

• IR produced by the vectoriser, and
• IR originating from hand written (vector) intrinsics, which can also trigger tail-predication.

loads/stores produced by the vectoriser (when it thinks tail-folding is good) will all be nicely masked, so this concerns the 2nd use case. Is that right?

560

nit: analyse -> analysis?

574

I appreciate this is bikeshedding, but here we go anyway: UnknownFalseLanes and KnownFalseZeros is a bit difficult to read (in the code below), but your descriptions/comments above are clear. Summarising that, we have a set of instructions with some behaviour that we (immediately) understand, and there are some instructions that needs to be further analysed.
So, was wondering KnownFalseZeros could be named something along the lines of just Predicated, and UnknownFalseLanes be the Worklist of things to check?

589

can you use one of your new RDA helper functions for this?

[samparker]

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

555

This pass shouldn't be concerned with whether the input is from the vectorizer or the intrinsics, as we can't know. On line 762 we enforce that all MVE memory operations are explicitly predicated.

574

I did have Predicated at some point, but I thought that was misleading because some instructions won't be using the VPR. I like KnownFalseZeros because it's explicit in what we expect and I'm much rather have code that doesn't require a big comment to make any sense of what is going on. It doesn't have to be that, but its name should convey what's important about the set. The naming of UnknownFalseLanes is definitely less important though.

589

There were just static helpers and I don't think we won't really do enough of this logic to warrant more helpers.

[SjoerdMeijer]

was just wondering about more corner cases for "negative" tests: do we have test where lanes are swapped (if that makes sense)?

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

582

Is there a test for this?

llvm/test/CodeGen/Thumb2/LowOverheadLoops/unpredicated-min.mir

2

nit: rename file unpredicated-min.mir -> unpredicated-max.mir?

[samparker]

do we have test where lanes are swapped (if that makes sense)?

I don't think there's one for this pass... but that was the reason for having a unit test for validForTailPredication, so we don't have to write a whole loop to test an instruction.

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

582

The only instructions that I can think of would be indexed load/stores, but they'll be caught by isVectorPredicated. If/When we can enable reductions, we could have an instruction producing two scalar values and then this logic will need to be reorganised a bit.

[SjoerdMeijer]

LGTM, with one nit inline.

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

582

okidoki, then probably just remove this, or replace it with an assert?

llvm/lib/Target/ARM/ARMLowOverheadLoops.cpp

                  for (auto *MI : ElementChain)
                    dbgs() << " - " << *MI);
       ToRemove.insert(ElementChain.begin(), ElementChain.end());
     }
   }
   return true;
 }
 
+static bool isVectorPredicated(MachineInstr *MI) {
+  int PIdx = llvm::findFirstVPTPredOperandIdx(*MI);
+  return PIdx != -1 && MI->getOperand(PIdx + 1).getReg() == ARM::VPR;
+}
+
+static bool isRegInClass(const MachineOperand &MO,
+                         const TargetRegisterClass *Class) {
+  return MO.isReg() && MO.getReg() && Class->contains(MO.getReg());
+}
+
 bool LowOverheadLoop::ValidateLiveOuts() const {
   // Collect Q-regs that are live in the exit blocks. We don't collect scalars
   // because they won't be affected by lane predication.
   const TargetRegisterClass *QPRs = TRI.getRegClass(ARM::MQPRRegClassID);
   SmallSet<Register, 2> LiveOuts;
-  SmallVector<MachineBasicBlock*, 2> ExitBlocks;
+  SmallVector<MachineBasicBlock *, 2> ExitBlocks;
   ML.getExitBlocks(ExitBlocks);
   for (auto *MBB : ExitBlocks)
     for (const MachineBasicBlock::RegisterMaskPair &RegMask : MBB->liveins())
       if (QPRs->contains(RegMask.PhysReg))
         LiveOuts.insert(RegMask.PhysReg);
 
   // Collect the instructions in the loop body that define the live-out values.
-  SmallPtrSet<MachineInstr*, 2> LiveMIs;
+  SmallPtrSet<MachineInstr *, 2> LiveMIs;
+  assert(ML.getNumBlocks() == 1 && "Expected single block loop!");
   MachineBasicBlock *MBB = ML.getHeader();
   for (auto Reg : LiveOuts)
     if (auto *MI = RDA.getLocalLiveOutMIDef(MBB, Reg))
       LiveMIs.insert(MI);
 
   LLVM_DEBUG(dbgs() << "ARM Loops: Found loop live-outs:\n";
              for (auto *MI : LiveMIs)
                dbgs() << " - " << *MI);
   // We've already validated that any VPT predication within the loop will be
   // equivalent when we perform the predication transformation; so we know that
   // any VPT predicated instruction is predicated upon VCTP. Any live-out
   // instruction needs to be predicated, so check this here.
-  for (auto *MI : LiveMIs) {
-    int PIdx = llvm::findFirstVPTPredOperandIdx(*MI);
-    if (PIdx == -1 || MI->getOperand(PIdx+1).getReg() != ARM::VPR)
+  for (auto *MI : LiveMIs)
+    if (!isVectorPredicated(MI))
       return false;
+
+  // We want to find out if the tail-predicated version of this loop will
+  // produce the same values as the loop in its original form. For this to
+  // be true, the newly inserted implicit predication must not change the
+  // the results. All MVE loads and stores have to be predicated, so we know
+  // that any load operands, or stored results are equivalent already.
+  // Other explicitly predicated instructions will perform the same operation
+  // in the original loop and the tail-predicated form too. Here, we call
+  // predicated instructions 'Known' and their users are also Known if it
+  // overwrites an Known operand. This is because tail predication will mean
+  // that the false lanes are not updated, but we know the output because we
+  // know the input.
+  // For any 'Unknown' instructions, we can check that they're only consumed by
+  // Known instructions because it means that the unknown false lane(s) are
+  // replaced with known lane(s).
+  // What this should result in is that all instructions are dependent upon
+  // predicated inputs. It also means that for an iteration of the loop, a
+  // register should hold the same value whether tail predication has happened
+  // or not, with the exception when the differences are masked away by their
+  // user(s) and not observable elsewhere.
+  SetVector<MachineInstr *> Unknowns;
+  SmallPtrSet<MachineInstr *, 4> Knowns;
+  for (auto &MI : *MBB) {
+    if (isVectorPredicated(&MI)) {
+      Knowns.insert(&MI);
+      continue;
+    }
+
+    // Only evaluate instructions which produce a single value.
+    if (MI.getNumDefs() != 1 || !MI.defs().begin()->isReg()) {
+      Unknowns.insert(&MI);
+      continue;
+    }
+
+    Register DefReg = MI.defs().begin()->getReg();
+    for (auto &MO : MI.operands()) {
+      if (!isRegInClass(MO, QPRs) || !MO.isUse() || MO.getReg() != DefReg)
+        continue;
+
+      // If this instruction overwrites one of its operands, and that register
+      // has known lanes, then this instruction also has known predicated false
+      // lanes.
+      if (auto *OpDef = RDA.getMIOperand(&MI, MO)) {
+        if (Knowns.count(OpDef)) {
+          Knowns.insert(&MI);
+          break;
+        }
+      }
+    }
+    if (!Knowns.count(&MI))
+      Unknowns.insert(&MI);
   }
 
+  auto HasKnownUsers = [this](MachineInstr *MI, const MachineOperand &MO,
+                              SmallPtrSetImpl<MachineInstr *> &Knowns) {
+    SmallPtrSet<MachineInstr *, 2> Uses;
+    RDA.getGlobalUses(MI, MO.getReg(), Uses);
+    for (auto *Use : Uses) {
+      if (Use != MI && !Knowns.count(Use))
+        return false;
+    }
+    return true;
+  };
+
+  // Now for all the unknown values, see if they're only consumed by known
+  // instructions.
+  for (auto *MI : reverse(Unknowns)) {
+    for (auto &MO : MI->operands()) {
+      if (!isRegInClass(MO, QPRs) || !MO.isDef())
+        continue;
+      if (!HasKnownUsers(MI, MO, Knowns)) {
+        LLVM_DEBUG(dbgs() << "ARM Loops: Found an unknown def of : "
+                          << TRI.getRegAsmName(MO.getReg()) << " at " << *MI);
+        return false;
+      }
+    }
+    Knowns.insert(MI);
+  }
   return true;
 }
 
 void LowOverheadLoop::CheckLegality(ARMBasicBlockUtils *BBUtils) {
   if (Revert)
     return;
 
   if (!End->getOperand(1).isMBB())

llvm/test/CodeGen/Thumb2/LowOverheadLoops/invariant-qreg.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -run-pass=arm-low-overhead-loops %s -o - | FileCheck %s
+
+--- |
+  define dso_local <4 x i32> @invariant_use_store(i16* nocapture readonly %a, i32* %c, i32 %N, <4 x i32> %pass) {
+  entry:
+    %cmp9 = icmp eq i32 %N, 0
+    %tmp = add i32 %N, 3
+    %tmp1 = lshr i32 %tmp, 2
+    %tmp2 = shl nuw i32 %tmp1, 2
+    %tmp3 = add i32 %tmp2, -4
+    %tmp4 = lshr i32 %tmp3, 2
+    %tmp5 = add nuw nsw i32 %tmp4, 1
+    br i1 %cmp9, label %exit, label %vector.ph
+
+  vector.ph:                                        ; preds = %entry
+    call void @llvm.set.loop.iterations.i32(i32 %tmp5)
+    br label %vector.body
+
+  vector.body:                                      ; preds = %vector.body, %vector.ph
+    %lsr.iv1 = phi i32 [ %lsr.iv.next, %vector.body ], [ %tmp5, %vector.ph ]
+    %lsr.iv20 = phi i32* [ %scevgep20, %vector.body ], [ %c, %vector.ph ]
+    %lsr.iv = phi i16* [ %scevgep, %vector.body ], [ %a, %vector.ph ]
+    %vec.phi = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %tmp13, %vector.body ]
+    %tmp7 = phi i32 [ %N, %vector.ph ], [ %tmp9, %vector.body ]
+    %lsr.iv17 = bitcast i16* %lsr.iv to <4 x i16>*
+    %lsr.store = bitcast i32* %lsr.iv20 to <4 x i32>*
+    %tmp8 = call <4 x i1> @llvm.arm.mve.vctp32(i32 %tmp7)
+    %tmp9 = sub i32 %tmp7, 4
+    %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %lsr.iv17, i32 2, <4 x i1> %tmp8, <4 x i16> undef)
+    %tmp10 = sext <4 x i16> %wide.masked.load to <4 x i32>
+    %tmp12 = mul nsw <4 x i32> %pass, %tmp10
+    %tmp13 = add <4 x i32> %tmp12, %vec.phi
+    call void @llvm.masked.store.v4i32.p0v4i32(<4 x i32> %tmp13, <4 x i32>* %lsr.store, i32 4, <4 x i1> %tmp8)
+    %scevgep = getelementptr i16, i16* %lsr.iv, i32 4
+    %scevgep20 = getelementptr i32, i32* %lsr.iv20, i32 4
+    %tmp14 = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %lsr.iv1, i32 1)
+    %tmp15 = icmp ne i32 %tmp14, 0
+    %lsr.iv.next = add nsw i32 %lsr.iv1, -1
+    br i1 %tmp15, label %vector.body, label %exit
+
+  exit:                                             ; preds = %vector.body, %entry
+    ret <4 x i32> %pass
+  }
+
+  define dso_local i32 @invariant_mul_use_reduce(i16* nocapture readonly %a, i32* %c, i32 %N, <4 x i32> %pass) {
+  entry:
+    %cmp9 = icmp eq i32 %N, 0
+    %tmp = add i32 %N, 3
+    %tmp1 = lshr i32 %tmp, 2
+    %tmp2 = shl nuw i32 %tmp1, 2
+    %tmp3 = add i32 %tmp2, -4
+    %tmp4 = lshr i32 %tmp3, 2
+    %tmp5 = add nuw nsw i32 %tmp4, 1
+    br i1 %cmp9, label %exit, label %vector.ph
+
+  vector.ph:                                        ; preds = %entry
+    call void @llvm.set.loop.iterations.i32(i32 %tmp5)
+    br label %vector.body
+
+  vector.body:                                      ; preds = %vector.body, %vector.ph
+    %lsr.iv1 = phi i32 [ %lsr.iv.next, %vector.body ], [ %tmp5, %vector.ph ]
+    %lsr.iv = phi i16* [ %scevgep, %vector.body ], [ %a, %vector.ph ]
+    %tmp7 = phi i32 [ %N, %vector.ph ], [ %tmp9, %vector.body ]
+    %lsr.iv17 = bitcast i16* %lsr.iv to <4 x i16>*
+    %tmp8 = call <4 x i1> @llvm.arm.mve.vctp32(i32 %tmp7)
+    %tmp9 = sub i32 %tmp7, 4
+    %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %lsr.iv17, i32 2, <4 x i1> %tmp8, <4 x i16> undef)
+    %tmp10 = sext <4 x i16> %wide.masked.load to <4 x i32>
+    %tmp12 = mul nsw <4 x i32> %pass, %tmp10
+    %tmp13 = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> %tmp12)
+    %scevgep = getelementptr i16, i16* %lsr.iv, i32 4
+    %tmp15 = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %lsr.iv1, i32 1)
+    %tmp16 = icmp ne i32 %tmp15, 0
+    %lsr.iv.next = add nsw i32 %lsr.iv1, -1
+    br i1 %tmp16, label %vector.body, label %exit
+
+  exit:                                             ; preds = %vector.body, %entry
+    %res = phi i32 [ 0, %entry ], [ %tmp13, %vector.body ]
+    ret i32 %res
+  }
+
+  define dso_local i32 @invariant_add_use_reduce(i16* nocapture readonly %a, i32* %c, i32 %N, <4 x i32> %pass) {
+  entry:
+    %cmp9 = icmp eq i32 %N, 0
+    %tmp = add i32 %N, 3
+    %tmp1 = lshr i32 %tmp, 2
+    %tmp2 = shl nuw i32 %tmp1, 2
+    %tmp3 = add i32 %tmp2, -4
+    %tmp4 = lshr i32 %tmp3, 2
+    %tmp5 = add nuw nsw i32 %tmp4, 1
+    br i1 %cmp9, label %exit, label %vector.ph
+
+  vector.ph:                                        ; preds = %entry
+    call void @llvm.set.loop.iterations.i32(i32 %tmp5)
+    br label %vector.body
+
+  vector.body:                                      ; preds = %vector.body, %vector.ph
+    %lsr.iv1 = phi i32 [ %lsr.iv.next, %vector.body ], [ %tmp5, %vector.ph ]
+    %lsr.iv = phi i16* [ %scevgep, %vector.body ], [ %a, %vector.ph ]
+    %tmp7 = phi i32 [ %N, %vector.ph ], [ %tmp9, %vector.body ]
+    %lsr.iv17 = bitcast i16* %lsr.iv to <4 x i16>*
+    %tmp8 = call <4 x i1> @llvm.arm.mve.vctp32(i32 %tmp7)
+    %tmp9 = sub i32 %tmp7, 4
+    %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %lsr.iv17, i32 2, <4 x i1> %tmp8, <4 x i16> undef)
+    %tmp10 = sext <4 x i16> %wide.masked.load to <4 x i32>
+    %tmp12 = add nsw <4 x i32> %pass, %tmp10
+    %tmp13 = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> %tmp12)
+    %scevgep = getelementptr i16, i16* %lsr.iv, i32 4
+    %tmp15 = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %lsr.iv1, i32 1)
+    %tmp16 = icmp ne i32 %tmp15, 0
+    %lsr.iv.next = add nsw i32 %lsr.iv1, -1
+    br i1 %tmp16, label %vector.body, label %exit
+
+  exit:                                             ; preds = %vector.body, %entry
+    %res = phi i32 [ 0, %entry ], [ %tmp13, %vector.body ]
+    ret i32 %res
+  }
+
+  declare i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32>)
+  declare <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>*, i32 immarg, <4 x i1>, <4 x i16>)
+  declare void @llvm.masked.store.v4i32.p0v4i32(<4 x i32>, <4 x i32>*, i32 immarg, <4 x i1>)
+  declare void @llvm.set.loop.iterations.i32(i32)
+  declare i32 @llvm.loop.decrement.reg.i32.i32.i32(i32, i32)
+  declare <4 x i1> @llvm.arm.mve.vctp32(i32)
+
+...
+---
+name:            invariant_use_store
+alignment:       2
+tracksRegLiveness: true
+registers:       []
+liveins:
+  - { reg: '$r0', virtual-reg: '' }
+  - { reg: '$r1', virtual-reg: '' }
+  - { reg: '$r2', virtual-reg: '' }
+frameInfo:
+  stackSize:       8
+  offsetAdjustment: 0
+  maxAlignment:    8
+fixedStack:
+  - { id: 0, type: default, offset: 0, size: 16, alignment: 8, stack-id: default,
+      isImmutable: true, isAliased: false, callee-saved-register: '', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+stack:
+  - { id: 0, name: '', type: spill-slot, offset: -4, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$lr', callee-saved-restored: false,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+  - { id: 1, name: '', type: spill-slot, offset: -8, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$r7', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+callSites:       []
+constants:       []
+machineFunctionInfo: {}
+body:             |
+  ; CHECK-LABEL: name: invariant_use_store
+  ; CHECK: bb.0.entry:
+  ; CHECK:   successors: %bb.3(0x30000000), %bb.1(0x50000000)
+  ; CHECK:   liveins: $lr, $r0, $r1, $r2, $r7
+  ; CHECK:   frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+  ; CHECK:   frame-setup CFI_INSTRUCTION def_cfa_offset 8
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $lr, -4
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $r7, -8
+  ; CHECK:   renamable $r3 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $q0 = MVE_VLDRWU32 killed renamable $r3, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+  ; CHECK:   tCBZ $r2, %bb.3
+  ; CHECK: bb.1.vector.ph:
+  ; CHECK:   successors: %bb.2(0x80000000)
+  ; CHECK:   liveins: $q0, $r0, $r1, $r2
+  ; CHECK:   renamable $q1 = MVE_VMOVimmi32 0, 0, $noreg, undef renamable $q1
+  ; CHECK:   $lr = MVE_DLSTP_32 killed renamable $r2
+  ; CHECK: bb.2.vector.body:
+  ; CHECK:   successors: %bb.2(0x7c000000), %bb.3(0x04000000)
+  ; CHECK:   liveins: $lr, $q0, $q1, $r0, $r1
+  ; CHECK:   renamable $r0, renamable $q2 = MVE_VLDRHS32_post killed renamable $r0, 8, 0, $noreg :: (load 8 from %ir.lsr.iv17, align 2)
+  ; CHECK:   renamable $q2 = nsw MVE_VMULi32 renamable $q0, killed renamable $q2, 0, $noreg, undef renamable $q2
+  ; CHECK:   renamable $q1 = MVE_VADDi32 killed renamable $q2, killed renamable $q1, 0, $noreg, undef renamable $q1
+  ; CHECK:   renamable $r1 = MVE_VSTRWU32_post renamable $q1, killed renamable $r1, 16, 0, killed $noreg :: (store 16 into %ir.lsr.store, align 4)
+  ; CHECK:   $lr = MVE_LETP killed renamable $lr, %bb.2
+  ; CHECK: bb.3.exit:
+  ; CHECK:   liveins: $q0
+  ; CHECK:   renamable $r0, renamable $r1 = VMOVRRD renamable $d0, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r2, renamable $r3 = VMOVRRD killed renamable $d1, 14 /* CC::al */, $noreg, implicit killed $q0
+  ; CHECK:   tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0, implicit killed $r1, implicit killed $r2, implicit killed $r3
+  bb.0.entry:
+    successors: %bb.3(0x30000000), %bb.1(0x50000000)
+    liveins: $r0, $r1, $r2, $r7, $lr
+
+    frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+    frame-setup CFI_INSTRUCTION def_cfa_offset 8
+    frame-setup CFI_INSTRUCTION offset $lr, -4
+    frame-setup CFI_INSTRUCTION offset $r7, -8
+    renamable $r3 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+    renamable $q0 = MVE_VLDRWU32 killed renamable $r3, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+    tCBZ $r2, %bb.3
+
+  bb.1.vector.ph:
+    successors: %bb.2(0x80000000)
+    liveins: $q0, $r0, $r1, $r2
+
+    renamable $r3, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+    renamable $q1 = MVE_VMOVimmi32 0, 0, $noreg, undef renamable $q1
+    renamable $r3 = t2BICri killed renamable $r3, 3, 14 /* CC::al */, $noreg, $noreg
+    renamable $r12 = t2SUBri killed renamable $r3, 4, 14 /* CC::al */, $noreg, $noreg
+    renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+    renamable $r12 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r12, 19, 14 /* CC::al */, $noreg, $noreg
+    t2DoLoopStart renamable $r12
+    $r3 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+
+  bb.2.vector.body:
+    successors: %bb.2(0x7c000000), %bb.3(0x04000000)
+    liveins: $q0, $q1, $r0, $r1, $r2, $r3
+
+    renamable $vpr = MVE_VCTP32 renamable $r2, 0, $noreg
+    MVE_VPST 8, implicit $vpr
+    renamable $r0, renamable $q2 = MVE_VLDRHS32_post killed renamable $r0, 8, 1, renamable $vpr :: (load 8 from %ir.lsr.iv17, align 2)
+    $lr = tMOVr $r3, 14 /* CC::al */, $noreg
+    renamable $q2 = nsw MVE_VMULi32 renamable $q0, killed renamable $q2, 0, $noreg, undef renamable $q2
+    renamable $r3, dead $cpsr = nsw tSUBi8 killed $r3, 1, 14 /* CC::al */, $noreg
+    renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 4, 14 /* CC::al */, $noreg
+    renamable $q1 = MVE_VADDi32 killed renamable $q2, killed renamable $q1, 0, $noreg, undef renamable $q1
+    renamable $lr = t2LoopDec killed renamable $lr, 1
+    MVE_VPST 8, implicit $vpr
+    renamable $r1 = MVE_VSTRWU32_post renamable $q1, killed renamable $r1, 16, 1, killed renamable $vpr :: (store 16 into %ir.lsr.store, align 4)
+    t2LoopEnd killed renamable $lr, %bb.2, implicit-def dead $cpsr
+    tB %bb.3, 14 /* CC::al */, $noreg
+
+  bb.3.exit:
+    liveins: $q0
+
+    renamable $r0, renamable $r1 = VMOVRRD renamable $d0, 14 /* CC::al */, $noreg
+    renamable $r2, renamable $r3 = VMOVRRD killed renamable $d1, 14 /* CC::al */, $noreg, implicit $q0
+    tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0, implicit killed $r1, implicit killed $r2, implicit killed $r3
+
+...
+---
+name:            invariant_mul_use_reduce
+alignment:       2
+tracksRegLiveness: true
+registers:       []
+liveins:
+  - { reg: '$r0', virtual-reg: '' }
+  - { reg: '$r2', virtual-reg: '' }
+frameInfo:
+  stackSize:       8
+  offsetAdjustment: 0
+  maxAlignment:    8
+fixedStack:
+  - { id: 0, type: default, offset: 0, size: 16, alignment: 8, stack-id: default,
+      isImmutable: true, isAliased: false, callee-saved-register: '', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+stack:
+  - { id: 0, name: '', type: spill-slot, offset: -4, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$lr', callee-saved-restored: false,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+  - { id: 1, name: '', type: spill-slot, offset: -8, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$r7', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+callSites:       []
+constants:       []
+machineFunctionInfo: {}
+body:             |
+  ; CHECK-LABEL: name: invariant_mul_use_reduce
+  ; CHECK: bb.0.entry:
+  ; CHECK:   successors: %bb.4(0x30000000), %bb.1(0x50000000)
+  ; CHECK:   liveins: $lr, $r0, $r2, $r7
+  ; CHECK:   frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+  ; CHECK:   frame-setup CFI_INSTRUCTION def_cfa_offset 8
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $lr, -4
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $r7, -8
+  ; CHECK:   tCBZ $r2, %bb.4
+  ; CHECK: bb.1.vector.ph:
+  ; CHECK:   successors: %bb.2(0x80000000)
+  ; CHECK:   liveins: $r0, $r2
+  ; CHECK:   renamable $r1, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r1 = t2BICri killed renamable $r1, 3, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   renamable $r1, dead $cpsr = tSUBi8 killed renamable $r1, 4, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r3 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r1, 19, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   renamable $r1 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $q0 = MVE_VLDRWU32 killed renamable $r1, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+  ; CHECK:   dead $lr = t2DLS renamable $r3
+  ; CHECK:   $r1 = tMOVr killed $r3, 14 /* CC::al */, $noreg
+  ; CHECK: bb.2.vector.body:
+  ; CHECK:   successors: %bb.2(0x7c000000), %bb.3(0x04000000)
+  ; CHECK:   liveins: $q0, $r0, $r1, $r2
+  ; CHECK:   renamable $vpr = MVE_VCTP32 renamable $r2, 0, $noreg
+  ; CHECK:   $lr = tMOVr $r1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r1, dead $cpsr = nsw tSUBi8 killed $r1, 1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 4, 14 /* CC::al */, $noreg
+  ; CHECK:   MVE_VPST 8, implicit $vpr
+  ; CHECK:   renamable $r0, renamable $q1 = MVE_VLDRHS32_post killed renamable $r0, 8, 1, killed renamable $vpr :: (load 8 from %ir.lsr.iv17, align 2)
+  ; CHECK:   renamable $r12 = MVE_VMLADAVu32 renamable $q0, killed renamable $q1, 0, $noreg
+  ; CHECK:   dead $lr = t2LEUpdate killed renamable $lr, %bb.2
+  ; CHECK: bb.3.exit:
+  ; CHECK:   liveins: $r12
+  ; CHECK:   $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+  ; CHECK:   tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+  ; CHECK: bb.4:
+  ; CHECK:   renamable $r12 = t2MOVi 0, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+  ; CHECK:   tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+  bb.0.entry:
+    successors: %bb.4(0x30000000), %bb.1(0x50000000)
+    liveins: $r0, $r2, $r7, $lr
+
+    frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+    frame-setup CFI_INSTRUCTION def_cfa_offset 8
+    frame-setup CFI_INSTRUCTION offset $lr, -4
+    frame-setup CFI_INSTRUCTION offset $r7, -8
+    tCBZ $r2, %bb.4
+
+  bb.1.vector.ph:
+    successors: %bb.2(0x80000000)
+    liveins: $r0, $r2
+
+    renamable $r1, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+    renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+    renamable $r1 = t2BICri killed renamable $r1, 3, 14 /* CC::al */, $noreg, $noreg
+    renamable $r1, dead $cpsr = tSUBi8 killed renamable $r1, 4, 14 /* CC::al */, $noreg
+    renamable $r3 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r1, 19, 14 /* CC::al */, $noreg, $noreg
+    renamable $r1 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+    renamable $q0 = MVE_VLDRWU32 killed renamable $r1, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+    t2DoLoopStart renamable $r3
+    $r1 = tMOVr killed $r3, 14 /* CC::al */, $noreg
+
+  bb.2.vector.body:
+    successors: %bb.2(0x7c000000), %bb.3(0x04000000)
+    liveins: $q0, $r0, $r1, $r2
+
+    renamable $vpr = MVE_VCTP32 renamable $r2, 0, $noreg
+    $lr = tMOVr $r1, 14 /* CC::al */, $noreg
+    renamable $r1, dead $cpsr = nsw tSUBi8 killed $r1, 1, 14 /* CC::al */, $noreg
+    renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 4, 14 /* CC::al */, $noreg
+    MVE_VPST 8, implicit $vpr
+    renamable $r0, renamable $q1 = MVE_VLDRHS32_post killed renamable $r0, 8, 1, killed renamable $vpr :: (load 8 from %ir.lsr.iv17, align 2)
+    renamable $r12 = MVE_VMLADAVu32 renamable $q0, killed renamable $q1, 0, $noreg
+    renamable $lr = t2LoopDec killed renamable $lr, 1
+    t2LoopEnd killed renamable $lr, %bb.2, implicit-def dead $cpsr
+    tB %bb.3, 14 /* CC::al */, $noreg
+
+  bb.3.exit:
+    liveins: $r12
+
+    $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+    tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+
+  bb.4:
+    renamable $r12 = t2MOVi 0, 14 /* CC::al */, $noreg, $noreg
+    $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+    tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+
+...
+---
+name:            invariant_add_use_reduce
+alignment:       2
+tracksRegLiveness: true
+registers:       []
+liveins:
+  - { reg: '$r0', virtual-reg: '' }
+  - { reg: '$r2', virtual-reg: '' }
+frameInfo:
+  stackSize:       8
+  offsetAdjustment: 0
+  maxAlignment:    8
+fixedStack:
+  - { id: 0, type: default, offset: 0, size: 16, alignment: 8, stack-id: default,
+      isImmutable: true, isAliased: false, callee-saved-register: '', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+stack:
+  - { id: 0, name: '', type: spill-slot, offset: -4, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$lr', callee-saved-restored: false,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+  - { id: 1, name: '', type: spill-slot, offset: -8, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$r7', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+callSites:       []
+constants:       []
+machineFunctionInfo: {}
+body:             |
+  ; CHECK-LABEL: name: invariant_add_use_reduce
+  ; CHECK: bb.0.entry:
+  ; CHECK:   successors: %bb.4(0x30000000), %bb.1(0x50000000)
+  ; CHECK:   liveins: $lr, $r0, $r2, $r7
+  ; CHECK:   frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+  ; CHECK:   frame-setup CFI_INSTRUCTION def_cfa_offset 8
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $lr, -4
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $r7, -8
+  ; CHECK:   tCBZ $r2, %bb.4
+  ; CHECK: bb.1.vector.ph:
+  ; CHECK:   successors: %bb.2(0x80000000)
+  ; CHECK:   liveins: $r0, $r2
+  ; CHECK:   renamable $r1, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r1 = t2BICri killed renamable $r1, 3, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   renamable $r1, dead $cpsr = tSUBi8 killed renamable $r1, 4, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r3 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r1, 19, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   renamable $r1 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $q0 = MVE_VLDRWU32 killed renamable $r1, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+  ; CHECK:   dead $lr = t2DLS renamable $r3
+  ; CHECK:   $r1 = tMOVr killed $r3, 14 /* CC::al */, $noreg
+  ; CHECK: bb.2.vector.body:
+  ; CHECK:   successors: %bb.2(0x7c000000), %bb.3(0x04000000)
+  ; CHECK:   liveins: $q0, $r0, $r1, $r2
+  ; CHECK:   renamable $vpr = MVE_VCTP32 renamable $r2, 0, $noreg
+  ; CHECK:   MVE_VPST 8, implicit $vpr
+  ; CHECK:   renamable $r0, renamable $q1 = MVE_VLDRHS32_post killed renamable $r0, 8, 1, killed renamable $vpr :: (load 8 from %ir.lsr.iv17, align 2)
+  ; CHECK:   $lr = tMOVr $r1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $q1 = nsw MVE_VADDi32 renamable $q0, killed renamable $q1, 0, $noreg, undef renamable $q1
+  ; CHECK:   renamable $r1, dead $cpsr = nsw tSUBi8 killed $r1, 1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 4, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r12 = MVE_VADDVu32no_acc killed renamable $q1, 0, $noreg
+  ; CHECK:   dead $lr = t2LEUpdate killed renamable $lr, %bb.2
+  ; CHECK: bb.3.exit:
+  ; CHECK:   liveins: $r12
+  ; CHECK:   $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+  ; CHECK:   tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+  ; CHECK: bb.4:
+  ; CHECK:   renamable $r12 = t2MOVi 0, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+  ; CHECK:   tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+  bb.0.entry:
+    successors: %bb.4(0x30000000), %bb.1(0x50000000)
+    liveins: $r0, $r2, $r7, $lr
+
+    frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+    frame-setup CFI_INSTRUCTION def_cfa_offset 8
+    frame-setup CFI_INSTRUCTION offset $lr, -4
+    frame-setup CFI_INSTRUCTION offset $r7, -8
+    tCBZ $r2, %bb.4
+
+  bb.1.vector.ph:
+    successors: %bb.2(0x80000000)
+    liveins: $r0, $r2
+
+    renamable $r1, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+    renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+    renamable $r1 = t2BICri killed renamable $r1, 3, 14 /* CC::al */, $noreg, $noreg
+    renamable $r1, dead $cpsr = tSUBi8 killed renamable $r1, 4, 14 /* CC::al */, $noreg
+    renamable $r3 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r1, 19, 14 /* CC::al */, $noreg, $noreg
+    renamable $r1 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+    renamable $q0 = MVE_VLDRWU32 killed renamable $r1, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+    t2DoLoopStart renamable $r3
+    $r1 = tMOVr killed $r3, 14 /* CC::al */, $noreg
+
+  bb.2.vector.body:
+    successors: %bb.2(0x7c000000), %bb.3(0x04000000)
+    liveins: $q0, $r0, $r1, $r2
+
+    renamable $vpr = MVE_VCTP32 renamable $r2, 0, $noreg
+    MVE_VPST 8, implicit $vpr
+    renamable $r0, renamable $q1 = MVE_VLDRHS32_post killed renamable $r0, 8, 1, killed renamable $vpr :: (load 8 from %ir.lsr.iv17, align 2)
+    $lr = tMOVr $r1, 14 /* CC::al */, $noreg
+    renamable $q1 = nsw MVE_VADDi32 renamable $q0, killed renamable $q1, 0, $noreg, undef renamable $q1
+    renamable $r1, dead $cpsr = nsw tSUBi8 killed $r1, 1, 14 /* CC::al */, $noreg
+    renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 4, 14 /* CC::al */, $noreg
+    renamable $r12 = MVE_VADDVu32no_acc killed renamable $q1, 0, $noreg
+    renamable $lr = t2LoopDec killed renamable $lr, 1
+    t2LoopEnd killed renamable $lr, %bb.2, implicit-def dead $cpsr
+    tB %bb.3, 14 /* CC::al */, $noreg
+
+  bb.3.exit:
+    liveins: $r12
+
+    $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+    tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+
+  bb.4:
+    renamable $r12 = t2MOVi 0, 14 /* CC::al */, $noreg, $noreg
+    $r0 = tMOVr killed $r12, 14 /* CC::al */, $noreg
+    tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0
+
+...

llvm/test/CodeGen/Thumb2/LowOverheadLoops/predicated-invariant.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -run-pass=arm-low-overhead-loops %s -o - | FileCheck %s
+
+--- |
+  define dso_local <4 x i32> @invariant_predicated_add_use(i16* nocapture readonly %a, i32* %c, i32 %N, <4 x i32> %pass) #0 {
+  entry:
+    %cmp9 = icmp eq i32 %N, 0
+    %tmp = add i32 %N, 3
+    %tmp1 = lshr i32 %tmp, 2
+    %tmp2 = shl nuw i32 %tmp1, 2
+    %tmp3 = add i32 %tmp2, -4
+    %tmp4 = lshr i32 %tmp3, 2
+    %tmp5 = add nuw nsw i32 %tmp4, 1
+    br i1 %cmp9, label %exit, label %vector.ph
+
+  vector.ph:                                        ; preds = %entry
+    call void @llvm.set.loop.iterations.i32(i32 %tmp5)
+    br label %vector.body
+
+  vector.body:                                      ; preds = %vector.body, %vector.ph
+    %lsr.iv1 = phi i32 [ %lsr.iv.next, %vector.body ], [ %tmp5, %vector.ph ]
+    %lsr.iv = phi i16* [ %scevgep, %vector.body ], [ %a, %vector.ph ]
+    %tmp7 = phi i32 [ %N, %vector.ph ], [ %tmp9, %vector.body ]
+    %lsr.iv17 = bitcast i16* %lsr.iv to <4 x i16>*
+    %tmp8 = call <4 x i1> @llvm.arm.mve.vctp32(i32 %tmp7)
+    %tmp9 = sub i32 %tmp7, 4
+    %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %lsr.iv17, i32 2, <4 x i1> %tmp8, <4 x i16> undef)
+    %tmp10 = sext <4 x i16> %wide.masked.load to <4 x i32>
+    %acc.next = tail call <4 x i32> @llvm.arm.mve.add.predicated.v4i32.v4i1(<4 x i32> %pass, <4 x i32> %tmp10, <4 x i1> %tmp8, <4 x i32> undef)
+    %scevgep = getelementptr i16, i16* %lsr.iv, i32 4
+    %tmp11 = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %lsr.iv1, i32 1)
+    %tmp12 = icmp ne i32 %tmp11, 0
+    %lsr.iv.next = add nsw i32 %lsr.iv1, -1
+    br i1 %tmp12, label %vector.body, label %exit
+
+  exit:                                             ; preds = %vector.body, %entry
+    %res = phi <4 x i32> [ zeroinitializer, %entry ], [ %acc.next, %vector.body ]
+    ret <4 x i32> %res
+  }
+
+  declare <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>*, i32 immarg, <4 x i1>, <4 x i16>)
+  declare void @llvm.set.loop.iterations.i32(i32)
+  declare i32 @llvm.loop.decrement.reg.i32.i32.i32(i32, i32)
+  declare <4 x i1> @llvm.arm.mve.vctp32(i32)
+  declare <4 x i32> @llvm.arm.mve.add.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, <4 x i1>, <4 x i32>)
+
+...
+---
+name:            invariant_predicated_add_use
+alignment:       2
+tracksRegLiveness: true
+registers:       []
+liveins:
+  - { reg: '$r0', virtual-reg: '' }
+  - { reg: '$r2', virtual-reg: '' }
+frameInfo:
+  stackSize:       8
+  offsetAdjustment: 0
+  maxAlignment:    8
+fixedStack:
+  - { id: 0, type: default, offset: 0, size: 16, alignment: 8, stack-id: default,
+      isImmutable: true, isAliased: false, callee-saved-register: '', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+stack:
+  - { id: 0, name: '', type: spill-slot, offset: -4, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$lr', callee-saved-restored: false,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+  - { id: 1, name: '', type: spill-slot, offset: -8, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$r7', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+callSites:       []
+constants:       []
+machineFunctionInfo: {}
+body:             |
+  ; CHECK-LABEL: name: invariant_predicated_add_use
+  ; CHECK: bb.0.entry:
+  ; CHECK:   successors: %bb.3(0x30000000), %bb.1(0x50000000)
+  ; CHECK:   liveins: $lr, $r0, $r2, $r7
+  ; CHECK:   frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+  ; CHECK:   frame-setup CFI_INSTRUCTION def_cfa_offset 8
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $lr, -4
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $r7, -8
+  ; CHECK:   tCBZ $r2, %bb.3
+  ; CHECK: bb.1.vector.ph:
+  ; CHECK:   successors: %bb.2(0x80000000)
+  ; CHECK:   liveins: $r0, $r2
+  ; CHECK:   renamable $r1 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $q0 = MVE_VLDRWU32 killed renamable $r1, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+  ; CHECK:   $lr = MVE_DLSTP_32 killed renamable $r2
+  ; CHECK: bb.2.vector.body:
+  ; CHECK:   successors: %bb.2(0x7c000000), %bb.4(0x04000000)
+  ; CHECK:   liveins: $lr, $q0, $r0
+  ; CHECK:   renamable $r0, renamable $q1 = MVE_VLDRHS32_post killed renamable $r0, 8, 0, $noreg :: (load 8 from %ir.lsr.iv17, align 2)
+  ; CHECK:   renamable $q1 = MVE_VADDi32 renamable $q0, killed renamable $q1, 0, killed $noreg, undef renamable $q1
+  ; CHECK:   $lr = MVE_LETP killed renamable $lr, %bb.2
+  ; CHECK:   tB %bb.4, 14 /* CC::al */, $noreg
+  ; CHECK: bb.3:
+  ; CHECK:   successors: %bb.4(0x80000000)
+  ; CHECK:   renamable $q1 = MVE_VMOVimmi32 0, 0, $noreg, undef renamable $q1
+  ; CHECK: bb.4.exit:
+  ; CHECK:   liveins: $q1
+  ; CHECK:   renamable $r0, renamable $r1 = VMOVRRD renamable $d2, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r2, renamable $r3 = VMOVRRD killed renamable $d3, 14 /* CC::al */, $noreg, implicit killed $q1
+  ; CHECK:   tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0, implicit killed $r1, implicit killed $r2, implicit killed $r3
+  bb.0.entry:
+    successors: %bb.3(0x30000000), %bb.1(0x50000000)
+    liveins: $r0, $r2, $r7, $lr
+
+    frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r7, killed $lr, implicit-def $sp, implicit $sp
+    frame-setup CFI_INSTRUCTION def_cfa_offset 8
+    frame-setup CFI_INSTRUCTION offset $lr, -4
+    frame-setup CFI_INSTRUCTION offset $r7, -8
+    tCBZ $r2, %bb.3
+
+  bb.1.vector.ph:
+    successors: %bb.2(0x80000000)
+    liveins: $r0, $r2
+
+    renamable $r1, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+    renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+    renamable $r1 = t2BICri killed renamable $r1, 3, 14 /* CC::al */, $noreg, $noreg
+    renamable $r1, dead $cpsr = tSUBi8 killed renamable $r1, 4, 14 /* CC::al */, $noreg
+    renamable $r3 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r1, 19, 14 /* CC::al */, $noreg, $noreg
+    renamable $r1 = tADDrSPi $sp, 2, 14 /* CC::al */, $noreg
+    renamable $q0 = MVE_VLDRWU32 killed renamable $r1, 0, 0, $noreg :: (load 16 from %fixed-stack.0, align 8)
+    t2DoLoopStart renamable $r3
+    $r1 = tMOVr killed $r3, 14 /* CC::al */, $noreg
+
+  bb.2.vector.body:
+    successors: %bb.2(0x7c000000), %bb.4(0x04000000)
+    liveins: $q0, $r0, $r1, $r2
+
+    renamable $vpr = MVE_VCTP32 renamable $r2, 0, $noreg
+    $lr = tMOVr $r1, 14 /* CC::al */, $noreg
+    renamable $r1, dead $cpsr = nsw tSUBi8 killed $r1, 1, 14 /* CC::al */, $noreg
+    renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 4, 14 /* CC::al */, $noreg
+    renamable $lr = t2LoopDec killed renamable $lr, 1
+    MVE_VPST 4, implicit $vpr
+    renamable $r0, renamable $q1 = MVE_VLDRHS32_post killed renamable $r0, 8, 1, renamable $vpr :: (load 8 from %ir.lsr.iv17, align 2)
+    renamable $q1 = MVE_VADDi32 renamable $q0, killed renamable $q1, 1, killed renamable $vpr, undef renamable $q1
+    t2LoopEnd killed renamable $lr, %bb.2, implicit-def dead $cpsr
+    tB %bb.4, 14 /* CC::al */, $noreg
+
+  bb.3:
+    successors: %bb.4(0x80000000)
+
+    renamable $q1 = MVE_VMOVimmi32 0, 0, $noreg, undef renamable $q1
+
+  bb.4.exit:
+    liveins: $q1
+
+    renamable $r0, renamable $r1 = VMOVRRD renamable $d2, 14 /* CC::al */, $noreg
+    renamable $r2, renamable $r3 = VMOVRRD killed renamable $d3, 14 /* CC::al */, $noreg, implicit $q1
+    tPOP_RET 14 /* CC::al */, $noreg, def $r7, def $pc, implicit killed $r0, implicit killed $r1, implicit killed $r2, implicit killed $r3
+
+...

llvm/test/CodeGen/Thumb2/LowOverheadLoops/unpredicated-min.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve %s -run-pass=arm-low-overhead-loops -o - | FileCheck %s
+
+--- |
+  define dso_local void @variant_max_use(i16* nocapture readonly %a, i16* %c, i32 %N) #0 {
+  entry:
+    %cmp9 = icmp eq i32 %N, 0
+    %tmp = add i32 %N, 3
+    %tmp1 = lshr i32 %tmp, 2
+    %tmp2 = shl nuw i32 %tmp1, 2
+    %tmp3 = add i32 %tmp2, -4
+    %tmp4 = lshr i32 %tmp3, 2
+    %tmp5 = add nuw nsw i32 %tmp4, 1
+    br i1 %cmp9, label %exit, label %vector.ph
+
+  vector.ph:                                        ; preds = %entry
+    call void @llvm.set.loop.iterations.i32(i32 %tmp5)
+    br label %vector.body
+
+  vector.body:                                      ; preds = %vector.body, %vector.ph
+    %lsr.iv1 = phi i32 [ %lsr.iv.next, %vector.body ], [ %tmp5, %vector.ph ]
+    %lsr.iv = phi i16* [ %scevgep, %vector.body ], [ %a, %vector.ph ]
+    %lsr.iv.2 = phi i16* [ %scevgep.2, %vector.body ], [ %c, %vector.ph ]
+    %tmp7 = phi i32 [ %N, %vector.ph ], [ %tmp9, %vector.body ]
+    %lsr.iv17 = bitcast i16* %lsr.iv to <8 x i16>*
+    %tmp8 = call <8 x i1> @llvm.arm.mve.vctp16(i32 %tmp7)
+    %tmp9 = sub i32 %tmp7, 8
+    %wide.masked.load = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %lsr.iv17, i32 2, <8 x i1> %tmp8, <8 x i16> undef)
+    %min = tail call i16 @llvm.experimental.vector.reduce.smax.v8i16(<8 x i16> %wide.masked.load)
+    store i16 %min, i16* %lsr.iv.2
+    %scevgep = getelementptr i16, i16* %lsr.iv, i32 8
+    %scevgep.2 = getelementptr i16, i16* %lsr.iv.2, i32 1
+    %tmp10 = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %lsr.iv1, i32 1)
+    %tmp11 = icmp ne i32 %tmp10, 0
+    %lsr.iv.next = add nsw i32 %lsr.iv1, -1
+    br i1 %tmp11, label %vector.body, label %exit
+
+  exit:                                             ; preds = %vector.body, %entry
+    ret void
+  }
+
+  declare <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>*, i32 immarg, <8 x i1>, <8 x i16>)
+  declare void @llvm.set.loop.iterations.i32(i32)
+  declare i32 @llvm.loop.decrement.reg.i32.i32.i32(i32, i32)
+  declare <8 x i1> @llvm.arm.mve.vctp16(i32)
+  declare i16 @llvm.experimental.vector.reduce.smax.v8i16(<8 x i16>)
+
+...
+---
+name:            variant_max_use
+alignment:       2
+tracksRegLiveness: true
+registers:       []
+liveins:
+  - { reg: '$r0', virtual-reg: '' }
+  - { reg: '$r1', virtual-reg: '' }
+  - { reg: '$r2', virtual-reg: '' }
+frameInfo:
+  stackSize:       8
+  offsetAdjustment: 0
+  maxAlignment:    4
+fixedStack:      []
+stack:
+  - { id: 0, name: '', type: spill-slot, offset: -4, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$lr', callee-saved-restored: false,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+  - { id: 1, name: '', type: spill-slot, offset: -8, size: 4, alignment: 4,
+      stack-id: default, callee-saved-register: '$r5', callee-saved-restored: true,
+      debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
+callSites:       []
+constants:       []
+machineFunctionInfo: {}
+body:             |
+  ; CHECK-LABEL: name: variant_max_use
+  ; CHECK: bb.0.entry:
+  ; CHECK:   successors: %bb.1(0x80000000)
+  ; CHECK:   liveins: $lr, $r0, $r1, $r2, $r5
+  ; CHECK:   frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r5, killed $lr, implicit-def $sp, implicit $sp
+  ; CHECK:   frame-setup CFI_INSTRUCTION def_cfa_offset 8
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $lr, -4
+  ; CHECK:   frame-setup CFI_INSTRUCTION offset $r5, -8
+  ; CHECK:   tCMPi8 renamable $r2, 0, 14 /* CC::al */, $noreg, implicit-def $cpsr
+  ; CHECK:   t2IT 0, 8, implicit-def $itstate
+  ; CHECK:   tPOP_RET 0 /* CC::eq */, killed $cpsr, def $r5, def $pc, implicit killed $itstate
+  ; CHECK:   renamable $r3, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r3 = t2BICri killed renamable $r3, 3, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   renamable $r12 = t2SUBri killed renamable $r3, 4, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r3 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r12, 19, 14 /* CC::al */, $noreg, $noreg
+  ; CHECK:   $r12 = t2MOVi16 32768, 14 /* CC::al */, $noreg
+  ; CHECK:   $r12 = t2MOVTi16 killed $r12, 65535, 14 /* CC::al */, $noreg
+  ; CHECK:   dead $lr = t2DLS renamable $r3
+  ; CHECK:   $r5 = tMOVr killed $r3, 14 /* CC::al */, $noreg
+  ; CHECK: bb.1.vector.body:
+  ; CHECK:   successors: %bb.1(0x7c000000), %bb.2(0x04000000)
+  ; CHECK:   liveins: $r0, $r1, $r2, $r5, $r12
+  ; CHECK:   $r3 = tMOVr $r12, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $vpr = MVE_VCTP16 renamable $r2, 0, $noreg
+  ; CHECK:   MVE_VPST 8, implicit $vpr
+  ; CHECK:   renamable $r0, renamable $q0 = MVE_VLDRHU16_post killed renamable $r0, 16, 1, killed renamable $vpr :: (load 16 from %ir.lsr.iv17, align 2)
+  ; CHECK:   renamable $r3 = MVE_VMAXVs16 killed renamable $r3, killed renamable $q0, 0, $noreg
+  ; CHECK:   $lr = tMOVr $r5, 14 /* CC::al */, $noreg
+  ; CHECK:   early-clobber renamable $r1 = t2STRH_POST killed renamable $r3, killed renamable $r1, 2, 14 /* CC::al */, $noreg :: (store 2 into %ir.lsr.iv.2)
+  ; CHECK:   renamable $r5, dead $cpsr = nsw tSUBi8 killed $r5, 1, 14 /* CC::al */, $noreg
+  ; CHECK:   renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 8, 14 /* CC::al */, $noreg
+  ; CHECK:   dead $lr = t2LEUpdate killed renamable $lr, %bb.1
+  ; CHECK: bb.2.exit:
+  ; CHECK:   tPOP_RET 14 /* CC::al */, $noreg, def $r5, def $pc
+  bb.0.entry:
+    successors: %bb.1(0x80000000)
+    liveins: $r0, $r1, $r2, $r5, $lr
+
+    frame-setup tPUSH 14 /* CC::al */, $noreg, killed $r5, killed $lr, implicit-def $sp, implicit $sp
+    frame-setup CFI_INSTRUCTION def_cfa_offset 8
+    frame-setup CFI_INSTRUCTION offset $lr, -4
+    frame-setup CFI_INSTRUCTION offset $r5, -8
+    tCMPi8 renamable $r2, 0, 14 /* CC::al */, $noreg, implicit-def $cpsr
+    t2IT 0, 8, implicit-def $itstate
+    tPOP_RET 0 /* CC::eq */, killed $cpsr, def $r5, def $pc, implicit killed $itstate
+    renamable $r3, dead $cpsr = tADDi3 renamable $r2, 3, 14 /* CC::al */, $noreg
+    renamable $r3 = t2BICri killed renamable $r3, 3, 14 /* CC::al */, $noreg, $noreg
+    renamable $r12 = t2SUBri killed renamable $r3, 4, 14 /* CC::al */, $noreg, $noreg
+    renamable $r3, dead $cpsr = tMOVi8 1, 14 /* CC::al */, $noreg
+    renamable $r3 = nuw nsw t2ADDrs killed renamable $r3, killed renamable $r12, 19, 14 /* CC::al */, $noreg, $noreg
+    $r12 = t2MOVi16 32768, 14 /* CC::al */, $noreg
+    $r12 = t2MOVTi16 killed $r12, 65535, 14 /* CC::al */, $noreg
+    t2DoLoopStart renamable $r3
+    $r5 = tMOVr killed $r3, 14 /* CC::al */, $noreg
+
+  bb.1.vector.body:
+    successors: %bb.1(0x7c000000), %bb.2(0x04000000)
+    liveins: $r0, $r1, $r2, $r5, $r12
+
+    $r3 = tMOVr $r12, 14 /* CC::al */, $noreg
+    renamable $vpr = MVE_VCTP16 renamable $r2, 0, $noreg
+    MVE_VPST 8, implicit $vpr
+    renamable $r0, renamable $q0 = MVE_VLDRHU16_post killed renamable $r0, 16, 1, killed renamable $vpr :: (load 16 from %ir.lsr.iv17, align 2)
+    renamable $r3 = MVE_VMAXVs16 killed renamable $r3, killed renamable $q0, 0, $noreg
+    $lr = tMOVr $r5, 14 /* CC::al */, $noreg
+    early-clobber renamable $r1 = t2STRH_POST killed renamable $r3, killed renamable $r1, 2, 14 /* CC::al */, $noreg :: (store 2 into %ir.lsr.iv.2)
+    renamable $r5, dead $cpsr = nsw tSUBi8 killed $r5, 1, 14 /* CC::al */, $noreg
+    renamable $r2, dead $cpsr = tSUBi8 killed renamable $r2, 8, 14 /* CC::al */, $noreg
+    renamable $lr = t2LoopDec killed renamable $lr, 1
+    t2LoopEnd killed renamable $lr, %bb.1, implicit-def dead $cpsr
+    tB %bb.2, 14 /* CC::al */, $noreg
+
+  bb.2.exit:
+    tPOP_RET 14 /* CC::al */, $noreg, def $r5, def $pc
+
+...