[ModuloSchedule] Peel out prologs and epilogs, generate actual code

Summary:
This extends the PeelingModuloScheduleExpander to generate prolog and epilog code,
and correctly stitch uses through the prolog, kernel, epilog DAG.

The key concept in this patch is to ensure that all transforms are *local*; only a
function of a block and its immediate predecessor and successor. By defining the problem in this way
we can inductively rewrite the entire DAG using only local knowledge that is easy to
reason about.

For example, we assume that all prologs and epilogs are near-perfect clones of the
steady-state kernel. This means that if a block has an instruction that is predicated out,
we can redirect all users of that instruction to that equivalent instruction in our
immediate predecessor. As all blocks are clones, every instruction must have an equivalent in
every other block.

Similarly we can make the assumption by construction that if a value defined in a block is used
outside that block, the only possible user is its immediate successors. We maintain this
even for values that are used outside the loop by creating a limited form of LCSSA.

This code isn't small, but it isn't complex.

Enabled a bunch of testing from Hexagon. There are a couple of tests not enabled yet;
I'm about 80% sure there isn't buggy codegen but the tests are checking for patterns
that we don't produce. Those still need a bit more investigation. In the meantime we
(Google) are happy with the code produced by this on our downstream SMS implementation,
and believe it generates correct code.

Subscribers: mgorny, hiraditya, jsji, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D68205

llvm-svn: 373462

llvm/include/llvm/CodeGen/MachineLoopUtils.h

@@ -0,0 +1,41 @@
+//=- MachineLoopUtils.h - Helper functions for manipulating loops -*- C++ -*-=//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_CODEGEN_MACHINELOOPUTILS_H
+#define LLVM_LIB_CODEGEN_MACHINELOOPUTILS_H
+
+namespace llvm {
+class MachineBasicBlock;
+class MachineRegisterInfo;
+class TargetInstrInfo;
+
+enum LoopPeelDirection {
+  LPD_Front, ///< Peel the first iteration of the loop.
+  LPD_Back   ///< Peel the last iteration of the loop.
+};
+
+/// Peels a single block loop. Loop must have two successors, one of which
+/// must be itself. Similarly it must have two predecessors, one of which must
+/// be itself.
+///
+/// The loop block is copied and inserted into the CFG such that two copies of
+/// the loop follow on from each other. The copy is inserted either before or
+/// after the loop based on Direction.
+///
+/// Phis are updated and an unconditional branch inserted at the end of the
+/// clone so as to execute a single iteration.
+///
+/// The trip count of Loop is not updated.
+MachineBasicBlock *PeelSingleBlockLoop(LoopPeelDirection Direction,
+                                       MachineBasicBlock *Loop,
+                                       MachineRegisterInfo &MRI,
+                                       const TargetInstrInfo *TII);
+
+} // namespace llvm
+
+#endif // LLVM_LIB_CODEGEN_MACHINELOOPUTILS_H

llvm/include/llvm/CodeGen/ModuloSchedule.h

@@ -62,8 +62,10 @@
 
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineLoopUtils.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/CodeGen/TargetSubtargetInfo.h"
+#include <deque>
 #include <vector>
 
 namespace llvm {
@@ -142,9 +144,7 @@ public:
   /// Return the rescheduled instructions in order.
   ArrayRef<MachineInstr *> getInstructions() { return ScheduledInstrs; }
 
-  void dump() {
-    print(dbgs());
-  }
+  void dump() { print(dbgs()); }
   void print(raw_ostream &OS);
 };
 
@@ -270,9 +270,6 @@ public:
 
 /// A reimplementation of ModuloScheduleExpander. It works by generating a
 /// standalone kernel loop and peeling out the prologs and epilogs.
-///
-/// FIXME: This implementation cannot yet generate valid code. It can generate
-/// a correct kernel but cannot peel out prologs and epilogs.
 class PeelingModuloScheduleExpander {
   ModuloSchedule &Schedule;
   MachineFunction &MF;
@@ -281,17 +278,70 @@ class PeelingModuloScheduleExpander {
   const TargetInstrInfo *TII;
   LiveIntervals *LIS;
 
+  /// The original loop block that gets rewritten in-place.
   MachineBasicBlock *BB;
+  /// The original loop preheader.
   MachineBasicBlock *Preheader;
+  /// All prolog and epilog blocks.
+  SmallVector<MachineBasicBlock *, 4> Prologs, Epilogs;
+  /// For every block, the stages that are produced.
+  DenseMap<MachineBasicBlock *, BitVector> LiveStages;
+  /// For every block, the stages that are available. A stage can be available
+  /// but not produced (in the epilog) or produced but not available (in the
+  /// prolog).
+  DenseMap<MachineBasicBlock *, BitVector> AvailableStages;
+
+  /// CanonicalMIs and BlockMIs form a bidirectional map between any of the
+  /// loop kernel clones.
+  DenseMap<MachineInstr *, MachineInstr *> CanonicalMIs;
+  DenseMap<std::pair<MachineBasicBlock *, MachineInstr *>, MachineInstr *>
+      BlockMIs;
+
+  /// State passed from peelKernel to peelPrologAndEpilogs().
+  std::deque<MachineBasicBlock *> PeeledFront, PeeledBack;
+
 public:
   PeelingModuloScheduleExpander(MachineFunction &MF, ModuloSchedule &S,
                                 LiveIntervals *LIS)
       : Schedule(S), MF(MF), ST(MF.getSubtarget()), MRI(MF.getRegInfo()),
         TII(ST.getInstrInfo()), LIS(LIS) {}
 
+  void expand();
+
   /// Runs ModuloScheduleExpander and treats it as a golden input to validate
   /// aspects of the code generated by PeelingModuloScheduleExpander.
   void validateAgainstModuloScheduleExpander();
+
+protected:
+  /// Converts BB from the original loop body to the rewritten, pipelined
+  /// steady-state.
+  void rewriteKernel();
+
+private:
+  /// Peels one iteration of the rewritten kernel (BB) in the specified
+  /// direction.
+  MachineBasicBlock *peelKernel(LoopPeelDirection LPD);
+  /// Peel the kernel forwards and backwards to produce prologs and epilogs,
+  /// and stitch them together.
+  void peelPrologAndEpilogs();
+  /// All prolog and epilog blocks are clones of the kernel, so any produced
+  /// register in one block has an corollary in all other blocks.
+  Register getEquivalentRegisterIn(Register Reg, MachineBasicBlock *BB);
+  /// Change all users of MI, if MI is predicated out
+  /// (LiveStages[MI->getParent()] == false).
+  void rewriteUsesOf(MachineInstr *MI);
+  /// Insert branches between prologs, kernel and epilogs.
+  void fixupBranches();
+  /// Create a poor-man's LCSSA by cloning only the PHIs from the kernel block
+  /// to a block dominated by all prologs and epilogs. This allows us to treat
+  /// the loop exiting block as any other kernel clone.
+  MachineBasicBlock *CreateLCSSAExitingBlock();
+  /// Helper to get the stage of an instruction in the schedule.
+  unsigned getStage(MachineInstr *MI) {
+    if (CanonicalMIs.count(MI))
+      MI = CanonicalMIs[MI];
+    return Schedule.getStage(MI);
+  }
 };
 
 /// Expander that simply annotates each scheduled instruction with a post-instr

llvm/lib/CodeGen/CMakeLists.txt

@@ -80,6 +80,7 @@ add_llvm_library(LLVMCodeGen
   MachineInstr.cpp
   MachineLICM.cpp
   MachineLoopInfo.cpp
+  MachineLoopUtils.cpp
   MachineModuleInfo.cpp
   MachineModuleInfoImpls.cpp
   MachineOperand.cpp

llvm/lib/CodeGen/MachineLoopUtils.cpp

@@ -0,0 +1,132 @@
+//=- MachineLoopUtils.cpp - Functions for manipulating loops ----------------=//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/MachineLoopUtils.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/TargetInstrInfo.h"
+using namespace llvm;
+
+namespace {
+// MI's parent and BB are clones of each other. Find the equivalent copy of MI
+// in BB.
+MachineInstr &findEquivalentInstruction(MachineInstr &MI,
+                                        MachineBasicBlock *BB) {
+  MachineBasicBlock *PB = MI.getParent();
+  unsigned Offset = std::distance(PB->instr_begin(), MachineBasicBlock::instr_iterator(MI));
+  return *std::next(BB->instr_begin(), Offset);
+}
+} // namespace
+
+MachineBasicBlock *llvm::PeelSingleBlockLoop(LoopPeelDirection Direction,
+                                             MachineBasicBlock *Loop,
+                                             MachineRegisterInfo &MRI,
+                                             const TargetInstrInfo *TII) {
+  MachineFunction &MF = *Loop->getParent();
+  MachineBasicBlock *Preheader = *Loop->pred_begin();
+  if (Preheader == Loop)
+    Preheader = *std::next(Loop->pred_begin());
+  MachineBasicBlock *Exit = *Loop->succ_begin();
+  if (Exit == Loop)
+    Exit = *std::next(Loop->succ_begin());
+
+  MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(Loop->getBasicBlock());
+  if (Direction == LPD_Front)
+    MF.insert(Loop->getIterator(), NewBB);
+  else
+    MF.insert(std::next(Loop->getIterator()), NewBB);
+
+  // FIXME: Add DenseMapInfo trait for Register so we can use it as a key.
+  DenseMap<unsigned, Register> Remaps;
+  auto InsertPt = NewBB->end();
+  for (MachineInstr &MI : *Loop) {
+    MachineInstr *NewMI = MF.CloneMachineInstr(&MI);
+    NewBB->insert(InsertPt, NewMI);
+    for (MachineOperand &MO : NewMI->defs()) {
+      Register OrigR = MO.getReg();
+      if (OrigR.isPhysical())
+        continue;
+      Register &R = Remaps[OrigR];
+      R = MRI.createVirtualRegister(MRI.getRegClass(OrigR));
+      MO.setReg(R);
+
+      if (Direction == LPD_Back) {
+        // Replace all uses outside the original loop with the new register.
+        // FIXME: is the use_iterator stable enough to mutate register uses
+        // while iterating?
+        SmallVector<MachineOperand *, 4> Uses;
+        for (auto &Use : MRI.use_operands(OrigR))
+          if (Use.getParent()->getParent() != Loop)
+            Uses.push_back(&Use);
+        for (auto *Use : Uses) {
+          MRI.constrainRegClass(R, MRI.getRegClass(Use->getReg()));
+          Use->setReg(R);
+        }
+      }
+    }
+  }
+
+  for (auto I = NewBB->getFirstNonPHI(); I != NewBB->end(); ++I)
+    for (MachineOperand &MO : I->uses())
+      if (MO.isReg() && Remaps.count(MO.getReg()))
+        MO.setReg(Remaps[MO.getReg()]);
+
+  for (auto I = NewBB->begin(); I->isPHI(); ++I) {
+    MachineInstr &MI = *I;
+    unsigned LoopRegIdx = 3, InitRegIdx = 1;
+    if (MI.getOperand(2).getMBB() != Preheader)
+      std::swap(LoopRegIdx, InitRegIdx);
+    MachineInstr &OrigPhi = findEquivalentInstruction(MI, Loop);
+    assert(OrigPhi.isPHI());
+    if (Direction == LPD_Front) {
+      // When peeling front, we are only left with the initial value from the
+      // preheader.
+      Register R = MI.getOperand(LoopRegIdx).getReg();
+      if (Remaps.count(R))
+        R = Remaps[R];
+      OrigPhi.getOperand(InitRegIdx).setReg(R);
+      MI.RemoveOperand(LoopRegIdx + 1);
+      MI.RemoveOperand(LoopRegIdx + 0);
+    } else {
+      // When peeling back, the initial value is the loop-carried value from
+      // the original loop.
+      Register LoopReg = OrigPhi.getOperand(LoopRegIdx).getReg();
+      MI.getOperand(LoopRegIdx).setReg(LoopReg);
+      MI.RemoveOperand(InitRegIdx + 1);
+      MI.RemoveOperand(InitRegIdx + 0);
+    }
+  }
+
+  DebugLoc DL;
+  if (Direction == LPD_Front) {
+    Preheader->replaceSuccessor(Loop, NewBB);
+    NewBB->addSuccessor(Loop);
+    Loop->replacePhiUsesWith(Preheader, NewBB);
+    if (TII->removeBranch(*Preheader) > 0)
+      TII->insertBranch(*Preheader, NewBB, nullptr, {}, DL);
+    TII->removeBranch(*NewBB);
+    TII->insertBranch(*NewBB, Loop, nullptr, {}, DL);
+  } else {
+    Loop->replaceSuccessor(Exit, NewBB);
+    Exit->replacePhiUsesWith(Loop, NewBB);
+    NewBB->addSuccessor(Exit);
+
+    MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
+    SmallVector<MachineOperand, 4> Cond;
+    bool CanAnalyzeBr = !TII->analyzeBranch(*Loop, TBB, FBB, Cond);
+    (void)CanAnalyzeBr;
+    assert(CanAnalyzeBr && "Must be able to analyze the loop branch!");
+    TII->removeBranch(*Loop);
+    TII->insertBranch(*Loop, TBB == Exit ? NewBB : TBB,
+                      FBB == Exit ? NewBB : FBB, Cond, DL);
+    if (TII->removeBranch(*NewBB) > 0)
+      TII->insertBranch(*NewBB, Exit, nullptr, {}, DL);
+  }
+
+  return NewBB;
+}

llvm/lib/CodeGen/MachinePipeliner.cpp

@@ -557,10 +557,7 @@ void SwingSchedulerDAG::schedule() {
   // The experimental code generator can't work if there are InstChanges.
   if (ExperimentalCodeGen && NewInstrChanges.empty()) {
     PeelingModuloScheduleExpander MSE(MF, MS, &LIS);
-    // Experimental code generation isn't complete yet, but it can partially
-    // validate the code it generates against the original
-    // ModuloScheduleExpander.
-    MSE.validateAgainstModuloScheduleExpander();
+    MSE.expand();
   } else {
     ModuloScheduleExpander MSE(MF, MS, LIS, std::move(NewInstrChanges));
     MSE.expand();

llvm/lib/CodeGen/ModuloSchedule.cpp

@@ -10,6 +10,7 @@
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/CodeGen/LiveIntervals.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineLoopUtils.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/MC/MCContext.h"
@@ -1564,6 +1565,266 @@ private:
 };
 } // namespace
 
+MachineBasicBlock *
+PeelingModuloScheduleExpander::peelKernel(LoopPeelDirection LPD) {
+  MachineBasicBlock *NewBB = PeelSingleBlockLoop(LPD, BB, MRI, TII);
+  if (LPD == LPD_Front)
+    PeeledFront.push_back(NewBB);
+  else
+    PeeledBack.push_front(NewBB);
+  for (auto I = BB->begin(), NI = NewBB->begin(); !I->isTerminator();
+       ++I, ++NI) {
+    CanonicalMIs[&*I] = &*I;
+    CanonicalMIs[&*NI] = &*I;
+    BlockMIs[{NewBB, &*I}] = &*NI;
+    BlockMIs[{BB, &*I}] = &*I;
+  }
+  return NewBB;
+}
+
+void PeelingModuloScheduleExpander::peelPrologAndEpilogs() {
+  BitVector LS(Schedule.getNumStages(), true);
+  BitVector AS(Schedule.getNumStages(), true);
+  LiveStages[BB] = LS;
+  AvailableStages[BB] = AS;
+
+  // Peel out the prologs.
+  LS.reset();
+  for (int I = 0; I < Schedule.getNumStages() - 1; ++I) {
+    LS[I] = 1;
+    Prologs.push_back(peelKernel(LPD_Front));
+    LiveStages[Prologs.back()] = LS;
+    AvailableStages[Prologs.back()] = LS;
+  }
+
+  // Create a block that will end up as the new loop exiting block (dominated by
+  // all prologs and epilogs). It will only contain PHIs, in the same order as
+  // BB's PHIs. This gives us a poor-man's LCSSA with the inductive property
+  // that the exiting block is a (sub) clone of BB. This in turn gives us the
+  // property that any value deffed in BB but used outside of BB is used by a
+  // PHI in the exiting block.
+  MachineBasicBlock *ExitingBB = CreateLCSSAExitingBlock();
+
+  // Push out the epilogs, again in reverse order.
+  // We can't assume anything about the minumum loop trip count at this point,
+  // so emit a fairly complex epilog:
+  //  K[0, 1, 2]     // Kernel runs stages 0, 1, 2
+  //  E0[2] <- P1    // Epilog runs stage 2 only, so the state after is [0].
+  //  E1[1, 2] <- P0 // Epilog 1 moves the last item from stage 0 to stage 2.
+  //
+  // This creates a single-successor single-predecessor sequence of blocks for
+  // each epilog, which are kept this way for simplicity at this stage and
+  // cleaned up by the optimizer later.
+  for (int I = 1; I <= Schedule.getNumStages() - 1; ++I) {
+    Epilogs.push_back(nullptr);
+    for (int J = Schedule.getNumStages() - 1; J >= I; --J) {
+      LS.reset();
+      LS[J] = 1;
+      Epilogs.back() = peelKernel(LPD_Back);
+      LiveStages[Epilogs.back()] = LS;
+      AvailableStages[Epilogs.back()] = AS;
+    }
+  }
+
+  // Now we've defined all the prolog and epilog blocks as a fallthrough
+  // sequence, add the edges that will be followed if the loop trip count is
+  // lower than the number of stages (connecting prologs directly with epilogs).
+  auto PI = Prologs.begin();
+  auto EI = Epilogs.begin();
+  assert(Prologs.size() == Epilogs.size());
+  for (; PI != Prologs.end(); ++PI, ++EI) {
+    MachineBasicBlock *Pred = *(*EI)->pred_begin();
+    (*PI)->addSuccessor(*EI);
+    for (MachineInstr &MI : (*EI)->phis()) {
+      Register Reg = MI.getOperand(1).getReg();
+      MachineInstr *Use = MRI.getUniqueVRegDef(Reg);
+      if (Use && Use->getParent() == Pred)
+        Reg = getEquivalentRegisterIn(Reg, *PI);
+      MI.addOperand(MachineOperand::CreateReg(Reg, /*isDef=*/false));
+      MI.addOperand(MachineOperand::CreateMBB(*PI));
+    }
+  }
+
+  // Create a list of all blocks in order.
+  SmallVector<MachineBasicBlock *, 8> Blocks;
+  llvm::copy(PeeledFront, std::back_inserter(Blocks));
+  Blocks.push_back(BB);
+  llvm::copy(PeeledBack, std::back_inserter(Blocks));
+
+  // Iterate in reverse order over all instructions, remapping as we go.
+  for (MachineBasicBlock *B : reverse(Blocks)) {
+    for (auto I = B->getFirstInstrTerminator()->getReverseIterator();
+         I != std::next(B->getFirstNonPHI()->getReverseIterator());) {
+      MachineInstr *MI = &*I++;
+      rewriteUsesOf(MI);
+    }
+  }
+  // Now all remapping has been done, we're free to optimize the generated code.
+  for (MachineBasicBlock *B : reverse(Blocks))
+    EliminateDeadPhis(B, MRI, LIS);
+  EliminateDeadPhis(ExitingBB, MRI, LIS);
+}
+
+MachineBasicBlock *PeelingModuloScheduleExpander::CreateLCSSAExitingBlock() {
+  MachineFunction &MF = *BB->getParent();
+  MachineBasicBlock *Exit = *BB->succ_begin();
+  if (Exit == BB)
+    Exit = *std::next(BB->succ_begin());
+
+  MachineBasicBlock *NewBB = MF.CreateMachineBasicBlock(BB->getBasicBlock());
+  MF.insert(std::next(BB->getIterator()), NewBB);
+
+  // Clone all phis in BB into NewBB and rewrite.
+  for (MachineInstr &MI : BB->phis()) {
+    auto RC = MRI.getRegClass(MI.getOperand(0).getReg());
+    Register OldR = MI.getOperand(3).getReg();
+    Register R = MRI.createVirtualRegister(RC);
+    SmallVector<MachineInstr *, 4> Uses;
+    for (MachineInstr &Use : MRI.use_instructions(OldR))
+      if (Use.getParent() != BB)
+        Uses.push_back(&Use);
+    for (MachineInstr *Use : Uses)
+      Use->substituteRegister(OldR, R, /*SubIdx=*/0,
+                              *MRI.getTargetRegisterInfo());
+    MachineInstr *NI = BuildMI(NewBB, DebugLoc(), TII->get(TargetOpcode::PHI), R)
+        .addReg(OldR)
+        .addMBB(BB);
+    BlockMIs[{NewBB, &MI}] = NI;
+    CanonicalMIs[NI] = &MI;
+  }
+  BB->replaceSuccessor(Exit, NewBB);
+  Exit->replacePhiUsesWith(BB, NewBB);
+  NewBB->addSuccessor(Exit);
+
+  MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
+  SmallVector<MachineOperand, 4> Cond;
+  bool CanAnalyzeBr = !TII->analyzeBranch(*BB, TBB, FBB, Cond);
+  (void)CanAnalyzeBr;
+  assert(CanAnalyzeBr && "Must be able to analyze the loop branch!");
+  TII->removeBranch(*BB);
+  TII->insertBranch(*BB, TBB == Exit ? NewBB : TBB, FBB == Exit ? NewBB : FBB,
+                    Cond, DebugLoc());
+  TII->insertUnconditionalBranch(*NewBB, Exit, DebugLoc());
+  return NewBB;
+}
+
+Register
+PeelingModuloScheduleExpander::getEquivalentRegisterIn(Register Reg,
+                                                       MachineBasicBlock *BB) {
+  MachineInstr *MI = MRI.getUniqueVRegDef(Reg);
+  unsigned OpIdx = MI->findRegisterDefOperandIdx(Reg);
+  return BlockMIs[{BB, CanonicalMIs[MI]}]->getOperand(OpIdx).getReg();
+}
+
+void PeelingModuloScheduleExpander::rewriteUsesOf(MachineInstr *MI) {
+  if (MI->isPHI()) {
+    // This is an illegal PHI. The loop-carried (desired) value is operand 3,
+    // and it is produced by this block.
+    Register PhiR = MI->getOperand(0).getReg();
+    Register R = MI->getOperand(3).getReg();
+    int RMIStage = getStage(MRI.getUniqueVRegDef(R));
+    if (RMIStage != -1 && !AvailableStages[MI->getParent()].test(RMIStage))
+      R = MI->getOperand(1).getReg();
+    MRI.setRegClass(R, MRI.getRegClass(PhiR));
+    MRI.replaceRegWith(PhiR, R);
+    if (LIS)
+      LIS->RemoveMachineInstrFromMaps(*MI);
+    MI->eraseFromParent();
+    return;
+  }
+
+  int Stage = getStage(MI);
+  if (Stage == -1 || LiveStages.count(MI->getParent()) == 0 ||
+      LiveStages[MI->getParent()].test(Stage))
+    // Instruction is live, no rewriting to do.
+    return;
+
+  for (MachineOperand &DefMO : MI->defs()) {
+    SmallVector<std::pair<MachineInstr *, Register>, 4> Subs;
+    for (MachineInstr &UseMI : MRI.use_instructions(DefMO.getReg())) {
+      // Only PHIs can use values from this block by construction.
+      // Match with the equivalent PHI in B.
+      assert(UseMI.isPHI());
+      Register Reg = getEquivalentRegisterIn(UseMI.getOperand(0).getReg(),
+                                             MI->getParent());
+      Subs.emplace_back(&UseMI, Reg);
+    }
+    for (auto &Sub : Subs)
+      Sub.first->substituteRegister(DefMO.getReg(), Sub.second, /*SubIdx=*/0,
+                                    *MRI.getTargetRegisterInfo());
+  }
+  if (LIS)
+    LIS->RemoveMachineInstrFromMaps(*MI);
+  MI->eraseFromParent();
+}
+
+void PeelingModuloScheduleExpander::fixupBranches() {
+  std::unique_ptr<TargetInstrInfo::PipelinerLoopInfo> Info =
+      TII->analyzeLoopForPipelining(BB);
+  assert(Info);
+
+  // Work outwards from the kernel.
+  bool KernelDisposed = false;
+  int TC = Schedule.getNumStages() - 1;
+  for (auto PI = Prologs.rbegin(), EI = Epilogs.rbegin(); PI != Prologs.rend();
+       ++PI, ++EI, --TC) {
+    MachineBasicBlock *Prolog = *PI;
+    MachineBasicBlock *Fallthrough = *Prolog->succ_begin();
+    MachineBasicBlock *Epilog = *EI;
+    SmallVector<MachineOperand, 4> Cond;
+    Optional<bool> StaticallyGreater =
+        Info->createTripCountGreaterCondition(TC, *Prolog, Cond);
+    if (!StaticallyGreater.hasValue()) {
+      LLVM_DEBUG(dbgs() << "Dynamic: TC > " << TC << "\n");
+      // Dynamically branch based on Cond.
+      TII->removeBranch(*Prolog);
+      TII->insertBranch(*Prolog, Epilog, Fallthrough, Cond, DebugLoc());
+    } else if (*StaticallyGreater == false) {
+      LLVM_DEBUG(dbgs() << "Static-false: TC > " << TC << "\n");
+      // Prolog never falls through; branch to epilog and orphan interior
+      // blocks. Leave it to unreachable-block-elim to clean up.
+      Prolog->removeSuccessor(Fallthrough);
+      for (MachineInstr &P : Fallthrough->phis()) {
+        P.RemoveOperand(2);
+        P.RemoveOperand(1);
+      }
+      TII->removeBranch(*Prolog);
+      TII->insertUnconditionalBranch(*Prolog, Epilog, DebugLoc());
+      KernelDisposed = true;
+    } else {
+      LLVM_DEBUG(dbgs() << "Static-true: TC > " << TC << "\n");
+      // Prolog always falls through; remove incoming values in epilog.
+      Prolog->removeSuccessor(Epilog);
+      for (MachineInstr &P : Epilog->phis()) {
+        P.RemoveOperand(4);
+        P.RemoveOperand(3);
+      }
+    }
+  }
+
+  if (!KernelDisposed) {
+    Info->adjustTripCount(-(Schedule.getNumStages() - 1));
+    Info->setPreheader(Prologs.back());
+  } else {
+    Info->disposed();
+  }
+}
+
+void PeelingModuloScheduleExpander::rewriteKernel() {
+  KernelRewriter KR(*Schedule.getLoop(), Schedule);
+  KR.rewrite();
+}
+
+void PeelingModuloScheduleExpander::expand() {
+  BB = Schedule.getLoop()->getTopBlock();
+  Preheader = Schedule.getLoop()->getLoopPreheader();
+  LLVM_DEBUG(Schedule.dump());
+
+  rewriteKernel();
+  peelPrologAndEpilogs();
+  fixupBranches();
+}
+
 void PeelingModuloScheduleExpander::validateAgainstModuloScheduleExpander() {
   BB = Schedule.getLoop()->getTopBlock();
   Preheader = Schedule.getLoop()->getLoopPreheader();
@@ -1593,6 +1854,7 @@ void PeelingModuloScheduleExpander::validateAgainstModuloScheduleExpander() {
   // Now run the new expansion algorithm.
   KernelRewriter KR(*Schedule.getLoop(), Schedule);
   KR.rewrite();
+  peelPrologAndEpilogs();
 
   // Collect all illegal phis that the new algorithm created. We'll give these
   // to KernelOperandInfo.

llvm/test/CodeGen/Hexagon/pipeliner/swp-phi-start.mir

@@ -1,4 +1,4 @@
-# RUN: llc < %s -x mir -march=hexagon -run-pass=modulo-schedule-test | FileCheck %s
+# RUN: llc < %s -x mir -march=hexagon -run-pass=modulo-schedule-test -pipeliner-experimental-cg=true | FileCheck %s
 
 # Simple check for this sanity test; ensure all instructions are in stage 0 in
 # the prolog and stage 3 in the epilog.

llvm/test/CodeGen/Hexagon/swp-art-deps-rec.ll

@@ -1,7 +1,7 @@
 ; REQUIRES: asserts
 
 ; RUN: llc -march=hexagon -mcpu=hexagonv65 -O3 -debug-only=pipeliner \
-; RUN: < %s 2>&1 | FileCheck %s
+; RUN: < %s 2>&1 -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the artificial dependences are ignored while computing the
 ; circuits.

llvm/test/CodeGen/Hexagon/swp-bad-sched.ll

@@ -1,5 +1,5 @@
 ; REQUIRES: asserts
-; RUN: llc -march=hexagon -enable-pipeliner -enable-aa-sched-mi < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -enable-aa-sched-mi < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; CHECK: loop0(
 ; CHECK: loop0(.LBB0_[[LOOP:.]],

llvm/test/CodeGen/Hexagon/swp-carried-1.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -rdf-opt=0 -disable-hexagon-misched -hexagon-initial-cfg-cleanup=0 -lsr-setupcost-depth-limit=1 < %s | FileCheck %s
+; RUN: llc -march=hexagon -rdf-opt=0 -disable-hexagon-misched -hexagon-initial-cfg-cleanup=0 -lsr-setupcost-depth-limit=1 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we generate the correct code when a loop carried value
 ; is scheduled one stage earlier than it's use. The code in

llvm/test/CodeGen/Hexagon/swp-carried-dep1.mir

@@ -1,4 +1,4 @@
-# RUN: llc -mtriple=hexagon -run-pass pipeliner -debug-only=pipeliner %s -o /dev/null 2>&1 | FileCheck %s
+# RUN: llc -mtriple=hexagon -run-pass pipeliner -debug-only=pipeliner %s -o /dev/null 2>&1 -pipeliner-experimental-cg=true | FileCheck %s
 # REQUIRES: asserts
 
 # Test that the loop carried dependence check correctly identifies a recurrence.

llvm/test/CodeGen/Hexagon/swp-carried-dep2.mir

@@ -1,4 +1,4 @@
-# RUN: llc -mtriple=hexagon -run-pass pipeliner -debug-only=pipeliner %s -o /dev/null 2>&1 | FileCheck %s
+# RUN: llc -mtriple=hexagon -run-pass pipeliner -debug-only=pipeliner %s -o /dev/null 2>&1 -pipeliner-experimental-cg=true | FileCheck %s
 # REQUIRES: asserts
 
 # Test that the loop carried dependence check correctly identifies a recurrence

llvm/test/CodeGen/Hexagon/swp-chain-refs.ll

@@ -1,5 +1,5 @@
 ; RUN: llc -march=hexagon -enable-pipeliner=true -stats -o /dev/null < %s \
-; RUN:      2>&1 | FileCheck %s --check-prefix=STATS
+; RUN:      2>&1 -pipeliner-experimental-cg=true | FileCheck %s --check-prefix=STATS
 ; REQUIRES: asserts
 
 ; Test that we do not schedule chained references too far apart,

llvm/test/CodeGen/Hexagon/swp-change-dep1.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=1 < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=1 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we update the offset correctly for loads that are
 ; moved past stores. In these cases, we change the dependences

llvm/test/CodeGen/Hexagon/swp-change-deps.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -hexagon-initial-cfg-cleanup=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -hexagon-initial-cfg-cleanup=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we generate the correct offsets for loads in the prolog
 ; after removing dependences on a post-increment instructions of the

llvm/test/CodeGen/Hexagon/swp-check-offset.ll

@@ -1,6 +1,6 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s | FileCheck %s
-; RUN: llc -march=hexagon -mcpu=hexagonv62 -enable-pipeliner < %s | FileCheck --check-prefix=CHECK-V62 %s
-; RUN: llc -march=hexagon -mcpu=hexagonv65 -enable-pipeliner < %s | FileCheck --check-prefix=CHECK-V65 %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv62 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck --check-prefix=CHECK-V62 %s
+; RUN: llc -march=hexagon -mcpu=hexagonv65 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck --check-prefix=CHECK-V65 %s
 
 ;
 ; Make sure we pipeline the loop and that we generate the correct

llvm/test/CodeGen/Hexagon/swp-const-tc1.ll

@@ -1,7 +1,7 @@
 ; RUN: llc -march=hexagon -enable-pipeliner -enable-pipeliner-opt-size \
 ; RUN:     -verify-machineinstrs -hexagon-initial-cfg-cleanup=0 \
 ; RUN:     -enable-aa-sched-mi=false -hexagon-expand-condsets=0 \
-; RUN:     < %s | FileCheck %s
+; RUN:     < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Disable expand-condsets because it will assert on undefined registers.
 

llvm/test/CodeGen/Hexagon/swp-const-tc2.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -rdf-opt=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -rdf-opt=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we fixup a pipelined loop correctly when the number of
 ; stages is greater than the compile-time loop trip count. In this

llvm/test/CodeGen/Hexagon/swp-const-tc3.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the pipeliner correctly fixes up the pipelined CFG when the loop
 ; has a constant trip count, and the trip count is less than the number of

llvm/test/CodeGen/Hexagon/swp-conv3x3-nested.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 ; XFAIL: *
 ; LSR changes required.
 

llvm/test/CodeGen/Hexagon/swp-copytophi-dag.ll

@@ -1,7 +1,7 @@
 ; REQUIRES: asserts
 ;
 ; RUN: llc -march=hexagon -enable-pipeliner=true -debug-only=pipeliner < %s \
-; RUN: 2>&1 | FileCheck %s
+; RUN: 2>&1 -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the artificial dependence is created as a result of
 ; CopyToPhi DAG mutation.

llvm/test/CodeGen/Hexagon/swp-dep-neg-offset.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -hexagon-initial-cfg-cleanup=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -hexagon-initial-cfg-cleanup=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the code that changes the dependences does not allow
 ; a load with a negative offset to be overlapped with the post

llvm/test/CodeGen/Hexagon/swp-disable-Os.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 ; CHECK: loop0(.LBB0_{{[0-9]+}},#347)
 
 target triple = "hexagon"

llvm/test/CodeGen/Hexagon/swp-epilog-numphis.ll

@@ -1,6 +1,6 @@
 ; XFAIL: *
 ; Needs some fixed in the pipeliner.
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; CHECK: endloop0
 ; CHECK: vmem

llvm/test/CodeGen/Hexagon/swp-epilog-phi2.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=3 < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=3 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 %s.0 = type { i16, i8, i8, i16, i8, i8, i16, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i32, i16, i8, i8, %s.1, [2 x [16 x %s.2]], i32 (i8*, i8*, i8*, i8*, i8*)*, %s.3*, %s.3*, [120 x i8], i8, i8, %s.4*, [2 x [120 x [8 x i8]]], [56 x i8], [2 x [121 x %s.5]], [2 x %s.5], %s.5*, %s.5*, i32, i32, i16, i8, i8, %s.7, %s.9, %s.11, %s.8*, %s.8* }
 %s.1 = type { i8, i8, i8, i8, i8, i8, i8, i8, i32, i8, [16 x i8], i8, [4 x i8], [32 x i16], [32 x i16], [2 x i8], [4 x i8], [2 x [4 x i8]], [2 x [4 x i8]], i32, i32, i16, i8 }

llvm/test/CodeGen/Hexagon/swp-epilog-phi4.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we generate the correct value for a Phi in the epilog
 ; that is for a value defined two stages earlier. An extra copy in the

llvm/test/CodeGen/Hexagon/swp-epilog-phi5.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we use the correct name in an epilog phi for a phi value
 ; that is defined for the last time in the kernel. Previously, we

llvm/test/CodeGen/Hexagon/swp-epilog-phi8.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -mno-pairing -mno-compound -hexagon-initial-cfg-cleanup=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -mno-pairing -mno-compound -hexagon-initial-cfg-cleanup=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 ; XFAIL: *
 
 ; Test that we generate the correct phi names in the epilog when the pipeliner

llvm/test/CodeGen/Hexagon/swp-kernel-phi1.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner-opt-size -hexagon-initial-cfg-cleanup=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner-opt-size -hexagon-initial-cfg-cleanup=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we generate the correct names for the phis in the kernel for the
 ; incoming values. In this case, the loop contains a phi and has another phi

llvm/test/CodeGen/Hexagon/swp-large-rec.ll

@@ -1,6 +1,6 @@
 ; RUN: llc -march=hexagon -enable-pipeliner -stats \
 ; RUN:     -pipeliner-prune-loop-carried=false -fp-contract=fast \
-; RUN:     -o /dev/null < %s 2>&1 | FileCheck %s --check-prefix=STATS
+; RUN:     -o /dev/null < %s 2>&1 -pipeliner-experimental-cg=true | FileCheck %s --check-prefix=STATS
 ; REQUIRES: asserts
 
 ; That that we do not pipeline this loop. The recurrence is too large. If

llvm/test/CodeGen/Hexagon/swp-listen-loop3.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -pipeliner-ignore-recmii -pipeliner-max-stages=2 -enable-pipeliner < %s | FileCheck %s
+; RUN: llc -march=hexagon -pipeliner-ignore-recmii -pipeliner-max-stages=2 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; This is a loop we pipeline to three packets, though we could do bettter.
 

llvm/test/CodeGen/Hexagon/swp-loop-carried-unknown.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -hexagon-initial-cfg-cleanup=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -hexagon-initial-cfg-cleanup=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the pipeliner schedules a store before the load in which there is a
 ; loop carried dependence. Previously, the loop carried dependence wasn't added

llvm/test/CodeGen/Hexagon/swp-lots-deps.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -stats -o /dev/null < %s 2>&1 | FileCheck %s --check-prefix=STATS
+; RUN: llc -march=hexagon -enable-pipeliner -stats -o /dev/null < %s 2>&1 -pipeliner-experimental-cg=true | FileCheck %s --check-prefix=STATS
 ; REQUIRES: asserts
 
 ; STATS: 1 pipeliner        - Number of loops software pipelined

llvm/test/CodeGen/Hexagon/swp-max.ll

@@ -1,5 +1,5 @@
 ; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner \
-; RUN:     -pipeliner-max-stages=2 < %s | FileCheck %s
+; RUN:     -pipeliner-max-stages=2 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 @A = global [8 x i32] [i32 4, i32 -3, i32 5, i32 -2, i32 -1, i32 2, i32 6, i32 -2], align 8
 

llvm/test/CodeGen/Hexagon/swp-maxstart.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -O3 < %s | FileCheck %s
+; RUN: llc -march=hexagon -O3 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the MinStart computation, which is based upon the length
 ; of the chain edges, is computed correctly. A bug in the code allowed

llvm/test/CodeGen/Hexagon/swp-memrefs-epilog.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -O2 -fp-contract=fast < %s | FileCheck %s
+; RUN: llc -march=hexagon -O2 -fp-contract=fast < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the memoperands for instructions in the epilog are updated
 ; correctly. Previously, the pipeliner updated the offset for the memoperands

llvm/test/CodeGen/Hexagon/swp-multi-loops.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Make sure we attempt to pipeline all inner most loops.
 

llvm/test/CodeGen/Hexagon/swp-new-phi.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=2 < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=2 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the generatePhi code doesn't rename a a Phi instruction that's defined
 ; in the same block.  The bug causes a Phi to incorrectly depend on another Phi.

llvm/test/CodeGen/Hexagon/swp-order-copies.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the instruction ordering code in the pipeliner fixes up dependences
 ; between post-increment register definitions and uses so that the register

llvm/test/CodeGen/Hexagon/swp-order-deps7.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the pipeliner cause an assert and correctly pipelines the
 ; loop.

llvm/test/CodeGen/Hexagon/swp-order.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that when we order instructions in a packet we check for
 ; order dependences so that the source of an order dependence

llvm/test/CodeGen/Hexagon/swp-phi-ch-offset.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=2 < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -pipeliner-max-stages=2 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we generate the correct offsets after we removed unneeded
 ; chain dependences between Phis and generated a better pipeline.

llvm/test/CodeGen/Hexagon/swp-phi-chains.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -debug-only=pipeliner < %s -o - 2>&1 | FileCheck %s
+; RUN: llc -march=hexagon -debug-only=pipeliner < %s -o - 2>&1 -pipeliner-experimental-cg=true | FileCheck %s
 ; REQUIRES: asserts
 
 ; Test that there is a chain edge between two dependent Phis.

llvm/test/CodeGen/Hexagon/swp-phi-dep.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -hexagon-initial-cfg-cleanup=0 -enable-pipeliner -pipeliner-max-stages=2 < %s | FileCheck %s
+; RUN: llc -march=hexagon -hexagon-initial-cfg-cleanup=0 -enable-pipeliner -pipeliner-max-stages=2 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Check that the pipelined code uses the proper address in the
 ; prolog and the kernel. The bug occurs when the address computation

llvm/test/CodeGen/Hexagon/swp-phi-ref.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -enable-bsb-sched=0 -join-liveintervals=false < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -enable-bsb-sched=0 -join-liveintervals=false < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; XFAIL: *
 ; This test is failing after post-ra machine sinking.

llvm/test/CodeGen/Hexagon/swp-pragma-disable.ii

@@ -1,5 +1,5 @@
 ; RUN: llc -disable-lsr -march=hexagon -enable-pipeliner  \
-; RUN:     -debug-only=pipeliner < %s 2>&1 > /dev/null | FileCheck %s
+; RUN:     -debug-only=pipeliner < %s 2>&1 > /dev/null -pipeliner-experimental-cg=true | FileCheck %s
 ; REQUIRES: asserts
 ;
 ; Test that checks if pipeliner disabled by pragma 

llvm/test/CodeGen/Hexagon/swp-pragma-initiation-interval.ii

@@ -1,5 +1,5 @@
 ; RUN: llc -disable-lsr -march=hexagon -enable-pipeliner  \
-; RUN:     -debug-only=pipeliner < %s 2>&1 > /dev/null | FileCheck %s
+; RUN:     -debug-only=pipeliner < %s 2>&1 > /dev/null -pipeliner-experimental-cg=true | FileCheck %s
 ; REQUIRES: asserts
 ;
 ; Test that checks if the II set by pragma was taken by pipeliner.

llvm/test/CodeGen/Hexagon/swp-prolog-phi.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -rdf-opt=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -rdf-opt=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we generate the correct name for a value in a prolog block. The
 ; pipeliner was using an incorrect value for an instruction in the 2nd prolog

llvm/test/CodeGen/Hexagon/swp-rename.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner < %s | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; A test that the Phi rewrite logic is correct.
 

llvm/test/CodeGen/Hexagon/swp-resmii-1.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -debug-only=pipeliner < %s -o - 2>&1 > /dev/null | FileCheck %s
+; RUN: llc -march=hexagon -enable-pipeliner -debug-only=pipeliner < %s -o - 2>&1 > /dev/null -pipeliner-experimental-cg=true | FileCheck %s
 ; REQUIRES: asserts
 
 ; Test that checks that we compute the correct ResMII for haar.

llvm/test/CodeGen/Hexagon/swp-resmii.ll

@@ -1,5 +1,5 @@
 ; RUN: llc -disable-lsr -march=hexagon -enable-pipeliner  \
-; RUN:     -debug-only=pipeliner < %s 2>&1 > /dev/null | FileCheck %s
+; RUN:     -debug-only=pipeliner < %s 2>&1 > /dev/null -pipeliner-experimental-cg=true | FileCheck %s
 ; REQUIRES: asserts
 ;
 ; Test that checks if the ResMII is 1.

llvm/test/CodeGen/Hexagon/swp-reuse-phi-6.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon < %s | FileCheck %s
+; RUN: llc -march=hexagon < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that the pipeliner generates correct code when attempting to reuse
 ; an existing phi. This test case contains a phi that references another

llvm/test/CodeGen/Hexagon/swp-sigma.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -O2 < %s | FileCheck %s
+; RUN: llc -march=hexagon -O2 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; We do not pipeline sigma yet, but the non-pipelined version
 ; with good scheduling is pretty fast. The compiler generates

llvm/test/CodeGen/Hexagon/swp-stages4.ll

@@ -1,11 +1,11 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner -pipeliner-max-stages=2 -disable-block-placement=0 -hexagon-bit=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner -pipeliner-max-stages=2 -disable-block-placement=0 -hexagon-bit=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Test that we rename registers correctly for multiple stages when there is a
 ; Phi and depends upon another Phi.
 
 ; CHECK: = and
 ; CHECK: = and
-; CHECK: = and
+; CHECK: r[[REGA:[0-9]+]] = memub(r{{[0-9]+}}+#1)
 ; CHECK: r[[REG0:[0-9]+]] = and(r[[REG1:[0-9]+]],#255)
 ; CHECK-NOT: r[[REG0]] = and(r[[REG1]],#255)
 ; CHECK: loop0(.LBB0_[[LOOP:.]],

llvm/test/CodeGen/Hexagon/swp-stages5.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner -pipeliner-max-stages=2 -hexagon-bit=0 < %s | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner -pipeliner-max-stages=2 -hexagon-bit=0 < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Very similar to swp-stages4.ll, but the pipelined schedule is a little
 ; different.

llvm/test/CodeGen/Hexagon/swp-subreg.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -stats -o /dev/null < %s 2>&1 | FileCheck %s --check-prefix=STATS
+; RUN: llc -march=hexagon -enable-pipeliner -stats -o /dev/null < %s 2>&1 -pipeliner-experimental-cg=true | FileCheck %s --check-prefix=STATS
 ; REQUIRES: asserts
 
 ; We're unable to pipeline a loop with a subreg as an operand of a Phi.

llvm/test/CodeGen/Hexagon/swp-swap.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -stats -o /dev/null < %s 2>&1 | FileCheck %s --check-prefix=STATS
+; RUN: llc -march=hexagon -enable-pipeliner -stats -o /dev/null < %s 2>&1 -pipeliner-experimental-cg=true | FileCheck %s --check-prefix=STATS
 ; REQUIRES: asserts
 
 ; Test that we don't pipeline, incorrectly, the swap operation.

llvm/test/CodeGen/Hexagon/swp-tfri.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=hexagon -enable-pipeliner -hexagon-initial-cfg-cleanup=0 -stats -o /dev/null < %s 2>&1 | FileCheck %s --check-prefix=STATS
+; RUN: llc -march=hexagon -enable-pipeliner -hexagon-initial-cfg-cleanup=0 -stats -o /dev/null < %s 2>&1 -pipeliner-experimental-cg=true | FileCheck %s --check-prefix=STATS
 ; REQUIRES: asserts
 
 ; Check that we handle the case when a value is first defined in the loop.

llvm/test/CodeGen/Hexagon/swp-vect-dotprod.ll

@@ -1,6 +1,6 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s | FileCheck %s
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -O2 < %s | FileCheck %s
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -O3 < %s | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -O2 < %s -pipeliner-experimental-cg=true | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -O3 < %s -pipeliner-experimental-cg=true | FileCheck %s
 ;
 ; Check that we pipeline a vectorized dot product in a single packet.
 ;

llvm/test/CodeGen/Hexagon/swp-vmult.ll

@@ -1,5 +1,5 @@
 ; REQUIRES: to-be-fixed
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
 
 ; Multiply and accumulate
 ; CHECK: mpyi([[REG0:r([0-9]+)]],[[REG1:r([0-9]+)]])

llvm/test/CodeGen/Hexagon/swp-vsum.ll

@@ -1,5 +1,5 @@
-; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s | FileCheck %s
-; RUN: llc -march=hexagon -mcpu=hexagonv60 -enable-pipeliner < %s | FileCheck %s --check-prefix=CHECKV60
+; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv60 -enable-pipeliner < %s -pipeliner-experimental-cg=true | FileCheck %s --check-prefix=CHECKV60
 
 ; Simple vector total.
 ; CHECK: loop0(.LBB0_[[LOOP:.]],