Hexagon: Pass to replace tranfer/copy instructions into combine instruction
authorJyotsna Verma <jverma@codeaurora.org>
Tue, 14 May 2013 18:54:06 +0000 (18:54 +0000)
committerJyotsna Verma <jverma@codeaurora.org>
Tue, 14 May 2013 18:54:06 +0000 (18:54 +0000)
where possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181817 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/Hexagon/CMakeLists.txt
lib/Target/Hexagon/Hexagon.h
lib/Target/Hexagon/HexagonCopyToCombine.cpp [new file with mode: 0644]
lib/Target/Hexagon/HexagonInstrInfo.td
lib/Target/Hexagon/HexagonTargetMachine.cpp
test/CodeGen/Hexagon/args.ll
test/CodeGen/Hexagon/tfr-to-combine.ll [new file with mode: 0644]

index 99d7a3a96394ab4debf98fde2b7393358269de85..57044b27d67faa1ac35985346ccb86d042fa1688 100644 (file)
@@ -35,6 +35,7 @@ add_llvm_target(HexagonCodeGen
   HexagonTargetObjectFile.cpp
   HexagonVLIWPacketizer.cpp
   HexagonNewValueJump.cpp
+  HexagonCopyToCombine.cpp
 )
 
 add_subdirectory(TargetInfo)
index 8e19c61f4003de712bd073cd418143d21e6958f2..b88637ad57248d0cf00eb662a5a98b96e5cd3acd 100644 (file)
@@ -44,6 +44,8 @@ namespace llvm {
   FunctionPass *createHexagonHardwareLoops();
   FunctionPass *createHexagonPeephole();
   FunctionPass *createHexagonFixupHwLoops();
+  FunctionPass *createHexagonNewValueJump();
+  FunctionPass *createHexagonCopyToCombine();
   FunctionPass *createHexagonPacketizer();
   FunctionPass *createHexagonNewValueJump();
 
diff --git a/lib/Target/Hexagon/HexagonCopyToCombine.cpp b/lib/Target/Hexagon/HexagonCopyToCombine.cpp
new file mode 100644 (file)
index 0000000..dd63523
--- /dev/null
@@ -0,0 +1,676 @@
+//===------- HexagonCopyToCombine.cpp - Hexagon Copy-To-Combine Pass ------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+// This pass replaces transfer instructions by combine instructions.
+// We walk along a basic block and look for two combinable instructions and try
+// to move them together. If we can move them next to each other we do so and
+// replace them with a combine instruction.
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "hexagon-copy-combine"
+
+#include "llvm/PassSupport.h"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Support/CodeGen.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include "Hexagon.h"
+#include "HexagonInstrInfo.h"
+#include "HexagonRegisterInfo.h"
+#include "HexagonSubtarget.h"
+#include "HexagonTargetMachine.h"
+#include "HexagonMachineFunctionInfo.h"
+
+using namespace llvm;
+
+static
+cl::opt<bool> IsCombinesDisabled("disable-merge-into-combines",
+                                 cl::Hidden, cl::ZeroOrMore,
+                                 cl::init(false),
+                                 cl::desc("Disable merging into combines"));
+static
+cl::opt<unsigned>
+MaxNumOfInstsBetweenNewValueStoreAndTFR("max-num-inst-between-tfr-and-nv-store",
+                   cl::Hidden, cl::init(4),
+                   cl::desc("Maximum distance between a tfr feeding a store we "
+                            "consider the store still to be newifiable"));
+
+namespace llvm {
+  void initializeHexagonCopyToCombinePass(PassRegistry&);
+}
+
+
+namespace {
+
+class HexagonCopyToCombine : public MachineFunctionPass  {
+  const HexagonInstrInfo *TII;
+  const TargetRegisterInfo *TRI;
+  bool ShouldCombineAggressively;
+
+  DenseSet<MachineInstr *> PotentiallyNewifiableTFR;
+public:
+  static char ID;
+
+  HexagonCopyToCombine() : MachineFunctionPass(ID) {
+    initializeHexagonCopyToCombinePass(*PassRegistry::getPassRegistry());
+  }
+
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  const char *getPassName() const {
+    return "Hexagon Copy-To-Combine Pass";
+  }
+
+  virtual bool runOnMachineFunction(MachineFunction &Fn);
+
+private:
+  MachineInstr *findPairable(MachineInstr *I1, bool &DoInsertAtI1);
+
+  void findPotentialNewifiableTFRs(MachineBasicBlock &);
+
+  void combine(MachineInstr *I1, MachineInstr *I2,
+               MachineBasicBlock::iterator &MI, bool DoInsertAtI1);
+
+  bool isSafeToMoveTogether(MachineInstr *I1, MachineInstr *I2,
+                            unsigned I1DestReg, unsigned I2DestReg,
+                            bool &DoInsertAtI1);
+
+  void emitCombineRR(MachineBasicBlock::iterator &Before, unsigned DestReg,
+                     MachineOperand &HiOperand, MachineOperand &LoOperand);
+
+  void emitCombineRI(MachineBasicBlock::iterator &Before, unsigned DestReg,
+                     MachineOperand &HiOperand, MachineOperand &LoOperand);
+
+  void emitCombineIR(MachineBasicBlock::iterator &Before, unsigned DestReg,
+                     MachineOperand &HiOperand, MachineOperand &LoOperand);
+
+  void emitCombineII(MachineBasicBlock::iterator &Before, unsigned DestReg,
+                     MachineOperand &HiOperand, MachineOperand &LoOperand);
+};
+
+} // End anonymous namespace.
+
+char HexagonCopyToCombine::ID = 0;
+
+INITIALIZE_PASS(HexagonCopyToCombine, "hexagon-copy-combine",
+                "Hexagon Copy-To-Combine Pass", false, false)
+
+static bool isCombinableInstType(MachineInstr *MI,
+                                 const HexagonInstrInfo *TII,
+                                 bool ShouldCombineAggressively) {
+  switch(MI->getOpcode()) {
+  case Hexagon::TFR: {
+    // A COPY instruction can be combined if its arguments are IntRegs (32bit).
+    assert(MI->getOperand(0).isReg() && MI->getOperand(1).isReg());
+
+    unsigned DestReg = MI->getOperand(0).getReg();
+    unsigned SrcReg = MI->getOperand(1).getReg();
+    return Hexagon::IntRegsRegClass.contains(DestReg) &&
+      Hexagon::IntRegsRegClass.contains(SrcReg);
+  }
+
+  case Hexagon::TFRI: {
+    // A transfer-immediate can be combined if its argument is a signed 8bit
+    // value.
+    assert(MI->getOperand(0).isReg() && MI->getOperand(1).isImm());
+    unsigned DestReg = MI->getOperand(0).getReg();
+
+    // Only combine constant extended TFRI if we are in aggressive mode.
+    return Hexagon::IntRegsRegClass.contains(DestReg) &&
+      (ShouldCombineAggressively || isInt<8>(MI->getOperand(1).getImm()));
+  }
+
+  case Hexagon::TFRI_V4: {
+    if (!ShouldCombineAggressively)
+      return false;
+    assert(MI->getOperand(0).isReg() && MI->getOperand(1).isGlobal());
+
+    // Ensure that TargetFlags are MO_NO_FLAG for a global. This is a
+    // workaround for an ABI bug that prevents GOT relocations on combine
+    // instructions
+    if (MI->getOperand(1).getTargetFlags() != HexagonII::MO_NO_FLAG)
+      return false;
+
+    unsigned DestReg = MI->getOperand(0).getReg();
+    return Hexagon::IntRegsRegClass.contains(DestReg);
+  }
+
+  default:
+    break;
+  }
+
+  return false;
+}
+
+static bool isGreaterThan8BitTFRI(MachineInstr *I) {
+  return I->getOpcode() == Hexagon::TFRI &&
+    !isInt<8>(I->getOperand(1).getImm());
+}
+static bool isGreaterThan6BitTFRI(MachineInstr *I) {
+  return I->getOpcode() == Hexagon::TFRI &&
+    !isUInt<6>(I->getOperand(1).getImm());
+}
+
+/// areCombinableOperations - Returns true if the two instruction can be merge
+/// into a combine (ignoring register constraints).
+static bool areCombinableOperations(const TargetRegisterInfo *TRI,
+                                    MachineInstr *HighRegInst,
+                                    MachineInstr *LowRegInst) {
+  assert((HighRegInst->getOpcode() == Hexagon::TFR ||
+          HighRegInst->getOpcode() == Hexagon::TFRI ||
+          HighRegInst->getOpcode() == Hexagon::TFRI_V4) &&
+         (LowRegInst->getOpcode() == Hexagon::TFR ||
+          LowRegInst->getOpcode() == Hexagon::TFRI ||
+          LowRegInst->getOpcode() == Hexagon::TFRI_V4) &&
+         "Assume individual instructions are of a combinable type");
+
+  const HexagonRegisterInfo *QRI =
+    static_cast<const HexagonRegisterInfo *>(TRI);
+
+  // V4 added some combine variations (mixed immediate and register source
+  // operands), if we are on < V4 we can only combine 2 register-to-register
+  // moves and 2 immediate-to-register moves. We also don't have
+  // constant-extenders.
+  if (!QRI->Subtarget.hasV4TOps())
+    return HighRegInst->getOpcode() == LowRegInst->getOpcode() &&
+      !isGreaterThan8BitTFRI(HighRegInst) &&
+      !isGreaterThan6BitTFRI(LowRegInst);
+
+  // There is no combine of two constant extended values.
+  if ((HighRegInst->getOpcode() == Hexagon::TFRI_V4 ||
+       isGreaterThan8BitTFRI(HighRegInst)) &&
+      (LowRegInst->getOpcode() == Hexagon::TFRI_V4 ||
+       isGreaterThan6BitTFRI(LowRegInst)))
+    return false;
+
+  return true;
+}
+
+static bool isEvenReg(unsigned Reg) {
+  assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
+         Hexagon::IntRegsRegClass.contains(Reg));
+  return (Reg - Hexagon::R0) % 2 == 0;
+}
+
+static void removeKillInfo(MachineInstr *MI, unsigned RegNotKilled) {
+  for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
+    MachineOperand &Op = MI->getOperand(I);
+    if (!Op.isReg() || Op.getReg() != RegNotKilled || !Op.isKill())
+      continue;
+    Op.setIsKill(false);
+  }
+}
+
+/// isUnsafeToMoveAccross - Returns true if it is unsafe to move a copy
+/// instruction from \p UseReg to \p DestReg over the instruction \p I.
+bool isUnsafeToMoveAccross(MachineInstr *I, unsigned UseReg, unsigned DestReg,
+                            const TargetRegisterInfo *TRI) {
+  return (UseReg && (I->modifiesRegister(UseReg, TRI))) ||
+          I->modifiesRegister(DestReg, TRI) ||
+          I->readsRegister(DestReg, TRI) ||
+          I->hasUnmodeledSideEffects() ||
+          I->isInlineAsm() || I->isDebugValue();
+}
+
+/// isSafeToMoveTogether - Returns true if it is safe to move I1 next to I2 such
+/// that the two instructions can be paired in a combine.
+bool HexagonCopyToCombine::isSafeToMoveTogether(MachineInstr *I1,
+                                                MachineInstr *I2,
+                                                unsigned I1DestReg,
+                                                unsigned I2DestReg,
+                                                bool &DoInsertAtI1) {
+
+  bool IsImmUseReg = I2->getOperand(1).isImm() || I2->getOperand(1).isGlobal();
+  unsigned I2UseReg = IsImmUseReg ? 0 : I2->getOperand(1).getReg();
+
+  // It is not safe to move I1 and I2 into one combine if I2 has a true
+  // dependence on I1.
+  if (I2UseReg && I1->modifiesRegister(I2UseReg, TRI))
+    return false;
+
+  bool isSafe = true;
+
+  // First try to move I2 towards I1.
+  {
+    // A reverse_iterator instantiated like below starts before I2, and I1
+    // respectively.
+    // Look at instructions I in between I2 and (excluding) I1.
+    MachineBasicBlock::reverse_iterator I(I2),
+      End = --(MachineBasicBlock::reverse_iterator(I1));
+    // At 03 we got better results (dhrystone!) by being more conservative.
+    if (!ShouldCombineAggressively)
+      End = MachineBasicBlock::reverse_iterator(I1);
+    // If I2 kills its operand and we move I2 over an instruction that also
+    // uses I2's use reg we need to modify that (first) instruction to now kill
+    // this reg.
+    unsigned KilledOperand = 0;
+    if (I2->killsRegister(I2UseReg))
+      KilledOperand = I2UseReg;
+    MachineInstr *KillingInstr = 0;
+
+    for (; I != End; ++I) {
+      // If the intervening instruction I:
+      //   * modifies I2's use reg
+      //   * modifies I2's def reg
+      //   * reads I2's def reg
+      //   * or has unmodelled side effects
+      // we can't move I2 across it.
+      if (isUnsafeToMoveAccross(&*I, I2UseReg, I2DestReg, TRI)) {
+        isSafe = false;
+        break;
+      }
+
+      // Update first use of the killed operand.
+      if (!KillingInstr && KilledOperand &&
+          I->readsRegister(KilledOperand, TRI))
+        KillingInstr = &*I;
+    }
+    if (isSafe) {
+      // Update the intermediate instruction to with the kill flag.
+      if (KillingInstr) {
+        bool Added = KillingInstr->addRegisterKilled(KilledOperand, TRI, true);
+        (void)Added; // supress compiler warning
+        assert(Added && "Must successfully update kill flag");
+        removeKillInfo(I2, KilledOperand);
+      }
+      DoInsertAtI1 = true;
+      return true;
+    }
+  }
+
+  // Try to move I1 towards I2.
+  {
+    // Look at instructions I in between I1 and (excluding) I2.
+    MachineBasicBlock::iterator I(I1), End(I2);
+    // At O3 we got better results (dhrystone) by being more conservative here.
+    if (!ShouldCombineAggressively)
+      End = llvm::next(MachineBasicBlock::iterator(I2));
+    IsImmUseReg = I1->getOperand(1).isImm() || I1->getOperand(1).isGlobal();
+    unsigned I1UseReg = IsImmUseReg ? 0 : I1->getOperand(1).getReg();
+    // Track killed operands. If we move accross an instruction that kills our
+    // operand, we need to update the kill information on the moved I1. It kills
+    // the operand now.
+    MachineInstr *KillingInstr = 0;
+    unsigned KilledOperand = 0;
+
+    while(++I != End) {
+      // If the intervening instruction I:
+      //   * modifies I1's use reg
+      //   * modifies I1's def reg
+      //   * reads I1's def reg
+      //   * or has unmodelled side effects
+      //   We introduce this special case because llvm has no api to remove a
+      //   kill flag for a register (a removeRegisterKilled() analogous to
+      //   addRegisterKilled) that handles aliased register correctly.
+      //   * or has a killed aliased register use of I1's use reg
+      //           %D4<def> = TFRI64 16
+      //           %R6<def> = TFR %R9
+      //           %R8<def> = KILL %R8, %D4<imp-use,kill>
+      //      If we want to move R6 = across the KILL instruction we would have
+      //      to remove the %D4<imp-use,kill> operand. For now, we are
+      //      conservative and disallow the move.
+      // we can't move I1 across it.
+      if (isUnsafeToMoveAccross(I, I1UseReg, I1DestReg, TRI) ||
+          // Check for an aliased register kill. Bail out if we see one.
+          (!I->killsRegister(I1UseReg) && I->killsRegister(I1UseReg, TRI)))
+        return false;
+
+      // Check for an exact kill (registers match).
+      if (I1UseReg && I->killsRegister(I1UseReg)) {
+        assert(KillingInstr == 0 && "Should only see one killing instruction");
+        KilledOperand = I1UseReg;
+        KillingInstr = &*I;
+      }
+    }
+    if (KillingInstr) {
+      removeKillInfo(KillingInstr, KilledOperand);
+      // Update I1 to set the kill flag. This flag will later be picked up by
+      // the new COMBINE instruction.
+      bool Added = I1->addRegisterKilled(KilledOperand, TRI);
+      (void)Added; // supress compiler warning
+      assert(Added && "Must successfully update kill flag");
+    }
+    DoInsertAtI1 = false;
+  }
+
+  return true;
+}
+
+/// findPotentialNewifiableTFRs - Finds tranfers that feed stores that could be
+/// newified. (A use of a 64 bit register define can not be newified)
+void
+HexagonCopyToCombine::findPotentialNewifiableTFRs(MachineBasicBlock &BB) {
+  DenseMap<unsigned, MachineInstr *> LastDef;
+  for (MachineBasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) {
+    MachineInstr *MI = I;
+    // Mark TFRs that feed a potential new value store as such.
+    if(TII->mayBeNewStore(MI)) {
+      // Look for uses of TFR instructions.
+      for (unsigned OpdIdx = 0, OpdE = MI->getNumOperands(); OpdIdx != OpdE;
+           ++OpdIdx) {
+        MachineOperand &Op = MI->getOperand(OpdIdx);
+
+        // Skip over anything except register uses.
+        if (!Op.isReg() || !Op.isUse() || !Op.getReg())
+          continue;
+
+        // Look for the defining instruction.
+        unsigned Reg = Op.getReg();
+        MachineInstr *DefInst = LastDef[Reg];
+        if (!DefInst)
+          continue;
+        if (!isCombinableInstType(DefInst, TII, ShouldCombineAggressively))
+          continue;
+
+        // Only close newifiable stores should influence the decision.
+        MachineBasicBlock::iterator It(DefInst);
+        unsigned NumInstsToDef = 0;
+        while (&*It++ != MI)
+          ++NumInstsToDef;
+
+        if (NumInstsToDef > MaxNumOfInstsBetweenNewValueStoreAndTFR)
+          continue;
+
+        PotentiallyNewifiableTFR.insert(DefInst);
+      }
+      // Skip to next instruction.
+      continue;
+    }
+
+    // Put instructions that last defined integer or double registers into the
+    // map.
+    for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
+      MachineOperand &Op = MI->getOperand(I);
+      if (!Op.isReg() || !Op.isDef() || !Op.getReg())
+        continue;
+      unsigned Reg = Op.getReg();
+      if (Hexagon::DoubleRegsRegClass.contains(Reg)) {
+        for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) {
+          LastDef[*SubRegs] = MI;
+        }
+      } else if (Hexagon::IntRegsRegClass.contains(Reg))
+        LastDef[Reg] = MI;
+    }
+  }
+}
+
+bool HexagonCopyToCombine::runOnMachineFunction(MachineFunction &MF) {
+
+  if (IsCombinesDisabled) return false;
+
+  bool HasChanged = false;
+
+  // Get target info.
+  TRI = MF.getTarget().getRegisterInfo();
+  TII = static_cast<const HexagonInstrInfo *>(MF.getTarget().getInstrInfo());
+
+  // Combine aggressively (for code size)
+  ShouldCombineAggressively =
+    MF.getTarget().getOptLevel() <= CodeGenOpt::Default;
+
+  // Traverse basic blocks.
+  for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE;
+       ++BI) {
+    PotentiallyNewifiableTFR.clear();
+    findPotentialNewifiableTFRs(*BI);
+
+    // Traverse instructions in basic block.
+    for(MachineBasicBlock::iterator MI = BI->begin(), End = BI->end();
+        MI != End;) {
+      MachineInstr *I1 = MI++;
+      // Don't combine a TFR whose user could be newified (instructions that
+      // define double registers can not be newified - Programmer's Ref Manual
+      // 5.4.2 New-value stores).
+      if (ShouldCombineAggressively && PotentiallyNewifiableTFR.count(I1))
+        continue;
+
+      // Ignore instructions that are not combinable.
+      if (!isCombinableInstType(I1, TII, ShouldCombineAggressively))
+        continue;
+
+      // Find a second instruction that can be merged into a combine
+      // instruction.
+      bool DoInsertAtI1 = false;
+      MachineInstr *I2 = findPairable(I1, DoInsertAtI1);
+      if (I2) {
+        HasChanged = true;
+        combine(I1, I2, MI, DoInsertAtI1);
+      }
+    }
+  }
+
+  return HasChanged;
+}
+
+/// findPairable - Returns an instruction that can be merged with \p I1 into a
+/// COMBINE instruction or 0 if no such instruction can be found. Returns true
+/// in \p DoInsertAtI1 if the combine must be inserted at instruction \p I1
+/// false if the combine must be inserted at the returned instruction.
+MachineInstr *HexagonCopyToCombine::findPairable(MachineInstr *I1,
+                                                 bool &DoInsertAtI1) {
+  MachineBasicBlock::iterator I2 = llvm::next(MachineBasicBlock::iterator(I1));
+  unsigned I1DestReg = I1->getOperand(0).getReg();
+
+  for (MachineBasicBlock::iterator End = I1->getParent()->end(); I2 != End;
+       ++I2) {
+    // Bail out early if we see a second definition of I1DestReg.
+    if (I2->modifiesRegister(I1DestReg, TRI))
+      break;
+
+    // Ignore non-combinable instructions.
+    if (!isCombinableInstType(I2, TII, ShouldCombineAggressively))
+      continue;
+
+    // Don't combine a TFR whose user could be newified.
+    if (ShouldCombineAggressively && PotentiallyNewifiableTFR.count(I2))
+      continue;
+
+    unsigned I2DestReg = I2->getOperand(0).getReg();
+
+    // Check that registers are adjacent and that the first destination register
+    // is even.
+    bool IsI1LowReg = (I2DestReg - I1DestReg) == 1;
+    bool IsI2LowReg = (I1DestReg - I2DestReg) == 1;
+    unsigned FirstRegIndex = IsI1LowReg ? I1DestReg : I2DestReg;
+    if ((!IsI1LowReg && !IsI2LowReg) || !isEvenReg(FirstRegIndex))
+      continue;
+
+    // Check that the two instructions are combinable. V4 allows more
+    // instructions to be merged into a combine.
+    // The order matters because in a TFRI we might can encode a int8 as the
+    // hi reg operand but only a uint6 as the low reg operand.
+    if ((IsI2LowReg && !areCombinableOperations(TRI, I1, I2)) ||
+        (IsI1LowReg && !areCombinableOperations(TRI, I2, I1)))
+      break;
+
+    if (isSafeToMoveTogether(I1, I2, I1DestReg, I2DestReg,
+                             DoInsertAtI1))
+      return I2;
+
+    // Not safe. Stop searching.
+    break;
+  }
+  return 0;
+}
+
+void HexagonCopyToCombine::combine(MachineInstr *I1, MachineInstr *I2,
+                                   MachineBasicBlock::iterator &MI,
+                                   bool DoInsertAtI1) {
+  // We are going to delete I2. If MI points to I2 advance it to the next
+  // instruction.
+  if ((MachineInstr *)MI == I2) ++MI;
+
+  // Figure out whether I1 or I2 goes into the lowreg part.
+  unsigned I1DestReg = I1->getOperand(0).getReg();
+  unsigned I2DestReg = I2->getOperand(0).getReg();
+  bool IsI1Loreg = (I2DestReg - I1DestReg) == 1;
+  unsigned LoRegDef = IsI1Loreg ? I1DestReg : I2DestReg;
+
+  // Get the double word register.
+  unsigned DoubleRegDest =
+    TRI->getMatchingSuperReg(LoRegDef, Hexagon::subreg_loreg,
+                             &Hexagon::DoubleRegsRegClass);
+  assert(DoubleRegDest != 0 && "Expect a valid register");
+
+
+  // Setup source operands.
+  MachineOperand &LoOperand = IsI1Loreg ? I1->getOperand(1) :
+    I2->getOperand(1);
+  MachineOperand &HiOperand = IsI1Loreg ? I2->getOperand(1) :
+    I1->getOperand(1);
+
+  // Figure out which source is a register and which a constant.
+  bool IsHiReg = HiOperand.isReg();
+  bool IsLoReg = LoOperand.isReg();
+
+  MachineBasicBlock::iterator InsertPt(DoInsertAtI1 ? I1 : I2);
+  // Emit combine.
+  if (IsHiReg && IsLoReg)
+    emitCombineRR(InsertPt, DoubleRegDest, HiOperand, LoOperand);
+  else if (IsHiReg)
+    emitCombineRI(InsertPt, DoubleRegDest, HiOperand, LoOperand);
+  else if (IsLoReg)
+    emitCombineIR(InsertPt, DoubleRegDest, HiOperand, LoOperand);
+  else
+    emitCombineII(InsertPt, DoubleRegDest, HiOperand, LoOperand);
+
+  I1->eraseFromParent();
+  I2->eraseFromParent();
+}
+
+void HexagonCopyToCombine::emitCombineII(MachineBasicBlock::iterator &InsertPt,
+                                         unsigned DoubleDestReg,
+                                         MachineOperand &HiOperand,
+                                         MachineOperand &LoOperand) {
+  DebugLoc DL = InsertPt->getDebugLoc();
+  MachineBasicBlock *BB = InsertPt->getParent();
+
+  // Handle  globals.
+  if (HiOperand.isGlobal()) {
+    BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg)
+      .addGlobalAddress(HiOperand.getGlobal(), HiOperand.getOffset(),
+                        HiOperand.getTargetFlags())
+      .addImm(LoOperand.getImm());
+    return;
+  }
+  if (LoOperand.isGlobal()) {
+    BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_iI_V4), DoubleDestReg)
+      .addImm(HiOperand.getImm())
+      .addGlobalAddress(LoOperand.getGlobal(), LoOperand.getOffset(),
+                        LoOperand.getTargetFlags());
+    return;
+  }
+
+  // Handle constant extended immediates.
+  if (!isInt<8>(HiOperand.getImm())) {
+    assert(isInt<8>(LoOperand.getImm()));
+    BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg)
+      .addImm(HiOperand.getImm())
+      .addImm(LoOperand.getImm());
+    return;
+  }
+
+  if (!isUInt<6>(LoOperand.getImm())) {
+    assert(isInt<8>(HiOperand.getImm()));
+    BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_iI_V4), DoubleDestReg)
+      .addImm(HiOperand.getImm())
+      .addImm(LoOperand.getImm());
+    return;
+  }
+
+  // Insert new combine instruction.
+  //  DoubleRegDest = combine #HiImm, #LoImm
+  BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg)
+    .addImm(HiOperand.getImm())
+    .addImm(LoOperand.getImm());
+}
+
+void HexagonCopyToCombine::emitCombineIR(MachineBasicBlock::iterator &InsertPt,
+                                         unsigned DoubleDestReg,
+                                         MachineOperand &HiOperand,
+                                         MachineOperand &LoOperand) {
+  unsigned LoReg = LoOperand.getReg();
+  unsigned LoRegKillFlag = getKillRegState(LoOperand.isKill());
+
+  DebugLoc DL = InsertPt->getDebugLoc();
+  MachineBasicBlock *BB = InsertPt->getParent();
+
+  // Handle global.
+  if (HiOperand.isGlobal()) {
+    BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ir_V4), DoubleDestReg)
+      .addGlobalAddress(HiOperand.getGlobal(), HiOperand.getOffset(),
+                        HiOperand.getTargetFlags())
+      .addReg(LoReg, LoRegKillFlag);
+    return;
+  }
+  // Insert new combine instruction.
+  //  DoubleRegDest = combine #HiImm, LoReg
+  BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ir_V4), DoubleDestReg)
+    .addImm(HiOperand.getImm())
+    .addReg(LoReg, LoRegKillFlag);
+}
+
+void HexagonCopyToCombine::emitCombineRI(MachineBasicBlock::iterator &InsertPt,
+                                         unsigned DoubleDestReg,
+                                         MachineOperand &HiOperand,
+                                         MachineOperand &LoOperand) {
+  unsigned HiRegKillFlag = getKillRegState(HiOperand.isKill());
+  unsigned HiReg = HiOperand.getReg();
+
+  DebugLoc DL = InsertPt->getDebugLoc();
+  MachineBasicBlock *BB = InsertPt->getParent();
+
+  // Handle global.
+  if (LoOperand.isGlobal()) {
+    BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rI_V4), DoubleDestReg)
+      .addReg(HiReg, HiRegKillFlag)
+      .addGlobalAddress(LoOperand.getGlobal(), LoOperand.getOffset(),
+                        LoOperand.getTargetFlags());
+    return;
+  }
+
+  // Insert new combine instruction.
+  //  DoubleRegDest = combine HiReg, #LoImm
+  BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rI_V4), DoubleDestReg)
+    .addReg(HiReg, HiRegKillFlag)
+    .addImm(LoOperand.getImm());
+}
+
+void HexagonCopyToCombine::emitCombineRR(MachineBasicBlock::iterator &InsertPt,
+                                         unsigned DoubleDestReg,
+                                         MachineOperand &HiOperand,
+                                         MachineOperand &LoOperand) {
+  unsigned LoRegKillFlag = getKillRegState(LoOperand.isKill());
+  unsigned HiRegKillFlag = getKillRegState(HiOperand.isKill());
+  unsigned LoReg = LoOperand.getReg();
+  unsigned HiReg = HiOperand.getReg();
+
+  DebugLoc DL = InsertPt->getDebugLoc();
+  MachineBasicBlock *BB = InsertPt->getParent();
+
+  // Insert new combine instruction.
+  //  DoubleRegDest = combine HiReg, LoReg
+  BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rr), DoubleDestReg)
+    .addReg(HiReg, HiRegKillFlag)
+    .addReg(LoReg, LoRegKillFlag);
+}
+
+FunctionPass *llvm::createHexagonCopyToCombine() {
+  return new HexagonCopyToCombine();
+}
index d1e32c65adc34f7a100654ac857f7a2124c0b8a6..c96aaca8f8d2514294e253b35b45a11ac66271e4 100644 (file)
@@ -384,6 +384,12 @@ def TFCR : CRInst<(outs CRRegs:$dst), (ins IntRegs:$src1),
 // ALU32/PERM +
 //===----------------------------------------------------------------------===//
 
+let neverHasSideEffects = 1 in
+def COMBINE_ii : ALU32_ii<(outs DoubleRegs:$dst),
+            (ins s8Imm:$src1, s8Imm:$src2),
+            "$dst = combine(#$src1, #$src2)",
+            []>;
+
 // Mux.
 def VMUX_prr64 : ALU64_rr<(outs DoubleRegs:$dst), (ins PredRegs:$src1,
                                                    DoubleRegs:$src2,
index f08da4def37b78fa2f7c800705d575a871f89ee5..676dff2a4a9d5ac1649614d9c359cd00d74e743b 100644 (file)
@@ -162,6 +162,7 @@ bool HexagonPassConfig::addPreSched2() {
   HexagonTargetObjectFile &TLOF =
     (HexagonTargetObjectFile&)(getTargetLowering()->getObjFileLowering());
 
+  addPass(createHexagonCopyToCombine());
   if (getOptLevel() != CodeGenOpt::None)
     addPass(&IfConverterID);
   if (!TLOF.IsSmallDataEnabled()) {
index f8c9e44c831d80531c82dc778c57725e7a913ff7..aea4ffe2eee522bf395a672fe2612c252e8235f6 100644 (file)
@@ -1,11 +1,8 @@
 ; RUN: llc -march=hexagon -mcpu=hexagonv4 -disable-dfa-sched -disable-hexagon-misched < %s | FileCheck %s
 ; CHECK: memw(r29{{ *}}+{{ *}}#0){{ *}}={{ *}}#7
-; CHECK: r0 = #1
-; CHECK: r1 = #2
-; CHECK: r2 = #3
-; CHECK: r3 = #4
-; CHECK: r4 = #5
-; CHECK: r5 = #6
+; CHECK: r1:0 = combine(#2, #1)
+; CHECK: r3:2 = combine(#4, #3)
+; CHECK: r5:4 = combine(#6, #5)
 
 
 define void @foo() nounwind {
diff --git a/test/CodeGen/Hexagon/tfr-to-combine.ll b/test/CodeGen/Hexagon/tfr-to-combine.ll
new file mode 100644 (file)
index 0000000..e3057cd
--- /dev/null
@@ -0,0 +1,35 @@
+; RUN: llc -march=hexagon -mcpu=hexagonv5  -O3 < %s | FileCheck %s
+
+; Check that we combine TFRs and TFRIs into COMBINEs.
+
+@a = external global i16
+@b = external global i16
+@c = external global i16
+
+; Function Attrs: nounwind
+define i64 @test1() #0 {
+; CHECK: combine(#10, #0)
+entry:
+  store i16 0, i16* @a, align 2
+  store i16 10, i16* @b, align 2
+  ret i64 10
+}
+
+; Function Attrs: nounwind
+define i64 @test2() #0 {
+; CHECK: combine(#0, r{{[0-9]+}})
+entry:
+  store i16 0, i16* @a, align 2
+  %0 = load i16* @c, align 2
+  %conv2 = zext i16 %0 to i64
+  ret i64 %conv2
+}
+
+; Function Attrs: nounwind
+define i64 @test4() #0 {
+; CHECK: combine(#0, ##100)
+entry:
+  store i16 100, i16* @b, align 2
+  store i16 0, i16* @a, align 2
+  ret i64 0
+}