[mips] Implement the following optimizations using dominance information to
make PIC calls a little more efficient:

1. Remove instructions setting up $gp if it is known that a function has been
   called at least once.
2. Save the address of a called function in a register instead of loading
   it from the GOT at every call site.

llvm-svn: 195892
diff --git a/llvm/lib/Target/Mips/MipsOptimizePICCall.cpp b/llvm/lib/Target/Mips/MipsOptimizePICCall.cpp
new file mode 100644
index 0000000..8718e04
--- /dev/null
+++ b/llvm/lib/Target/Mips/MipsOptimizePICCall.cpp
@@ -0,0 +1,297 @@
+//===--------- MipsOptimizePICCall.cpp - Optimize PIC Calls ---------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass eliminates unnecessary instructions that set up $gp and replace
+// instructions that load target function addresses with copy instructions.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "optimize-mips-pic-call"
+
+#include "Mips.h"
+#include "MipsTargetMachine.h"
+#include "MipsMachineFunction.h"
+#include "MCTargetDesc/MipsBaseInfo.h"
+#include "llvm/ADT/ScopedHashTable.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Support/CommandLine.h"
+
+using namespace llvm;
+
+static cl::opt<bool> LoadTargetFromGOT("mips-load-target-from-got",
+                                       cl::init(true),
+                                       cl::desc("Load target address from GOT"),
+                                       cl::Hidden);
+
+static cl::opt<bool> EraseGPOpnd("mips-erase-gp-opnd",
+                                 cl::init(true), cl::desc("Erase GP Operand"),
+                                 cl::Hidden);
+
+namespace {
+typedef std::pair<unsigned, unsigned> CntRegP;
+typedef RecyclingAllocator<BumpPtrAllocator,
+                           ScopedHashTableVal<const Value *, CntRegP> >
+AllocatorTy;
+typedef ScopedHashTable<const Value *, CntRegP, DenseMapInfo<const Value *>,
+                        AllocatorTy> ScopedHTType;
+
+class MBBInfo {
+public:
+  MBBInfo(MachineDomTreeNode *N);
+  const MachineDomTreeNode *getNode() const;
+  bool isVisited() const;
+  void preVisit(ScopedHTType &ScopedHT);
+  void postVisit();
+
+private:
+  MachineDomTreeNode *Node;
+  ScopedHTType::ScopeTy *HTScope;
+};
+
+class OptimizePICCall : public MachineFunctionPass {
+public:
+  OptimizePICCall(TargetMachine &tm) : MachineFunctionPass(ID) {}
+
+  virtual const char *getPassName() const { return "Mips OptimizePICCall"; }
+
+  bool runOnMachineFunction(MachineFunction &F);
+
+  void getAnalysisUsage(AnalysisUsage &AU) const {
+    AU.addRequired<MachineDominatorTree>();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+private:
+  /// \brief Visit MBB.
+  bool visitNode(MBBInfo &MBBI);
+
+  /// \brief Test if MI jumps to a function via a register.
+  ///
+  /// Also, return the virtual register containing the target function's address
+  /// and the underlying object in Reg and Val respectively, if the function's
+  /// address can be resolved lazily.
+  bool isCallViaRegister(MachineInstr &MI, unsigned &Reg,
+                         const Value *&Val) const;
+
+  /// \brief Return the number of instructions that dominate the current
+  /// instruction and load the function address from object Entry.
+  unsigned getCount(const Value *Entry);
+
+  /// \brief Return the destination virtual register of the last instruction
+  /// that loads from object Entry.
+  unsigned getReg(const Value *Entry);
+
+  /// \brief Update ScopedHT.
+  void incCntAndSetReg(const Value *Entry, unsigned Reg);
+
+  ScopedHTType ScopedHT;
+  static char ID;
+};
+
+char OptimizePICCall::ID = 0;
+} // end of anonymous namespace
+
+/// Return the first MachineOperand of MI if it is a used virtual register.
+static MachineOperand *getCallTargetRegOpnd(MachineInstr &MI) {
+  if (MI.getNumOperands() == 0)
+    return 0;
+
+  MachineOperand &MO = MI.getOperand(0);
+
+  if (!MO.isReg() || !MO.isUse() ||
+      !TargetRegisterInfo::isVirtualRegister(MO.getReg()))
+    return 0;
+
+  return &MO;
+}
+
+/// Return type of register Reg.
+static MVT::SimpleValueType getRegTy(unsigned Reg, MachineFunction &MF) {
+  const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(Reg);
+  assert(RC->vt_end() - RC->vt_begin() == 1);
+  return *RC->vt_begin();
+}
+
+/// Do the following transformation:
+///
+/// jalr $vreg
+/// =>
+/// copy $t9, $vreg
+/// jalr $t9
+static void setCallTargetReg(MachineBasicBlock *MBB,
+                             MachineBasicBlock::iterator I) {
+  MachineFunction &MF = *MBB->getParent();
+  const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
+  unsigned SrcReg = I->getOperand(0).getReg();
+  unsigned DstReg = getRegTy(SrcReg, MF) == MVT::i32 ? Mips::T9 : Mips::T9_64;
+  BuildMI(*MBB, I, I->getDebugLoc(), TII.get(TargetOpcode::COPY), DstReg)
+      .addReg(SrcReg);
+  I->getOperand(0).setReg(DstReg);
+}
+
+/// Search MI's operands for register GP and erase it.
+static void eraseGPOpnd(MachineInstr &MI) {
+  if (!EraseGPOpnd)
+    return;
+
+  MachineFunction &MF = *MI.getParent()->getParent();
+  MVT::SimpleValueType Ty = getRegTy(MI.getOperand(0).getReg(), MF);
+  unsigned Reg = Ty == MVT::i32 ? Mips::GP : Mips::GP_64;
+
+  for (unsigned I = 0; I < MI.getNumOperands(); ++I) {
+    MachineOperand &MO = MI.getOperand(I);
+    if (MO.isReg() && MO.getReg() == Reg) {
+      MI.RemoveOperand(I);
+      return;
+    }
+  }
+
+  llvm_unreachable(0);
+}
+
+MBBInfo::MBBInfo(MachineDomTreeNode *N) : Node(N), HTScope(0) {}
+
+const MachineDomTreeNode *MBBInfo::getNode() const { return Node; }
+
+bool MBBInfo::isVisited() const { return HTScope; }
+
+void MBBInfo::preVisit(ScopedHTType &ScopedHT) {
+  HTScope = new ScopedHTType::ScopeTy(ScopedHT);
+}
+
+void MBBInfo::postVisit() {
+  delete HTScope;
+}
+
+// OptimizePICCall methods.
+bool OptimizePICCall::runOnMachineFunction(MachineFunction &F) {
+  if (F.getTarget().getSubtarget<MipsSubtarget>().inMips16Mode())
+    return false;
+
+  // Do a pre-order traversal of the dominator tree.
+  MachineDominatorTree *MDT = &getAnalysis<MachineDominatorTree>();
+  bool Changed = false;
+
+  SmallVector<MBBInfo, 8> WorkList(1, MBBInfo(MDT->getRootNode()));
+
+  while (!WorkList.empty()) {
+    MBBInfo &MBBI = WorkList.back();
+
+    // If this MBB has already been visited, destroy the scope for the MBB and
+    // pop it from the work list.
+    if (MBBI.isVisited()) {
+      MBBI.postVisit();
+      WorkList.pop_back();
+      continue;
+    }
+
+    // Visit the MBB and add its children to the work list.
+    MBBI.preVisit(ScopedHT);
+    Changed |= visitNode(MBBI);
+    const MachineDomTreeNode *Node = MBBI.getNode();
+    const std::vector<MachineDomTreeNode *> &Children = Node->getChildren();
+    WorkList.append(Children.begin(), Children.end());
+  }
+
+  return Changed;
+}
+
+bool OptimizePICCall::visitNode(MBBInfo &MBBI) {
+  bool Changed = false;
+  MachineBasicBlock *MBB = MBBI.getNode()->getBlock();
+
+  for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
+       ++I) {
+    unsigned Reg;
+    const Value *Entry;
+
+    // Skip instructions that are not call instructions via registers.
+    if (!isCallViaRegister(*I, Reg, Entry))
+      continue;
+
+    Changed = true;
+    unsigned N = getCount(Entry);
+
+    if (N != 0) {
+      // If a function has been called more than twice, we do not have to emit a
+      // load instruction to get the function address from the GOT, but can
+      // instead reuse the address that has been loaded before.
+      if (N >= 2 && !LoadTargetFromGOT)
+        getCallTargetRegOpnd(*I)->setReg(getReg(Entry));
+
+      // Erase the $gp operand if this isn't the first time a function has
+      // been called. $gp needs to be set up only if the function call can go
+      // through a lazy binding stub.
+      eraseGPOpnd(*I);
+    }
+
+    if (Entry)
+      incCntAndSetReg(Entry, Reg);
+
+    setCallTargetReg(MBB, I);
+  }
+
+  return Changed;
+}
+
+bool OptimizePICCall::isCallViaRegister(MachineInstr &MI, unsigned &Reg,
+                                        const Value *&Val) const {
+  if (!MI.isCall())
+    return false;
+
+  MachineOperand *MO = getCallTargetRegOpnd(MI);
+
+  // Return if MI is not a function call via a register.
+  if (!MO)
+    return false;
+
+  // Get the instruction that loads the function address from the GOT.
+  Reg = MO->getReg();
+  Val = 0;
+  MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
+  MachineInstr *DefMI = MRI.getVRegDef(Reg);
+
+  assert(DefMI);
+
+  // See if DefMI is an instruction that loads from a GOT entry that holds the
+  // address of a lazy binding stub.
+  if (!DefMI->mayLoad() || DefMI->getNumOperands() < 3)
+    return true;
+
+  unsigned Flags = DefMI->getOperand(2).getTargetFlags();
+
+  if (Flags != MipsII::MO_GOT_CALL && Flags != MipsII::MO_CALL_LO16)
+    return true;
+
+  // Return the underlying object for the GOT entry in Val.
+  assert(DefMI->hasOneMemOperand());
+  Val = (*DefMI->memoperands_begin())->getValue();
+  return true;
+}
+
+unsigned OptimizePICCall::getCount(const Value *Entry) {
+  return ScopedHT.lookup(Entry).first;
+}
+
+unsigned OptimizePICCall::getReg(const Value *Entry) {
+  unsigned Reg = ScopedHT.lookup(Entry).second;
+  assert(Reg);
+  return Reg;
+}
+
+void OptimizePICCall::incCntAndSetReg(const Value *Entry, unsigned Reg) {
+  CntRegP P = ScopedHT.lookup(Entry);
+  ScopedHT.insert(Entry, std::make_pair(P.first + 1, Reg));
+}
+
+/// Return an OptimizeCall object.
+FunctionPass *llvm::createMipsOptimizePICCallPass(MipsTargetMachine &TM) {
+  return new OptimizePICCall(TM);
+}