Frame lowering optimizations

llvm/lib/Target/Mips/MipsMachineFunction.cpp

@@ -201,3 +201,19 @@ int MipsFunctionInfo::getMoveF64ViaSpillFI(MachineFunction &MF,
 }
 
 void MipsFunctionInfo::anchor() {}
+
+unsigned MipsFunctionInfo::getCalleeSavedStackSize() {
+  return CalleeSavedStackSize;
+}
+
+void MipsFunctionInfo::setCalleeSavedStackSize(unsigned Size) {
+  CalleeSavedStackSize = Size;
+}
+
+bool MipsFunctionInfo::isTwoStepStackSetup(MachineFunction &MF) {
+
+  const MipsSubtarget &STI =
+      *static_cast<const MipsSubtarget *>(&MF.getSubtarget());
+
+  return (MF.getFrameInfo().getStackSize() > 4096) && STI.hasNanoMips();
+}

llvm/lib/Target/Mips/MipsMachineFunction.h

@@ -108,6 +108,9 @@ class MipsFunctionInfo : public MachineFunctionInfo {
       JumpTableEntryInfo[Idx]->Signed = Sign;
     }
   }
+  unsigned getCalleeSavedStackSize();
+  void setCalleeSavedStackSize(unsigned Size);
+  bool isTwoStepStackSetup(MachineFunction &MF);
 
 private:
   virtual void anchor();
@@ -168,6 +171,7 @@ class MipsFunctionInfo : public MachineFunctionInfo {
   };
 
   SmallVector<NanoMipsJumpTableInfo *, 2> JumpTableEntryInfo;
+  unsigned CalleeSavedStackSize = 0;
 };
 
 } // end namespace llvm

llvm/lib/Target/Mips/MipsRegisterInfo.cpp

@@ -303,11 +303,21 @@ Register MipsRegisterInfo::
 getFrameRegister(const MachineFunction &MF) const {
   const MipsSubtarget &Subtarget = MF.getSubtarget<MipsSubtarget>();
   const TargetFrameLowering *TFI = Subtarget.getFrameLowering();
+  const MachineFrameInfo &MFI = MF.getFrameInfo();
+  const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
   bool IsN64 =
       static_cast<const MipsTargetMachine &>(MF.getTarget()).getABI().IsN64();
   bool IsP32 =
       static_cast<const MipsTargetMachine &>(MF.getTarget()).getABI().IsP32();
 
+  // If function doesn't have var-sized objects and function doesn't need stack
+  // realignment but frame pointer elimination is disabled we want offsets to be
+  // relative to sp instead of fp
+  if (Subtarget.hasNanoMips())
+    if (!MFI.hasVarSizedObjects() && !TRI->hasStackRealignment(MF) &&
+        MF.getTarget().Options.DisableFramePointerElim(MF))
+      return Mips::SP_NM;
+
   if (Subtarget.inMips16Mode())
     return TFI->hasFP(MF) ? Mips::S0 : Mips::SP;
   else

llvm/lib/Target/Mips/MipsSEFrameLowering.cpp

@@ -432,19 +432,61 @@ void MipsSEFrameLowering::emitPrologue(MachineFunction &MF,
   MachineModuleInfo &MMI = MF.getMMI();
   const MCRegisterInfo *MRI = MMI.getContext().getRegisterInfo();
 
-  // Adjust stack.
-  TII.adjustStackPtr(SP, -StackSize, MBB, MBBI);
+  const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
+
+  int64_t CalleeSavedStackSize;
+  int64_t LoaclStackSize;
+  // If we have two-step stack setup MBBI_2 will point to the
+  // first instruction after calle-saves store sequence
+  MachineBasicBlock::iterator MBBI_2 = MBBI;
+
+  if (MipsFI->isTwoStepStackSetup(MF)) {
+
+    CalleeSavedStackSize = MipsFI->getCalleeSavedStackSize();
+    unsigned NumOfCSI = MFI.getCalleeSavedInfo().size();
+
+    // Move MBBI_2 to point to the first instruction after
+    // calle-saves store sequence. That's the place for the second
+    // steck pointer adjustment.
+    std::advance(MBBI_2, NumOfCSI);
+
+    // The first stack pointer adjustment to cover space needed
+    // to spill callee-saved registers on stack.
+    TII.adjustStackPtr(SP, -CalleeSavedStackSize, MBB, MBBI);
+
+    LoaclStackSize = StackSize - CalleeSavedStackSize;
 
-  // emit ".cfi_def_cfa_offset StackSize"
-  unsigned CFIIndex =
-      MF.addFrameInst(MCCFIInstruction::cfiDefCfaOffset(nullptr, StackSize));
-  BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
-      .addCFIIndex(CFIIndex);
+    // The second stack pointer adjustment to cover space needed
+    // to spill local objects on stack.
+    TII.adjustStackPtr(SP, -LoaclStackSize, MBB, MBBI_2);
+
+  } else
+    // Adjust stack.
+    TII.adjustStackPtr(SP, -StackSize, MBB, MBBI);
+
+  if (MipsFI->isTwoStepStackSetup(MF)) {
+    // emit ".cfi_def_cfa_offset CalleeSavedStackSize"
+    // emit ".cfi_def_cfa_offset StackSize = CalleeSavedStackSize +
+    // LoaclStackSize"
+    unsigned CFIIndex_1 = MF.addFrameInst(
+        MCCFIInstruction::cfiDefCfaOffset(nullptr, CalleeSavedStackSize));
+    unsigned CFIIndex_2 =
+        MF.addFrameInst(MCCFIInstruction::cfiDefCfaOffset(nullptr, StackSize));
+    BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
+        .addCFIIndex(CFIIndex_1);
+    BuildMI(MBB, MBBI_2, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
+        .addCFIIndex(CFIIndex_2);
+  } else {
+    // emit ".cfi_def_cfa_offset StackSize"
+    unsigned CFIIndex =
+        MF.addFrameInst(MCCFIInstruction::cfiDefCfaOffset(nullptr, StackSize));
+    BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
+        .addCFIIndex(CFIIndex);
+  }
 
   if (MF.getFunction().hasFnAttribute("interrupt"))
     emitInterruptPrologueStub(MF, MBB);
 
-  const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
 
   if (!CSI.empty()) {
     // Find the instruction past the last instruction that saves a callee-saved
@@ -527,15 +569,40 @@ void MipsSEFrameLowering::emitPrologue(MachineFunction &MF,
 
   // if framepointer enabled, set it to point to the stack pointer.
   if (hasFP(MF)) {
-    // Insert instruction "move $fp, $sp" at this location.
-    BuildMI(MBB, MBBI, dl, TII.get(MOVE), FP).addReg(SP).addReg(ZERO)
-      .setMIFlag(MachineInstr::FrameSetup);
 
-    // emit ".cfi_def_cfa_register $fp"
-    unsigned CFIIndex = MF.addFrameInst(MCCFIInstruction::createDefCfaRegister(
-        nullptr, MRI->getDwarfRegNum(FP, true)));
-    BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
-        .addCFIIndex(CFIIndex);
+    if (STI.hasNanoMips()) {
+
+      BuildMI(MBB, MBBI_2, dl, TII.get(ADDiu), FP)
+          .addReg(SP)
+          .addImm(-4096 + StackSize);
+
+      // emit ".cfi_def_cfa_register $fp"
+      unsigned CFIIndex =
+          MF.addFrameInst(MCCFIInstruction::createDefCfaRegister(
+              nullptr, MRI->getDwarfRegNum(FP, true)));
+      BuildMI(MBB, MBBI_2, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
+          .addCFIIndex(CFIIndex);
+
+      // emit ".cfi_def_cfa_offset 4096"
+      unsigned CFIIndex_1 =
+          MF.addFrameInst(MCCFIInstruction::cfiDefCfaOffset(nullptr, 4096));
+      BuildMI(MBB, MBBI_2, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
+          .addCFIIndex(CFIIndex_1);
+
+    } else {
+      // Insert instruction "move $fp, $sp" at this location.
+      BuildMI(MBB, MBBI, dl, TII.get(MOVE), FP)
+          .addReg(SP)
+          .addReg(ZERO)
+          .setMIFlag(MachineInstr::FrameSetup);
+
+      // emit ".cfi_def_cfa_register $fp"
+      unsigned CFIIndex =
+          MF.addFrameInst(MCCFIInstruction::createDefCfaRegister(
+              nullptr, MRI->getDwarfRegNum(FP, true)));
+      BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
+          .addCFIIndex(CFIIndex);
+    }
 
     if (RegInfo.hasStackRealignment(MF)) {
       // addiu $Reg, $zero, -MaxAlignment
@@ -545,13 +612,19 @@ void MipsSEFrameLowering::emitPrologue(MachineFunction &MF,
              "Function's alignment size requirement is not supported.");
       int64_t MaxAlign = -(int64_t)MFI.getMaxAlign().value();
 
-      if (ABI.IsP32())
-        BuildMI(MBB, MBBI, dl, TII.get(Mips::Li_NM), VR).addImm(MaxAlign);
-      else
+      if (ABI.IsP32()) {
+        uint64_t MaxAlignment = MFI.getMaxAlign().value();
+        BuildMI(MBB, MBBI, dl, TII.get(Mips::INS_NM), SP)
+            .addReg(ZERO)
+            .addImm(0)
+            .addImm(Log2_64(MaxAlignment))
+            .addReg(SP);
+      } else {
         BuildMI(MBB, MBBI, dl, TII.get(ADDiu), VR)
             .addReg(ZERO)
             .addImm(MaxAlign);
-      BuildMI(MBB, MBBI, dl, TII.get(AND), SP).addReg(SP).addReg(VR);
+        BuildMI(MBB, MBBI, dl, TII.get(AND), SP).addReg(SP).addReg(VR);
+      }
 
       if (hasBP(MF)) {
         // move $s7, $sp
@@ -709,7 +782,7 @@ void MipsSEFrameLowering::emitEpilogue(MachineFunction &MF,
   unsigned MOVE = ABI.GetGPRMoveOp();
 
   // if framepointer enabled, restore the stack pointer.
-  if (hasFP(MF)) {
+  if (hasFP(MF) && !STI.hasNanoMips()) {
     // Find the first instruction that restores a callee-saved register.
     MachineBasicBlock::iterator I = MBBI;
 
@@ -747,8 +820,26 @@ void MipsSEFrameLowering::emitEpilogue(MachineFunction &MF,
   if (!StackSize)
     return;
 
-  // Adjust stack.
-  TII.adjustStackPtr(SP, StackSize, MBB, MBBI);
+  if (MipsFI->isTwoStepStackSetup(MF)) {
+
+    int64_t CalleeSavedStackSize = MipsFI->getCalleeSavedStackSize();
+    int64_t LoaclStackSize = StackSize - CalleeSavedStackSize;
+
+    int64_t NumOfCSI = MFI.getCalleeSavedInfo().size();
+
+    MachineBasicBlock::iterator MBBI_2 = MBBI;
+    // Move MBBI_2 to point to the first instruction in
+    // calle-saved load sequence. That's the place where we
+    // need to undo the second stack adjustment
+    std::advance(MBBI_2, (-1) * NumOfCSI);
+
+    // Undo the second stack pointer adjustment
+    TII.adjustStackPtr(SP, LoaclStackSize, MBB, MBBI_2);
+    // Undo the first stack pointer adjustment
+    TII.adjustStackPtr(SP, CalleeSavedStackSize, MBB, MBBI);
+  } else
+    // Adjust stack.
+    TII.adjustStackPtr(SP, StackSize, MBB, MBBI);
 }
 
 void MipsSEFrameLowering::emitInterruptEpilogueStub(
@@ -935,22 +1026,60 @@ bool MipsSEFrameLowering::assignCalleeSavedSpillSlots(
   if (!STI.hasNanoMips())
     return false;
 
+  static const std::unordered_map<unsigned, unsigned> Regs = {
+      {Mips::GP_NM, 0}, {Mips::FP_NM, 1}, {Mips::RA_NM, 2},  {Mips::S0_NM, 3},
+      {Mips::S1_NM, 4}, {Mips::S2_NM, 5}, {Mips::S3_NM, 6},  {Mips::S4_NM, 7},
+      {Mips::S5_NM, 8}, {Mips::S6_NM, 9}, {Mips::S7_NM, 10},
+  };
+
+  static const std::unordered_map<unsigned, unsigned> CSNumToReg = {
+      {0, Mips::GP_NM}, {1, Mips::FP_NM}, {2, Mips::RA_NM},  {3, Mips::S0_NM},
+      {4, Mips::S1_NM}, {5, Mips::S2_NM}, {6, Mips::S3_NM},  {7, Mips::S4_NM},
+      {8, Mips::S5_NM}, {9, Mips::S6_NM}, {10, Mips::S7_NM},
+  };
+
   // nanoMIPS save and restore instructions require callee-saved registers to be
   // saved in particular order on the stack.
-  auto SortCalleeSaves = [](CalleeSavedInfo First, CalleeSavedInfo Second) {
-    std::unordered_map<unsigned, unsigned> Regs{
-        {Mips::GP_NM, 0}, {Mips::FP_NM, 1}, {Mips::RA_NM, 2},  {Mips::S0_NM, 3},
-        {Mips::S1_NM, 4}, {Mips::S2_NM, 5}, {Mips::S3_NM, 6},  {Mips::S4_NM, 7},
-        {Mips::S5_NM, 8}, {Mips::S6_NM, 9}, {Mips::S7_NM, 10},
-    };
-
+  auto CompareCalleeSaves = [](CalleeSavedInfo First, CalleeSavedInfo Second) {
     // There should be no callee-saved registers that are not part of the list.
     assert(Regs.find(First.getReg()) != Regs.end() &&
            Regs.find(Second.getReg()) != Regs.end());
 
-    return Regs[First.getReg()] < Regs[Second.getReg()];
+    return Regs.at(First.getReg()) < Regs.at(Second.getReg());
   };
-  std::sort(CSI.begin(), CSI.end(), SortCalleeSaves);
+
+  // If CSI list has less than two callee-saved registers
+  // no insertions nor sorting is needed
+  if (CSI.size() >= 2) {
+
+    SmallBitVector CSNumBitVector(11);
+    for (CalleeSavedInfo &CS : CSI)
+      CSNumBitVector.set(Regs.at(CS.getReg()));
+
+    int MinCSNum = CSNumBitVector.find_first();
+    int MaxCSNum = CSNumBitVector.find_last();
+
+    // Inserting all of the missing callee-saved registers between min and max
+    // in order to allow further load-store optimizations
+    for (int i = MinCSNum + 1; i < MaxCSNum; ++i)
+      if (!CSNumBitVector.test(i))
+        CSI.push_back(CalleeSavedInfo(CSNumToReg.at(i)));
+
+    std::sort(CSI.begin(), CSI.end(), CompareCalleeSaves);
+  }
+
+  MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
+  const MachineRegisterInfo &MRI = MF.getRegInfo();
+
+  unsigned CalleeSavedOffsetSize = 0;
+  for (CalleeSavedInfo &CS : CSI) {
+    Register Reg = CS.getReg();
+    auto RegSize = TRI->getRegSizeInBits(Reg, MRI) / 8;
+    CalleeSavedOffsetSize += RegSize;
+  }
+  uint64_t AlignedCSStackSize = alignTo(CalleeSavedOffsetSize, 16);
+  MipsFI->setCalleeSavedStackSize(AlignedCSStackSize);
+
   return false;
 }
 

llvm/lib/Target/Mips/MipsSEISelDAGToDAG.cpp

@@ -461,8 +461,8 @@ bool MipsSEDAGToDAGISel::selectIntAddrLSL2MM(SDValue Addr, SDValue &Base,
 
 bool MipsSEDAGToDAGISel::selectIntAddrSImm9(SDValue Addr, SDValue &Base,
                                             SDValue &Offset) const {
-  return selectAddrFrameIndex(Addr, Base, Offset) ||
-    selectAddrFrameIndexOffset(Addr, Base, Offset, 9);
+  return selectAddrFrameIndexOffset(Addr, Base, Offset, 9) &&
+         !isa<FrameIndexSDNode>(Base);
 }
 
 bool MipsSEDAGToDAGISel::selectIntAddrSImm10(SDValue Addr, SDValue &Base,
@@ -541,7 +541,8 @@ bool MipsSEDAGToDAGISel::selectAddrFrameIndexUOffset(
 
 bool MipsSEDAGToDAGISel::selectIntAddrUImm12(SDValue Addr, SDValue &Base,
                                              SDValue &Offset) const {
-  return selectAddrFrameIndexUOffset(Addr, Base, Offset, 12, 0);
+  return selectAddrFrameIndex(Addr, Base, Offset) ||
+         selectAddrFrameIndexUOffset(Addr, Base, Offset, 12, 0);
 }
 
 // A load/store 'x' indexed (reg + reg)

llvm/lib/Target/Mips/MipsSERegisterInfo.cpp

@@ -175,6 +175,8 @@ void MipsSERegisterInfo::eliminateFI(MachineBasicBlock::iterator II,
       static_cast<const MipsTargetMachine &>(MF.getTarget()).getABI();
   const MipsRegisterInfo *RegInfo =
     static_cast<const MipsRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
+  const MipsSubtarget &STI =
+      *static_cast<const MipsSubtarget *>(&MF.getSubtarget());
 
   const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
   int MinCSFI = 0;
@@ -220,7 +222,29 @@ void MipsSERegisterInfo::eliminateFI(MachineBasicBlock::iterator II,
   bool IsKill = false;
   int64_t Offset;
 
-  Offset = SPOffset + (int64_t)StackSize;
+  if (STI.hasNanoMips()) {
+
+    if (MipsFI->isTwoStepStackSetup(MF)) {
+
+      int64_t CalleeSavedStackSize = MipsFI->getCalleeSavedStackSize();
+      if (FrameIndex >= MinCSFI && FrameIndex <= MaxCSFI)
+        Offset = SPOffset + (int64_t)CalleeSavedStackSize;
+      else if (FrameReg == Mips::FP_NM)
+        Offset = SPOffset + 4096;
+      else
+        Offset = SPOffset + StackSize;
+
+    } else {
+
+      if (FrameReg == Mips::FP_NM)
+        Offset = SPOffset + 4096;
+      else
+        Offset = SPOffset + StackSize;
+    }
+
+  } else
+    Offset = SPOffset + (int64_t)StackSize;
+
   Offset += MI.getOperand(OpNo + 1).getImm();
 
   LLVM_DEBUG(errs() << "Offset     : " << Offset << "\n"
@@ -247,7 +271,9 @@ void MipsSERegisterInfo::eliminateFI(MachineBasicBlock::iterator II,
     // TODO: This doesn't work well for nanoMIPS, because it has unsigned
     // offsets and this check assumes signed.
     if (OffsetBitSize < 16 && isInt<16>(Offset) &&
-        (!isIntN(OffsetBitSize, Offset) || !isAligned(OffsetAlign, Offset))) {
+        (STI.hasNanoMips() ? !isUIntN(OffsetBitSize, Offset)
+                           : !isIntN(OffsetBitSize, Offset) ||
+                                 !isAligned(OffsetAlign, Offset))) {
       // If we have an offset that needs to fit into a signed n-bit immediate
       // (where n < 16) and doesn't, but does fit into 16-bits then use an ADDiu
       MachineBasicBlock &MBB = *MI.getParent();