[Xtensa] Implement branch relaxation.

Implement insertIndirectBranch function and other functions needed by Branch Relaxation pass.
Also implement estimateFunctionSizeInBytes function in  MachineFunction class.

Author: andreisfr

llvm/include/llvm/CodeGen/MachineFunction.h

@@ -868,6 +868,10 @@ class LLVM_ABI MachineFunction {
   /// it are renumbered.
   void RenumberBlocks(MachineBasicBlock *MBBFrom = nullptr);
 
+  /// Return an estimate of the function's code size,
+  /// taking into account block and function alignment
+  int64_t estimateFunctionSizeInBytes();
+
   /// print - Print out the MachineFunction in a format suitable for debugging
   /// to the specified stream.
   void print(raw_ostream &OS, const SlotIndexes* = nullptr) const;

llvm/lib/CodeGen/MachineFunction.cpp

@@ -378,6 +378,37 @@ void MachineFunction::RenumberBlocks(MachineBasicBlock *MBB) {
   MBBNumberingEpoch++;
 }
 
+int64_t MachineFunction::estimateFunctionSizeInBytes() {
+  const TargetInstrInfo &TII = *getSubtarget().getInstrInfo();
+  const Align FunctionAlignment = getAlignment();
+  MachineFunction::iterator MBBI = begin(), E = end();
+  /// Offset - Distance from the beginning of the function to the end
+  /// of the basic block.
+  int64_t Offset = 0;
+
+  for (; MBBI != E; ++MBBI) {
+    const Align Alignment = MBBI->getAlignment();
+    int64_t BlockSize = 0;
+
+    for (auto &MI : *MBBI) {
+      BlockSize += TII.getInstSizeInBytes(MI);
+    }
+
+    int64_t OffsetBB;
+    if (Alignment <= FunctionAlignment) {
+      OffsetBB = alignTo(Offset, Alignment);
+    } else {
+      // The alignment of this MBB is larger than the function's alignment, so
+      // we can't tell whether or not it will insert nops. Assume that it will.
+      OffsetBB = alignTo(Offset, Alignment) + Alignment.value() -
+                 FunctionAlignment.value();
+    }
+    Offset = OffsetBB + BlockSize;
+  }
+
+  return Offset;
+}
+
 /// This method iterates over the basic blocks and assigns their IsBeginSection
 /// and IsEndSection fields. This must be called after MBB layout is finalized
 /// and the SectionID's are assigned to MBBs.

llvm/lib/Target/Xtensa/XtensaAsmPrinter.cpp

@@ -69,6 +69,9 @@ void XtensaAsmPrinter::emitMachineConstantPoolValue(
     const BlockAddress *BA =
         cast<XtensaConstantPoolConstant>(ACPV)->getBlockAddress();
     MCSym = GetBlockAddressSymbol(BA);
+  } else if (ACPV->isMachineBasicBlock()) {
+    const MachineBasicBlock *MBB = cast<XtensaConstantPoolMBB>(ACPV)->getMBB();
+    MCSym = MBB->getSymbol();
   } else if (ACPV->isJumpTable()) {
     unsigned Idx = cast<XtensaConstantPoolJumpTable>(ACPV)->getIndex();
     MCSym = this->GetJTISymbol(Idx, false);

llvm/lib/Target/Xtensa/XtensaFrameLowering.cpp

@@ -12,6 +12,7 @@
 
 #include "XtensaFrameLowering.h"
 #include "XtensaInstrInfo.h"
+#include "XtensaMachineFunctionInfo.h"
 #include "XtensaSubtarget.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
@@ -260,14 +261,26 @@ void XtensaFrameLowering::processFunctionBeforeFrameFinalized(
   // Set scavenging frame index if necessary.
   MachineFrameInfo &MFI = MF.getFrameInfo();
   uint64_t MaxSPOffset = MFI.estimateStackSize(MF);
+  auto *XtensaFI = MF.getInfo<XtensaMachineFunctionInfo>();
+  unsigned ScavSlotsNum = 0;
 
-  if (isInt<12>(MaxSPOffset))
-    return;
+  if (!isInt<12>(MaxSPOffset))
+    ScavSlotsNum = 1;
+
+  // Far branches over 18-bit offset require a spill slot for scratch register.
+  bool IsLargeFunction = !isInt<18>(MF.estimateFunctionSizeInBytes());
+  if (IsLargeFunction)
+    ScavSlotsNum = std::max(ScavSlotsNum, 1u);
 
   const TargetRegisterClass &RC = Xtensa::ARRegClass;
   unsigned Size = TRI->getSpillSize(RC);
   Align Alignment = TRI->getSpillAlign(RC);
-  int FI = MF.getFrameInfo().CreateStackObject(Size, Alignment, false);
+  for (unsigned I = 0; I < ScavSlotsNum; I++) {
+    int FI = MFI.CreateStackObject(Size, Alignment, false);
+    RS->addScavengingFrameIndex(FI);
 
-  RS->addScavengingFrameIndex(FI);
+    if (IsLargeFunction &&
+        XtensaFI->getBranchRelaxationScratchFrameIndex() == -1)
+      XtensaFI->setBranchRelaxationScratchFrameIndex(FI);
+  }
 }

llvm/lib/Target/Xtensa/XtensaInstrInfo.cpp

@@ -13,11 +13,14 @@
 //===----------------------------------------------------------------------===//
 
 #include "XtensaInstrInfo.h"
+#include "XtensaConstantPoolValue.h"
+#include "XtensaMachineFunctionInfo.h"
 #include "XtensaTargetMachine.h"
 #include "llvm/CodeGen/MachineConstantPool.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/RegisterScavenging.h"
 
 #define GET_INSTRINFO_CTOR_DTOR
 #include "XtensaGenInstrInfo.inc"
@@ -186,6 +189,18 @@ void XtensaInstrInfo::loadImmediate(MachineBasicBlock &MBB,
   }
 }
 
+unsigned XtensaInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
+  switch (MI.getOpcode()) {
+  case TargetOpcode::INLINEASM: { // Inline Asm: Variable size.
+    const MachineFunction *MF = MI.getParent()->getParent();
+    const char *AsmStr = MI.getOperand(0).getSymbolName();
+    return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo());
+  }
+  default:
+    return MI.getDesc().getSize();
+  }
+}
+
 bool XtensaInstrInfo::reverseBranchCondition(
     SmallVectorImpl<MachineOperand> &Cond) const {
   assert(Cond.size() <= 4 && "Invalid branch condition!");
@@ -244,6 +259,74 @@ bool XtensaInstrInfo::reverseBranchCondition(
   }
 }
 
+MachineBasicBlock *
+XtensaInstrInfo::getBranchDestBlock(const MachineInstr &MI) const {
+  unsigned OpCode = MI.getOpcode();
+  switch (OpCode) {
+  case Xtensa::BR_JT:
+  case Xtensa::JX:
+    return nullptr;
+  case Xtensa::J:
+    return MI.getOperand(0).getMBB();
+  case Xtensa::BEQ:
+  case Xtensa::BNE:
+  case Xtensa::BLT:
+  case Xtensa::BLTU:
+  case Xtensa::BGE:
+  case Xtensa::BGEU:
+    return MI.getOperand(2).getMBB();
+  case Xtensa::BEQI:
+  case Xtensa::BNEI:
+  case Xtensa::BLTI:
+  case Xtensa::BLTUI:
+  case Xtensa::BGEI:
+  case Xtensa::BGEUI:
+    return MI.getOperand(2).getMBB();
+  case Xtensa::BEQZ:
+  case Xtensa::BNEZ:
+  case Xtensa::BLTZ:
+  case Xtensa::BGEZ:
+    return MI.getOperand(1).getMBB();
+  default:
+    llvm_unreachable("Unknown branch opcode");
+  }
+}
+
+bool XtensaInstrInfo::isBranchOffsetInRange(unsigned BranchOp,
+                                            int64_t BrOffset) const {
+  switch (BranchOp) {
+  case Xtensa::J:
+    BrOffset -= 4;
+    return isIntN(18, BrOffset);
+  case Xtensa::JX:
+    return true;
+  case Xtensa::BR_JT:
+    return true;
+  case Xtensa::BEQ:
+  case Xtensa::BNE:
+  case Xtensa::BLT:
+  case Xtensa::BLTU:
+  case Xtensa::BGE:
+  case Xtensa::BGEU:
+  case Xtensa::BEQI:
+  case Xtensa::BNEI:
+  case Xtensa::BLTI:
+  case Xtensa::BLTUI:
+  case Xtensa::BGEI:
+  case Xtensa::BGEUI:
+    BrOffset -= 4;
+    return isIntN(8, BrOffset);
+  case Xtensa::BEQZ:
+  case Xtensa::BNEZ:
+  case Xtensa::BLTZ:
+  case Xtensa::BGEZ:
+    BrOffset -= 4;
+    return isIntN(12, BrOffset);
+  default:
+    llvm_unreachable("Unknown branch opcode");
+  }
+}
+
 bool XtensaInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
                                     MachineBasicBlock *&TBB,
                                     MachineBasicBlock *&FBB,
@@ -376,6 +459,130 @@ unsigned XtensaInstrInfo::insertBranch(
   return Count;
 }
 
+void XtensaInstrInfo::insertIndirectBranch(MachineBasicBlock &MBB,
+                                           MachineBasicBlock &DestBB,
+                                           MachineBasicBlock &RestoreBB,
+                                           const DebugLoc &DL, int64_t BrOffset,
+                                           RegScavenger *RS) const {
+  assert(RS && "RegScavenger required for long branching");
+  assert(MBB.empty() &&
+         "new block should be inserted for expanding unconditional branch");
+  assert(MBB.pred_size() == 1);
+
+  MachineFunction *MF = MBB.getParent();
+  MachineRegisterInfo &MRI = MF->getRegInfo();
+  MachineConstantPool *ConstantPool = MF->getConstantPool();
+  auto *XtensaFI = MF->getInfo<XtensaMachineFunctionInfo>();
+  MachineBasicBlock *JumpToMBB = &DestBB;
+
+  if (!isInt<32>(BrOffset))
+    report_fatal_error(
+        "Branch offsets outside of the signed 32-bit range not supported");
+
+  Register ScratchReg = MRI.createVirtualRegister(&Xtensa::ARRegClass);
+  auto II = MBB.end();
+
+  // Create l32r without last operand. We will add this operand later when
+  // JumpToMMB will be calculated and placed to the ConstantPool.
+  MachineInstr &L32R = *BuildMI(MBB, II, DL, get(Xtensa::L32R), ScratchReg);
+  BuildMI(MBB, II, DL, get(Xtensa::JX)).addReg(ScratchReg, RegState::Kill);
+
+  RS->enterBasicBlockEnd(MBB);
+  Register ScavRegister =
+      RS->scavengeRegisterBackwards(Xtensa::ARRegClass, L32R.getIterator(),
+                                    /*RestoreAfter=*/false, /*SpAdj=*/0,
+                                    /*AllowSpill=*/false);
+  if (ScavRegister != Xtensa::NoRegister)
+    RS->setRegUsed(ScavRegister);
+  else {
+    // The case when there is no scavenged register needs special handling.
+    // Pick A8 because it doesn't make a difference
+    ScavRegister = Xtensa::A12;
+
+    int FrameIndex = XtensaFI->getBranchRelaxationScratchFrameIndex();
+    if (FrameIndex == -1)
+      report_fatal_error(
+          "Unable to properly handle scavenged register for indirect jump, "
+          "function code size is significantly larger than estimated");
+
+    storeRegToStackSlot(MBB, L32R, ScavRegister, /*IsKill=*/true, FrameIndex,
+                        &Xtensa::ARRegClass, &RI, Register());
+    RI.eliminateFrameIndex(std::prev(L32R.getIterator()),
+                           /*SpAdj=*/0, /*FIOperandNum=*/1);
+
+    loadRegFromStackSlot(RestoreBB, RestoreBB.end(), ScavRegister, FrameIndex,
+                         &Xtensa::ARRegClass, &RI, Register());
+    RI.eliminateFrameIndex(RestoreBB.back(),
+                           /*SpAdj=*/0, /*FIOperandNum=*/1);
+    JumpToMBB = &RestoreBB;
+  }
+
+  XtensaConstantPoolValue *C = XtensaConstantPoolMBB::Create(
+      MF->getFunction().getContext(), JumpToMBB, 0);
+  unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align(4));
+  L32R.addOperand(MachineOperand::CreateCPI(Idx, 0));
+
+  MRI.replaceRegWith(ScratchReg, ScavRegister);
+  MRI.clearVirtRegs();
+}
+
+unsigned XtensaInstrInfo::insertConstBranchAtInst(
+    MachineBasicBlock &MBB, MachineInstr *I, int64_t offset,
+    ArrayRef<MachineOperand> Cond, DebugLoc DL, int *BytesAdded) const {
+  // Shouldn't be a fall through.
+  assert(&MBB && "InsertBranch must not be told to insert a fallthrough");
+  assert(Cond.size() <= 4 &&
+         "Xtensa branch conditions have less than four components!");
+
+  if (Cond.empty() || (Cond[0].getImm() == Xtensa::J)) {
+    // Unconditional branch
+    MachineInstr *MI = BuildMI(MBB, I, DL, get(Xtensa::J)).addImm(offset);
+    if (BytesAdded && MI)
+      *BytesAdded += getInstSizeInBytes(*MI);
+    return 1;
+  }
+
+  unsigned Count = 0;
+  unsigned BR_C = Cond[0].getImm();
+  MachineInstr *MI = nullptr;
+  switch (BR_C) {
+  case Xtensa::BEQ:
+  case Xtensa::BNE:
+  case Xtensa::BLT:
+  case Xtensa::BLTU:
+  case Xtensa::BGE:
+  case Xtensa::BGEU:
+    MI = BuildMI(MBB, I, DL, get(BR_C))
+             .addImm(offset)
+             .addReg(Cond[1].getReg())
+             .addReg(Cond[2].getReg());
+    break;
+  case Xtensa::BEQI:
+  case Xtensa::BNEI:
+  case Xtensa::BLTI:
+  case Xtensa::BLTUI:
+  case Xtensa::BGEI:
+  case Xtensa::BGEUI:
+    MI = BuildMI(MBB, I, DL, get(BR_C))
+             .addImm(offset)
+             .addReg(Cond[1].getReg())
+             .addImm(Cond[2].getImm());
+    break;
+  case Xtensa::BEQZ:
+  case Xtensa::BNEZ:
+  case Xtensa::BLTZ:
+  case Xtensa::BGEZ:
+    MI = BuildMI(MBB, I, DL, get(BR_C)).addImm(offset).addReg(Cond[1].getReg());
+    break;
+  default:
+    llvm_unreachable("Invalid branch type!");
+  }
+  if (BytesAdded && MI)
+    *BytesAdded += getInstSizeInBytes(*MI);
+  ++Count;
+  return Count;
+}
+
 unsigned XtensaInstrInfo::insertBranchAtInst(MachineBasicBlock &MBB,
                                              MachineBasicBlock::iterator I,
                                              MachineBasicBlock *TBB,

llvm/lib/Target/Xtensa/XtensaInstrInfo.h

@@ -38,6 +38,8 @@ class XtensaInstrInfo : public XtensaGenInstrInfo {
   void adjustStackPtr(unsigned SP, int64_t Amount, MachineBasicBlock &MBB,
                       MachineBasicBlock::iterator I) const;
 
+  unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
+
   // Return the XtensaRegisterInfo, which this class owns.
   const XtensaRegisterInfo &getRegisterInfo() const { return RI; }
 
@@ -77,6 +79,11 @@ class XtensaInstrInfo : public XtensaGenInstrInfo {
   bool
   reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const override;
 
+  MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const override;
+
+  bool isBranchOffsetInRange(unsigned BranchOpc,
+                             int64_t BrOffset) const override;
+
   bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
                      MachineBasicBlock *&FBB,
                      SmallVectorImpl<MachineOperand> &Cond,
@@ -90,12 +97,22 @@ class XtensaInstrInfo : public XtensaGenInstrInfo {
                         const DebugLoc &DL,
                         int *BytesAdded = nullptr) const override;
 
+  void insertIndirectBranch(MachineBasicBlock &MBB, MachineBasicBlock &DestBB,
+                            MachineBasicBlock &RestoreBB, const DebugLoc &DL,
+                            int64_t BrOffset = 0,
+                            RegScavenger *RS = nullptr) const override;
+
   unsigned insertBranchAtInst(MachineBasicBlock &MBB,
                               MachineBasicBlock::iterator I,
                               MachineBasicBlock *TBB,
                               ArrayRef<MachineOperand> Cond, const DebugLoc &DL,
                               int *BytesAdded) const;
 
+  unsigned insertConstBranchAtInst(MachineBasicBlock &MBB, MachineInstr *I,
+                                   int64_t offset,
+                                   ArrayRef<MachineOperand> Cond, DebugLoc DL,
+                                   int *BytesAdded) const;
+
   // Return true if MI is a conditional or unconditional branch.
   // When returning true, set Cond to the mask of condition-code
   // values on which the instruction will branch, and set Target