+
+bool ARMConstantIslands::OptimizeThumb2Instructions(MachineFunction &MF) {
+ bool MadeChange = false;
+
+ // Shrink ADR and LDR from constantpool.
+ for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) {
+ CPUser &U = CPUsers[i];
+ unsigned Opcode = U.MI->getOpcode();
+ unsigned NewOpc = 0;
+ unsigned Scale = 1;
+ unsigned Bits = 0;
+ switch (Opcode) {
+ default: break;
+ case ARM::t2LEApcrel:
+ if (isARMLowRegister(U.MI->getOperand(0).getReg())) {
+ NewOpc = ARM::tLEApcrel;
+ Bits = 8;
+ Scale = 4;
+ }
+ break;
+ case ARM::t2LDRpci:
+ if (isARMLowRegister(U.MI->getOperand(0).getReg())) {
+ NewOpc = ARM::tLDRpci;
+ Bits = 8;
+ Scale = 4;
+ }
+ break;
+ }
+
+ if (!NewOpc)
+ continue;
+
+ unsigned UserOffset = GetOffsetOf(U.MI) + 4;
+ unsigned MaxOffs = ((1 << Bits) - 1) * Scale;
+ // FIXME: Check if offset is multiple of scale if scale is not 4.
+ if (CPEIsInRange(U.MI, UserOffset, U.CPEMI, MaxOffs, false, true)) {
+ U.MI->setDesc(TII->get(NewOpc));
+ MachineBasicBlock *MBB = U.MI->getParent();
+ BBSizes[MBB->getNumber()] -= 2;
+ AdjustBBOffsetsAfter(MBB, -2);
+ ++NumT2CPShrunk;
+ MadeChange = true;
+ }
+ }
+
+ MadeChange |= OptimizeThumb2Branches(MF);
+ MadeChange |= OptimizeThumb2JumpTables(MF);
+ return MadeChange;
+}
+
+bool ARMConstantIslands::OptimizeThumb2Branches(MachineFunction &MF) {
+ bool MadeChange = false;
+
+ for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) {
+ ImmBranch &Br = ImmBranches[i];
+ unsigned Opcode = Br.MI->getOpcode();
+ unsigned NewOpc = 0;
+ unsigned Scale = 1;
+ unsigned Bits = 0;
+ switch (Opcode) {
+ default: break;
+ case ARM::t2B:
+ NewOpc = ARM::tB;
+ Bits = 11;
+ Scale = 2;
+ break;
+ case ARM::t2Bcc: {
+ NewOpc = ARM::tBcc;
+ Bits = 8;
+ Scale = 2;
+ break;
+ }
+ }
+ if (NewOpc) {
+ unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale;
+ MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB();
+ if (BBIsInRange(Br.MI, DestBB, MaxOffs)) {
+ Br.MI->setDesc(TII->get(NewOpc));
+ MachineBasicBlock *MBB = Br.MI->getParent();
+ BBSizes[MBB->getNumber()] -= 2;
+ AdjustBBOffsetsAfter(MBB, -2);
+ ++NumT2BrShrunk;
+ MadeChange = true;
+ }
+ }
+
+ Opcode = Br.MI->getOpcode();
+ if (Opcode != ARM::tBcc)
+ continue;
+
+ NewOpc = 0;
+ unsigned PredReg = 0;
+ ARMCC::CondCodes Pred = llvm::getInstrPredicate(Br.MI, PredReg);
+ if (Pred == ARMCC::EQ)
+ NewOpc = ARM::tCBZ;
+ else if (Pred == ARMCC::NE)
+ NewOpc = ARM::tCBNZ;
+ if (!NewOpc)
+ continue;
+ MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB();
+ // Check if the distance is within 126. Subtract starting offset by 2
+ // because the cmp will be eliminated.
+ unsigned BrOffset = GetOffsetOf(Br.MI) + 4 - 2;
+ unsigned DestOffset = BBOffsets[DestBB->getNumber()];
+ if (BrOffset < DestOffset && (DestOffset - BrOffset) <= 126) {
+ MachineBasicBlock::iterator CmpMI = Br.MI; --CmpMI;
+ if (CmpMI->getOpcode() == ARM::tCMPzi8) {
+ unsigned Reg = CmpMI->getOperand(0).getReg();
+ Pred = llvm::getInstrPredicate(CmpMI, PredReg);
+ if (Pred == ARMCC::AL &&
+ CmpMI->getOperand(1).getImm() == 0 &&
+ isARMLowRegister(Reg)) {
+ MachineBasicBlock *MBB = Br.MI->getParent();
+ MachineInstr *NewBR =
+ BuildMI(*MBB, CmpMI, Br.MI->getDebugLoc(), TII->get(NewOpc))
+ .addReg(Reg).addMBB(DestBB, Br.MI->getOperand(0).getTargetFlags());
+ CmpMI->eraseFromParent();
+ Br.MI->eraseFromParent();
+ Br.MI = NewBR;
+ BBSizes[MBB->getNumber()] -= 2;
+ AdjustBBOffsetsAfter(MBB, -2);
+ ++NumCBZ;
+ MadeChange = true;
+ }
+ }
+ }
+ }
+
+ return MadeChange;
+}
+
+/// OptimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller
+/// jumptables when it's possible.
+bool ARMConstantIslands::OptimizeThumb2JumpTables(MachineFunction &MF) {
+ bool MadeChange = false;
+
+ // FIXME: After the tables are shrunk, can we get rid some of the
+ // constantpool tables?
+ MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
+ if (MJTI == 0) return false;
+
+ const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
+ for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
+ MachineInstr *MI = T2JumpTables[i];
+ const TargetInstrDesc &TID = MI->getDesc();
+ unsigned NumOps = TID.getNumOperands();
+ unsigned JTOpIdx = NumOps - (TID.isPredicable() ? 3 : 2);
+ MachineOperand JTOP = MI->getOperand(JTOpIdx);
+ unsigned JTI = JTOP.getIndex();
+ assert(JTI < JT.size());
+
+ bool ByteOk = true;
+ bool HalfWordOk = true;
+ unsigned JTOffset = GetOffsetOf(MI) + 4;
+ const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs;
+ for (unsigned j = 0, ee = JTBBs.size(); j != ee; ++j) {
+ MachineBasicBlock *MBB = JTBBs[j];
+ unsigned DstOffset = BBOffsets[MBB->getNumber()];
+ // Negative offset is not ok. FIXME: We should change BB layout to make
+ // sure all the branches are forward.
+ if (ByteOk && (DstOffset - JTOffset) > ((1<<8)-1)*2)
+ ByteOk = false;
+ unsigned TBHLimit = ((1<<16)-1)*2;
+ if (HalfWordOk && (DstOffset - JTOffset) > TBHLimit)
+ HalfWordOk = false;
+ if (!ByteOk && !HalfWordOk)
+ break;
+ }
+
+ if (ByteOk || HalfWordOk) {
+ MachineBasicBlock *MBB = MI->getParent();
+ unsigned BaseReg = MI->getOperand(0).getReg();
+ bool BaseRegKill = MI->getOperand(0).isKill();
+ if (!BaseRegKill)
+ continue;
+ unsigned IdxReg = MI->getOperand(1).getReg();
+ bool IdxRegKill = MI->getOperand(1).isKill();
+
+ // Scan backwards to find the instruction that defines the base
+ // register. Due to post-RA scheduling, we can't count on it
+ // immediately preceding the branch instruction.
+ MachineBasicBlock::iterator PrevI = MI;
+ MachineBasicBlock::iterator B = MBB->begin();
+ while (PrevI != B && !PrevI->definesRegister(BaseReg))
+ --PrevI;
+
+ // If for some reason we didn't find it, we can't do anything, so
+ // just skip this one.
+ if (!PrevI->definesRegister(BaseReg))
+ continue;
+
+ MachineInstr *AddrMI = PrevI;
+ bool OptOk = true;
+ // Examine the instruction that calculates the jumptable entry address.
+ // Make sure it only defines the base register and kills any uses
+ // other than the index register.
+ for (unsigned k = 0, eee = AddrMI->getNumOperands(); k != eee; ++k) {
+ const MachineOperand &MO = AddrMI->getOperand(k);
+ if (!MO.isReg() || !MO.getReg())
+ continue;
+ if (MO.isDef() && MO.getReg() != BaseReg) {
+ OptOk = false;
+ break;
+ }
+ if (MO.isUse() && !MO.isKill() && MO.getReg() != IdxReg) {
+ OptOk = false;
+ break;
+ }
+ }
+ if (!OptOk)
+ continue;
+
+ // Now scan back again to find the tLEApcrel or t2LEApcrelJT instruction
+ // that gave us the initial base register definition.
+ for (--PrevI; PrevI != B && !PrevI->definesRegister(BaseReg); --PrevI)
+ ;
+
+ // The instruction should be a tLEApcrel or t2LEApcrelJT; we want
+ // to delete it as well.
+ MachineInstr *LeaMI = PrevI;
+ if ((LeaMI->getOpcode() != ARM::tLEApcrelJT &&
+ LeaMI->getOpcode() != ARM::t2LEApcrelJT) ||
+ LeaMI->getOperand(0).getReg() != BaseReg)
+ OptOk = false;
+
+ if (!OptOk)
+ continue;
+
+ unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT;
+ MachineInstr *NewJTMI = BuildMI(MBB, MI->getDebugLoc(), TII->get(Opc))
+ .addReg(IdxReg, getKillRegState(IdxRegKill))
+ .addJumpTableIndex(JTI, JTOP.getTargetFlags())
+ .addImm(MI->getOperand(JTOpIdx+1).getImm());
+ // FIXME: Insert an "ALIGN" instruction to ensure the next instruction
+ // is 2-byte aligned. For now, asm printer will fix it up.
+ unsigned NewSize = TII->GetInstSizeInBytes(NewJTMI);
+ unsigned OrigSize = TII->GetInstSizeInBytes(AddrMI);
+ OrigSize += TII->GetInstSizeInBytes(LeaMI);
+ OrigSize += TII->GetInstSizeInBytes(MI);
+
+ AddrMI->eraseFromParent();
+ LeaMI->eraseFromParent();
+ MI->eraseFromParent();
+
+ int delta = OrigSize - NewSize;
+ BBSizes[MBB->getNumber()] -= delta;
+ AdjustBBOffsetsAfter(MBB, -delta);
+
+ ++NumTBs;
+ MadeChange = true;
+ }
+ }
+
+ return MadeChange;
+}
+
+/// ReorderThumb2JumpTables - Adjust the function's block layout to ensure that
+/// jump tables always branch forwards, since that's what tbb and tbh need.
+bool ARMConstantIslands::ReorderThumb2JumpTables(MachineFunction &MF) {
+ bool MadeChange = false;
+
+ MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
+ if (MJTI == 0) return false;
+
+ const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
+ for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) {
+ MachineInstr *MI = T2JumpTables[i];
+ const TargetInstrDesc &TID = MI->getDesc();
+ unsigned NumOps = TID.getNumOperands();
+ unsigned JTOpIdx = NumOps - (TID.isPredicable() ? 3 : 2);
+ MachineOperand JTOP = MI->getOperand(JTOpIdx);
+ unsigned JTI = JTOP.getIndex();
+ assert(JTI < JT.size());
+
+ // We prefer if target blocks for the jump table come after the jump
+ // instruction so we can use TB[BH]. Loop through the target blocks
+ // and try to adjust them such that that's true.
+ int JTNumber = MI->getParent()->getNumber();
+ const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs;
+ for (unsigned j = 0, ee = JTBBs.size(); j != ee; ++j) {
+ MachineBasicBlock *MBB = JTBBs[j];
+ int DTNumber = MBB->getNumber();
+
+ if (DTNumber < JTNumber) {
+ // The destination precedes the switch. Try to move the block forward
+ // so we have a positive offset.
+ MachineBasicBlock *NewBB =
+ AdjustJTTargetBlockForward(MBB, MI->getParent());
+ if (NewBB)
+ MJTI->ReplaceMBBInJumpTable(JTI, JTBBs[j], NewBB);
+ MadeChange = true;
+ }
+ }
+ }
+
+ return MadeChange;
+}
+
+MachineBasicBlock *ARMConstantIslands::
+AdjustJTTargetBlockForward(MachineBasicBlock *BB, MachineBasicBlock *JTBB)
+{
+ MachineFunction &MF = *BB->getParent();
+
+ // If the destination block is terminated by an unconditional branch,
+ // try to move it; otherwise, create a new block following the jump
+ // table that branches back to the actual target. This is a very simple
+ // heuristic. FIXME: We can definitely improve it.
+ MachineBasicBlock *TBB = 0, *FBB = 0;
+ SmallVector<MachineOperand, 4> Cond;
+ SmallVector<MachineOperand, 4> CondPrior;
+ MachineFunction::iterator BBi = BB;
+ MachineFunction::iterator OldPrior = prior(BBi);
+
+ // If the block terminator isn't analyzable, don't try to move the block
+ bool B = TII->AnalyzeBranch(*BB, TBB, FBB, Cond);
+
+ // If the block ends in an unconditional branch, move it. The prior block
+ // has to have an analyzable terminator for us to move this one. Be paranoid
+ // and make sure we're not trying to move the entry block of the function.
+ if (!B && Cond.empty() && BB != MF.begin() &&
+ !TII->AnalyzeBranch(*OldPrior, TBB, FBB, CondPrior)) {
+ BB->moveAfter(JTBB);
+ OldPrior->updateTerminator();
+ BB->updateTerminator();
+ // Update numbering to account for the block being moved.
+ MF.RenumberBlocks();
+ ++NumJTMoved;
+ return NULL;
+ }
+
+ // Create a new MBB for the code after the jump BB.
+ MachineBasicBlock *NewBB =
+ MF.CreateMachineBasicBlock(JTBB->getBasicBlock());
+ MachineFunction::iterator MBBI = JTBB; ++MBBI;
+ MF.insert(MBBI, NewBB);
+
+ // Add an unconditional branch from NewBB to BB.
+ // There doesn't seem to be meaningful DebugInfo available; this doesn't
+ // correspond directly to anything in the source.
+ assert (isThumb2 && "Adjusting for TB[BH] but not in Thumb2?");
+ BuildMI(NewBB, DebugLoc(), TII->get(ARM::t2B)).addMBB(BB);
+
+ // Update internal data structures to account for the newly inserted MBB.
+ MF.RenumberBlocks(NewBB);
+
+ // Update the CFG.
+ NewBB->addSuccessor(BB);
+ JTBB->removeSuccessor(BB);
+ JTBB->addSuccessor(NewBB);
+
+ ++NumJTInserted;
+ return NewBB;
+}